July 29, 2026

The Death of the Chatbot: Why Your Dataverse Strategy Is Broken

You pour significant resources into your enterprise data infrastructure, yet you often find your investment does not yield the promised value. This frustration is common, but the issue often resides not within the technology itself—your Dataverse Strategy Broken approach is failing to deliver the scalable governance and security your organization requires. This flawed methodology impacts your entire digital transformation journey, leading to governance debt and unmonitored shadow AI.

Failures in these environments are multifaceted, involving technical challenges, strategic oversights, and human factors such as uncontrolled "ghost agents" and ambiguous permissions. This deep dive uncovers why a Dataverse Strategy Broken model fails to support modern, autonomous workflows. To achieve true digital transformation, you need clear insights, robust identity management, and precise token optimization.

Instead of relying on fragmented management, transitioning to a centralized control plane like Agent 365 by M365 FM resolves the issues of a Dataverse Strategy Broken framework. By providing full operational visibility, dedicated agent identities, and seamless integration into an Agent Mesh architecture, Agent 365 ensures your enterprise can scale hundreds of autonomous AI agents safely, responsibly, and cost-effectively.

Key Takeaways

Clear business goals help you measure success and get real value from your Dataverse investment.
Strong data governance creates clear ownership rules and keeps your company information accurate and secure.
Out-of-the-box features prevent system rigidness and lower long-term maintenance costs for your team.
Continuous user training and clear support improve adoption rates across your entire organization.
Centralized control platforms protect your system from unmanaged AI tools and security risks.

Why Your Dataverse Strategy Is Broken: Vision & Governance Gaps

Your Dataverse strategy often falters at its very foundation. You invest in powerful technology, but without a clear vision and robust governance, your efforts become a house built on sand. This section explores these critical, often overlooked, foundational issues.

Undefined Business Objectives

You implement Dataverse, but do you truly know why? Many organizations adopt Dataverse simply because it comes with their Microsoft ecosystem. This approach is a common pitfall. You must define clear business objectives for your Dataverse implementation. Without a precise "why," you cannot measure success. You cannot prioritize features. You cannot even tell if your Dataverse is delivering value. This lack of direction means you often build solutions without a clear purpose. You end up with features nobody uses and data that serves no strategic goal. Your investment then yields little return, leaving you frustrated and questioning the platform's utility.

Weak Data Governance & Ownership

A robust Dataverse implementation demands strong data governance. You need clear rules for how your data is created, used, and maintained. Without these rules, your data becomes inconsistent and unreliable. Who owns the customer records? Who decides what data fields are mandatory? These questions require definitive answers. When you lack clear data ownership, accountability disappears. This leads to data quality issues, compliance risks, and a general distrust in your information. Effective data management is impossible without a solid governance framework. This framework includes defining a data retention strategy and ensuring proper retention policies are in place. It also covers crucial aspects like governance and data leak protection and implementing robust data encryption for sensitive information.

Consider these essential components for a robust Dataverse data governance framework:

Governance Component	Key Responsibilities & Capabilities
Ownership & Stewardship	Designate Data Stewards responsible for core entities like Accounts, Contacts, and Products.
Standards & Policies	Establish clear naming conventions, duplicate detection rules, and field validation criteria.
Operational Processes	Standardize procedures for validating, enriching, and rectifying data across connected systems.
Data Lineage & Source Mapping	Differentiate Systems of Record from secondary sources while maintaining lineage maps to track data movement.
Platform Integrity Controls	Leverage native Dataverse features like Business Rules, Field-Level Security, Required Fields, Rollup/Calculated Columns, and Power Automate flows/Plug-ins for automated enforcement.

Flawed Data Modeling Practices

Your Dataverse's foundation rests on its data model. If you build this model poorly, your entire system suffers. Tightly coupling business logic directly with the underlying data schema complicates refactoring and solution scaling. Any alteration to the data structure risks unexpectedly breaking business rules, escalating maintenance complexity, and hindering system administrators from auditing compliance policies effectively.

Flawed data modeling creates numerous problems for your Dataverse:

Performance Bottlenecks: High data volume leads to reduced responsiveness, which degrades user satisfaction and delays processing.
Increased Maintenance Workload: Managing intricate relationship structures like complex hierarchies creates long-term organizational overhead.
High Resource Consumption: Frequent updates to tables require continuous redesign of UI forms, views, and business rules.
Data Inconsistencies and Unexpected Costs: Discrepancies across departmental reports erode trust in operational insights and lead to unforeseen IT expenditures.

These defects directly impact your business operations:

Data Defect	Consequence on Business Operations
Data Inaccuracy	Triggers operational missteps and damages customer satisfaction.
Missing Information	Yields incomplete analytical insights, hindering decision-making.
Cross-System Discrepancies	Causes workflow friction and operational confusion.
Outdated Records	Leads to flawed strategic assumptions and incorrect business actions.

Relying on overly simplified, flat data models works initially but fails when handling complex cross-cutting queries or AI contextualization. This causes list-based architectures to break down under system limits, forcing stressful data migrations driven by audit failures or strategic shifts.

Your Dataverse strategy broken by these fundamental design flaws will never reach its full potential. You need to address these issues head-on to build a stable and scalable platform.

Technical Missteps & Integration Failures

Your Dataverse strategy often stumbles on technical hurdles and integration challenges. These issues can turn a powerful platform into a source of frustration. You must address these common pitfalls to unlock your Dataverse's full potential.

Over-Customization vs. Out-of-the-Box

You frequently face a critical choice: build everything from scratch or use existing features. Over-customizing your Dataverse makes it rigid and hard to maintain. It complicates future updates and increases costs. Instead, you should leverage the robust out-of-the-box capabilities Dataverse offers. These features provide significant advantages:

Standardized Data Model: You get ready-to-use tables for common scenarios like managing contacts, accounts, and activities.
Built-in Integration Interfaces: Dataverse provides out-of-the-box APIs and an OData v4 provider to simplify data connectivity.

This balanced approach ensures your Dataverse remains agile and scalable.

Poor Data Quality & Migration

"Garbage in, garbage out" perfectly describes data migration into your Dataverse. If you move bad data, your entire system suffers. You must prioritize data cleansing and validation before any migration. High data quality is non-negotiable. Consider these practices for successful migration:

Migration Phase	Recommended Quality Practice	Key Validation Actions
Pre-Migration	Pre-migration Profiling	Check for duplicates, missing values, and formatting errors to define cleansing scope.
Cleansing & Preparation	Format Standardization & Enrichment	Conform formats (dates, currencies, addresses) to Dynamics 365 standards and involve business users to fill data gaps.
Testing & Rehearsal	Entity & Integration Testing	Conduct unit and integration tests across linked entities; execute full-scale rehearsals with production data volumes.
Post-Migration	Continuous Monitoring & Governance	Implement quality dashboards to track error metrics and refine governance policies using migration lessons learned.

You also need to manage entity relationships carefully and plan for robust data security and data encryption.

Siloed Integrations & Vendor Lock-in

Your Dataverse should not become an isolated island. Poor integration limits its value and creates data silos. You need seamless connections with your other business systems. Strategies for achieving this include:

Data Consolidation & Migration: Use tools like Azure Data Factory for one-time or continuous data movement.
Master Data Node Pattern: Connect Dataverse to Master Data Management (MDM) platforms for centralized "golden records."

You can also use Virtual Tables for real-time external data access and Event-Driven Architecture to automate workflows. This approach prevents vendor lock-in and expands your Dataverse's reach.

Performance & Operational Overhead

Slow performance frustrates users and wastes valuable resources. High-load database queries and inefficient designs often cause these issues in your Dataverse. You can optimize your Dataverse performance and reduce operational overhead. Key strategies include:

Table Architecture: Select Elastic tables for rapid data loading and high throughput.
Plugin Execution: Transition synchronous logic to asynchronous processes like Power Automate.
Bulk Operations API: Utilize efficient APIs such as CreateMultiple for synchronous plugin logic.

You should also archive cold transactional data for long term data retention, clean system execution logs, and offload telemetry to external platforms. This helps you gain better insights into your system's health.

Neglecting Schema & API Management

Poor schema management leads to application failures and instability. You must follow best practices for your Dataverse schema and API governance. This includes:

Solution Header Usage: Include the MSCRM.SolutionUniqueName header when creating components.
Typed Payload Requirements: Explicitly assign @odata.type in attribute creation requests.
Primary Name Attribute Constraint: Ensure each new table has one primary name attribute.
Publishing Requirement: Trigger PublishXml after updating forms, views, or sitemaps.
XML Alignment: Maintain full attribute consistency between a view's layoutxml and fetchxml.

Establish standard naming conventions for all entities and attributes. This improves data discoverability and streamlines structural understanding. You must also track API call performance and perform routine health audits. This ensures operational stability and helps with data leak protection and managing the data lifecycle.

People Problems: Adoption & Change Management

Your Dataverse strategy often fails not because of the technology, but because of the people using it. If your team does not embrace the system, its potential value disappears. You must address these human elements to ensure your Dataverse investment truly pays off.

Low User Adoption Rates

You invest heavily in Dataverse, but are your users actually using it? Low adoption rates mean your powerful Dataverse system sits underutilized. Users often revert to old habits or manual processes if the new system feels difficult or irrelevant. This leads to wasted resources and a significant gap between your intended Dataverse capabilities and actual usage. You cannot achieve your goals if your team avoids the very tools you provide.

Insufficient Training & Support

You cannot expect users to simply "figure it out." Insufficient training and ongoing support are major roadblocks for any Dataverse implementation. When you launch Dataverse without proper guidance, users become frustrated. They encounter issues, cannot find solutions, and quickly lose confidence. This lack of support directly impacts their ability to leverage Dataverse effectively. You need clear, accessible training and continuous help to empower your team to master the Dataverse platform. This support ensures a smoother digital transformation.

Resistance to Organizational Change

People naturally resist change. Your Dataverse implementation represents a significant shift in how your organization operates. Without proactive change management, you will face pushback. Users may fear new processes, worry about job security, or simply prefer the familiar. You must communicate the benefits of Dataverse clearly and involve users in the process. Address their concerns directly. This proactive approach helps overcome resistance, fostering a positive environment for your Dataverse adoption and overall organizational transformation.

Strategic Blind Spots: Measuring & Evolving Your Dataverse

You often overlook crucial strategic elements when implementing Dataverse. These blind spots prevent you from truly understanding your investment's value and adapting to changing needs. You must address these oversights to ensure your Dataverse delivers sustained success.

Lack of Clear KPIs & ROI

You cannot manage what you do not measure. Many organizations deploy Dataverse without defining clear Key Performance Indicators (KPIs) or a method to track Return on Investment (ROI). You need to establish specific metrics from the start. How will you know if your Dataverse is saving time, reducing costs, or improving customer satisfaction? Without these benchmarks, your Dataverse becomes a black box. You lack the concrete insights to justify further investment or demonstrate its value to stakeholders. This absence of measurable success leaves your Dataverse vulnerable to budget cuts and skepticism.

Ignoring Feedback & Iteration

Your Dataverse implementation is not a one-time project; it is an ongoing journey. You often make the mistake of deploying Dataverse and then moving on, ignoring the critical need for continuous feedback and iteration. Establishing continuous user feedback loops alongside ongoing testing is essential for refining custom pages and enhancing the overall user experience. You must remain flexible to refine and modify the application based on incoming user input. Key strategies for gathering and integrating user feedback throughout the app creation process include:

Early Stakeholder Involvement: Engage your team and end-users during the initial planning phase to define accurate requirements.
Prototyping & Early Testing: Share early app builds with users to collect actionable feedback and identify issues prior to deployment.
Iterative Adaptation: Remain flexible to refine and modify the application based on incoming user input.
Centralized Tracking: Utilize a shared system to log and manage user feedback and bugs efficiently. You must evolve your Dataverse based on real-world usage and changing business demands.

Absence of Executive Sponsorship

Your Dataverse initiative needs a champion at the top. Without strong executive sponsorship, your Dataverse project often falters. Executive sponsors provide the necessary authority, resources, and strategic alignment. They remove roadblocks and ensure the Dataverse strategy remains a priority. When you lack this high-level backing, your Dataverse efforts can lose momentum, struggle for funding, and fail to gain organizational buy-in. This absence of leadership leaves your Dataverse vulnerable and undermines its potential impact.

Uncontrolled "Shadow AI" & Governance Debt

A poor data strategy leads to significant risks, including uncontrolled "shadow AI" and mounting "governance debt." You might find unmanaged AI agents or data flows operating outside your official Dataverse governance. These "ghost agents" create security vulnerabilities and compliance risks. They consume resources without oversight, leading to unexpected costs and a lack of data encryption. You accumulate governance debt when you neglect proper data management, security, and retention policies for your Dataverse. This debt manifests as increased risk, reduced trust in your data, and a complex, unmanageable environment. Your Dataverse strategy broken by these issues cannot provide reliable insights or secure operations.

Fixing Your Dataverse Strategy: A Path Forward

You can turn your Dataverse challenges into triumphs. This section provides actionable steps. You will learn how to re-evaluate and realign your Dataverse implementation. You will ensure it meets your core business objectives.

Realigning with Business Goals

You must start by revisiting your fundamental business goals. Ask yourself: "What problems does this Dataverse solve?" "How does it support our overall digital transformation?" Without clear answers, your Dataverse efforts will drift. Define specific, measurable objectives.

For example, aim to reduce customer service response times by 15% using Dataverse. Or, seek to automate 20% of manual data entry tasks.

These clear goals guide your configuration choices. They help you prioritize features. They ensure every part of your Dataverse contributes to real business value. This strategic alignment prevents your Dataverse from becoming just another unused tool. It makes your investment meaningful.

Establishing Robust Governance

A strong foundation requires robust governance. You need clear rules for your data. Define who owns what data. Establish processes for data creation, updates, and deletion. This includes a formal data strategy. Implement strict data quality standards. Ensure data is accurate and consistent. You must also define your data retention policies. Decide how long you keep different types of data. This ensures compliance and reduces clutter. Crucially, protect sensitive information with strong data encryption. This safeguards your Dataverse and builds trust. Good data management practices prevent future problems. They make your Dataverse reliable.

Prioritizing User Experience & Adoption

Your Dataverse is only as good as its adoption. You must make it easy and valuable for your users. Involve users early in the design process. Listen to their needs and feedback. Provide comprehensive training. Do not just offer a one-time session. Offer ongoing support and resources. Create clear documentation. Address user resistance proactively. Explain the benefits of the new system. Show them how Dataverse makes their jobs easier. A positive user experience drives high adoption rates. This ensures your Dataverse investment delivers its full potential. It fuels your digital transformation.

Continuous Improvement & Measurement

Your Dataverse strategy is not a static plan. It needs constant evolution. Define Key Performance Indicators (KPIs) for your Dataverse. Track these metrics regularly. Measure the Return on Investment (ROI). This shows the value your Dataverse brings. Establish feedback loops. Collect user input continuously. Use this feedback to make iterative improvements. Regularly review your Dataverse performance. Look for areas to optimize. Strong executive sponsorship is vital here. It ensures resources and support for ongoing enhancements. This continuous cycle of measurement and improvement ensures your Dataverse remains relevant. It provides valuable insights. It prevents your dataverse strategy broken from happening again.

Your dataverse strategy broken is not a dead end. It is a clear call for a more thoughtful approach to your Dataverse. You have seen how vision, technical, people, and strategic issues intertwine. These problems affect your Dataverse implementation. By addressing these fundamental challenges, you can transform your Dataverse. It moves from a source of frustration into a powerful, value-generating asset. This ensures true digital transformation. A well-executed Dataverse strategy unlocks incredible insights. It drives your organization forward. Your Dataverse will then deliver immense value. This makes your Dataverse a success. Your Dataverse can truly shine. Your Dataverse potential is vast. Your Dataverse journey starts now.

FAQ

What makes my Dataverse strategy ineffective?

Your Dataverse strategy often fails due to unclear business objectives. You also have weak data governance. Poor data modeling practices undermine your Dataverse foundation. These issues prevent your Dataverse from delivering expected value.

How can I prevent over-customization in Dataverse?

You should leverage out-of-the-box Dataverse features. This approach keeps your Dataverse agile. It simplifies maintenance and future updates. Over-customization makes your Dataverse rigid.

What is the main reason for low Dataverse user adoption?

Insufficient training and support are major hurdles. You cannot expect users to just figure it out. Provide clear guidance. Offer continuous help. This empowers your team to master the Dataverse platform.

How do I measure the success of my Dataverse implementation?

You must define clear Key Performance Indicators (KPIs). Track your Return on Investment (ROI). Without these metrics, you cannot prove your Dataverse value. Measure time saved or costs reduced. This shows the impact of your Dataverse.

What is "governance debt" in Dataverse?

Governance debt happens when you neglect proper data management. You also ignore security and retention policies for your Dataverse. This leads to increased risk. It reduces trust in your data. It creates a complex, unmanageable environment.

🚀 Want to be part of m365.fm?

Then stop just listening… and start showing up.

👉 Connect with me on LinkedIn and let’s make something happen:

🎙️ Be a podcast guest and share your story
🎧 Host your own episode (yes, seriously)
💡 Pitch topics the community actually wants to hear
🌍 Build your personal brand in the Microsoft 365 space

This isn’t just a podcast — it’s a platform for people who take action.

🔥 Most people wait. The best ones don’t.

👉 Connect with me on LinkedIn and send me a message:
"I want in"

Let’s build something awesome 👊

1
00:00:00,000 --> 00:00:02,320
Most organizations today have made the same bet.

2
00:00:02,320 --> 00:00:05,160
They've implemented co-pilot or built a chatbot

3
00:00:05,160 --> 00:00:07,680
or deployed some flavor of AI assistant

4
00:00:07,680 --> 00:00:10,440
that lets employees ask questions and get answers back.

5
00:00:10,440 --> 00:00:13,240
It works, employees use it, it feels productive,

6
00:00:13,240 --> 00:00:14,720
but here's what they don't see yet.

7
00:00:14,720 --> 00:00:17,400
That initial success is masking a structural problem.

8
00:00:17,400 --> 00:00:19,080
The problem isn't that co-pilot is bad.

9
00:00:19,080 --> 00:00:21,000
The problem is that chat is incomplete.

10
00:00:21,000 --> 00:00:24,480
Treating AI as a series of isolated conversational interfaces,

11
00:00:24,480 --> 00:00:27,440
one agent here, another tool there, a workflow somewhere else,

12
00:00:27,440 --> 00:00:29,840
feels fast because you're shipping individual pieces,

13
00:00:29,840 --> 00:00:33,120
but as you add more agents, more integrations and more autonomy,

14
00:00:33,120 --> 00:00:35,960
the hidden costs start appearing governance-debt.

15
00:00:35,960 --> 00:00:38,040
You've got agents borrowing human identities

16
00:00:38,040 --> 00:00:40,480
because nobody set up proper identity infrastructure.

17
00:00:40,480 --> 00:00:42,720
You've got models being called directly by every team

18
00:00:42,720 --> 00:00:44,600
because there's no central gateway.

19
00:00:44,600 --> 00:00:46,880
You've got data moving across regions and clouds

20
00:00:46,880 --> 00:00:48,480
with no policy enforcement.

21
00:00:48,480 --> 00:00:49,840
You've got agents still running two years

22
00:00:49,840 --> 00:00:51,440
after their original project ended,

23
00:00:51,440 --> 00:00:53,440
consuming tokens nobody's paying attention to.

24
00:00:53,440 --> 00:00:54,920
You've got no way to answer the question,

25
00:00:54,920 --> 00:00:57,320
which agent accessed this data and when.

26
00:00:57,320 --> 00:00:59,400
This is the moment most organizations realize

27
00:00:59,400 --> 00:01:00,760
that chat isn't a platform.

28
00:01:00,760 --> 00:01:03,120
It's a feature sitting on top of a missing architecture,

29
00:01:03,120 --> 00:01:05,120
the shift from co-pilot to agent mesh,

30
00:01:05,120 --> 00:01:06,880
from chat to agent fabric,

31
00:01:06,880 --> 00:01:08,760
requires building four structural layers

32
00:01:08,760 --> 00:01:10,200
that most organizations skip.

33
00:01:10,200 --> 00:01:11,800
You need a traceable identity plane.

34
00:01:11,800 --> 00:01:14,040
You need a reasoning layer that understands business logic

35
00:01:14,040 --> 00:01:15,680
instead of guessing at documents.

36
00:01:15,680 --> 00:01:18,600
You need a governance hub that centralizes policy enforcement

37
00:01:18,600 --> 00:01:20,640
and you need observability that lets you trace

38
00:01:20,640 --> 00:01:22,280
every action back to its source.

39
00:01:22,280 --> 00:01:24,920
Without these layers, you're not scaling AI.

40
00:01:24,920 --> 00:01:26,120
You're scaling risk.

41
00:01:26,120 --> 00:01:27,600
The co-pilot illusion.

42
00:01:27,600 --> 00:01:29,200
Let's be clear about what co-pilot is

43
00:01:29,200 --> 00:01:30,680
because the confusion starts here.

44
00:01:30,680 --> 00:01:32,040
Co-pilot is a feature.

45
00:01:32,040 --> 00:01:33,800
It's an incredibly useful feature.

46
00:01:33,800 --> 00:01:36,520
A conversational interface layered on top of a language model.

47
00:01:36,520 --> 00:01:38,800
It feels like you're building an AI platform

48
00:01:38,800 --> 00:01:41,640
because it's easy to stand up, easy to show results with

49
00:01:41,640 --> 00:01:43,880
and easy to get executives excited about.

50
00:01:43,880 --> 00:01:46,120
You spin it up on a Friday, show it to your team on Monday

51
00:01:46,120 --> 00:01:48,640
and suddenly everyone's talking about AI transformation,

52
00:01:48,640 --> 00:01:49,840
but a feature is not a platform.

53
00:01:49,840 --> 00:01:52,280
A platform is what you need to actually run agents at scale.

54
00:01:52,280 --> 00:01:53,520
A platform is the infrastructure

55
00:01:53,520 --> 00:01:55,080
that lets you deploy govern monitor

56
00:01:55,080 --> 00:01:58,200
and retire dozens or hundreds of agents without creating chaos.

57
00:01:58,200 --> 00:01:59,240
Here's the seduction.

58
00:01:59,240 --> 00:02:01,440
In isolation, co-pilot works beautifully.

59
00:02:01,440 --> 00:02:04,120
One team builds an agent for customer support.

60
00:02:04,120 --> 00:02:05,920
It integrates with their knowledge base.

61
00:02:05,920 --> 00:02:07,400
It handles routine questions

62
00:02:07,400 --> 00:02:09,080
and it escalates complex issues.

63
00:02:09,080 --> 00:02:10,080
Everyone's happy.

64
00:02:10,080 --> 00:02:11,280
Token consumption is low.

65
00:02:11,280 --> 00:02:13,080
Nobody's asking hard governance questions

66
00:02:13,080 --> 00:02:14,440
and the team ships fast.

67
00:02:14,440 --> 00:02:15,880
Then another team wants an agent.

68
00:02:15,880 --> 00:02:17,840
Then another suddenly you've got a support agent,

69
00:02:17,840 --> 00:02:20,360
a finance agent, a sales agent, and a recruiting agent.

70
00:02:20,360 --> 00:02:21,760
They're built in different places.

71
00:02:21,760 --> 00:02:25,120
Some in co-pilot studio, some in Azure AI, some in Foundry.

72
00:02:25,120 --> 00:02:26,720
Each team chose their own approach

73
00:02:26,720 --> 00:02:29,400
because there was no central architecture to guide them.

74
00:02:29,400 --> 00:02:30,880
And that's when the seam starts showing.

75
00:02:30,880 --> 00:02:33,440
The support agent is borrowing the IT manager's identity

76
00:02:33,440 --> 00:02:36,480
because nobody defined what an agent identity should look like.

77
00:02:36,480 --> 00:02:39,240
The finance agent is calling Azure OpenAI directly

78
00:02:39,240 --> 00:02:40,560
because there's no gateway.

79
00:02:40,560 --> 00:02:42,040
The sales agent is accessing data

80
00:02:42,040 --> 00:02:44,560
from multiple regions without any policy enforcement.

81
00:02:44,560 --> 00:02:45,880
The recruiting agent is still running

82
00:02:45,880 --> 00:02:48,800
on the same credentials the original builder set up three years ago.

83
00:02:48,800 --> 00:02:51,720
Nobody knows which agent is consuming the most tokens.

84
00:02:51,720 --> 00:02:54,360
Nobody knows which agent has access to which data.

85
00:02:54,360 --> 00:02:55,280
Nobody knows what happens

86
00:02:55,280 --> 00:02:56,800
if one of them starts behaving badly.

87
00:02:56,800 --> 00:02:59,560
This is the scaling problem that co-pilot doesn't solve.

88
00:02:59,560 --> 00:03:01,920
When you're running one agent, governance feels optional.

89
00:03:01,920 --> 00:03:04,560
When you're running 50, governance becomes survival.

90
00:03:04,560 --> 00:03:07,560
And by then your retrofitting controls onto a system

91
00:03:07,560 --> 00:03:09,440
that was never built to support them.

92
00:03:09,440 --> 00:03:11,400
The real cost isn't in the technology.

93
00:03:11,400 --> 00:03:12,760
It's in the organizational chaos.

94
00:03:12,760 --> 00:03:14,360
Finance runs their own model access

95
00:03:14,360 --> 00:03:17,440
because IT was too slow, marketing integrates their own tools

96
00:03:17,440 --> 00:03:19,640
because they couldn't get on the central catalog.

97
00:03:19,640 --> 00:03:22,280
Sales builds agents that move data across clouds

98
00:03:22,280 --> 00:03:24,400
because nobody enforced regional policies.

99
00:03:24,400 --> 00:03:26,160
And when compliance comes knocking

100
00:03:26,160 --> 00:03:29,360
with a question about which agent accessed this customer record,

101
00:03:29,360 --> 00:03:30,240
you don't have an answer.

102
00:03:30,240 --> 00:03:31,120
You can't trace it.

103
00:03:31,120 --> 00:03:32,000
You can't audit it.

104
00:03:32,000 --> 00:03:33,160
You can't even shut it down

105
00:03:33,160 --> 00:03:35,240
without breaking the process it was supporting.

106
00:03:35,240 --> 00:03:36,320
This is governance debt.

107
00:03:36,320 --> 00:03:38,440
You're borrowing control against future certainty

108
00:03:38,440 --> 00:03:39,560
that you won't have it.

109
00:03:39,560 --> 00:03:41,440
The hidden assumption underneath co-pilot,

110
00:03:41,440 --> 00:03:43,440
the one nobody voices but everyone makes,

111
00:03:43,440 --> 00:03:45,720
is that chat interfaces are governance enough

112
00:03:45,720 --> 00:03:48,320
that if employees can ask questions and get answers,

113
00:03:48,320 --> 00:03:49,720
that's platform behavior.

114
00:03:49,720 --> 00:03:51,400
But chat has no identity layer.

115
00:03:51,400 --> 00:03:52,880
Chat has no policy enforcement.

116
00:03:52,880 --> 00:03:54,680
Chat has no reasoning about business logic.

117
00:03:54,680 --> 00:03:55,600
Chat is stateless.

118
00:03:55,600 --> 00:03:57,080
Each conversation is isolated

119
00:03:57,080 --> 00:03:59,360
with no understanding of the full context.

120
00:03:59,360 --> 00:04:01,480
Chat scales to chaos remarkably fast.

121
00:04:01,480 --> 00:04:03,280
The move from co-pilot to agent mesh

122
00:04:03,280 --> 00:04:05,080
isn't about replacing co-pilot.

123
00:04:05,080 --> 00:04:06,680
It's about building the architecture

124
00:04:06,680 --> 00:04:09,320
that co-pilot should have been sitting on top of from the start.

125
00:04:09,320 --> 00:04:10,480
And that architecture starts

126
00:04:10,480 --> 00:04:13,200
with the problem most organizations don't even know they have.

127
00:04:13,200 --> 00:04:15,040
The moment your first agent logs in,

128
00:04:15,040 --> 00:04:16,840
you need to answer the question of who,

129
00:04:16,840 --> 00:04:20,160
or what is actually logging in, the identity problem.

130
00:04:20,160 --> 00:04:21,080
That question has a name.

131
00:04:21,080 --> 00:04:22,120
It's called identity.

132
00:04:22,120 --> 00:04:24,880
And identity is where most agent architectures fall apart.

133
00:04:24,880 --> 00:04:26,200
Here is what happens in practice.

134
00:04:26,200 --> 00:04:27,320
A team builds an agent.

135
00:04:27,320 --> 00:04:30,920
It needs to access data, call APIs, or trigger workflows.

136
00:04:30,920 --> 00:04:32,720
So someone asks a simple question,

137
00:04:32,720 --> 00:04:34,440
what account should this agent use?

138
00:04:34,440 --> 00:04:36,520
The answer is usually whatever is easiest,

139
00:04:36,520 --> 00:04:37,720
a shared service account,

140
00:04:37,720 --> 00:04:39,000
a developer's personal login.

141
00:04:39,000 --> 00:04:40,920
Sometimes they just use the identity of whoever

142
00:04:40,920 --> 00:04:43,560
creates the agent and leave it that way forever.

143
00:04:43,560 --> 00:04:45,480
The thinking is simple, the agent needs to do work

144
00:04:45,480 --> 00:04:49,160
so you give it credentials that let it do work problem solved.

145
00:04:49,160 --> 00:04:51,240
Except the problem isn't solved, it's hidden.

146
00:04:51,240 --> 00:04:53,440
The moment you do this, you create an accountability void.

147
00:04:53,440 --> 00:04:55,280
An agent running under a shared service account

148
00:04:55,280 --> 00:04:58,120
looks exactly like a human running under that same account.

149
00:04:58,120 --> 00:05:00,600
When that agent accesses customer data at 3 AM

150
00:05:00,600 --> 00:05:02,080
and moves it somewhere it shouldn't,

151
00:05:02,080 --> 00:05:04,880
your logs will show that the account access the data.

152
00:05:04,880 --> 00:05:06,320
But you cannot see that an agent did it.

153
00:05:06,320 --> 00:05:09,440
You cannot trace the decision back to the logic of the system.

154
00:05:09,440 --> 00:05:11,040
You cannot understand what it was trying to do.

155
00:05:11,040 --> 00:05:13,160
You cannot even prove it was an automated system

156
00:05:13,160 --> 00:05:15,120
and not the person whose credentials it's using.

157
00:05:15,120 --> 00:05:16,680
Now, scale that problem.

158
00:05:16,680 --> 00:05:19,320
You have 30 agents running under five shared accounts.

159
00:05:19,320 --> 00:05:20,920
One of them is accessing financial records

160
00:05:20,920 --> 00:05:22,720
without authorization, which one?

161
00:05:22,720 --> 00:05:23,920
When? Why?

162
00:05:23,920 --> 00:05:26,240
The audit log shows the account name and nothing else.

163
00:05:26,240 --> 00:05:28,560
The person who owns that account is confused

164
00:05:28,560 --> 00:05:31,080
because they don't remember authorizing that access.

165
00:05:31,080 --> 00:05:32,640
The developer who built the offending agent

166
00:05:32,640 --> 00:05:34,360
left the company months ago.

167
00:05:34,360 --> 00:05:36,720
You are stuck trying to reverse engineer a crisis

168
00:05:36,720 --> 00:05:39,320
by looking at a timestamp and a shared password.

169
00:05:39,320 --> 00:05:42,160
This is when regulators start asking hard questions.

170
00:05:42,160 --> 00:05:44,720
GDPR says you need to know who accessed personal data.

171
00:05:44,720 --> 00:05:47,400
ASOKI2 says you need to demonstrate access controls.

172
00:05:47,400 --> 00:05:49,920
HIPPA says you need to prove that only authorized systems

173
00:05:49,920 --> 00:05:51,200
touched health information.

174
00:05:51,200 --> 00:05:53,080
Your audit trail shows a service account.

175
00:05:53,080 --> 00:05:54,280
It doesn't show an agent.

176
00:05:54,280 --> 00:05:55,320
It doesn't show intent.

177
00:05:55,320 --> 00:05:57,680
It doesn't show whether the access was a legitimate task

178
00:05:57,680 --> 00:05:58,920
or a massive mistake.

179
00:05:58,920 --> 00:06:00,600
You cannot pass an audit with that.

180
00:06:00,600 --> 00:06:03,240
But there is a deeper structural reason identity matters.

181
00:06:03,240 --> 00:06:04,800
Agents are not passive tools.

182
00:06:04,800 --> 00:06:06,200
They are autonomous systems.

183
00:06:06,200 --> 00:06:07,120
They make decisions.

184
00:06:07,120 --> 00:06:08,520
They execute actions.

185
00:06:08,520 --> 00:06:09,720
They access data.

186
00:06:09,720 --> 00:06:10,880
And when something goes wrong,

187
00:06:10,880 --> 00:06:12,960
you need to know more than just what happened.

188
00:06:12,960 --> 00:06:14,480
You need to know why it happened.

189
00:06:14,480 --> 00:06:16,760
You need to trace back through the decision-making chain

190
00:06:16,760 --> 00:06:19,080
to see what data it saw, what policies it followed,

191
00:06:19,080 --> 00:06:20,360
and what its reasoning was.

192
00:06:20,360 --> 00:06:22,600
If the agent is hiding behind a human identity,

193
00:06:22,600 --> 00:06:23,720
that trace is broken.

194
00:06:23,720 --> 00:06:26,080
You lose the distinction between what a person authorized

195
00:06:26,080 --> 00:06:27,960
and what an agent decided to do on its own.

196
00:06:27,960 --> 00:06:30,880
The solution is what EntraID calls an agent ID.

197
00:06:30,880 --> 00:06:33,040
It is an identity object in your directory,

198
00:06:33,040 --> 00:06:34,720
just like a user or an app registration,

199
00:06:34,720 --> 00:06:36,640
but it is built specifically for agents.

200
00:06:36,640 --> 00:06:38,200
Each agent gets its own identity.

201
00:06:38,200 --> 00:06:40,040
Each agent has its own permissions,

202
00:06:40,040 --> 00:06:41,880
scope to exactly what it needs.

203
00:06:41,880 --> 00:06:44,000
Each action is logged under the agent's own name

204
00:06:44,000 --> 00:06:46,360
instead of being borrowed from a human.

205
00:06:46,360 --> 00:06:49,120
When that finance agent accesses a record at 3am,

206
00:06:49,120 --> 00:06:51,120
the log shows exactly which agent did it.

207
00:06:51,120 --> 00:06:53,360
You can trace the event back to the configuration,

208
00:06:53,360 --> 00:06:55,440
the permissions and the approval chain.

209
00:06:55,440 --> 00:06:58,280
You can ask if this agent was authorized to touch this data

210
00:06:58,280 --> 00:06:59,880
and the answer is a clear yes or no.

211
00:06:59,880 --> 00:07:00,960
That isn't just compliance.

212
00:07:00,960 --> 00:07:02,280
That's control.

213
00:07:02,280 --> 00:07:03,760
The moment you have traceable identities,

214
00:07:03,760 --> 00:07:04,560
the model changes.

215
00:07:04,560 --> 00:07:06,200
You can audit what agents are doing.

216
00:07:06,200 --> 00:07:07,480
You can enforce lease privilege

217
00:07:07,480 --> 00:07:09,880
so each agent gets only the permissions it requires.

218
00:07:09,880 --> 00:07:11,920
You can revoke access instantly

219
00:07:11,920 --> 00:07:13,320
if an agent misbehaves.

220
00:07:13,320 --> 00:07:16,400
You finally understand the full chain of accountability

221
00:07:16,400 --> 00:07:19,240
from who created the agent to what data it touches.

222
00:07:19,240 --> 00:07:22,480
Without it, you're flying blind, the reasoning floor,

223
00:07:22,480 --> 00:07:23,840
identity solved one problem,

224
00:07:23,840 --> 00:07:26,000
but it didn't solve the problem you actually hit

225
00:07:26,000 --> 00:07:27,960
when you try to make an agent useful.

226
00:07:27,960 --> 00:07:30,000
That problem is thinking.

227
00:07:30,000 --> 00:07:33,400
Here is a scenario that plays out in organizations every day.

228
00:07:33,400 --> 00:07:35,360
A team builds an agent to answer questions

229
00:07:35,360 --> 00:07:36,840
about company policy.

230
00:07:36,840 --> 00:07:40,040
Someone asks, what is our revenue recognition policy?

231
00:07:40,040 --> 00:07:41,880
The agent processes the question,

232
00:07:41,880 --> 00:07:43,200
searches through documents,

233
00:07:43,200 --> 00:07:45,440
finds something relevant and generates an answer.

234
00:07:45,440 --> 00:07:47,320
It looks right, it sounds confident.

235
00:07:47,320 --> 00:07:48,880
The person who asked the question

236
00:07:48,880 --> 00:07:50,840
reads the answer and acts on it.

237
00:07:50,840 --> 00:07:53,200
Six months later, that decision cost the company money

238
00:07:53,200 --> 00:07:54,320
because the agent was wrong.

239
00:07:54,320 --> 00:07:56,000
What actually happened?

240
00:07:56,000 --> 00:07:57,480
The agent found a document.

241
00:07:57,480 --> 00:08:00,240
It searched for text related to revenue and policy.

242
00:08:00,240 --> 00:08:01,640
It stitched together fragments

243
00:08:01,640 --> 00:08:03,720
that seemed to match the words in the question.

244
00:08:03,720 --> 00:08:05,560
It generated something plausible.

245
00:08:05,560 --> 00:08:07,600
But it had no idea if the answer was correct

246
00:08:07,600 --> 00:08:09,600
because it wasn't thinking about revenue recognition

247
00:08:09,600 --> 00:08:10,640
as a business concept.

248
00:08:10,640 --> 00:08:11,880
It was matching text patterns.

249
00:08:11,880 --> 00:08:14,440
This is what happens when you build agents on top of Ragn.

250
00:08:14,440 --> 00:08:16,440
Retrieval augmented generation is a technique

251
00:08:16,440 --> 00:08:18,560
where you take a question, search a document store

252
00:08:18,560 --> 00:08:20,680
for passages and feed those to a model.

253
00:08:20,680 --> 00:08:23,040
The goal is to ground the answer in your actual data

254
00:08:23,040 --> 00:08:24,920
instead of letting the model hallucinate.

255
00:08:24,920 --> 00:08:27,520
In theory, this solves the hallucination problem.

256
00:08:27,520 --> 00:08:30,240
In practice, Ragn just changes the shape of the hallucination

257
00:08:30,240 --> 00:08:32,560
because Ragn is document-centric.

258
00:08:32,560 --> 00:08:35,600
It treats your knowledge base as a pile of text chunks.

259
00:08:35,600 --> 00:08:36,880
When someone asks the question,

260
00:08:36,880 --> 00:08:38,960
the system searches for words that look relevant

261
00:08:38,960 --> 00:08:40,280
and hands them to the model.

262
00:08:40,280 --> 00:08:42,280
Then the model tries to guess what the question means

263
00:08:42,280 --> 00:08:44,600
and builds an answer from the fragments it was given.

264
00:08:44,600 --> 00:08:46,640
But text fragments are not meaning.

265
00:08:46,640 --> 00:08:48,280
They are just words.

266
00:08:48,280 --> 00:08:49,920
An agent that answers a policy question

267
00:08:49,920 --> 00:08:52,560
by retrieving a document is not understanding the policy,

268
00:08:52,560 --> 00:08:54,400
it is pattern matching against text.

269
00:08:54,400 --> 00:08:56,000
It doesn't know that revenue policy connects

270
00:08:56,000 --> 00:08:58,680
to sales processes, accounting systems, contract terms

271
00:08:58,680 --> 00:08:59,720
and regional regulations.

272
00:08:59,720 --> 00:09:02,280
It doesn't know which policies contradict each other

273
00:09:02,280 --> 00:09:04,720
or which one takes precedence in a specific case.

274
00:09:04,720 --> 00:09:07,080
It doesn't know the business logic underneath the words.

275
00:09:07,080 --> 00:09:09,720
So it hallucinates.

276
00:09:09,720 --> 00:09:11,640
Not in the way people usually think.

277
00:09:11,640 --> 00:09:13,280
It isn't making up facts from thin air.

278
00:09:13,280 --> 00:09:14,880
It is doing something much more dangerous.

279
00:09:14,880 --> 00:09:16,400
It is taking real text fragments

280
00:09:16,400 --> 00:09:18,520
and assembling them into something that sounds coherent

281
00:09:18,520 --> 00:09:21,640
but has no relationship to how your business actually works.

282
00:09:21,640 --> 00:09:23,720
The gap here is the gap between finding text

283
00:09:23,720 --> 00:09:25,040
and understanding context.

284
00:09:25,040 --> 00:09:27,680
The solution is what Microsoft calls fabric ontologies.

285
00:09:27,680 --> 00:09:29,320
And it is a completely different way of thinking

286
00:09:29,320 --> 00:09:30,880
about how agents reason.

287
00:09:30,880 --> 00:09:33,280
Instead of treating your knowledge as a document store,

288
00:09:33,280 --> 00:09:35,400
you treat it as a model of your business.

289
00:09:35,400 --> 00:09:37,360
You define the entities that matter.

290
00:09:37,360 --> 00:09:40,840
Customer, order, product, root, asset.

291
00:09:40,840 --> 00:09:42,360
You define how they relate to each other.

292
00:09:42,360 --> 00:09:44,080
You define the rules that govern them.

293
00:09:44,080 --> 00:09:45,560
A customer places orders.

294
00:09:45,560 --> 00:09:48,320
Orders contain products, products belong to categories.

295
00:09:48,320 --> 00:09:51,960
Root service regions, assets break down and need maintenance.

296
00:09:51,960 --> 00:09:53,560
Then when an agent needs to answer a question,

297
00:09:53,560 --> 00:09:54,840
it doesn't search for text.

298
00:09:54,840 --> 00:09:56,680
It traverses that model.

299
00:09:56,680 --> 00:09:58,080
When asked about revenue policy,

300
00:09:58,080 --> 00:10:00,320
the agent understands revenue as a concept

301
00:10:00,320 --> 00:10:01,720
connected to orders and products.

302
00:10:01,720 --> 00:10:03,560
It understands policy as a rule set

303
00:10:03,560 --> 00:10:05,640
connected to contract types and regions.

304
00:10:05,640 --> 00:10:07,520
It can reason across those relationships

305
00:10:07,520 --> 00:10:09,920
instead of guessing at text fragments.

306
00:10:09,920 --> 00:10:11,880
The difference is structural.

307
00:10:11,880 --> 00:10:14,280
Rags says, find the words that match.

308
00:10:14,280 --> 00:10:17,760
Ontologies say, find the entities and relationships that matter.

309
00:10:17,760 --> 00:10:20,280
With Rags, an agent is just a sophisticated search

310
00:10:20,280 --> 00:10:21,320
and copy machine.

311
00:10:21,320 --> 00:10:23,280
With ontologies, an agent becomes something

312
00:10:23,280 --> 00:10:25,320
that actually reasons about your business.

313
00:10:25,320 --> 00:10:27,240
This is where the thinking happens.

314
00:10:27,240 --> 00:10:29,120
The governance debt accumulates.

315
00:10:29,120 --> 00:10:30,800
Identity and reasoning only matter

316
00:10:30,800 --> 00:10:32,680
if you can actually manage what you've built.

317
00:10:32,680 --> 00:10:33,960
But for most organizations,

318
00:10:33,960 --> 00:10:35,800
that's where the architecture breaks down.

319
00:10:35,800 --> 00:10:37,160
Here is what typically happens.

320
00:10:37,160 --> 00:10:39,960
You've solved the chat problem, you have agents working.

321
00:10:39,960 --> 00:10:41,320
Identity is starting to make sense.

322
00:10:41,320 --> 00:10:43,520
But now you have three different teams

323
00:10:43,520 --> 00:10:44,760
that all want to build agents.

324
00:10:44,760 --> 00:10:46,000
They all need to connect to models.

325
00:10:46,000 --> 00:10:47,080
They all need policies.

326
00:10:47,080 --> 00:10:48,160
They all need logging.

327
00:10:48,160 --> 00:10:50,080
And each team does what makes sense to them.

328
00:10:50,080 --> 00:10:52,440
The support team goes directly to Azure OpenAI

329
00:10:52,440 --> 00:10:53,920
because they need fast answers

330
00:10:53,920 --> 00:10:55,920
and don't want to wait for central approval.

331
00:10:55,920 --> 00:10:59,320
So they set up their own keys and monitoring and cost tracking.

332
00:10:59,320 --> 00:11:01,520
The finance team uses co-pilot studio

333
00:11:01,520 --> 00:11:03,240
because that's what they already know

334
00:11:03,240 --> 00:11:05,080
and they use it to build out their own version

335
00:11:05,080 --> 00:11:06,640
of workflow automation.

336
00:11:06,640 --> 00:11:08,960
The data team experiments with open source models

337
00:11:08,960 --> 00:11:10,600
because they want more control

338
00:11:10,600 --> 00:11:12,880
which leads them to spin up their own separate infrastructure.

339
00:11:12,880 --> 00:11:13,880
Nobody coordinated.

340
00:11:13,880 --> 00:11:14,880
Nobody asked for permission.

341
00:11:14,880 --> 00:11:16,360
Nobody followed a standard.

342
00:11:16,360 --> 00:11:17,560
And suddenly you have a problem

343
00:11:17,560 --> 00:11:20,120
that identity and reasoning alone cannot fix

344
00:11:20,120 --> 00:11:22,320
because you have no idea what is actually running.

345
00:11:22,320 --> 00:11:23,400
This is shadow AI.

346
00:11:23,400 --> 00:11:24,200
It isn't malicious.

347
00:11:24,200 --> 00:11:25,720
It's just what happens when governance

348
00:11:25,720 --> 00:11:27,960
doesn't exist at the point where decisions get made.

349
00:11:27,960 --> 00:11:28,880
Teams need to ship.

350
00:11:28,880 --> 00:11:30,200
They need to move fast.

351
00:11:30,200 --> 00:11:31,560
Central processes are slow.

352
00:11:31,560 --> 00:11:32,720
So they go around them.

353
00:11:32,720 --> 00:11:34,400
They configure their own model access.

354
00:11:34,400 --> 00:11:36,360
They set up their own approval workflows

355
00:11:36,360 --> 00:11:37,880
or they just skip them entirely.

356
00:11:37,880 --> 00:11:39,360
They implement their own logging

357
00:11:39,360 --> 00:11:40,960
or they don't log anything at all.

358
00:11:40,960 --> 00:11:44,400
The result is that your organization now has multiple control planes.

359
00:11:44,400 --> 00:11:46,720
Or more accurately, it has no control plane.

360
00:11:46,720 --> 00:11:48,280
The support team's direct integration

361
00:11:48,280 --> 00:11:51,320
has no relationship to what finance is doing in co-pilot studio

362
00:11:51,320 --> 00:11:53,400
and the data team's open source setup

363
00:11:53,400 --> 00:11:55,600
has no connection to either of them.

364
00:11:55,600 --> 00:11:56,720
You have three different teams

365
00:11:56,720 --> 00:11:57,960
using three different models

366
00:11:57,960 --> 00:11:59,920
with three different sets of tools and policies

367
00:11:59,920 --> 00:12:03,080
and there is no governance layer that understands how they relate.

368
00:12:03,080 --> 00:12:05,520
If the data team's model needs to access data

369
00:12:05,520 --> 00:12:07,480
that the support team is currently monitoring,

370
00:12:07,480 --> 00:12:10,120
there is no central policy to say whether that's even allowed.

371
00:12:10,120 --> 00:12:12,560
You have compliance risks spread across three different systems

372
00:12:12,560 --> 00:12:13,720
with no way to manage it.

373
00:12:13,720 --> 00:12:15,360
And then there are the costs.

374
00:12:15,360 --> 00:12:17,560
When you don't have centralized visibility,

375
00:12:17,560 --> 00:12:18,600
you cannot measure.

376
00:12:18,600 --> 00:12:21,400
Each team is consuming tokens against their own budget

377
00:12:21,400 --> 00:12:24,240
or a general cloud spend that nobody is tracking anymore.

378
00:12:24,240 --> 00:12:26,440
The support agent might be calling an expensive model

379
00:12:26,440 --> 00:12:27,640
for every single question

380
00:12:27,640 --> 00:12:30,080
when it could use a cheaper one for routine queries

381
00:12:30,080 --> 00:12:32,560
and the finance agent might be running duplicate logic

382
00:12:32,560 --> 00:12:34,640
that the data team already built.

383
00:12:34,640 --> 00:12:36,440
The recruiting agent might be fetching data

384
00:12:36,440 --> 00:12:39,360
from three different sources when one unified source would work

385
00:12:39,360 --> 00:12:40,320
but nobody knows.

386
00:12:40,320 --> 00:12:42,880
Nobody can see, nobody can optimize.

387
00:12:42,880 --> 00:12:45,360
A poorly optimized agent using an expensive model

388
00:12:45,360 --> 00:12:48,960
with inefficient prompts can easily burn $50,000 a month

389
00:12:48,960 --> 00:12:52,120
and you won't find out until the cloud bill shows up.

390
00:12:52,120 --> 00:12:53,280
By the time you see the bill,

391
00:12:53,280 --> 00:12:54,920
the agent has been running for six months

392
00:12:54,920 --> 00:12:57,800
and the team that built it has already moved on to the next project

393
00:12:57,800 --> 00:13:00,200
when you finally asked them to optimize the costs,

394
00:13:00,200 --> 00:13:02,240
they just shrug because they were never given the tools

395
00:13:02,240 --> 00:13:04,160
to measure consumption in the first place.

396
00:13:04,160 --> 00:13:07,200
This is the financial consequence of distributed governance.

397
00:13:07,200 --> 00:13:08,640
You are paying for invisibility

398
00:13:08,640 --> 00:13:10,760
but the real problem is deeper than cost.

399
00:13:10,760 --> 00:13:11,800
It's compliance.

400
00:13:11,800 --> 00:13:16,000
Compliance requires a single, traceable answer to one question

401
00:13:16,000 --> 00:13:18,440
which systems accessed what data and when.

402
00:13:18,440 --> 00:13:20,800
With three independent teams running independent setups,

403
00:13:20,800 --> 00:13:22,920
that question becomes impossible to answer.

404
00:13:22,920 --> 00:13:24,840
The finance team's logging goes to one place

405
00:13:24,840 --> 00:13:26,520
the support teams go somewhere else.

406
00:13:26,520 --> 00:13:27,920
The data team isn't logging at all.

407
00:13:27,920 --> 00:13:31,280
If a regulator asks if an agent accessed a sensitive record,

408
00:13:31,280 --> 00:13:34,480
you don't have a way to search across all three systems at once

409
00:13:34,480 --> 00:13:36,640
and you don't have a standard for what access even means

410
00:13:36,640 --> 00:13:37,920
or how it gets recorded.

411
00:13:37,920 --> 00:13:39,440
You cannot pass an audit with that.

412
00:13:39,440 --> 00:13:41,360
This is what people mean by governance debt.

413
00:13:41,360 --> 00:13:43,800
You are borrowing against future control.

414
00:13:43,800 --> 00:13:45,560
You are accepting the chaos now in exchange

415
00:13:45,560 --> 00:13:47,080
for the ability to ship fast.

416
00:13:47,080 --> 00:13:49,320
You're trading short term speed for the long term certainty

417
00:13:49,320 --> 00:13:51,080
that you won't be able to govern what you've built.

418
00:13:51,080 --> 00:13:54,080
An unlike financial debt, governance debt compounds.

419
00:13:54,080 --> 00:13:56,080
Every new team that stands up their own system

420
00:13:56,080 --> 00:13:57,400
makes the problem worse.

421
00:13:57,400 --> 00:13:59,400
And every month, you delay centralizing governance

422
00:13:59,400 --> 00:14:01,880
makes the eventual cleanup work much larger.

423
00:14:01,880 --> 00:14:04,000
Eventually, you hit a wall, a compliance failure,

424
00:14:04,000 --> 00:14:07,520
a security incident, a cost overrun, you cannot explain.

425
00:14:07,520 --> 00:14:09,000
And at that point, you have to go back

426
00:14:09,000 --> 00:14:11,120
and rebuild everything you should have built from the start.

427
00:14:11,120 --> 00:14:13,160
The way out starts with a single question,

428
00:14:13,160 --> 00:14:14,760
what if governance happened first?

429
00:14:14,760 --> 00:14:17,720
What a landing zone actually is.

430
00:14:17,720 --> 00:14:19,760
To understand the solution, we need to reframe

431
00:14:19,760 --> 00:14:21,440
what a landing zone actually means.

432
00:14:21,440 --> 00:14:25,080
Most people hear the term and think infrastructure, networking,

433
00:14:25,080 --> 00:14:26,960
cloud subscriptions, the plumbing.

434
00:14:26,960 --> 00:14:29,400
And yes, a landing zone includes those things.

435
00:14:29,400 --> 00:14:31,160
But if that's what you think a landing zone is.

436
00:14:31,160 --> 00:14:32,600
You're looking at it backwards.

437
00:14:32,600 --> 00:14:34,640
A landing zone isn't primarily about networks.

438
00:14:34,640 --> 00:14:37,320
It's about how AI operates in your organization.

439
00:14:37,320 --> 00:14:39,400
It's the operational model that prevents the problems

440
00:14:39,400 --> 00:14:40,480
we just talked about.

441
00:14:40,480 --> 00:14:42,480
The identity collapse, the reasoning gaps,

442
00:14:42,480 --> 00:14:43,680
and the governance debt.

443
00:14:43,680 --> 00:14:45,320
It is the framework that says,

444
00:14:45,320 --> 00:14:46,680
this is how we build agents here.

445
00:14:46,680 --> 00:14:47,880
This is who controls access.

446
00:14:47,880 --> 00:14:48,920
This is what gets logged.

447
00:14:48,920 --> 00:14:51,840
A landing zone is the moment your organization stops treating

448
00:14:51,840 --> 00:14:54,160
agents as one of projects and starts treating them

449
00:14:54,160 --> 00:14:55,640
as a managed capability.

450
00:14:55,640 --> 00:14:58,480
It is the shift from teams building their own fragmented

451
00:14:58,480 --> 00:15:01,400
infrastructure to the organization having a standard way agents

452
00:15:01,400 --> 00:15:03,360
get built, deployed, and retired.

453
00:15:03,360 --> 00:15:06,760
That shift matters because standardization is how you scale safely.

454
00:15:06,760 --> 00:15:08,000
That standard has three layers.

455
00:15:08,000 --> 00:15:09,480
The first is the identity plane.

456
00:15:09,480 --> 00:15:11,720
This is where you solve the accountability problem.

457
00:15:11,720 --> 00:15:13,760
Every agent that runs in your organization

458
00:15:13,760 --> 00:15:16,840
gets a traceable identity, which is an agent ID registered

459
00:15:16,840 --> 00:15:18,880
in your directory just like a user account,

460
00:15:18,880 --> 00:15:20,880
but built for autonomous systems.

461
00:15:20,880 --> 00:15:22,440
That identity carries permissions

462
00:15:22,440 --> 00:15:24,800
scoped to exactly what the agent needs,

463
00:15:24,800 --> 00:15:26,800
including which data it can access,

464
00:15:26,800 --> 00:15:29,520
and which specific models it is allowed to use.

465
00:15:29,520 --> 00:15:30,960
When that agent acts,

466
00:15:30,960 --> 00:15:32,840
the action is attributed to the agent itself

467
00:15:32,840 --> 00:15:34,880
rather than being borrowed from a human account,

468
00:15:34,880 --> 00:15:36,760
so the audit trail stays clean.

469
00:15:36,760 --> 00:15:38,720
The second layer is the governance hub.

470
00:15:38,720 --> 00:15:40,000
This is centralization.

471
00:15:40,000 --> 00:15:42,080
It is the single place where all approved models live,

472
00:15:42,080 --> 00:15:44,760
all tools are catalogued and all approvals flow through.

473
00:15:44,760 --> 00:15:46,000
When a team wants to build an agent,

474
00:15:46,000 --> 00:15:48,840
they don't go configure Azure OpenAI directly.

475
00:15:48,840 --> 00:15:50,360
They request access through the hub.

476
00:15:50,360 --> 00:15:53,080
When they want to add a tool, they don't integrate it independently.

477
00:15:53,080 --> 00:15:55,200
They submitted to the catalog where it is reviewed

478
00:15:55,200 --> 00:15:56,600
and approved for everyone.

479
00:15:56,600 --> 00:15:59,920
If policies change and you need stricter data, residency rules,

480
00:15:59,920 --> 00:16:01,880
you update the policy once in the hub,

481
00:16:01,880 --> 00:16:03,120
and it applies everywhere.

482
00:16:03,120 --> 00:16:05,000
One place, one source of truth.

483
00:16:05,000 --> 00:16:07,080
Without the hub, you have distributed control

484
00:16:07,080 --> 00:16:08,160
that looks like governance,

485
00:16:08,160 --> 00:16:10,240
but is really just decentralized chaos.

486
00:16:10,240 --> 00:16:12,080
With the hub, you have the control plane

487
00:16:12,080 --> 00:16:15,600
that the support team, finance team, and data team all operate under.

488
00:16:15,600 --> 00:16:17,880
They aren't competing for access to different models

489
00:16:17,880 --> 00:16:19,320
or defining their own rules.

490
00:16:19,320 --> 00:16:21,240
They are all operating within the same framework

491
00:16:21,240 --> 00:16:22,560
with the same visibility.

492
00:16:22,560 --> 00:16:24,480
The third layer is the data plane.

493
00:16:24,480 --> 00:16:26,040
This is where ontologies live.

494
00:16:26,040 --> 00:16:27,880
It is the structured model of your business,

495
00:16:27,880 --> 00:16:29,440
including the entities and the rules

496
00:16:29,440 --> 00:16:31,480
that govern how those entities connect.

497
00:16:31,480 --> 00:16:33,400
It isn't stored in three different places

498
00:16:33,400 --> 00:16:34,560
by three different teams.

499
00:16:34,560 --> 00:16:36,160
It is centralized and versioned.

500
00:16:36,160 --> 00:16:37,400
When a new agent gets built,

501
00:16:37,400 --> 00:16:39,360
it doesn't have to figure out what a customer is

502
00:16:39,360 --> 00:16:40,880
or how customers relate to orders

503
00:16:40,880 --> 00:16:42,840
because those definitions already exist.

504
00:16:42,840 --> 00:16:44,480
The agent inherits that understanding

505
00:16:44,480 --> 00:16:46,080
it reasons over the ontology

506
00:16:46,080 --> 00:16:48,920
instead of trying to make sense of disconnected data sources.

507
00:16:48,920 --> 00:16:50,880
That is how you scale consistency

508
00:16:50,880 --> 00:16:52,360
across dozens of agents

509
00:16:52,360 --> 00:16:54,880
without them constantly contradicting each other.

510
00:16:54,880 --> 00:16:56,520
These three layers work together.

511
00:16:56,520 --> 00:16:58,240
Identity makes agents accountable.

512
00:16:58,240 --> 00:17:00,520
The governance hub makes decisions centralized.

513
00:17:00,520 --> 00:17:02,520
The ontology makes reasoning consistent.

514
00:17:02,520 --> 00:17:04,920
Individually, each one solves a structural problem.

515
00:17:04,920 --> 00:17:06,680
But together, they form the architecture

516
00:17:06,680 --> 00:17:08,520
that lets you deploy hundreds of agents

517
00:17:08,520 --> 00:17:09,960
without creating chaos.

518
00:17:09,960 --> 00:17:11,120
And this is the key shift.

519
00:17:11,120 --> 00:17:12,720
You aren't building infrastructure

520
00:17:12,720 --> 00:17:14,800
to support isolated chatbots anymore.

521
00:17:14,800 --> 00:17:17,280
You're building a platform and AI operating system.

522
00:17:17,280 --> 00:17:19,680
It's a foundation that any team can build on

523
00:17:19,680 --> 00:17:21,400
without having to recreate governance

524
00:17:21,400 --> 00:17:23,320
or identity every single time.

525
00:17:23,320 --> 00:17:24,840
That is what a landing zone actually is.

526
00:17:24,840 --> 00:17:26,440
It isn't just plumbing or a checklist.

527
00:17:26,440 --> 00:17:28,080
It is the moment your organization moves

528
00:17:28,080 --> 00:17:29,600
from experimenting with AI

529
00:17:29,600 --> 00:17:32,280
to operating AI responsibly at scale.

530
00:17:32,280 --> 00:17:33,800
The model gateway pattern,

531
00:17:33,800 --> 00:17:36,600
the identity plane and governance hub are foundational

532
00:17:36,600 --> 00:17:38,920
but they only work if you actually control

533
00:17:38,920 --> 00:17:42,160
how agents access models and most organizations don't.

534
00:17:42,160 --> 00:17:43,560
What typically happens is this.

535
00:17:43,560 --> 00:17:46,360
An agent needs to call Azure OpenAI or Anthropic.

536
00:17:46,360 --> 00:17:48,080
Someone generates an API key.

537
00:17:48,080 --> 00:17:49,480
They embed it in the agent's code.

538
00:17:49,480 --> 00:17:51,160
The agent calls the model directly.

539
00:17:51,160 --> 00:17:52,400
No intermediary, no filtering,

540
00:17:52,400 --> 00:17:54,480
no policy check, just agent to model.

541
00:17:54,480 --> 00:17:55,520
This feels efficient.

542
00:17:55,520 --> 00:17:57,640
The agent gets what it needs without extra layers.

543
00:17:57,640 --> 00:18:00,120
There's no latency penalty from rooting through a gateway.

544
00:18:00,120 --> 00:18:01,640
There's no extra service to operate.

545
00:18:01,640 --> 00:18:02,760
The team ships fast.

546
00:18:02,760 --> 00:18:03,880
But you've lost control.

547
00:18:03,880 --> 00:18:07,560
When every agent connects directly to models,

548
00:18:07,560 --> 00:18:09,800
you have no central way to enforce policy.

549
00:18:09,800 --> 00:18:11,760
One agent might be calling an expensive model

550
00:18:11,760 --> 00:18:13,640
when a cheaper one would work.

551
00:18:13,640 --> 00:18:15,520
Another might be sending data to a region

552
00:18:15,520 --> 00:18:17,560
that violates your compliance requirements.

553
00:18:17,560 --> 00:18:19,920
A third might be consuming tokens at a rate

554
00:18:19,920 --> 00:18:22,280
that nobody noticed until the bill arrived.

555
00:18:22,280 --> 00:18:24,080
Each agent is making independent decisions

556
00:18:24,080 --> 00:18:26,360
about which model to use and what data to expose.

557
00:18:26,360 --> 00:18:27,400
There's no enforcement.

558
00:18:27,400 --> 00:18:28,400
There's no visibility.

559
00:18:28,400 --> 00:18:29,640
There's no way to say no.

560
00:18:29,640 --> 00:18:31,160
This is where a model gateway comes in.

561
00:18:31,160 --> 00:18:34,160
A gateway is a single entry point for all AI traffic.

562
00:18:34,160 --> 00:18:35,720
Every agent that needs to call a model

563
00:18:35,720 --> 00:18:37,080
doesn't call it directly.

564
00:18:37,080 --> 00:18:38,440
The agent calls the gateway.

565
00:18:38,440 --> 00:18:40,320
The gateway then decides which model to root

566
00:18:40,320 --> 00:18:42,120
the request to based on policy.

567
00:18:42,120 --> 00:18:44,600
Think of it like a border checkpoint for AI traffic.

568
00:18:44,600 --> 00:18:45,960
What does the gateway do?

569
00:18:45,960 --> 00:18:47,920
At minimum, it handles authentication.

570
00:18:47,920 --> 00:18:50,080
Every request comes with credentials identifying

571
00:18:50,080 --> 00:18:51,920
which agent is making the request.

572
00:18:51,920 --> 00:18:53,920
The gateway validates those credentials

573
00:18:53,920 --> 00:18:55,360
against the agent's permissions.

574
00:18:55,360 --> 00:18:57,920
This is the identity layer you built in the landing zone,

575
00:18:57,920 --> 00:18:59,360
now enforced at runtime.

576
00:18:59,360 --> 00:19:00,760
When agent can't call a model,

577
00:19:00,760 --> 00:19:02,760
it doesn't have permission to use.

578
00:19:02,760 --> 00:19:04,680
But authentication is just the entry point.

579
00:19:04,680 --> 00:19:05,880
Once you have that checkpoint,

580
00:19:05,880 --> 00:19:08,040
you can add policy enforcement on top.

581
00:19:08,040 --> 00:19:09,520
Here's where it gets powerful.

582
00:19:09,520 --> 00:19:11,880
Let's say your organization has a policy.

583
00:19:11,880 --> 00:19:14,640
No customer data leaves the US region.

584
00:19:14,640 --> 00:19:17,480
An agent is built to process customer support tickets.

585
00:19:17,480 --> 00:19:19,880
It's configured to use a model that runs in Europe.

586
00:19:19,880 --> 00:19:22,560
Without a gateway, that agent calls the European model

587
00:19:22,560 --> 00:19:25,640
directly and sends customer data across the ocean.

588
00:19:25,640 --> 00:19:28,040
You find out six months later when an auditor asks

589
00:19:28,040 --> 00:19:29,560
where your customer data goes

590
00:19:29,560 --> 00:19:32,080
where the gateway that request hits the checkpoint.

591
00:19:32,080 --> 00:19:33,920
The gateway evaluates the policy.

592
00:19:33,920 --> 00:19:36,200
This agent is asking to call a European model.

593
00:19:36,200 --> 00:19:38,680
This policy says customer data can't leave the US.

594
00:19:38,680 --> 00:19:39,840
The gateway blocks the request

595
00:19:39,840 --> 00:19:42,240
and re-route it to a US-based model instead.

596
00:19:42,240 --> 00:19:44,120
Policy enforced, no data violation,

597
00:19:44,120 --> 00:19:45,440
no surprise audit failure.

598
00:19:45,440 --> 00:19:48,080
The gateway also handles cost attribution.

599
00:19:48,080 --> 00:19:50,360
Every model call goes through a single point.

600
00:19:50,360 --> 00:19:52,440
You can track which agent called which model,

601
00:19:52,440 --> 00:19:55,360
how many tokens it consumed, and what the cost was.

602
00:19:55,360 --> 00:19:57,000
Instead of discovering in your cloud bill

603
00:19:57,000 --> 00:19:59,560
that you're spending $50,000 a month on model calls,

604
00:19:59,560 --> 00:20:00,880
you have real-time visibility.

605
00:20:00,880 --> 00:20:02,560
You can see that one agent is responsible

606
00:20:02,560 --> 00:20:04,080
for 90% of the spend.

607
00:20:04,080 --> 00:20:04,920
You can optimize it.

608
00:20:04,920 --> 00:20:06,560
You can set budgets and alerts.

609
00:20:06,560 --> 00:20:08,720
You can route expensive queries to cheaper models

610
00:20:08,720 --> 00:20:11,360
or cache results to reduce redundant calls.

611
00:20:11,360 --> 00:20:13,440
Without a gateway, that visibility doesn't exist.

612
00:20:13,440 --> 00:20:16,040
You have 10 agents calling 10 different models

613
00:20:16,040 --> 00:20:17,960
with no way to aggregate the picture.

614
00:20:17,960 --> 00:20:19,560
The gateway also enables logging.

615
00:20:19,560 --> 00:20:22,520
Every call gets recorded, which agent called the model when?

616
00:20:22,520 --> 00:20:23,360
What was the input?

617
00:20:23,360 --> 00:20:24,280
What was the output?

618
00:20:24,280 --> 00:20:26,200
This creates an audit trail for compliance.

619
00:20:26,200 --> 00:20:28,040
It creates evidence if something goes wrong.

620
00:20:28,040 --> 00:20:29,640
It creates the visibility you need

621
00:20:29,640 --> 00:20:31,800
to understand what your agents are actually doing.

622
00:20:31,800 --> 00:20:33,320
And here's the scaling inside.

623
00:20:33,320 --> 00:20:35,400
With a gateway, you govern it the gateway,

624
00:20:35,400 --> 00:20:36,840
not at the application level.

625
00:20:36,840 --> 00:20:38,480
You don't have to reconfigure every agent

626
00:20:38,480 --> 00:20:39,800
every time policy changes.

627
00:20:39,800 --> 00:20:42,280
You update the policy once in one place.

628
00:20:42,280 --> 00:20:43,800
The next request hits the gateway

629
00:20:43,800 --> 00:20:47,160
and the new policy applies, one change, everywhere.

630
00:20:47,160 --> 00:20:48,600
This is why the gateway is essential.

631
00:20:48,600 --> 00:20:50,720
It's the enforcement layer that makes your governance hub

632
00:20:50,720 --> 00:20:52,760
actually work, the ontology layer.

633
00:20:52,760 --> 00:20:54,480
The gateway solves the control problem,

634
00:20:54,480 --> 00:20:55,920
but there's still one layer missing.

635
00:20:55,920 --> 00:20:58,080
Control over what agents access isn't the same

636
00:20:58,080 --> 00:20:59,920
as control over what agents understand.

637
00:20:59,920 --> 00:21:01,520
Let's go back to the data itself.

638
00:21:01,520 --> 00:21:03,680
Most organizations store their operational knowledge

639
00:21:03,680 --> 00:21:04,600
in two places.

640
00:21:04,600 --> 00:21:06,960
They have a data warehouse, tables and columns,

641
00:21:06,960 --> 00:21:09,280
schemas and joins, the relational model

642
00:21:09,280 --> 00:21:12,440
that's been the backbone of business computing for 40 years.

643
00:21:12,440 --> 00:21:15,600
And they have documents, PDFs, wikis, emails, contracts,

644
00:21:15,600 --> 00:21:17,560
the unstructured stuff that people write.

645
00:21:17,560 --> 00:21:19,760
When an agent needs to answer a business question,

646
00:21:19,760 --> 00:21:22,120
it usually has access to one or the other.

647
00:21:22,120 --> 00:21:24,280
Either it's running a query against your warehouse

648
00:21:24,280 --> 00:21:25,720
or it's searching documents

649
00:21:25,720 --> 00:21:27,560
or it's doing both independently

650
00:21:27,560 --> 00:21:29,480
and trying to reconcile the results.

651
00:21:29,480 --> 00:21:31,240
But neither of those is a business model.

652
00:21:31,240 --> 00:21:33,480
Neither one captures what your business actually is.

653
00:21:33,480 --> 00:21:34,840
Your business isn't a schema.

654
00:21:34,840 --> 00:21:36,760
Your business is a set of relationships.

655
00:21:36,760 --> 00:21:37,720
You have customers.

656
00:21:37,720 --> 00:21:40,480
Those customers place orders, orders contain products.

657
00:21:40,480 --> 00:21:41,920
Products come from suppliers.

658
00:21:41,920 --> 00:21:43,360
Suppliers have lead times.

659
00:21:43,360 --> 00:21:45,040
Customers have credit limits.

660
00:21:45,040 --> 00:21:48,320
Orders that exceed a customer's credit limit need approval.

661
00:21:48,320 --> 00:21:50,320
Some suppliers are marked as high risk,

662
00:21:50,320 --> 00:21:52,400
which affects how their orders are handled.

663
00:21:52,400 --> 00:21:53,880
Orders that go to certain regions

664
00:21:53,880 --> 00:21:55,880
trigger different compliance rules.

665
00:21:55,880 --> 00:21:59,280
A delayed shipment to a VIP customer is a priority incident.

666
00:21:59,280 --> 00:22:02,120
A routine delay to a standard customer is normal.

667
00:22:02,120 --> 00:22:03,400
None of that lives in a table.

668
00:22:03,400 --> 00:22:04,640
None of that is a document.

669
00:22:04,640 --> 00:22:05,880
All of it is business logic.

670
00:22:05,880 --> 00:22:07,040
All of it is meaning.

671
00:22:07,040 --> 00:22:08,960
An ontology is how you encode that meaning.

672
00:22:08,960 --> 00:22:10,680
Think of it as a map of your business.

673
00:22:10,680 --> 00:22:13,480
Not a map of your data, a map of your actual business.

674
00:22:13,480 --> 00:22:14,280
Entities.

675
00:22:14,280 --> 00:22:14,920
Customer.

676
00:22:14,920 --> 00:22:15,520
Order.

677
00:22:15,520 --> 00:22:16,160
Product.

678
00:22:16,160 --> 00:22:17,040
Supplier.

679
00:22:17,040 --> 00:22:17,680
Region.

680
00:22:17,680 --> 00:22:18,480
Acid.

681
00:22:18,480 --> 00:22:20,120
Each one exists as a concept.

682
00:22:20,120 --> 00:22:21,240
Each one has properties.

683
00:22:21,240 --> 00:22:24,120
A customer has a name, an account number, a credit limit,

684
00:22:24,120 --> 00:22:25,360
and a payment history.

685
00:22:25,360 --> 00:22:26,920
An order has an order number, a date,

686
00:22:26,920 --> 00:22:28,680
a customer, products, and a status.

687
00:22:28,680 --> 00:22:30,400
But here's what makes an ontology different

688
00:22:30,400 --> 00:22:31,400
from a spreadsheet.

689
00:22:31,400 --> 00:22:32,760
Relationships matter.

690
00:22:32,760 --> 00:22:35,480
A customer places orders, an order contains products.

691
00:22:35,480 --> 00:22:36,960
A product comes from a supplier.

692
00:22:36,960 --> 00:22:40,040
A supplier operates in regions and asks it belongs to a location.

693
00:22:40,040 --> 00:22:41,240
An order goes to a location.

694
00:22:41,240 --> 00:22:42,440
A location has inventory.

695
00:22:42,440 --> 00:22:44,440
Inventory of a product comes from a supplier.

696
00:22:44,440 --> 00:22:46,720
When you map these relationships explicitly,

697
00:22:46,720 --> 00:22:47,760
something changes.

698
00:22:47,760 --> 00:22:50,280
An agent doesn't have to guess at meaning anymore.

699
00:22:50,280 --> 00:22:52,680
It doesn't have to infer that a customer is high value

700
00:22:52,680 --> 00:22:54,680
by looking at aggregate spending numbers.

701
00:22:54,680 --> 00:22:56,160
The ontology can tell it directly.

702
00:22:56,160 --> 00:22:58,880
The ontology can say this customer has a VIP flag.

703
00:22:58,880 --> 00:23:01,920
When a VIP customer's order is delayed, this matters more.

704
00:23:01,920 --> 00:23:03,720
The agent reasons across the relationships

705
00:23:03,720 --> 00:23:05,360
instead of guessing at patterns.

706
00:23:05,360 --> 00:23:08,200
This matters because business logic lives in relationships.

707
00:23:08,200 --> 00:23:09,800
A standard query of your warehouse

708
00:23:09,800 --> 00:23:12,000
might tell you that a customer hasn't paid their invoice

709
00:23:12,000 --> 00:23:13,240
in 60 days.

710
00:23:13,240 --> 00:23:16,440
An ontology can tell you not just that, but the context.

711
00:23:16,440 --> 00:23:17,920
What region are they in?

712
00:23:17,920 --> 00:23:19,640
What's the payment term for that region?

713
00:23:19,640 --> 00:23:21,360
Are they a long-standing customer?

714
00:23:21,360 --> 00:23:22,720
With a history of late payment

715
00:23:22,720 --> 00:23:25,240
and what was their last interaction with support?

716
00:23:25,240 --> 00:23:27,200
All of that is implicit in the relationships.

717
00:23:27,200 --> 00:23:30,520
The agent doesn't have to construct it from disconnected facts.

718
00:23:30,520 --> 00:23:32,160
And when the agent needs to make a decision,

719
00:23:32,160 --> 00:23:34,200
the ontology acts as a constrained layer.

720
00:23:34,200 --> 00:23:37,040
In a normal data warehouse, an agent can pull whatever it wants

721
00:23:37,040 --> 00:23:38,680
and do whatever it wants with it.

722
00:23:38,680 --> 00:23:41,760
In an ontology, the agent understands what paths are valid.

723
00:23:41,760 --> 00:23:44,240
A marketing agent can see customer spend patterns.

724
00:23:44,240 --> 00:23:46,120
A security agent can see access logs,

725
00:23:46,120 --> 00:23:47,840
but neither one should be able to see data

726
00:23:47,840 --> 00:23:49,160
that's outside their scope.

727
00:23:49,160 --> 00:23:51,120
The ontology enforces that implicitly

728
00:23:51,120 --> 00:23:52,880
through the relationships it defines.

729
00:23:52,880 --> 00:23:54,680
If a path doesn't exist in the ontology,

730
00:23:54,680 --> 00:23:56,160
the agent can't traverse it.

731
00:23:56,160 --> 00:23:59,440
This is how hallucination gets prevented at a structural level.

732
00:23:59,440 --> 00:24:01,760
When an agent is reasoning over an ontology,

733
00:24:01,760 --> 00:24:03,480
it's not pattern matching against documents

734
00:24:03,480 --> 00:24:05,320
or guessing a table relationships.

735
00:24:05,320 --> 00:24:08,040
It's traversing an explicit model of your business.

736
00:24:08,040 --> 00:24:10,560
Every relationship has been intentionally defined.

737
00:24:10,560 --> 00:24:12,280
Every entity has been carefully modeled.

738
00:24:12,280 --> 00:24:13,880
The agent doesn't have creative freedom

739
00:24:13,880 --> 00:24:16,160
to interpret data in ways the business doesn't intend.

740
00:24:16,160 --> 00:24:17,520
It operates within the constraints

741
00:24:17,520 --> 00:24:19,160
of how you've defined your business.

742
00:24:19,160 --> 00:24:20,600
And here's the scaling insight.

743
00:24:20,600 --> 00:24:22,760
When you add a new agent to your organization,

744
00:24:22,760 --> 00:24:24,600
it doesn't have to rebuild the ontology.

745
00:24:24,600 --> 00:24:25,560
It inherits it.

746
00:24:25,560 --> 00:24:27,840
The same customer entity, the same VIP flag

747
00:24:27,840 --> 00:24:29,720
and the same relationships between order

748
00:24:29,720 --> 00:24:30,960
and product and supplier.

749
00:24:30,960 --> 00:24:32,960
10 agents, 50 agents, 100 agents.

750
00:24:32,960 --> 00:24:35,160
They all reason over the same business model.

751
00:24:35,160 --> 00:24:37,640
They all make decisions using the same definitions.

752
00:24:37,640 --> 00:24:39,680
Consistency is structural, not achieved

753
00:24:39,680 --> 00:24:41,400
through training or documentation

754
00:24:41,400 --> 00:24:44,040
or hoping everyone remembers how things work.

755
00:24:44,040 --> 00:24:48,000
The shift from chat to fabric, identity, reasoning, governance,

756
00:24:48,000 --> 00:24:50,480
the gateway and the ontology.

757
00:24:50,480 --> 00:24:52,920
Up until now, we've looked at these as individual layers,

758
00:24:52,920 --> 00:24:54,880
but these pieces only matter when they combine,

759
00:24:54,880 --> 00:24:57,280
when they merge, they change the fundamental nature

760
00:24:57,280 --> 00:24:58,800
of what an agent actually is.

761
00:24:58,800 --> 00:25:00,480
That shift is from chat to fabric.

762
00:25:00,480 --> 00:25:02,200
Most people think chat is the full picture.

763
00:25:02,200 --> 00:25:03,280
You ask a question.

764
00:25:03,280 --> 00:25:05,280
The agent processes it, you get an answer,

765
00:25:05,280 --> 00:25:07,680
then you repeat the process, but in reality,

766
00:25:07,680 --> 00:25:09,800
each of those conversations is isolated.

767
00:25:09,800 --> 00:25:12,480
The agent has no memory of what you asked two minutes ago,

768
00:25:12,480 --> 00:25:13,880
no context carries forward.

769
00:25:13,880 --> 00:25:15,600
Every prompt starts fresh.

770
00:25:15,600 --> 00:25:18,400
The agent is stateless, which means it has no understanding

771
00:25:18,400 --> 00:25:19,760
of your broader business landscape.

772
00:25:19,760 --> 00:25:20,960
It doesn't know your history.

773
00:25:20,960 --> 00:25:22,920
It doesn't see the relationship between a decision made

774
00:25:22,920 --> 00:25:24,960
yesterday and a decision being made today.

775
00:25:24,960 --> 00:25:26,720
This works fine for simple scenarios.

776
00:25:26,720 --> 00:25:27,600
What's the weather?

777
00:25:27,600 --> 00:25:29,440
What's the office policy on remote work?

778
00:25:29,440 --> 00:25:30,600
What are the products specs?

779
00:25:30,600 --> 00:25:32,240
One question, one answer done.

780
00:25:32,240 --> 00:25:33,720
But business doesn't work that way.

781
00:25:33,720 --> 00:25:35,400
Real business is about understanding

782
00:25:35,400 --> 00:25:36,640
chains of cause and effect.

783
00:25:36,640 --> 00:25:37,920
Why did revenue drop?

784
00:25:37,920 --> 00:25:40,720
Because a major customer didn't place their monthly order.

785
00:25:40,720 --> 00:25:42,160
Why didn't they place that order?

786
00:25:42,160 --> 00:25:43,600
Because their inventory is full.

787
00:25:43,600 --> 00:25:44,720
Why is their inventory full?

788
00:25:44,720 --> 00:25:47,600
Because a supplier delay backed up their entire supply chain.

789
00:25:47,600 --> 00:25:48,960
Why was there a delay?

790
00:25:48,960 --> 00:25:52,200
Because that supplier operates in a region hit by new tariffs.

791
00:25:52,200 --> 00:25:53,760
How does that affect our forecasting?

792
00:25:53,760 --> 00:25:55,920
What do we tell Wall Street about Q3?

793
00:25:55,920 --> 00:25:57,800
Each of those questions requires understanding

794
00:25:57,800 --> 00:25:58,680
the previous answer.

795
00:25:58,680 --> 00:26:00,720
They require context that accumulates over time.

796
00:26:00,720 --> 00:26:01,800
A chat agent can't do that.

797
00:26:01,800 --> 00:26:05,040
It can answer what is our revenue in total isolation.

798
00:26:05,040 --> 00:26:06,960
But it can't answer why did it drop?

799
00:26:06,960 --> 00:26:08,360
And what should we do about it?

800
00:26:08,360 --> 00:26:11,480
Answering that requires holding multiple pieces of context

801
00:26:11,480 --> 00:26:13,080
in relationship to each other.

802
00:26:13,080 --> 00:26:14,240
All at once.

803
00:26:14,240 --> 00:26:16,280
This is the difference between stateless and stateful.

804
00:26:16,280 --> 00:26:17,680
Chat is stateless.

805
00:26:17,680 --> 00:26:19,840
Fabric is stateful.

806
00:26:19,840 --> 00:26:21,680
With Fabric, the agent understands

807
00:26:21,680 --> 00:26:23,840
the business landscape surrounding the question.

808
00:26:23,840 --> 00:26:26,240
It knows this specific customer is high value.

809
00:26:26,240 --> 00:26:28,120
It knows their inventory patterns predict

810
00:26:28,120 --> 00:26:29,480
when they will order next.

811
00:26:29,480 --> 00:26:32,200
It knows the supplier risk in their region has changed.

812
00:26:32,200 --> 00:26:34,960
It understands the full chain that led to the revenue drop.

813
00:26:34,960 --> 00:26:37,040
It isn't answering questions in a vacuum.

814
00:26:37,040 --> 00:26:40,040
It is reasoning across a continuous model of your business.

815
00:26:40,040 --> 00:26:41,800
The operational consequence is profound.

816
00:26:41,800 --> 00:26:43,360
With chat agents are reactive.

817
00:26:43,360 --> 00:26:45,200
You point them at a question and they answer it.

818
00:26:45,200 --> 00:26:47,600
They are useful for someone looking for a quick response.

819
00:26:47,600 --> 00:26:48,880
But they can't be proactive.

820
00:26:48,880 --> 00:26:51,800
They can't detect patterns before those patterns become problems.

821
00:26:51,800 --> 00:26:54,160
They can't understand causation across multiple systems.

822
00:26:54,160 --> 00:26:57,000
They can't make the connection between a supplier delay in Asia

823
00:26:57,000 --> 00:26:59,120
and a cash flow problem in North America.

824
00:26:59,120 --> 00:27:01,120
With Fabric, agents become orchestrators.

825
00:27:01,120 --> 00:27:02,440
They see the full picture.

826
00:27:02,440 --> 00:27:04,520
They can monitor multiple signals simultaneously.

827
00:27:04,520 --> 00:27:06,280
They can detect when something is about to go wrong

828
00:27:06,280 --> 00:27:07,760
before it actually happens.

829
00:27:07,760 --> 00:27:10,360
They can recommend actions based on the complete chain.

830
00:27:10,360 --> 00:27:11,800
They aren't just responding to questions.

831
00:27:11,800 --> 00:27:14,440
They are actively maintaining the health of the system.

832
00:27:14,440 --> 00:27:17,720
Consider what happens when a major customer hasn't ordered in two months.

833
00:27:17,720 --> 00:27:20,560
A chat agent searches for recent orders, finds none,

834
00:27:20,560 --> 00:27:22,840
and reports no recent orders.

835
00:27:22,840 --> 00:27:23,760
That's the answer.

836
00:27:23,760 --> 00:27:25,080
Information delivered.

837
00:27:25,080 --> 00:27:28,000
A Fabric agent understands this customer's historical pattern.

838
00:27:28,000 --> 00:27:30,240
It detects they are three weeks past their normal cycle.

839
00:27:30,240 --> 00:27:33,160
It checks their inventory levels and correlates that with supplier data

840
00:27:33,160 --> 00:27:34,400
to see if a backup is coming.

841
00:27:34,400 --> 00:27:37,320
It recognizes this customer is a VIP and that delays

842
00:27:37,320 --> 00:27:39,440
now create cascading problems downstream.

843
00:27:39,440 --> 00:27:41,720
Then it escalates this as a priority incident

844
00:27:41,720 --> 00:27:44,160
and recommends outreach with a targeted offer

845
00:27:44,160 --> 00:27:45,720
before the problem becomes critical.

846
00:27:45,720 --> 00:27:49,760
Same agent, different architecture, one saw a fact, one understood a situation.

847
00:27:49,760 --> 00:27:51,640
This is where the infrastructure you've built

848
00:27:51,640 --> 00:27:53,960
finally becomes more than technical scaffolding.

849
00:27:53,960 --> 00:27:56,360
This is where you stop managing a collection of chatbots

850
00:27:56,360 --> 00:27:58,480
and start managing an intelligent system

851
00:27:58,480 --> 00:28:00,760
where identity isn't just compliance theatre,

852
00:28:00,760 --> 00:28:02,640
where the ontology isn't just a database,

853
00:28:02,640 --> 00:28:05,600
where governance isn't just control for the sake of control.

854
00:28:05,600 --> 00:28:07,960
The organization that moves from chat to fabric

855
00:28:07,960 --> 00:28:10,760
is fundamentally changing how intelligence operates.

856
00:28:10,760 --> 00:28:13,480
From isolated questions to coordinated reasoning,

857
00:28:13,480 --> 00:28:15,400
from waiting for problems to detecting them,

858
00:28:15,400 --> 00:28:17,040
from a feature to a platform,

859
00:28:17,040 --> 00:28:19,520
agent 365 as the control plane,

860
00:28:19,520 --> 00:28:21,720
but building identity, reasoning, governance,

861
00:28:21,720 --> 00:28:23,720
and a gateway separately doesn't scale.

862
00:28:23,720 --> 00:28:25,440
These are infrastructure components.

863
00:28:25,440 --> 00:28:26,840
They still need to be orchestrated.

864
00:28:26,840 --> 00:28:29,040
They still need to be operated as a single system

865
00:28:29,040 --> 00:28:31,400
without a unified layer that understands all four.

866
00:28:31,400 --> 00:28:33,240
You are left with plumbing that doesn't connect.

867
00:28:33,240 --> 00:28:36,200
You need a layer that can inventory agents in force identity,

868
00:28:36,200 --> 00:28:39,080
apply rules and surface observability all at once.

869
00:28:39,080 --> 00:28:40,960
This is what agent 365 does.

870
00:28:40,960 --> 00:28:44,480
Agent 365 is Microsoft's answer to a practical problem.

871
00:28:44,480 --> 00:28:46,680
How do you actually operate an agent architecture

872
00:28:46,680 --> 00:28:48,360
across an entire organization?

873
00:28:48,360 --> 00:28:51,400
It is the control plane that ties together everything we've discussed.

874
00:28:51,400 --> 00:28:54,200
It's a single place where you can see every agent that exists.

875
00:28:54,200 --> 00:28:55,720
You can trace every action it takes.

876
00:28:55,720 --> 00:28:57,840
You can enforce every policy you've defined.

877
00:28:57,840 --> 00:28:59,560
And you can manage its complete life cycle,

878
00:28:59,560 --> 00:29:01,800
think of it as the operating system for agents.

879
00:29:01,800 --> 00:29:04,040
Without it, agents scatter.

880
00:29:04,040 --> 00:29:06,560
A team in support builds an agent in co-pilot studio.

881
00:29:06,560 --> 00:29:09,120
A team in finance builds one in Azure AI Foundry.

882
00:29:09,120 --> 00:29:11,880
Data signs deploys a custom framework on Kubernetes,

883
00:29:11,880 --> 00:29:14,160
marketing experiments with an open source tool.

884
00:29:14,160 --> 00:29:17,520
In six months, you have agents running on five different platforms,

885
00:29:17,520 --> 00:29:20,320
built with different tools using different policies.

886
00:29:20,320 --> 00:29:22,920
And logging to different systems, you don't have an architecture,

887
00:29:22,920 --> 00:29:25,360
you have fragments that happen to do similar things.

888
00:29:25,360 --> 00:29:28,160
Agent 365 brings them into one visibility layer.

889
00:29:28,160 --> 00:29:30,040
Every agent gets registered in the control plane

890
00:29:30,040 --> 00:29:32,600
regardless of where it was built or what framework it uses.

891
00:29:32,600 --> 00:29:33,960
That agent gets an identity.

892
00:29:33,960 --> 00:29:36,080
That identity gets permissions.

893
00:29:36,080 --> 00:29:37,520
Those permissions get logged.

894
00:29:37,520 --> 00:29:39,240
The agent's behavior gets monitored.

895
00:29:39,240 --> 00:29:40,920
All of it flows through a single system.

896
00:29:40,920 --> 00:29:42,160
What does that actually accomplish?

897
00:29:42,160 --> 00:29:44,680
First, registry.

898
00:29:44,680 --> 00:29:46,200
You can answer the basic question.

899
00:29:46,200 --> 00:29:48,360
What agents exist in our organization?

900
00:29:48,360 --> 00:29:49,480
Not what we think exists.

901
00:29:49,480 --> 00:29:50,800
What actually exists?

902
00:29:50,800 --> 00:29:53,800
Agent 365 provides formal registration and discovery.

903
00:29:53,800 --> 00:29:55,720
It can scan your environment and find agents

904
00:29:55,720 --> 00:29:57,600
that are running without being registered.

905
00:29:57,600 --> 00:29:59,640
These are the agents that exist in the shadows

906
00:29:59,640 --> 00:30:02,800
because they were built quickly and never went through onboarding.

907
00:30:02,800 --> 00:30:04,640
Second, identity management.

908
00:30:04,640 --> 00:30:06,880
Every agent that registers gets an agent ID,

909
00:30:06,880 --> 00:30:09,200
this is a proper directory identity in Entra

910
00:30:09,200 --> 00:30:10,880
that agent can be assigned permissions

911
00:30:10,880 --> 00:30:12,720
scoped to what it actually needs.

912
00:30:12,720 --> 00:30:14,280
Those permissions are enforced by the gateway

913
00:30:14,280 --> 00:30:15,840
invalidated by the ontology.

914
00:30:15,840 --> 00:30:17,320
They are logged and auditable.

915
00:30:17,320 --> 00:30:20,120
When that agent acts, the action is attributed to that agent,

916
00:30:20,120 --> 00:30:21,840
not to a human or a shared account.

917
00:30:21,840 --> 00:30:24,080
Accountability becomes possible.

918
00:30:24,080 --> 00:30:25,840
Third, policy enforcement.

919
00:30:25,840 --> 00:30:28,040
Your governance hub defines the rules,

920
00:30:28,040 --> 00:30:30,520
which models can be used, which tools are approved,

921
00:30:30,520 --> 00:30:32,200
which data sources are accessible.

922
00:30:32,200 --> 00:30:34,160
Those policies live in agent 365.

923
00:30:34,160 --> 00:30:35,440
When an agent tries to do something,

924
00:30:35,440 --> 00:30:38,000
agent 365 checks if that action is permitted,

925
00:30:38,000 --> 00:30:40,880
does the policy allow this agent to call this specific model?

926
00:30:40,880 --> 00:30:42,360
Does it allow accessing this data?

927
00:30:42,360 --> 00:30:44,680
Is this operation happening in an approved region?

928
00:30:44,680 --> 00:30:46,520
These questions get answered at runtime.

929
00:30:46,520 --> 00:30:49,080
Policies are enforced consistently across every agent,

930
00:30:49,080 --> 00:30:50,320
no matter how it was built.

931
00:30:50,320 --> 00:30:52,080
Fourth, observability.

932
00:30:52,080 --> 00:30:54,600
Agent 365 doesn't just enforce, it watches.

933
00:30:54,600 --> 00:30:56,560
It logs every action, every API call,

934
00:30:56,560 --> 00:30:57,600
and every data access.

935
00:30:57,600 --> 00:30:59,800
It correlates those logs with the agent's identity

936
00:30:59,800 --> 00:31:01,560
and the policies that were enforced.

937
00:31:01,560 --> 00:31:03,520
When something goes wrong, you can trace it back.

938
00:31:03,520 --> 00:31:05,720
You can understand what the agent did and why it did it.

939
00:31:05,720 --> 00:31:07,200
You can see what data it accessed

940
00:31:07,200 --> 00:31:08,600
and what decisions it made.

941
00:31:08,600 --> 00:31:11,200
That is the audit trail that compliance requires.

942
00:31:11,200 --> 00:31:12,600
And here is the lifecycle benefit

943
00:31:12,600 --> 00:31:15,200
that nobody talks about until it becomes a crisis.

944
00:31:15,200 --> 00:31:17,120
Agents get built, they accomplish their purpose,

945
00:31:17,120 --> 00:31:18,960
the project ends.

946
00:31:18,960 --> 00:31:20,040
But the agent keeps running.

947
00:31:20,040 --> 00:31:22,360
Six months later, the person who built it is gone.

948
00:31:22,360 --> 00:31:25,320
A year later, nobody even remembers what it was supposed to do.

949
00:31:25,320 --> 00:31:26,760
It is still consuming tokens.

950
00:31:26,760 --> 00:31:28,320
It is still accessing data.

951
00:31:28,320 --> 00:31:30,120
It might be configured with permissions

952
00:31:30,120 --> 00:31:31,920
that were reasonable when it was built,

953
00:31:31,920 --> 00:31:33,120
but are now inappropriate.

954
00:31:33,120 --> 00:31:36,320
It's a ghost, it's a cost drain, it's a security risk.

955
00:31:36,320 --> 00:31:39,840
Agent 365 tracks ownership and usage.

956
00:31:39,840 --> 00:31:42,200
You can see which agents haven't been used in six months.

957
00:31:42,200 --> 00:31:44,360
You can see which ones don't have a named owner.

958
00:31:44,360 --> 00:31:46,360
You can see which ones are consuming resources.

959
00:31:46,360 --> 00:31:48,600
Disproportionately, you can make data-driven decisions

960
00:31:48,600 --> 00:31:51,280
about retirement instead of just letting them pile up.

961
00:31:51,280 --> 00:31:53,280
The shift here is subtle, but absolute.

962
00:31:53,280 --> 00:31:55,640
Without Agent 365, you are managing agents

963
00:31:55,640 --> 00:31:58,520
through individual conversations with individual teams.

964
00:31:58,520 --> 00:32:00,160
Have you checked on your agents lately?

965
00:32:00,160 --> 00:32:03,760
With Agent 365, you are managing agents as a portfolio.

966
00:32:03,760 --> 00:32:05,600
You have visibility, you have control,

967
00:32:05,600 --> 00:32:07,520
you have the operational backbone

968
00:32:07,520 --> 00:32:09,320
that lets governance actually work.

969
00:32:09,320 --> 00:32:12,880
Agent stop being shadow projects and they become managed assets.

970
00:32:12,880 --> 00:32:16,480
The governance hub, Agent 365 gives you visibility and identity,

971
00:32:16,480 --> 00:32:19,160
but visibility without enforcement is just information.

972
00:32:19,160 --> 00:32:21,640
You can see what agents are doing and know who is doing it,

973
00:32:21,640 --> 00:32:23,840
but that doesn't stop them from doing the wrong thing.

974
00:32:23,840 --> 00:32:25,800
That's where the governance hub comes in.

975
00:32:25,800 --> 00:32:27,760
If Agent 365 is the control plane,

976
00:32:27,760 --> 00:32:29,800
the governance hub is the policy engine.

977
00:32:29,800 --> 00:32:32,560
It is the central registry for every approved model,

978
00:32:32,560 --> 00:32:35,440
every vetted tool, every authorized data source,

979
00:32:35,440 --> 00:32:37,920
and every rule that governs how they are used.

980
00:32:37,920 --> 00:32:39,040
It isn't distributed.

981
00:32:39,040 --> 00:32:41,120
It isn't something each team maintains locally.

982
00:32:41,120 --> 00:32:43,000
It is one place, one source of truth.

983
00:32:43,000 --> 00:32:44,120
But here is the problem.

984
00:32:44,120 --> 00:32:46,840
Without a hub, developers make independent decisions

985
00:32:46,840 --> 00:32:48,160
about what they integrate.

986
00:32:48,160 --> 00:32:51,120
A team building an agent needs a tool that connects to Salesforce

987
00:32:51,120 --> 00:32:54,000
instead of checking a catalog of approved Salesforce connectors

988
00:32:54,000 --> 00:32:56,360
they build their own because there is no catalog.

989
00:32:56,360 --> 00:32:57,280
They build it fast.

990
00:32:57,280 --> 00:32:57,800
It works.

991
00:32:57,800 --> 00:32:58,520
It chips.

992
00:32:58,520 --> 00:33:00,680
Six months later, another team builds the same connector

993
00:33:00,680 --> 00:33:02,600
because they also didn't know it existed.

994
00:33:02,600 --> 00:33:06,000
Now you have duplicate functionality, inconsistent implementations,

995
00:33:06,000 --> 00:33:08,240
and two different connectors that might behave differently

996
00:33:08,240 --> 00:33:10,200
or have different security properties.

997
00:33:10,200 --> 00:33:12,640
Multiply that by dozens of teams and dozens of tools.

998
00:33:12,640 --> 00:33:14,440
You end up with a sprawl of integrations

999
00:33:14,440 --> 00:33:16,560
that nobody understands and nobody can maintain.

1000
00:33:16,560 --> 00:33:18,040
A governance hub prevents that.

1001
00:33:18,040 --> 00:33:21,080
Every tool that an agent is allowed to use gets registered there.

1002
00:33:21,080 --> 00:33:21,880
It has been reviewed.

1003
00:33:21,880 --> 00:33:22,720
It has been tested.

1004
00:33:22,720 --> 00:33:24,680
Its security properties are documented.

1005
00:33:24,680 --> 00:33:26,560
When a new team wants to integrate Salesforce,

1006
00:33:26,560 --> 00:33:28,040
they don't build a new connector.

1007
00:33:28,040 --> 00:33:29,800
They request access to the approved one.

1008
00:33:29,800 --> 00:33:31,360
The request goes through the hub.

1009
00:33:31,360 --> 00:33:33,240
The hub checks their agents permissions.

1010
00:33:33,240 --> 00:33:35,440
If their agent is authorized to use that tool,

1011
00:33:35,440 --> 00:33:36,280
they get access.

1012
00:33:36,280 --> 00:33:37,600
If not, they are denied.

1013
00:33:37,600 --> 00:33:41,320
But they are also told what other approved tools might solve their problem.

1014
00:33:41,320 --> 00:33:42,720
This shifts the dynamic.

1015
00:33:42,720 --> 00:33:44,440
Instead of every team building independently

1016
00:33:44,440 --> 00:33:45,480
and hoping nothing breaks,

1017
00:33:45,480 --> 00:33:47,800
every team operates within a control set of options.

1018
00:33:47,800 --> 00:33:49,600
Standardization becomes possible.

1019
00:33:49,600 --> 00:33:51,400
The same principle applies to models.

1020
00:33:51,400 --> 00:33:54,720
Without a hub, teams go directly to Azure Open AI or Anthropic

1021
00:33:54,720 --> 00:33:56,320
or whatever they decide to use.

1022
00:33:56,320 --> 00:33:58,560
Some use expensive models for simple tasks.

1023
00:33:58,560 --> 00:34:01,560
Some use models that are restricted to certain regions when they shouldn't be.

1024
00:34:01,560 --> 00:34:03,440
Some use models that have safety properties

1025
00:34:03,440 --> 00:34:04,920
their agents shouldn't violate.

1026
00:34:04,920 --> 00:34:07,400
Without central control, you get inconsistency

1027
00:34:07,400 --> 00:34:09,520
with the hub approved models are listed.

1028
00:34:09,520 --> 00:34:11,680
Teams request the right model for their use case.

1029
00:34:11,680 --> 00:34:14,000
The hub validates that the model is appropriate.

1030
00:34:14,000 --> 00:34:16,400
It validates that the agent has permission to use it.

1031
00:34:16,400 --> 00:34:18,720
It validates that the use case is compliant.

1032
00:34:18,720 --> 00:34:21,080
But here is where the hub becomes operationally essential.

1033
00:34:21,080 --> 00:34:22,280
Policy updates.

1034
00:34:22,280 --> 00:34:26,720
Let's say your compliance team decides that agents can no longer process customer data

1035
00:34:26,720 --> 00:34:27,880
in the European region.

1036
00:34:27,880 --> 00:34:30,840
Every agent needs to use models that run in North America only.

1037
00:34:30,840 --> 00:34:34,000
Without a hub, you email this requirement to all the teams.

1038
00:34:34,000 --> 00:34:35,920
You hope they update their configurations.

1039
00:34:35,920 --> 00:34:37,920
Some do, some don't, some miss the email,

1040
00:34:37,920 --> 00:34:39,640
some interpret the rule differently.

1041
00:34:39,640 --> 00:34:43,240
You end up with agents in compliance and agents that violate the requirement

1042
00:34:43,240 --> 00:34:45,160
and you won't find out until an audit.

1043
00:34:45,160 --> 00:34:49,360
With a hub that policy gets updated once every agent that requests access to a model

1044
00:34:49,360 --> 00:34:51,520
now gets evaluated against that policy.

1045
00:34:51,520 --> 00:34:54,160
An agent tries to call a European model.

1046
00:34:54,160 --> 00:34:56,080
The hub evaluates the request.

1047
00:34:56,080 --> 00:34:58,240
The agent's permissions allow European models,

1048
00:34:58,240 --> 00:35:01,920
but the organization's policy now forbids customer data in that region.

1049
00:35:01,920 --> 00:35:04,440
The hub blocks it and suggests an alternative.

1050
00:35:04,440 --> 00:35:05,640
Policy enforced.

1051
00:35:05,640 --> 00:35:08,840
Everywhere, immediately, no individual configurations,

1052
00:35:08,840 --> 00:35:12,200
no hope that teams remember to update, no surprise audit failures.

1053
00:35:12,200 --> 00:35:13,720
This is operational leverage.

1054
00:35:13,720 --> 00:35:14,960
One policy change.

1055
00:35:14,960 --> 00:35:16,440
Hundreds of agents affected.

1056
00:35:16,440 --> 00:35:17,840
Consistent enforcement.

1057
00:35:17,840 --> 00:35:21,880
The audit benefit is equally significant when a regulator asks which tools your agents

1058
00:35:21,880 --> 00:35:24,240
are using and whether they are approved, you have an answer.

1059
00:35:24,240 --> 00:35:26,480
The governance hub is the authoritative list.

1060
00:35:26,480 --> 00:35:28,440
Every agent using an unapproved tool stands out.

1061
00:35:28,440 --> 00:35:29,120
You can see it.

1062
00:35:29,120 --> 00:35:30,240
You can trace why.

1063
00:35:30,240 --> 00:35:31,080
You can fix it.

1064
00:35:31,080 --> 00:35:34,120
This is what transforms policy from a document that people maybe read into

1065
00:35:34,120 --> 00:35:38,520
infrastructure that actually prevents violations from happening in the first place.

1066
00:35:38,520 --> 00:35:40,080
Data residency and compliance.

1067
00:35:40,080 --> 00:35:43,320
There is a category of problem that most organizations don't think about until

1068
00:35:43,320 --> 00:35:45,320
a regulator asks a question they can't answer.

1069
00:35:45,320 --> 00:35:48,120
It isn't about visibility or governance or even identity.

1070
00:35:48,120 --> 00:35:50,240
It is about something simpler and more fundamental.

1071
00:35:50,240 --> 00:35:51,880
Where your data actually lives.

1072
00:35:51,880 --> 00:35:53,200
Compliance isn't optional.

1073
00:35:53,200 --> 00:35:55,960
It is the cost of operating in regulated industries.

1074
00:35:55,960 --> 00:36:00,280
GDPR says personal data from EU citizens can't be processed outside the EU

1075
00:36:00,280 --> 00:36:01,880
without explicit mechanisms.

1076
00:36:01,880 --> 00:36:04,520
HIPAA says health information can't leave the United States.

1077
00:36:04,520 --> 00:36:09,640
SOC2 says you need to maintain data residency policies and prove you are enforcing them.

1078
00:36:09,640 --> 00:36:13,280
PCI DSS restricts where payment card data can be stored and processed.

1079
00:36:13,280 --> 00:36:14,480
These aren't suggestions.

1080
00:36:14,480 --> 00:36:15,640
They are legal requirements.

1081
00:36:15,640 --> 00:36:17,480
Violate them and you don't just lose a contract.

1082
00:36:17,480 --> 00:36:18,240
You face fines.

1083
00:36:18,240 --> 00:36:19,400
You face criminal liability.

1084
00:36:19,400 --> 00:36:21,200
You face organizational collapse.

1085
00:36:21,200 --> 00:36:23,520
Now imagine what happens when you don't have a landing zone.

1086
00:36:23,520 --> 00:36:26,520
A team builds an agent that processes customer records.

1087
00:36:26,520 --> 00:36:28,920
The agent is built in Azure running in North America.

1088
00:36:28,920 --> 00:36:29,800
It works great.

1089
00:36:29,800 --> 00:36:33,920
But then it needs to analyze trends against a data set that lives in a European warehouse.

1090
00:36:33,920 --> 00:36:36,440
The agent makes an API call directly to that warehouse.

1091
00:36:36,440 --> 00:36:37,400
The data is retrieved.

1092
00:36:37,400 --> 00:36:38,720
The agent processes it.

1093
00:36:38,720 --> 00:36:40,040
The results are cashed locally.

1094
00:36:40,040 --> 00:36:41,560
Nobody thought about geography.

1095
00:36:41,560 --> 00:36:46,640
Six months later an auditor asks, where does this agent process personal data from EU citizens?

1096
00:36:46,640 --> 00:36:47,640
You search the logs.

1097
00:36:47,640 --> 00:36:49,200
The agent is running in North America.

1098
00:36:49,200 --> 00:36:51,320
But the data it is analyzing comes from Europe.

1099
00:36:51,320 --> 00:36:52,320
Is that compliant?

1100
00:36:52,320 --> 00:36:53,720
Actually it is unclear.

1101
00:36:53,720 --> 00:36:58,000
GDPR says personal data can't be processed outside the EU unless there is a data processing

1102
00:36:58,000 --> 00:36:59,000
agreement in place.

1103
00:36:59,000 --> 00:37:00,000
Do you have one?

1104
00:37:00,000 --> 00:37:01,000
You don't know.

1105
00:37:01,000 --> 00:37:03,160
The agent was built without compliance review.

1106
00:37:03,160 --> 00:37:05,080
Now multiply this across 50 agents.

1107
00:37:05,080 --> 00:37:06,440
Each one was built independently.

1108
00:37:06,440 --> 00:37:10,280
Each one makes independent decisions about which data sources to access and which regions

1109
00:37:10,280 --> 00:37:11,440
to process in.

1110
00:37:11,440 --> 00:37:14,200
Some agents are calling models in regions that violate your policy.

1111
00:37:14,200 --> 00:37:17,120
Some are moving data across borders that shouldn't cross borders.

1112
00:37:17,120 --> 00:37:20,080
Some are caching results in locations that aren't approved.

1113
00:37:20,080 --> 00:37:21,080
Nobody coordinated.

1114
00:37:21,080 --> 00:37:22,240
Nobody enforced a standard.

1115
00:37:22,240 --> 00:37:24,160
You can't pass an audit with that.

1116
00:37:24,160 --> 00:37:25,720
The structural problem is this.

1117
00:37:25,720 --> 00:37:28,160
Without a landing zone, data movement is reactive.

1118
00:37:28,160 --> 00:37:30,800
When agent decides it needs something, so it goes and gets it.

1119
00:37:30,800 --> 00:37:34,160
By the time you audit what happened, the violation is already in the logs.

1120
00:37:34,160 --> 00:37:36,000
You are investigating after the fact.

1121
00:37:36,000 --> 00:37:40,400
You are trying to figure out what went wrong, why it went wrong, and how to prove to a

1122
00:37:40,400 --> 00:37:42,080
regulator that you fixed it.

1123
00:37:42,080 --> 00:37:43,480
This is reactive compliance.

1124
00:37:43,480 --> 00:37:46,000
You are checking boxes after the damage is done.

1125
00:37:46,000 --> 00:37:49,440
A landing zone changes that to something much more powerful.

1126
00:37:49,440 --> 00:37:50,960
Preventive compliance.

1127
00:37:50,960 --> 00:37:54,600
Data movement is gated at the infrastructure level before the violation can happen.

1128
00:37:54,600 --> 00:37:55,800
Here is how it works.

1129
00:37:55,800 --> 00:37:58,040
Your compliance team defines a policy.

1130
00:37:58,040 --> 00:38:02,640
Agents can process customer data only in regions where that customer's data is domiciled.

1131
00:38:02,640 --> 00:38:06,160
An agent is built that needs to analyze European customer records.

1132
00:38:06,160 --> 00:38:09,040
Before that agent can execute the landing zone evaluates the policy.

1133
00:38:09,040 --> 00:38:10,600
The agent is running in North America.

1134
00:38:10,600 --> 00:38:12,560
It is trying to access European data.

1135
00:38:12,560 --> 00:38:14,240
The policy forbids that combination.

1136
00:38:14,240 --> 00:38:15,920
The landing zone blocks the request.

1137
00:38:15,920 --> 00:38:20,440
The agent can't proceed until it is reconfigured to run in a region that is compliant.

1138
00:38:20,440 --> 00:38:22,520
Policy enforced before the violation happens.

1139
00:38:22,520 --> 00:38:23,920
Not audited after.

1140
00:38:23,920 --> 00:38:25,480
This is the fundamental shift.

1141
00:38:25,480 --> 00:38:28,680
Without a landing zone, compliance is a backward looking investigation.

1142
00:38:28,680 --> 00:38:31,040
You analyze what happened and try to explain it.

1143
00:38:31,040 --> 00:38:34,600
With a landing zone, compliance is forward looking enforcement.

1144
00:38:34,600 --> 00:38:38,640
Vialations become impossible because the infrastructure prevents them before they occur.

1145
00:38:38,640 --> 00:38:41,200
The operational benefit is equally significant.

1146
00:38:41,200 --> 00:38:44,600
When a regulator asks whether your agents are compliant, you don't have to investigate.

1147
00:38:44,600 --> 00:38:45,600
You have evidence.

1148
00:38:45,600 --> 00:38:46,600
The landing zone enforces the policy.

1149
00:38:46,600 --> 00:38:47,960
And that enforcement is logged.

1150
00:38:47,960 --> 00:38:51,760
You can demonstrate that your infrastructure makes non-compliance impossible for any agent

1151
00:38:51,760 --> 00:38:52,880
that runs within it.

1152
00:38:52,880 --> 00:38:55,080
This changes the risk profile entirely.

1153
00:38:55,080 --> 00:38:58,760
I'm hoping teams remember compliance rules and configured them correctly to knowing that

1154
00:38:58,760 --> 00:39:00,920
compliance is built into the system.

1155
00:39:00,920 --> 00:39:03,200
From audit surprises to regulated certainty.

1156
00:39:03,200 --> 00:39:05,480
From governance debt to governance certainty.

1157
00:39:05,480 --> 00:39:07,680
Data residency isn't just a technical requirement.

1158
00:39:07,680 --> 00:39:11,240
It is the structural difference between organizations that can prove they are compliant

1159
00:39:11,240 --> 00:39:13,240
and organizations that hope they are.

1160
00:39:13,240 --> 00:39:15,160
Cost control and fin ops.

1161
00:39:15,160 --> 00:39:16,960
There is a problem that happens at scale.

1162
00:39:16,960 --> 00:39:19,240
Almost nobody sees it coming until it's too late.

1163
00:39:19,240 --> 00:39:21,120
And by then, it's impossible to untangle it.

1164
00:39:21,120 --> 00:39:22,200
It's about the money.

1165
00:39:22,200 --> 00:39:25,320
When you start experimenting with AI, cost feels theoretical.

1166
00:39:25,320 --> 00:39:26,400
You spin up a few agents.

1167
00:39:26,400 --> 00:39:27,400
They call models.

1168
00:39:27,400 --> 00:39:28,400
The bill comes once a month.

1169
00:39:28,400 --> 00:39:29,400
It's acceptable.

1170
00:39:29,400 --> 00:39:30,400
You ship more agents.

1171
00:39:30,400 --> 00:39:31,400
The bill grows.

1172
00:39:31,400 --> 00:39:32,600
It's still manageable.

1173
00:39:32,600 --> 00:39:35,960
But somewhere around agent 50 or 20, something changes.

1174
00:39:35,960 --> 00:39:37,760
The monthly spend crosses a line.

1175
00:39:37,760 --> 00:39:41,400
It goes from an investment to a, where is all this money going?

1176
00:39:41,400 --> 00:39:42,400
Problem.

1177
00:39:42,400 --> 00:39:45,040
And when you ask that question, you realize you don't have an answer.

1178
00:39:45,040 --> 00:39:48,720
An agent in support is calling GPT-4 for every single customer query.

1179
00:39:48,720 --> 00:39:50,600
It could use a cheaper model for most of them.

1180
00:39:50,600 --> 00:39:54,480
An agent in finance is running redundant logic that another team already solved.

1181
00:39:54,480 --> 00:39:58,600
A data team's agent is making three separate API calls to fetch one piece of info.

1182
00:39:58,600 --> 00:40:03,160
A marketing agent is hitting a model every 60 seconds just to check for updates.

1183
00:40:03,160 --> 00:40:05,480
Nobody coordinated, nobody measured, nobody optimized.

1184
00:40:05,480 --> 00:40:06,840
They just shipped.

1185
00:40:06,840 --> 00:40:09,400
Without visibility into consumption, you can't manage cost.

1186
00:40:09,400 --> 00:40:10,480
You can't even see it.

1187
00:40:10,480 --> 00:40:12,680
You get a cloud bill that's higher than last month.

1188
00:40:12,680 --> 00:40:14,200
You email the teams to ask why.

1189
00:40:14,200 --> 00:40:15,200
Nobody knows.

1190
00:40:15,200 --> 00:40:18,240
The money is disappearing into a black box of token consumption.

1191
00:40:18,240 --> 00:40:19,560
Dozens of agents are running.

1192
00:40:19,560 --> 00:40:23,720
There's no way to attribute spend to a specific team or a specific use case.

1193
00:40:23,720 --> 00:40:27,000
This is where the landing zone's model gateway becomes the solution.

1194
00:40:27,000 --> 00:40:30,160
Every single model call goes through that gateway, not to slow things down.

1195
00:40:30,160 --> 00:40:34,000
But because that gateway is the one place where you can see and control access, which agent

1196
00:40:34,000 --> 00:40:35,000
called which model.

1197
00:40:35,000 --> 00:40:36,400
How many tokens did it use?

1198
00:40:36,400 --> 00:40:37,960
What region did the call go to?

1199
00:40:37,960 --> 00:40:40,240
This data flows into a cost tracking system.

1200
00:40:40,240 --> 00:40:44,840
Not monthly, not quarterly, continuously, real time visibility.

1201
00:40:44,840 --> 00:40:47,240
Now, cost management is actually possible.

1202
00:40:47,240 --> 00:40:50,920
You can see that one single agent accounts for 30% of your spend.

1203
00:40:50,920 --> 00:40:52,880
You can drill down and understand why.

1204
00:40:52,880 --> 00:40:55,840
Is it calling an expensive model when a cheap one would work?

1205
00:40:55,840 --> 00:40:58,360
Is it asking for a longer context window than it needs?

1206
00:40:58,360 --> 00:40:59,720
Is it making redundant calls?

1207
00:40:59,720 --> 00:41:01,360
Once you see the pattern, you can fix it.

1208
00:41:01,360 --> 00:41:04,560
Maybe that agent switches to a smaller model for routine questions.

1209
00:41:04,560 --> 00:41:07,680
It only escalates to the expensive model when it's actually needed.

1210
00:41:07,680 --> 00:41:11,080
Maybe you add caching so it stops asking the same thing over and over.

1211
00:41:11,080 --> 00:41:13,400
Maybe you redesigned the prompt to use fewer tokens.

1212
00:41:13,400 --> 00:41:14,600
But here's the critical part.

1213
00:41:14,600 --> 00:41:18,040
You can do this optimization without losing any capability.

1214
00:41:18,040 --> 00:41:20,840
Without visibility, the choices binary, the agent works.

1215
00:41:20,840 --> 00:41:21,840
Or it doesn't.

1216
00:41:21,840 --> 00:41:22,840
There's nothing in between.

1217
00:41:22,840 --> 00:41:24,240
With visibility, you can optimize.

1218
00:41:24,240 --> 00:41:26,760
You keep the same outcome at a much lower cost.

1219
00:41:26,760 --> 00:41:30,920
That's the difference between paying $60,000 a month and paying $20,000 for the exact same

1220
00:41:30,920 --> 00:41:31,920
result.

1221
00:41:31,920 --> 00:41:33,960
The gateway also enables model routing.

1222
00:41:33,960 --> 00:41:35,320
Different models have different costs.

1223
00:41:35,320 --> 00:41:38,720
An expensive model might be brilliant, but it's overkill for most tasks.

1224
00:41:38,720 --> 00:41:41,280
A cheaper model handles routine work just fine.

1225
00:41:41,280 --> 00:41:44,600
But a gateway every agent decides on its own which model to use.

1226
00:41:44,600 --> 00:41:47,840
With a gateway, policy-based routing makes that decision for them.

1227
00:41:47,840 --> 00:41:49,240
An agent submits a request.

1228
00:41:49,240 --> 00:41:51,240
The gateway evaluates it based on your rules.

1229
00:41:51,240 --> 00:41:54,120
If it's a simple question, root it to the cheap model.

1230
00:41:54,120 --> 00:41:56,960
If it needs deep reasoning, root it to the expensive one.

1231
00:41:56,960 --> 00:42:01,240
Same agent, different models depending on the actual need, cost is optimized, performance

1232
00:42:01,240 --> 00:42:02,960
is maintained.

1233
00:42:02,960 --> 00:42:04,800
Budget controls finally become real.

1234
00:42:04,800 --> 00:42:07,400
You set a monthly budget for a team or a project.

1235
00:42:07,400 --> 00:42:11,480
As agents use tokens, the spend is tied to that budget when they hit 80%.

1236
00:42:11,480 --> 00:42:12,480
You get an alert.

1237
00:42:12,480 --> 00:42:14,320
At 100%, the call stops.

1238
00:42:14,320 --> 00:42:15,520
This forces a conversation.

1239
00:42:15,520 --> 00:42:17,240
Why did this team use their whole budget?

1240
00:42:17,240 --> 00:42:18,240
Was it necessary?

1241
00:42:18,240 --> 00:42:19,240
Is there waste?

1242
00:42:19,240 --> 00:42:21,080
Or is this a success that we need to scale up?

1243
00:42:21,080 --> 00:42:23,560
You can't have these conversations without attribution.

1244
00:42:23,560 --> 00:42:24,560
With it.

1245
00:42:24,560 --> 00:42:27,360
Budget management is how you align AI spending with business value.

1246
00:42:27,360 --> 00:42:28,520
The logic is clear.

1247
00:42:28,520 --> 00:42:32,280
Without a landing zone, cost grows randomly as teams ship, with one.

1248
00:42:32,280 --> 00:42:33,520
Cost is a controlled variable.

1249
00:42:33,520 --> 00:42:34,960
You understand where the money goes.

1250
00:42:34,960 --> 00:42:36,400
You optimize what wastes it.

1251
00:42:36,400 --> 00:42:37,880
You scale what works.

1252
00:42:37,880 --> 00:42:41,040
Finops stops being something you do after a shocking bill arrives.

1253
00:42:41,040 --> 00:42:42,960
It becomes how you operate from day one.

1254
00:42:42,960 --> 00:42:44,200
Cost isn't just a finance problem.

1255
00:42:44,200 --> 00:42:45,560
It's a structural problem.

1256
00:42:45,560 --> 00:42:47,200
And the landing zone solves it.

1257
00:42:47,200 --> 00:42:49,080
The ontology advantage in practice.

1258
00:42:49,080 --> 00:42:50,880
We've talked about ontologies in the abstract.

1259
00:42:50,880 --> 00:42:54,200
Now let's talk about what they actually do in a production system.

1260
00:42:54,200 --> 00:42:57,200
An ontology encodes relationships that are usually invisible.

1261
00:42:57,200 --> 00:42:59,480
Your data warehouse has a table of customers.

1262
00:42:59,480 --> 00:43:02,480
A column says segment with a value like enterprise.

1263
00:43:02,480 --> 00:43:03,480
That is a fact.

1264
00:43:03,480 --> 00:43:05,200
An ontology says the same thing differently.

1265
00:43:05,200 --> 00:43:06,760
This customer is a VIP.

1266
00:43:06,760 --> 00:43:08,240
That means their orders get priority.

1267
00:43:08,240 --> 00:43:10,680
That means a delay triggers an immediate escalation.

1268
00:43:10,680 --> 00:43:13,360
That means the operations team gets in a lurch right now.

1269
00:43:13,360 --> 00:43:16,040
Instead of finding out next week, those sound like the same thing.

1270
00:43:16,040 --> 00:43:17,040
They are not.

1271
00:43:17,040 --> 00:43:18,040
The table is data.

1272
00:43:18,040 --> 00:43:19,040
The ontology is meaning.

1273
00:43:19,040 --> 00:43:20,200
One is a field in a record.

1274
00:43:20,200 --> 00:43:22,840
The other is a relationship that carries a consequence.

1275
00:43:22,840 --> 00:43:25,720
When an agent knows a customer segment, it has information.

1276
00:43:25,720 --> 00:43:30,720
When an agent understands how VIP status changes in operational decision, it has context.

1277
00:43:30,720 --> 00:43:35,040
Context is what separates an agent that answers questions from an agent that makes decisions.

1278
00:43:35,040 --> 00:43:37,560
And this is where it becomes operationally significant.

1279
00:43:37,560 --> 00:43:39,400
Agents reason across the relationships that matter.

1280
00:43:39,400 --> 00:43:41,000
Not every possible relationship.

1281
00:43:41,000 --> 00:43:43,160
Think about an agent monitoring in order.

1282
00:43:43,160 --> 00:43:45,280
Without an ontology, its reasoning is a mess.

1283
00:43:45,280 --> 00:43:46,440
The customer placed in order.

1284
00:43:46,440 --> 00:43:47,720
The order has products.

1285
00:43:47,720 --> 00:43:48,960
Products need inventory.

1286
00:43:48,960 --> 00:43:50,240
Inventory is in a warehouse.

1287
00:43:50,240 --> 00:43:51,560
The warehouse has a supplier.

1288
00:43:51,560 --> 00:43:52,760
The supplier might be late.

1289
00:43:52,760 --> 00:43:54,680
That's a dozen different relationships to analyze.

1290
00:43:54,680 --> 00:43:55,680
Which one's matter right now?

1291
00:43:55,680 --> 00:43:57,080
The agent has to guess.

1292
00:43:57,080 --> 00:43:58,440
Maybe it looks at lead times.

1293
00:43:58,440 --> 00:43:59,880
Maybe it looks at stock levels.

1294
00:43:59,880 --> 00:44:02,800
Maybe it checks payment history because it thinks that matters.

1295
00:44:02,800 --> 00:44:06,000
The agent might miss the one link that actually determines priority.

1296
00:44:06,000 --> 00:44:09,640
Or it might waste compute looking at things that don't matter at all.

1297
00:44:09,640 --> 00:44:10,640
With an ontology.

1298
00:44:10,640 --> 00:44:11,960
The reasoning is structured.

1299
00:44:11,960 --> 00:44:14,320
The relationships that matter are already encoded.

1300
00:44:14,320 --> 00:44:16,160
VIP customers get priority.

1301
00:44:16,160 --> 00:44:17,920
Peak season increases that priority.

1302
00:44:17,920 --> 00:44:19,320
Certain suppliers are high-risk.

1303
00:44:19,320 --> 00:44:22,360
High-risk suppliers in peak season require an acceleration plan.

1304
00:44:22,360 --> 00:44:23,200
The agent doesn't guess.

1305
00:44:23,200 --> 00:44:25,560
It follows a path that has already been thought through.

1306
00:44:25,560 --> 00:44:28,520
It follows relationships that were explicitly defined as relevant.

1307
00:44:28,520 --> 00:44:30,960
The hallucination prevention here isn't about rules.

1308
00:44:30,960 --> 00:44:32,480
It's structural.

1309
00:44:32,480 --> 00:44:35,880
An agent can't invent a relationship that doesn't exist in the ontology.

1310
00:44:35,880 --> 00:44:37,320
It doesn't have the freedom to do that.

1311
00:44:37,320 --> 00:44:39,080
The ontology isn't a suggestion.

1312
00:44:39,080 --> 00:44:40,080
It's a boundary.

1313
00:44:40,080 --> 00:44:41,920
If an agent needs to reason, it has to follow a path.

1314
00:44:41,920 --> 00:44:44,000
If there is no path, it can't reason about it.

1315
00:44:44,000 --> 00:44:45,240
And that is exactly what you want.

1316
00:44:45,240 --> 00:44:48,080
You don't want an agent making up logic you haven't approved.

1317
00:44:48,080 --> 00:44:51,800
This scales in a way that documents never will when you have ten agents.

1318
00:44:51,800 --> 00:44:53,440
They could all be reasoning differently.

1319
00:44:53,440 --> 00:44:55,720
One agent defines VIP status one way.

1320
00:44:55,720 --> 00:44:57,200
Another agent defines it another way.

1321
00:44:57,200 --> 00:44:58,880
One prioritizes inventory.

1322
00:44:58,880 --> 00:45:00,720
One prioritizes speed.

1323
00:45:00,720 --> 00:45:04,840
You get inconsistent decisions across the company because every agent built its own mental

1324
00:45:04,840 --> 00:45:05,840
model.

1325
00:45:05,840 --> 00:45:07,800
With an ontology.

1326
00:45:07,800 --> 00:45:09,080
Consistency is structural.

1327
00:45:09,080 --> 00:45:11,840
All ten agents see the same definition of VIP.

1328
00:45:11,840 --> 00:45:15,080
All ten use the same criteria for supply or risk.

1329
00:45:15,080 --> 00:45:17,280
All ten optimize against the same business logic.

1330
00:45:17,280 --> 00:45:18,360
You don't have to train them.

1331
00:45:18,360 --> 00:45:20,840
You don't have to write policies and hope they understand.

1332
00:45:20,840 --> 00:45:22,280
The ontology is the policy.

1333
00:45:22,280 --> 00:45:23,600
It is executable.

1334
00:45:23,600 --> 00:45:24,760
It is consistent.

1335
00:45:24,760 --> 00:45:26,480
It is auditable.

1336
00:45:26,480 --> 00:45:28,960
And here is the shift that nobody anticipates.

1337
00:45:28,960 --> 00:45:31,680
An agent stops being about stopping bad outputs.

1338
00:45:31,680 --> 00:45:33,440
It becomes about enabling good reasoning.

1339
00:45:33,440 --> 00:45:34,760
Without an ontology.

1340
00:45:34,760 --> 00:45:35,760
Governance is reactive.

1341
00:45:35,760 --> 00:45:37,080
An agent makes a mistake.

1342
00:45:37,080 --> 00:45:38,080
You review it.

1343
00:45:38,080 --> 00:45:39,080
You block it.

1344
00:45:39,080 --> 00:45:40,880
You tell the agent, don't do that.

1345
00:45:40,880 --> 00:45:42,400
But you're just policing the output.

1346
00:45:42,400 --> 00:45:44,360
You aren't changing how the agent thinks.

1347
00:45:44,360 --> 00:45:46,440
With an ontology, governance is proactive.

1348
00:45:46,440 --> 00:45:47,760
Before the agent ever runs.

1349
00:45:47,760 --> 00:45:49,520
You've defined how it should reason.

1350
00:45:49,520 --> 00:45:50,520
You've encoded the logic.

1351
00:45:50,520 --> 00:45:51,840
You've constrained the paths.

1352
00:45:51,840 --> 00:45:55,880
The agent can't make an unreasonable decision because making a decision requires following

1353
00:45:55,880 --> 00:45:57,440
the ontology.

1354
00:45:57,440 --> 00:46:00,920
And the ontology only contains the paths you've decided are reasonable.

1355
00:46:00,920 --> 00:46:03,120
This is the shift from control to enablement.

1356
00:46:03,120 --> 00:46:04,920
You aren't restricting what agents can do.

1357
00:46:04,920 --> 00:46:07,080
You are giving them the right way to think.

1358
00:46:07,080 --> 00:46:08,880
Multi-cloud agent orchestration.

1359
00:46:08,880 --> 00:46:11,560
Most organizations don't run everything on one cloud.

1360
00:46:11,560 --> 00:46:14,040
If they did, the governance problem would be simpler.

1361
00:46:14,040 --> 00:46:17,480
You'd have one control plane, one set of policies, one identity system.

1362
00:46:17,480 --> 00:46:19,560
Everything would be integrated.

1363
00:46:19,560 --> 00:46:21,080
But that's not the world we live in.

1364
00:46:21,080 --> 00:46:24,800
You have a co-pilot studio agent built by the support team running in Azure.

1365
00:46:24,800 --> 00:46:29,160
You have an AWS bedrock agent built by the data team running in their separate cloud account.

1366
00:46:29,160 --> 00:46:32,120
You have an open source framework agent that a contractor built.

1367
00:46:32,120 --> 00:46:35,520
And it's now running on Kubernetes in your on-premises data center.

1368
00:46:35,520 --> 00:46:39,680
You have a SAS agent from a third-party vendor running in their cloud, not yours.

1369
00:46:39,680 --> 00:46:41,040
Each one does something useful.

1370
00:46:41,040 --> 00:46:43,560
Each one was the right choice for the team that built it.

1371
00:46:43,560 --> 00:46:48,160
But together, they create a governance problem that no single cloud provider can solve.

1372
00:46:48,160 --> 00:46:51,680
Because without a unified control plane, governance ends up in multiple places.

1373
00:46:51,680 --> 00:46:52,680
Or nowhere.

1374
00:46:52,680 --> 00:46:56,160
Azure agent has an enter identity because it's in the Microsoft ecosystem.

1375
00:46:56,160 --> 00:46:59,240
But does the AWS agent have those same identity standards applied?

1376
00:46:59,240 --> 00:47:00,240
Not really.

1377
00:47:00,240 --> 00:47:01,240
AWS uses IAM roles.

1378
00:47:01,240 --> 00:47:02,240
They aren't the same thing.

1379
00:47:02,240 --> 00:47:05,800
The SAS agent might use whatever authentication the vendor decided on.

1380
00:47:05,800 --> 00:47:10,120
While the open source agent might be using keys that nobody really knows how to rotate securely,

1381
00:47:10,120 --> 00:47:13,840
you're looking at four different identity models, four different access control patterns,

1382
00:47:13,840 --> 00:47:15,360
and four different audit trails.

1383
00:47:15,360 --> 00:47:18,040
Now, try to enforce a policy across all of them.

1384
00:47:18,040 --> 00:47:21,960
Let's say your compliance team decides that agents can't call models that run outside

1385
00:47:21,960 --> 00:47:23,080
the United States.

1386
00:47:23,080 --> 00:47:27,560
The Azure agent uses Azure OpenAI, which is easy to restrict the AWS agent connects to

1387
00:47:27,560 --> 00:47:29,920
Bedrock, which is also relatively straightforward.

1388
00:47:29,920 --> 00:47:33,000
But the SAS agent is calling a European provider's API.

1389
00:47:33,000 --> 00:47:36,240
The open source agent is using a model that someone deployed in Google Cloud.

1390
00:47:36,240 --> 00:47:38,680
You can't control those from an Azure policy engine.

1391
00:47:38,680 --> 00:47:39,760
They aren't in your cloud.

1392
00:47:39,760 --> 00:47:41,480
They aren't in your identity system.

1393
00:47:41,480 --> 00:47:43,440
The result is a fragmented governance model.

1394
00:47:43,440 --> 00:47:45,360
Some policies you can enforce technically.

1395
00:47:45,360 --> 00:47:47,040
Others you enforce through documentation.

1396
00:47:47,040 --> 00:47:48,520
Others you just hope nobody violates.

1397
00:47:48,520 --> 00:47:50,120
This is the fragmentation problem.

1398
00:47:50,120 --> 00:47:54,280
As soon as your agents cross cloud boundaries, centralized governance becomes exponentially

1399
00:47:54,280 --> 00:47:55,280
harder.

1400
00:47:55,280 --> 00:47:57,320
The challenge isn't that cloud providers are bad at governance.

1401
00:47:57,320 --> 00:47:59,360
It's that they're only good at governing their own cloud.

1402
00:47:59,360 --> 00:48:01,120
They can't govern beyond their boundaries.

1403
00:48:01,120 --> 00:48:04,920
This is where agent 365 and the model context protocol become essential.

1404
00:48:04,920 --> 00:48:07,440
Agent 365 operates above any specific cloud.

1405
00:48:07,440 --> 00:48:11,480
It doesn't care whether an agent is running on Azure or AWS or on your own servers.

1406
00:48:11,480 --> 00:48:13,880
It can register agents regardless of where they run.

1407
00:48:13,880 --> 00:48:17,600
It can assign them identities that stay consistent regardless of the cloud.

1408
00:48:17,600 --> 00:48:21,080
They can apply policies that work across clouds because those policies are enforced at the

1409
00:48:21,080 --> 00:48:24,080
agent 365 layer, not at the cloud provider layer.

1410
00:48:24,080 --> 00:48:27,680
The model context protocol does something similar for tools and data sources.

1411
00:48:27,680 --> 00:48:32,360
MCP is a standard that lets agents connect to tools and data regardless of where those tools

1412
00:48:32,360 --> 00:48:33,360
live.

1413
00:48:33,360 --> 00:48:37,200
Instead of each agent negotiating its own connection to each tool, agents connect to tools

1414
00:48:37,200 --> 00:48:38,200
through MCP.

1415
00:48:38,200 --> 00:48:42,600
That means the governance hub can define which tools are approved and which aren't.

1416
00:48:42,600 --> 00:48:47,000
And that decision applies whether the tool is running on Azure, AWS or anywhere else.

1417
00:48:47,000 --> 00:48:50,120
Here's what this looks like in practice.

1418
00:48:50,120 --> 00:48:53,120
Your organization has agents on four different clouds.

1419
00:48:53,120 --> 00:48:57,240
A compliance policy gets created stating that agents can't connect to tools that aren't

1420
00:48:57,240 --> 00:49:00,880
in the approved catalog that policy lives in agent 365.

1421
00:49:00,880 --> 00:49:05,840
Every agent regardless of where it runs validates its tool connections through agent 365.

1422
00:49:05,840 --> 00:49:09,640
An agent tries to invoke a tool agent 365 checks the tool against the policy.

1423
00:49:09,640 --> 00:49:10,640
Is it approved?

1424
00:49:10,640 --> 00:49:13,640
If yes, the agent can call it if no, the request is blocked.

1425
00:49:13,640 --> 00:49:18,120
This works for the Azure agent, the AWS agent, the on-premises agent and the SAS agent,

1426
00:49:18,120 --> 00:49:20,840
same policy, same enforcement, different clouds.

1427
00:49:20,840 --> 00:49:23,640
The operational benefit is that you're not locked into one vendor.

1428
00:49:23,640 --> 00:49:27,280
You can choose the best runtime for each job without sacrificing governance.

1429
00:49:27,280 --> 00:49:28,600
Need a specialized service?

1430
00:49:28,600 --> 00:49:30,080
Use a third party SAS?

1431
00:49:30,080 --> 00:49:31,360
Need cost efficiency?

1432
00:49:31,360 --> 00:49:33,280
Use spot instances on AWS?

1433
00:49:33,280 --> 00:49:34,880
Need deep Azure integration?

1434
00:49:34,880 --> 00:49:36,040
Use Azure.

1435
00:49:36,040 --> 00:49:37,840
Each decision is made independently.

1436
00:49:37,840 --> 00:49:40,920
Governance is still unified because agent 365 sits above all of them.

1437
00:49:40,920 --> 00:49:42,960
This is what vendor flexibility looks like.

1438
00:49:42,960 --> 00:49:46,640
It's not being forced to choose one cloud because that's the only way to govern.

1439
00:49:46,640 --> 00:49:50,480
It's being able to choose based on actual capability while maintaining a single governance

1440
00:49:50,480 --> 00:49:52,160
layer across everything.

1441
00:49:52,160 --> 00:49:57,480
That flexibility is what lets the landing zone scale across an organization's entire AI footprint.

1442
00:49:57,480 --> 00:49:59,560
Life cycle management and ghost agents.

1443
00:49:59,560 --> 00:50:03,400
Here's something that happens in every organization that scales agents to a certain point.

1444
00:50:03,400 --> 00:50:04,400
Nobody plans for it.

1445
00:50:04,400 --> 00:50:05,400
Nobody expects it.

1446
00:50:05,400 --> 00:50:09,240
But by the time you notice, it's already a problem, an agent gets built, a team has a three-month

1447
00:50:09,240 --> 00:50:10,240
project.

1448
00:50:10,240 --> 00:50:11,240
They need automation.

1449
00:50:11,240 --> 00:50:12,240
They request an agent.

1450
00:50:12,240 --> 00:50:13,240
It works.

1451
00:50:13,240 --> 00:50:14,240
The project ends.

1452
00:50:14,240 --> 00:50:15,240
The team moves on.

1453
00:50:15,240 --> 00:50:16,240
The agent is still there.

1454
00:50:16,240 --> 00:50:17,240
Still running.

1455
00:50:17,240 --> 00:50:18,240
Still consuming tokens.

1456
00:50:18,240 --> 00:50:19,240
Still holding onto access.

1457
00:50:19,240 --> 00:50:20,240
It no longer needs.

1458
00:50:20,240 --> 00:50:21,240
Six months later.

1459
00:50:21,240 --> 00:50:22,880
Nobody remembers what it was for.

1460
00:50:22,880 --> 00:50:24,600
The person who built it left the company.

1461
00:50:24,600 --> 00:50:27,000
The original sponsor has moved to a different role.

1462
00:50:27,000 --> 00:50:29,240
The project might not even exist anymore.

1463
00:50:29,240 --> 00:50:30,720
But the agent keeps running.

1464
00:50:30,720 --> 00:50:31,720
It's invisible.

1465
00:50:31,720 --> 00:50:35,160
Until you look at your monthly cloud bill and wonder why it's higher than expected.

1466
00:50:35,160 --> 00:50:37,200
You trace the spend and find this ghost.

1467
00:50:37,200 --> 00:50:38,400
It's costing you money.

1468
00:50:38,400 --> 00:50:39,880
It's accessing data.

1469
00:50:39,880 --> 00:50:41,440
Nobody authorized it to access anymore.

1470
00:50:41,440 --> 00:50:45,440
It might be using models or tools that your organization no longer approves.

1471
00:50:45,440 --> 00:50:47,600
It's a security risk nobody knew existed.

1472
00:50:47,600 --> 00:50:51,560
Now multiply that by 50 or 100 as organizations build more agents.

1473
00:50:51,560 --> 00:50:53,800
The graveyard of forgotten projects grows.

1474
00:50:53,800 --> 00:50:55,080
These aren't catastrophic failures.

1475
00:50:55,080 --> 00:50:56,680
They aren't costing you millions.

1476
00:50:56,680 --> 00:50:58,560
But they're the death of a thousand cuts.

1477
00:50:58,560 --> 00:50:59,560
Waste it spend.

1478
00:50:59,560 --> 00:51:00,920
Slow erosion of governance.

1479
00:51:00,920 --> 00:51:05,560
Accumulation of risk that nobody's actively managing because the agents are just there.

1480
00:51:05,560 --> 00:51:07,960
This is where life cycle governance becomes critical.

1481
00:51:07,960 --> 00:51:11,760
Not as a nice to have compliance exercise, but as an operational necessity.

1482
00:51:11,760 --> 00:51:14,960
Without life cycle management agents are created but never retired.

1483
00:51:14,960 --> 00:51:15,960
They exist in a weird limbo.

1484
00:51:15,960 --> 00:51:17,440
They're still running.

1485
00:51:17,440 --> 00:51:18,520
Still consuming resources.

1486
00:51:18,520 --> 00:51:20,240
Still holding permissions.

1487
00:51:20,240 --> 00:51:21,960
But nobody claims responsibility for them.

1488
00:51:21,960 --> 00:51:23,680
The person who sponsored them is gone.

1489
00:51:23,680 --> 00:51:25,360
The team that built them has moved on.

1490
00:51:25,360 --> 00:51:26,800
Nobody knows if they're still needed.

1491
00:51:26,800 --> 00:51:28,960
Nobody knows who should decide whether to turn them off.

1492
00:51:28,960 --> 00:51:30,440
So they just keep running.

1493
00:51:30,440 --> 00:51:33,840
Agent 365 solves this by making the life cycle explicit.

1494
00:51:33,840 --> 00:51:37,800
When an agent registers in the control plane, it doesn't just get an identity and permissions.

1495
00:51:37,800 --> 00:51:39,000
It gets an owner.

1496
00:51:39,000 --> 00:51:40,000
That owner is accountable.

1497
00:51:40,000 --> 00:51:43,680
They're responsible for maintaining that agent, reviewing its access and ensuring it's

1498
00:51:43,680 --> 00:51:45,560
still doing what it's supposed to do.

1499
00:51:45,560 --> 00:51:47,760
But ownership alone doesn't prevent ghost agents.

1500
00:51:47,760 --> 00:51:48,760
You need visibility.

1501
00:51:48,760 --> 00:51:50,720
Agent 365 tracks usage.

1502
00:51:50,720 --> 00:51:52,600
Every call the agent makes gets logged.

1503
00:51:52,600 --> 00:51:54,640
Not just what it does, but how often it does it.

1504
00:51:54,640 --> 00:51:57,280
You can see if an agent hasn't been called in six months.

1505
00:51:57,280 --> 00:51:59,360
You can see if its permissions haven't been reviewed in a year.

1506
00:51:59,360 --> 00:52:00,760
You can see patterns of decay.

1507
00:52:00,760 --> 00:52:04,200
An agent that used to be called a hundred times a day now gets called once a week.

1508
00:52:04,200 --> 00:52:05,200
Something changed.

1509
00:52:05,200 --> 00:52:06,840
Either the business need went away.

1510
00:52:06,840 --> 00:52:08,840
Or the agent broke and nobody noticed.

1511
00:52:08,840 --> 00:52:12,240
This visibility enables decisions that are otherwise impossible.

1512
00:52:12,240 --> 00:52:14,200
You can't retire something you don't know is running.

1513
00:52:14,200 --> 00:52:16,160
You can't optimize something you can't see.

1514
00:52:16,160 --> 00:52:18,080
But with Agent 365, you have the data.

1515
00:52:18,080 --> 00:52:21,200
An owner gets notified that their agent hasn't been used in 90 days.

1516
00:52:21,200 --> 00:52:24,120
They have to decide, does this agent still serve a purpose?

1517
00:52:24,120 --> 00:52:25,760
If yes, they recertify it.

1518
00:52:25,760 --> 00:52:27,720
If no, they request retirement.

1519
00:52:27,720 --> 00:52:28,720
That's the governance model.

1520
00:52:28,720 --> 00:52:29,720
Not force.

1521
00:52:29,720 --> 00:52:30,720
Not threats.

1522
00:52:30,720 --> 00:52:32,640
Clarity and accountability.

1523
00:52:32,640 --> 00:52:34,800
The retirement process itself becomes structured.

1524
00:52:34,800 --> 00:52:36,520
An agent gets marked for retirement.

1525
00:52:36,520 --> 00:52:39,880
There's a grace period where the team can appeal or recertify it.

1526
00:52:39,880 --> 00:52:42,080
After the grace period, access gets revoked.

1527
00:52:42,080 --> 00:52:43,880
The agent can't call models anymore.

1528
00:52:43,880 --> 00:52:45,120
It can't access data.

1529
00:52:45,120 --> 00:52:46,200
It's still in the system.

1530
00:52:46,200 --> 00:52:48,280
You keep records for audit purposes.

1531
00:52:48,280 --> 00:52:49,360
But it's no longer active.

1532
00:52:49,360 --> 00:52:50,520
Costs stops accumulating.

1533
00:52:50,520 --> 00:52:51,520
Risk stops growing.

1534
00:52:51,520 --> 00:52:53,520
The resources it was holding are freed up.

1535
00:52:53,520 --> 00:52:54,880
Here's what's significant about this.

1536
00:52:54,880 --> 00:52:56,280
It's not a one-time cleanup.

1537
00:52:56,280 --> 00:52:57,560
It's an ongoing process.

1538
00:52:57,560 --> 00:52:59,200
Every agent has a life cycle.

1539
00:52:59,200 --> 00:53:00,200
Creation.

1540
00:53:00,200 --> 00:53:01,200
Active use.

1541
00:53:01,200 --> 00:53:02,200
Review.

1542
00:53:02,200 --> 00:53:03,200
Retirement.

1543
00:53:03,200 --> 00:53:05,680
Organizations that build sustainable AI practices treat agent life cycle the way they

1544
00:53:05,680 --> 00:53:07,880
treat employee life cycle.

1545
00:53:07,880 --> 00:53:09,840
Onboarding with clear expectations.

1546
00:53:09,840 --> 00:53:10,680
Regular check-ins.

1547
00:53:10,680 --> 00:53:11,680
Annual reviews.

1548
00:53:11,680 --> 00:53:13,120
Offboarding when the need ends.

1549
00:53:13,120 --> 00:53:15,480
The financial impact is measurable.

1550
00:53:15,480 --> 00:53:19,520
Organizations that implement life cycle governance report 15 to 30% cost reduction from

1551
00:53:19,520 --> 00:53:23,120
shutting down ghost agents and decommissioning redundant ones.

1552
00:53:23,120 --> 00:53:24,880
But the security impact is more important.

1553
00:53:24,880 --> 00:53:27,920
Every retired agent is a surface you no longer have to monitor.

1554
00:53:27,920 --> 00:53:31,400
Every unused permission you revoke is a tax surface you've eliminated.

1555
00:53:31,400 --> 00:53:34,520
Every ghost you eliminate is a potential liability you've erased.

1556
00:53:34,520 --> 00:53:37,840
The shift here is from treating agents as permanent fixtures.

1557
00:53:37,840 --> 00:53:41,480
To treating them as temporary solutions with built-in expiration dates, that's what responsible

1558
00:53:41,480 --> 00:53:43,720
AI operations looks like.

1559
00:53:43,720 --> 00:53:45,080
The observability layer.

1560
00:53:45,080 --> 00:53:46,720
You can't govern what you can't see.

1561
00:53:46,720 --> 00:53:49,160
That is the principle that ties this entire system together.

1562
00:53:49,160 --> 00:53:50,320
You've built the identity.

1563
00:53:50,320 --> 00:53:51,840
You've built the policy enforcement.

1564
00:53:51,840 --> 00:53:53,520
You've built the life cycle management.

1565
00:53:53,520 --> 00:53:57,160
But if you can't trace what actually happened, you're still flying blind.

1566
00:53:57,160 --> 00:53:59,840
Observability is the thread that connects every decision an agent makes back to the

1567
00:53:59,840 --> 00:54:01,240
context that prompted it.

1568
00:54:01,240 --> 00:54:02,480
It's not just logging.

1569
00:54:02,480 --> 00:54:04,480
Logging is recording that something happened.

1570
00:54:04,480 --> 00:54:06,480
Observability is understanding why it happened.

1571
00:54:06,480 --> 00:54:07,880
Who initiated the request?

1572
00:54:07,880 --> 00:54:09,200
What data informed the choice?

1573
00:54:09,200 --> 00:54:10,840
What were the actual consequences?

1574
00:54:10,840 --> 00:54:14,120
When you have true observability and agent's action is never isolated.

1575
00:54:14,120 --> 00:54:17,840
It is always traceable back through the entire chain of events that led to that specific

1576
00:54:17,840 --> 00:54:18,840
moment.

1577
00:54:18,840 --> 00:54:19,920
Here's the problem in practice.

1578
00:54:19,920 --> 00:54:23,920
An agent in your organization makes a recommendation that turns out to be wrong.

1579
00:54:23,920 --> 00:54:26,640
A customer acts on that bad advice and loses money.

1580
00:54:26,640 --> 00:54:30,720
Now you have a massive problem and you need to understand exactly what broke.

1581
00:54:30,720 --> 00:54:33,120
Without observability, you're stuck.

1582
00:54:33,120 --> 00:54:35,800
You know the output was incorrect, but you have no idea why.

1583
00:54:35,800 --> 00:54:37,600
Was it using stale data from the warehouse?

1584
00:54:37,600 --> 00:54:39,280
Did the ontology have a structural flaw?

1585
00:54:39,280 --> 00:54:41,960
Did the agent follow a path that shouldn't have been able to access?

1586
00:54:41,960 --> 00:54:43,040
You're just guessing.

1587
00:54:43,040 --> 00:54:45,880
But with observability, you trace the logic backward.

1588
00:54:45,880 --> 00:54:50,040
You see the exact prompt that triggered the agent and the specific data sources it consulted.

1589
00:54:50,040 --> 00:54:54,000
You see which relationships in the ontology it traversed and which policy decisions the

1590
00:54:54,000 --> 00:54:56,040
gateway made before calling the model.

1591
00:54:56,040 --> 00:54:57,680
You see the intermediate reasoning steps.

1592
00:54:57,680 --> 00:54:59,600
You see where the logic actually broke down.

1593
00:54:59,600 --> 00:55:03,080
The result is that you understand not just that something went wrong, but where

1594
00:55:03,080 --> 00:55:04,080
and why.

1595
00:55:04,080 --> 00:55:07,800
The architecture here is a centralized system that ties every single event together.

1596
00:55:07,800 --> 00:55:10,840
Every agent action gets logged with its full context attached.

1597
00:55:10,840 --> 00:55:14,600
The agent's identity, the user who triggered it, the data it accessed, the policies that

1598
00:55:14,600 --> 00:55:18,960
were evaluated, all of this flows into one place where it can be analyzed.

1599
00:55:18,960 --> 00:55:21,920
This is fundamentally different from standard application logging.

1600
00:55:21,920 --> 00:55:25,840
A typical app logs simple actions like a user clicking a button or a page loading.

1601
00:55:25,840 --> 00:55:28,720
An agent observability system logs the entire decision chain.

1602
00:55:28,720 --> 00:55:33,960
It shows that a specific analysis agent was called by a specific user at 247 pm.

1603
00:55:33,960 --> 00:55:39,000
It records that the system evaluated a policy for sales roles, satisfied that policy, and

1604
00:55:39,000 --> 00:55:41,560
then queered the ontology for high value customers.

1605
00:55:41,560 --> 00:55:46,200
It shows the agent checked order history, invoked a pricing model, and returned 12 candidates

1606
00:55:46,200 --> 00:55:47,760
with specific confidence scores.

1607
00:55:47,760 --> 00:55:49,080
It even tracks the cost.

1608
00:55:49,080 --> 00:55:53,680
Noting that the 0.47 tokens used should be built to sales operations.

1609
00:55:53,680 --> 00:55:55,400
That is a massive amount of information.

1610
00:55:55,400 --> 00:55:58,040
And because it's all connected, you can verify the logic.

1611
00:55:58,040 --> 00:56:01,760
If a customer was excluded from a list, you can see exactly which rule bumped them out.

1612
00:56:01,760 --> 00:56:03,240
This changes incident response.

1613
00:56:03,240 --> 00:56:07,280
In organizations without this layer, investigations are slow because fragments of info are scattered

1614
00:56:07,280 --> 00:56:10,640
across cloud logs, database logs, and policy logs.

1615
00:56:10,640 --> 00:56:11,960
None of them talk to each other.

1616
00:56:11,960 --> 00:56:15,520
You're trying to reassemble a broken picture from disconnected pieces.

1617
00:56:15,520 --> 00:56:17,720
With observability, the picture is already assembled.

1618
00:56:17,720 --> 00:56:20,640
The investigation becomes a simple search instead of an archaeological dig.

1619
00:56:20,640 --> 00:56:22,520
The compliance benefit is just as big.

1620
00:56:22,520 --> 00:56:25,920
When a regulator asks if you're monitoring your agents, you can actually prove it.

1621
00:56:25,920 --> 00:56:29,440
You have the logs to show that governance wasn't just a policy on paper.

1622
00:56:29,440 --> 00:56:30,960
It was enforced in real time.

1623
00:56:30,960 --> 00:56:33,800
But beyond compliance, this enables constant improvement.

1624
00:56:33,800 --> 00:56:37,400
You can finally see which agents are making good decisions and where the gaps are in your

1625
00:56:37,400 --> 00:56:38,400
ontology.

1626
00:56:38,400 --> 00:56:41,120
You can't answer those questions if you can't see the work.

1627
00:56:41,120 --> 00:56:45,120
With data, you optimize based on evidence instead of just hoping for the best.

1628
00:56:45,120 --> 00:56:48,120
The shift is moving from hoping things work to knowing they work.

1629
00:56:48,120 --> 00:56:52,680
It's moving from compliance theatre to actual control.

1630
00:56:52,680 --> 00:56:54,760
The shift in organizational structure.

1631
00:56:54,760 --> 00:56:58,960
Everything we've discussed, identity, reasoning, governance, and observability, those are

1632
00:56:58,960 --> 00:57:00,440
infrastructure layers.

1633
00:57:00,440 --> 00:57:02,360
But infrastructure doesn't exist in a vacuum.

1634
00:57:02,360 --> 00:57:03,960
It changes how organizations work.

1635
00:57:03,960 --> 00:57:05,200
It changes what teams do.

1636
00:57:05,200 --> 00:57:06,960
It changes where the power lives.

1637
00:57:06,960 --> 00:57:10,480
When you don't have a landing zone, AI adoption is a mess.

1638
00:57:10,480 --> 00:57:13,560
Every team that wants to build an agent has to start from scratch.

1639
00:57:13,560 --> 00:57:15,280
They figure out identity on their own.

1640
00:57:15,280 --> 00:57:16,960
They decide how to connect to models.

1641
00:57:16,960 --> 00:57:18,600
They define their own policies.

1642
00:57:18,600 --> 00:57:19,720
Each project is an island.

1643
00:57:19,720 --> 00:57:21,720
Each team spends months reinventing the wheel.

1644
00:57:21,720 --> 00:57:26,040
You end up with fragments scattered across different clouds, using different standards

1645
00:57:26,040 --> 00:57:27,440
and enforcing different rules.

1646
00:57:27,440 --> 00:57:28,960
It might work in isolation.

1647
00:57:28,960 --> 00:57:33,120
But when you look at the company's total AI footprint, it's chaos.

1648
00:57:33,120 --> 00:57:35,560
With a landing zone, that structure inverts.

1649
00:57:35,560 --> 00:57:37,080
There is a central platform team.

1650
00:57:37,080 --> 00:57:38,440
Their job isn't to build the agents.

1651
00:57:38,440 --> 00:57:41,160
It's to build the platform that makes building agents possible.

1652
00:57:41,160 --> 00:57:42,560
They design the identity model.

1653
00:57:42,560 --> 00:57:44,000
They set up the governance hub.

1654
00:57:44,000 --> 00:57:45,520
They maintain the ontologies.

1655
00:57:45,520 --> 00:57:47,560
Everything else is built on top of that foundation.

1656
00:57:47,560 --> 00:57:51,000
This changes what it means to be a project team.

1657
00:57:51,000 --> 00:57:52,120
We have started with a blank slate.

1658
00:57:52,120 --> 00:57:53,960
A team starts with an inheritance.

1659
00:57:53,960 --> 00:57:55,560
The identity layer is ready.

1660
00:57:55,560 --> 00:57:57,800
And the policy framework already exists.

1661
00:57:57,800 --> 00:57:59,400
The team doesn't have to invent the plumbing.

1662
00:57:59,400 --> 00:58:03,800
They focus on what makes their agent unique, like the business logic or the specific workflow.

1663
00:58:03,800 --> 00:58:07,680
They get to market faster because they aren't rebuilding the governance layer every single

1664
00:58:07,680 --> 00:58:08,680
time.

1665
00:58:08,680 --> 00:58:10,560
The real shift is in who makes the decisions.

1666
00:58:10,560 --> 00:58:13,520
Without a landing zone, every team decides things independently.

1667
00:58:13,520 --> 00:58:16,920
One team lets agents call any model while another team stays restricted.

1668
00:58:16,920 --> 00:58:19,920
One team logs everything and the next logs nothing at all.

1669
00:58:19,920 --> 00:58:23,800
You get inconsistency because there's nobody there to enforce a standard.

1670
00:58:23,800 --> 00:58:26,040
With a landing zone, the big decisions move up.

1671
00:58:26,040 --> 00:58:31,200
The organization decides on identity standards, approved models and cost controls.

1672
00:58:31,200 --> 00:58:33,480
These become central rules, not suggestions.

1673
00:58:33,480 --> 00:58:34,480
But here's the thing.

1674
00:58:34,480 --> 00:58:36,160
This doesn't slow teams down.

1675
00:58:36,160 --> 00:58:39,760
It actually speeds them up because they aren't wasting time debating the fundamentals.

1676
00:58:39,760 --> 00:58:42,840
They can focus entirely on solving their actual business problem.

1677
00:58:42,840 --> 00:58:45,480
The structure that emerges is a hub and spoke model.

1678
00:58:45,480 --> 00:58:47,320
The central platform team is the hub.

1679
00:58:47,320 --> 00:58:50,600
They provide the shared infrastructure and enforce the baseline standards.

1680
00:58:50,600 --> 00:58:52,040
The business units are the spokes.

1681
00:58:52,040 --> 00:58:54,760
They use that infrastructure to build specific use cases.

1682
00:58:54,760 --> 00:58:58,520
There is finally a clear boundary between what is centralized and what is distributed.

1683
00:58:58,520 --> 00:59:01,720
This solves the coordination problem that kills most companies as they scale.

1684
00:59:01,720 --> 00:59:06,160
Without this model, coordination costs explode as you add more agents.

1685
00:59:06,160 --> 00:59:09,960
Policy discussions never end because there are no actual policies to follow.

1686
00:59:09,960 --> 00:59:13,280
Incidents become nightmares because you are trying to fix incompatible systems.

1687
00:59:13,280 --> 00:59:16,600
The whole organization grinds to a halt because everyone is working against different

1688
00:59:16,600 --> 00:59:17,760
infrastructure.

1689
00:59:17,760 --> 00:59:19,640
With a landing zone, you coordinate once.

1690
00:59:19,640 --> 00:59:22,960
The platform team designs the system and the spokes operate within it.

1691
00:59:22,960 --> 00:59:26,440
New teams onboard faster because they are using proven patterns.

1692
00:59:26,440 --> 00:59:30,040
Policy changes hit the entire company at once because there is one place to flip the switch.

1693
00:59:30,040 --> 00:59:32,280
The time to value metric changes completely.

1694
00:59:32,280 --> 00:59:35,640
A team without a platform might take three months to build an agent spending two of those

1695
00:59:35,640 --> 00:59:37,400
months just on the infrastructure.

1696
00:59:37,400 --> 00:59:39,480
A team with a platform can do it in two weeks.

1697
00:59:39,480 --> 00:59:41,160
That isn't a small improvement.

1698
00:59:41,160 --> 00:59:43,680
That is the difference between an experiment and a deployment.

1699
00:59:43,680 --> 00:59:46,440
It's the difference between a proof of concept and actual production.

1700
00:59:46,440 --> 00:59:51,400
This shift is what transforms AI from a technical toy into a real operational advantage.

1701
00:59:51,400 --> 00:59:53,640
You are no longer managing a bunch of isolated projects.

1702
00:59:53,640 --> 00:59:56,760
You are operating a platform, security and threat detection.

1703
00:59:56,760 --> 01:00:00,880
The moment you decentralize governance is the moment your security perimeter collapses.

1704
01:00:00,880 --> 01:00:01,880
That is the hard truth.

1705
01:00:01,880 --> 01:00:05,080
Most organizations do not grasp until something goes wrong.

1706
01:00:05,080 --> 01:00:06,840
Consider what happens without a landing zone.

1707
01:00:06,840 --> 01:00:10,880
You have an agent built by a contractor that is running somewhere in your environment.

1708
01:00:10,880 --> 01:00:15,080
It has access to a specific set of systems that access was granted months ago.

1709
01:00:15,080 --> 01:00:16,440
Security has reviewed it since.

1710
01:00:16,440 --> 01:00:19,520
The contractor relationship ended but the agent is still running.

1711
01:00:19,520 --> 01:00:21,400
One day that agent gets compromised.

1712
01:00:21,400 --> 01:00:26,080
Maybe the person who built it handed over credentials or perhaps a vulnerability in the code was exploited.

1713
01:00:26,080 --> 01:00:28,440
The point is, control is lost.

1714
01:00:28,440 --> 01:00:31,120
Now that compromised agent can move laterally, it has credentials.

1715
01:00:31,120 --> 01:00:32,520
It knows how to call your APIs.

1716
01:00:32,520 --> 01:00:33,720
It can access your data.

1717
01:00:33,720 --> 01:00:35,040
But here is the problem.

1718
01:00:35,040 --> 01:00:38,680
Because the agent was built independently, it was never wired into your threat detection

1719
01:00:38,680 --> 01:00:39,680
systems.

1720
01:00:39,680 --> 01:00:40,920
Defender does not know it exists.

1721
01:00:40,920 --> 01:00:45,240
Your security team has no baseline for what normal behavior looks like for this specific tool.

1722
01:00:45,240 --> 01:00:46,480
So it starts accessing data.

1723
01:00:46,480 --> 01:00:47,640
It should not touch.

1724
01:00:47,640 --> 01:00:49,520
It starts making calls to external systems.

1725
01:00:49,520 --> 01:00:51,720
It starts exfiltrating information.

1726
01:00:51,720 --> 01:00:55,200
And you do not detect it until weeks later when the damage is already done.

1727
01:00:55,200 --> 01:00:59,560
You discover it not because your system is courted, but because a customer reports a problem

1728
01:00:59,560 --> 01:01:02,720
or an auditor finds suspicious patterns in the logs.

1729
01:01:02,720 --> 01:01:06,120
This is a breach that happens in slow motion while you are looking somewhere else.

1730
01:01:06,120 --> 01:01:07,800
With a landing zone, the story changes.

1731
01:01:07,800 --> 01:01:12,920
That same agent exists, but it is integrated into your security infrastructure from day one.

1732
01:01:12,920 --> 01:01:14,600
Defender monitors it constantly.

1733
01:01:14,600 --> 01:01:19,520
Every API call the agent makes gets logged and evaluated against behavioral baselines.

1734
01:01:19,520 --> 01:01:24,160
Defender understands that this agent normally calls the revenue API and the customer database.

1735
01:01:24,160 --> 01:01:28,280
Those are normal patterns, but today it is trying to call the financial export service

1736
01:01:28,280 --> 01:01:30,480
and the employee records database.

1737
01:01:30,480 --> 01:01:32,000
Those are outside the baseline.

1738
01:01:32,000 --> 01:01:33,000
And normally detected.

1739
01:01:33,000 --> 01:01:34,640
The request gets blocked immediately.

1740
01:01:34,640 --> 01:01:36,320
The security team gets an alert.

1741
01:01:36,320 --> 01:01:38,760
They can investigate before any real damage occurs.

1742
01:01:38,760 --> 01:01:40,480
This is defense in depth in practice.

1743
01:01:40,480 --> 01:01:42,760
You do not rely on one single control.

1744
01:01:42,760 --> 01:01:43,760
Identity is the first layer.

1745
01:01:43,760 --> 01:01:48,240
The agent has its own identity, so you know it is the agent and not someone else using

1746
01:01:48,240 --> 01:01:49,240
its account.

1747
01:01:49,240 --> 01:01:52,840
Policy enforcement is the second layer, even if the agent has an identity, it can only

1748
01:01:52,840 --> 01:01:56,120
call approved APIs based on policies you have defined.

1749
01:01:56,120 --> 01:01:57,960
Threat detection is the third layer.

1750
01:01:57,960 --> 01:02:01,640
Even if the agent is allowed to call certain APIs if it calls them in abnormal ways, the

1751
01:02:01,640 --> 01:02:02,800
system notices.

1752
01:02:02,800 --> 01:02:06,400
Other trails are the fourth layer, even if something slips through, you have a complete record

1753
01:02:06,400 --> 01:02:08,040
of exactly what happened.

1754
01:02:08,040 --> 01:02:11,000
Most organizations build security as an afterthought.

1755
01:02:11,000 --> 01:02:13,640
You build the agent first, get it working and ship it.

1756
01:02:13,640 --> 01:02:17,480
Then you add monitoring, then you add policies, then you try to integrate it with your security

1757
01:02:17,480 --> 01:02:18,480
stack.

1758
01:02:18,480 --> 01:02:21,360
By that point, the agent has already been operating unsupervised.

1759
01:02:21,360 --> 01:02:24,880
It is already running, without proper oversight, you are playing catch up.

1760
01:02:24,880 --> 01:02:27,280
A landing zone inverts that model.

1761
01:02:27,280 --> 01:02:28,280
Security is not added.

1762
01:02:28,280 --> 01:02:29,280
It is built in.

1763
01:02:29,280 --> 01:02:30,480
An agent cannot run without an identity.

1764
01:02:30,480 --> 01:02:32,280
The identity comes with default policies.

1765
01:02:32,280 --> 01:02:35,080
These policies are evaluated by the gateway in real time.

1766
01:02:35,080 --> 01:02:36,680
That evaluation gets logged.

1767
01:02:36,680 --> 01:02:40,200
Threat detection systems monitor the agent behavior based on patterns you have defined.

1768
01:02:40,200 --> 01:02:41,720
The agent is not a special case.

1769
01:02:41,720 --> 01:02:45,840
It is a managed entity with the same level of security scrutiny as any user or application.

1770
01:02:45,840 --> 01:02:47,520
The operational shift is profound.

1771
01:02:47,520 --> 01:02:49,520
Without integration, security is reactive.

1772
01:02:49,520 --> 01:02:50,520
Something goes wrong.

1773
01:02:50,520 --> 01:02:51,520
You investigate.

1774
01:02:51,520 --> 01:02:52,720
You respond.

1775
01:02:52,720 --> 01:02:55,920
Organizations spend months recovering from these breaches.

1776
01:02:55,920 --> 01:02:58,240
With a landing zone security is proactive.

1777
01:02:58,240 --> 01:03:00,440
You detect anomalies before they become breaches.

1778
01:03:00,440 --> 01:03:04,400
You contain threats immediately because the infrastructure is already built to do it.

1779
01:03:04,400 --> 01:03:06,920
This changes the conversation with your security team.

1780
01:03:06,920 --> 01:03:10,360
Instead of arguing about whether agents should be monitored, you are discussing how they

1781
01:03:10,360 --> 01:03:11,560
should be monitored.

1782
01:03:11,560 --> 01:03:14,200
The infrastructure already assumes threats will be attempted.

1783
01:03:14,200 --> 01:03:17,440
Your job is to make sure they are detected and contained.

1784
01:03:17,440 --> 01:03:19,080
The transition strategy.

1785
01:03:19,080 --> 01:03:23,560
Everything we have discussed is conceptual until you actually move from one model to another.

1786
01:03:23,560 --> 01:03:25,480
And this is where organizations stumble.

1787
01:03:25,480 --> 01:03:27,560
Because the temptation is to do it all at once.

1788
01:03:27,560 --> 01:03:31,560
You can't even think about the old system.

1789
01:03:31,560 --> 01:03:35,360
The jump from co-pilot as a feature to agent mash as infrastructure requires deliberate

1790
01:03:35,360 --> 01:03:36,360
steps.

1791
01:03:36,360 --> 01:03:37,960
The sequence matters more than the speed.

1792
01:03:37,960 --> 01:03:41,560
Start with inventory, not design, not architecture and not planning.

1793
01:03:41,560 --> 01:03:42,560
Inventory.

1794
01:03:42,560 --> 01:03:43,880
You need to know what you actually have.

1795
01:03:43,880 --> 01:03:45,160
How many agents are running?

1796
01:03:45,160 --> 01:03:46,160
Where are they running?

1797
01:03:46,160 --> 01:03:47,160
What are they doing?

1798
01:03:47,160 --> 01:03:48,160
Who built them?

1799
01:03:48,160 --> 01:03:49,160
Who maintains them?

1800
01:03:49,160 --> 01:03:50,160
This is the discovery phase.

1801
01:03:50,160 --> 01:03:53,560
And it is uncomfortable because you will find things you did not know existed.

1802
01:03:53,560 --> 01:03:56,880
You will find agents running in development that were promoted to production two years

1803
01:03:56,880 --> 01:03:57,880
ago and forgotten.

1804
01:03:57,880 --> 01:04:02,120
You will find agents built by contractors that are still calling external APIs.

1805
01:04:02,120 --> 01:04:05,440
You will find agents that were supposed to be retired but keep running because nobody

1806
01:04:05,440 --> 01:04:06,440
pulled the plug.

1807
01:04:06,440 --> 01:04:08,200
You will find chaos.

1808
01:04:08,200 --> 01:04:11,960
But chaos that you can see is better than chaos you do not know about.

1809
01:04:11,960 --> 01:04:12,960
Document what you find.

1810
01:04:12,960 --> 01:04:18,200
Agent name, owner, location, access level, data sources and model dependencies.

1811
01:04:18,200 --> 01:04:20,280
You are building an inventory, not a solution yet.

1812
01:04:20,280 --> 01:04:22,320
This phase takes weeks, not months.

1813
01:04:22,320 --> 01:04:24,160
You are not fixing anything.

1814
01:04:24,160 --> 01:04:25,960
You are just seeing what is there.

1815
01:04:25,960 --> 01:04:28,320
Once you have visibility, design the landing zone.

1816
01:04:28,320 --> 01:04:30,560
This is where you define the target architecture.

1817
01:04:30,560 --> 01:04:31,560
Not all of it.

1818
01:04:31,560 --> 01:04:33,640
Just the version you are going to build first.

1819
01:04:33,640 --> 01:04:34,920
Decide on your identity model.

1820
01:04:34,920 --> 01:04:36,720
Will it be an agent ID exclusively?

1821
01:04:36,720 --> 01:04:38,960
Will you support hybrid models during the transition?

1822
01:04:38,960 --> 01:04:42,200
How will you handle legacy agents that cannot migrate immediately?

1823
01:04:42,200 --> 01:04:43,720
Design your governance hub.

1824
01:04:43,720 --> 01:04:45,120
What models will be approved?

1825
01:04:45,120 --> 01:04:46,120
What tools?

1826
01:04:46,120 --> 01:04:47,120
What policies?

1827
01:04:47,120 --> 01:04:48,120
Design your data plane.

1828
01:04:48,120 --> 01:04:49,680
Which ontologies will you build?

1829
01:04:49,680 --> 01:04:51,880
Which existing data sources will you expose?

1830
01:04:51,880 --> 01:04:54,360
What migration path makes sense for your data?

1831
01:04:54,360 --> 01:04:55,880
Design your compliance posture.

1832
01:04:55,880 --> 01:04:57,360
What regulations matter?

1833
01:04:57,360 --> 01:05:00,360
What policies need to be enforced at the infrastructure level?

1834
01:05:00,360 --> 01:05:01,720
Design your cost model.

1835
01:05:01,720 --> 01:05:03,160
How will you attribute spend?

1836
01:05:03,160 --> 01:05:04,480
What budgets will you set?

1837
01:05:04,480 --> 01:05:05,680
This is not construction.

1838
01:05:05,680 --> 01:05:07,000
It is specification.

1839
01:05:07,000 --> 01:05:11,200
You are defining the target state clearly, so teams know what they are moving toward.

1840
01:05:11,200 --> 01:05:12,520
This phase takes weeks.

1841
01:05:12,520 --> 01:05:14,120
It is collaborative.

1842
01:05:14,120 --> 01:05:18,080
Security compliance, IT and business teams should all have input because they will all be

1843
01:05:18,080 --> 01:05:19,080
affected.

1844
01:05:19,080 --> 01:05:21,320
Migration happens in phases, not all at once.

1845
01:05:21,320 --> 01:05:22,680
And the order matters.

1846
01:05:22,680 --> 01:05:24,200
Start with high risk agents.

1847
01:05:24,200 --> 01:05:28,000
The ones accessing sensitive data, the ones in regulated industries, the ones that are

1848
01:05:28,000 --> 01:05:32,280
business critical, not because they are easiest, but because they are most important.

1849
01:05:32,280 --> 01:05:33,600
These are your proof points.

1850
01:05:33,600 --> 01:05:35,120
This is where governance matters most.

1851
01:05:35,120 --> 01:05:40,040
Migrate a small batch, prove the pattern works, document the process, and then scale.

1852
01:05:40,040 --> 01:05:42,760
Real organizations do not move every agent simultaneously.

1853
01:05:42,760 --> 01:05:44,080
They move in waves.

1854
01:05:44,080 --> 01:05:46,160
Month one, critical financial agents.

1855
01:05:46,160 --> 01:05:48,880
Month two, compliance related agents.

1856
01:05:48,880 --> 01:05:50,320
Month three.

1857
01:05:50,320 --> 01:05:51,920
Customer facing agents.

1858
01:05:51,920 --> 01:05:53,360
Month four.

1859
01:05:53,360 --> 01:05:55,120
Personal operations agents.

1860
01:05:55,120 --> 01:05:56,440
Month five.

1861
01:05:56,440 --> 01:05:58,280
Experimental and lower risk agents.

1862
01:05:58,280 --> 01:05:59,280
Month six.

1863
01:05:59,280 --> 01:06:00,720
Clean up and sunset of legacy systems.

1864
01:06:00,720 --> 01:06:01,720
This is not slow.

1865
01:06:01,720 --> 01:06:02,720
It is methodical.

1866
01:06:02,720 --> 01:06:06,280
And it is essential because every batch of migrations teaches you something.

1867
01:06:06,280 --> 01:06:08,640
You will find problems you did not anticipate.

1868
01:06:08,640 --> 01:06:11,520
You will discover policy configurations that need adjustment.

1869
01:06:11,520 --> 01:06:13,280
You will identify tooling gaps.

1870
01:06:13,280 --> 01:06:15,960
You can only learn this at scale, not in pilots.

1871
01:06:15,960 --> 01:06:17,080
Establish governance as you move.

1872
01:06:17,080 --> 01:06:18,080
Not after.

1873
01:06:18,080 --> 01:06:19,600
This is the principle most organizations.

1874
01:06:19,600 --> 01:06:20,600
Missed.

1875
01:06:20,600 --> 01:06:21,800
They migrate the agents first and the governance later.

1876
01:06:21,800 --> 01:06:22,800
That is backward.

1877
01:06:22,800 --> 01:06:24,720
governance first, migration second.

1878
01:06:24,720 --> 01:06:27,600
Before an agent moves to the new system, it gets an identity.

1879
01:06:27,600 --> 01:06:28,920
It gets enrolled in policies.

1880
01:06:28,920 --> 01:06:30,560
It gets connected to observability.

1881
01:06:30,560 --> 01:06:32,160
It gets its life cycle managed.

1882
01:06:32,160 --> 01:06:33,880
That agent is not free floating.

1883
01:06:33,880 --> 01:06:36,920
From day one, it is operating within the constraints you have defined.

1884
01:06:36,920 --> 01:06:41,080
Yes, this slows down the initial migration, but it prevents the chaos of moving infrastructure

1885
01:06:41,080 --> 01:06:42,080
without control.

1886
01:06:42,080 --> 01:06:45,160
By the time you finish migration governance is already embedded.

1887
01:06:45,160 --> 01:06:46,960
You are not bolting it on afterward.

1888
01:06:46,960 --> 01:06:48,480
Here is what a real timeline looks like.

1889
01:06:48,480 --> 01:06:52,040
A large organization with 300 agents needs about six months.

1890
01:06:52,040 --> 01:06:54,320
One and two are inventory and design.

1891
01:06:54,320 --> 01:06:56,560
Months three through five are phased migration.

1892
01:06:56,560 --> 01:06:58,280
Month six is clean up in stabilization.

1893
01:06:58,280 --> 01:07:00,080
At the end, you have an agent mesh.

1894
01:07:00,080 --> 01:07:01,840
All agents have traceable identities.

1895
01:07:01,840 --> 01:07:03,760
All policies are enforced consistently.

1896
01:07:03,760 --> 01:07:05,400
All ontologies are shared.

1897
01:07:05,400 --> 01:07:06,720
All costs are visible.

1898
01:07:06,720 --> 01:07:09,400
All compliance is structural, not just documented.

1899
01:07:09,400 --> 01:07:13,720
The key principle underneath all of this is simple, but counter-intuitive.

1900
01:07:13,720 --> 01:07:15,480
Governance comes before innovation.

1901
01:07:15,480 --> 01:07:18,480
Not as a slogan or a policy document, but structurally.

1902
01:07:18,480 --> 01:07:20,160
You build the constraints first.

1903
01:07:20,160 --> 01:07:24,840
When teams innovate within those constraints, it feels restrictive, but it is actually liberating.

1904
01:07:24,840 --> 01:07:29,040
Because once the foundation is set, teams move fast knowing their innovation won't create

1905
01:07:29,040 --> 01:07:31,760
chaos downstream in the future.

1906
01:07:31,760 --> 01:07:32,920
Autonomous operations.

1907
01:07:32,920 --> 01:07:37,120
Once you have the foundation in place, identity, reasoning, governance and observability,

1908
01:07:37,120 --> 01:07:39,240
the math of what's possible changes.

1909
01:07:39,240 --> 01:07:41,120
Not in a theoretical way, in an operational way.

1910
01:07:41,120 --> 01:07:44,120
Imagine an operations agent in a manufacturing plant.

1911
01:07:44,120 --> 01:07:46,400
It's monitoring production KPIs in real time.

1912
01:07:46,400 --> 01:07:51,200
It understands the ontology that defines what normal looks like for this specific business.

1913
01:07:51,200 --> 01:07:54,000
Line 3 should be hitting 500 units every hour.

1914
01:07:54,000 --> 01:07:56,560
Right now it's at 4.20, that's below the threshold.

1915
01:07:56,560 --> 01:07:58,920
It isn't a catastrophe yet, but it's a deviation.

1916
01:07:58,920 --> 01:08:00,680
The agent's instructions are clear.

1917
01:08:00,680 --> 01:08:03,120
Flag the minor issues and escalate the major ones.

1918
01:08:03,120 --> 01:08:04,520
First it checks the context.

1919
01:08:04,520 --> 01:08:05,520
Is today a training day?

1920
01:08:05,520 --> 01:08:06,520
No.

1921
01:08:06,520 --> 01:08:07,520
Is there maintenance scheduled?

1922
01:08:07,520 --> 01:08:09,000
No.

1923
01:08:09,000 --> 01:08:12,720
Because the ontology defines the relationships between equipment and output, the agent can

1924
01:08:12,720 --> 01:08:14,840
traverse those parts to find a root cause.

1925
01:08:14,840 --> 01:08:19,160
It sees a 40% correlation between this specific slowdown and equipment degradation.

1926
01:08:19,160 --> 01:08:21,160
It pulls the recent maintenance records.

1927
01:08:21,160 --> 01:08:22,680
Everything looked fine during the last service.

1928
01:08:22,680 --> 01:08:24,360
It checks the raw material quality.

1929
01:08:24,360 --> 01:08:25,360
Those metrics are normal.

1930
01:08:25,360 --> 01:08:26,520
It looks at labor data.

1931
01:08:26,520 --> 01:08:27,520
No absences.

1932
01:08:27,520 --> 01:08:29,080
No schedule shifts.

1933
01:08:29,080 --> 01:08:32,920
Because the ontology constrains the agent to only relevant data, it can surface a specific

1934
01:08:32,920 --> 01:08:33,920
hypothesis.

1935
01:08:33,920 --> 01:08:36,520
There is a minor calibration drift on the material feeder.

1936
01:08:36,520 --> 01:08:38,960
But here's where autonomy actually starts to mean something.

1937
01:08:38,960 --> 01:08:41,160
The agent doesn't just send a report and wait.

1938
01:08:41,160 --> 01:08:45,080
It's authorized to act.

1939
01:08:45,080 --> 01:08:46,520
A calibration check.

1940
01:08:46,520 --> 01:08:48,800
That request doesn't just fire off into the void.

1941
01:08:48,800 --> 01:08:50,080
It flows through policy gates.

1942
01:08:50,080 --> 01:08:52,040
The system checks the agent's identity.

1943
01:08:52,040 --> 01:08:53,960
Does it have permission to request maintenance?

1944
01:08:53,960 --> 01:08:54,960
Yes.

1945
01:08:54,960 --> 01:08:55,960
Is it targeting an approved system?

1946
01:08:55,960 --> 01:08:56,960
Yes.

1947
01:08:56,960 --> 01:08:57,960
Is there sensitive data involved?

1948
01:08:57,960 --> 01:08:58,960
No.

1949
01:08:58,960 --> 01:08:59,960
The request passes policy.

1950
01:08:59,960 --> 01:09:00,960
It hits the maintenance system.

1951
01:09:00,960 --> 01:09:04,040
A technician reviews the reasoning sees the data and hits a prove.

1952
01:09:04,040 --> 01:09:05,560
The calibration happens.

1953
01:09:05,560 --> 01:09:08,480
Production climbs back up to 495 units per hour.

1954
01:09:08,480 --> 01:09:09,880
The problem is solved.

1955
01:09:09,880 --> 01:09:11,960
But here is the shift that isn't automation.

1956
01:09:11,960 --> 01:09:13,400
That's orchestration.

1957
01:09:13,400 --> 01:09:16,680
The agent didn't just decide things in a vacuum, but it also didn't sit around waiting for

1958
01:09:16,680 --> 01:09:18,520
a human to tell it where to look.

1959
01:09:18,520 --> 01:09:21,560
It reasoned through the problem space that the ontology provided.

1960
01:09:21,560 --> 01:09:23,640
It made a recommendation based on that logic.

1961
01:09:23,640 --> 01:09:26,280
It initiated an action within its authorized boundaries.

1962
01:09:26,280 --> 01:09:27,920
The human made the final call.

1963
01:09:27,920 --> 01:09:29,880
But the agent did the cognitive heavy lifting.

1964
01:09:29,880 --> 01:09:33,880
That is the efficiency you get when identity reasoning and governance finally align.

1965
01:09:33,880 --> 01:09:35,280
Now scale that.

1966
01:09:35,280 --> 01:09:37,680
Imagine 10 agents across 10 different facilities.

1967
01:09:37,680 --> 01:09:39,200
They monitor different KPIs.

1968
01:09:39,200 --> 01:09:42,920
The reason over the same ontologies, they follow the same policies they are autonomous within

1969
01:09:42,920 --> 01:09:43,920
their boundaries.

1970
01:09:43,920 --> 01:09:46,440
You don't need a central command center barking instruction.

1971
01:09:46,440 --> 01:09:48,920
You don't need a human reviewer for every tiny move.

1972
01:09:48,920 --> 01:09:50,000
The agents operate.

1973
01:09:50,000 --> 01:09:51,000
They detect.

1974
01:09:51,000 --> 01:09:52,000
They propose.

1975
01:09:52,000 --> 01:09:53,680
They act.

1976
01:09:53,680 --> 01:09:54,800
Humans oversee the system.

1977
01:09:54,800 --> 01:09:56,000
Not the individual decisions.

1978
01:09:56,000 --> 01:09:57,520
It's a different model of work.

1979
01:09:57,520 --> 01:09:59,720
And here is why this doesn't turn into chaos.

1980
01:09:59,720 --> 01:10:00,960
Everything is auditable.

1981
01:10:00,960 --> 01:10:02,280
Everything is reversible.

1982
01:10:02,280 --> 01:10:05,320
Every single move an agent makes is logged with full context.

1983
01:10:05,320 --> 01:10:06,680
You know who authorized it?

1984
01:10:06,680 --> 01:10:08,120
You see the reasoning that led to it.

1985
01:10:08,120 --> 01:10:09,120
You see the outcome.

1986
01:10:09,120 --> 01:10:11,400
When an agent makes a mistake, you undo the action.

1987
01:10:11,400 --> 01:10:14,840
If it oversteeps the boundary, the logs show you exactly how it happened.

1988
01:10:14,840 --> 01:10:17,080
The system isn't trusting these agents blindly.

1989
01:10:17,080 --> 01:10:20,120
It's trusting them within transparent, hard-coded constraints.

1990
01:10:20,120 --> 01:10:22,160
This also makes coordination possible.

1991
01:10:22,160 --> 01:10:25,800
When agents use the same ontologies and policies, they can coordinate without being told

1992
01:10:25,800 --> 01:10:26,800
to.

1993
01:10:26,800 --> 01:10:30,440
One agent finds a supply issue and updates the shared context about material levels.

1994
01:10:30,440 --> 01:10:34,760
Another agent on the production side sees that update and adjusts its recommendations automatically.

1995
01:10:34,760 --> 01:10:36,280
They aren't talking to each other.

1996
01:10:36,280 --> 01:10:39,280
They're both reading from and writing to the same business model.

1997
01:10:39,280 --> 01:10:41,680
That's how an intelligent organization actually works.

1998
01:10:41,680 --> 01:10:45,200
Not through top-down commands, but through a shared understanding of what is true.

1999
01:10:45,200 --> 01:10:46,520
The human role doesn't go away.

2000
01:10:46,520 --> 01:10:47,520
It just moves.

2001
01:10:47,520 --> 01:10:51,000
Instead of making every decision or checking every box, humans set the direction.

2002
01:10:51,000 --> 01:10:52,080
They write the policies.

2003
01:10:52,080 --> 01:10:53,080
They build the boundaries.

2004
01:10:53,080 --> 01:10:55,360
They monitor the health of the whole machine.

2005
01:10:55,360 --> 01:10:58,840
They step in when a decision requires judgment or human values.

2006
01:10:58,840 --> 01:11:00,280
The agents handle the execution.

2007
01:11:00,280 --> 01:11:01,280
They handle the speed.

2008
01:11:01,280 --> 01:11:02,920
They handle the data at scale.

2009
01:11:02,920 --> 01:11:05,480
This is what autonomous operations look like in practice.

2010
01:11:05,480 --> 01:11:07,120
It isn't agents running wild.

2011
01:11:07,120 --> 01:11:12,040
It's agents operating with total clarity inside a cage of constraints, guided by an ontology

2012
01:11:12,040 --> 01:11:14,200
that actually understands your business.

2013
01:11:14,200 --> 01:11:17,680
It's humans keeping oversight, while the system scales far beyond what manual management

2014
01:11:17,680 --> 01:11:19,320
could ever touch.

2015
01:11:19,320 --> 01:11:20,640
Copilot is the gateway drug.

2016
01:11:20,640 --> 01:11:21,640
It works.

2017
01:11:21,640 --> 01:11:22,640
It feels good.

2018
01:11:22,640 --> 01:11:23,640
It makes you feel productive.

2019
01:11:23,640 --> 01:11:24,960
But Copilot isn't the platform.

2020
01:11:24,960 --> 01:11:27,120
It's just a feature that proves you need a platform.

2021
01:11:27,120 --> 01:11:30,800
The real shift moving from a chat box to an agent fabric requires infrastructure.

2022
01:11:30,800 --> 01:11:32,000
You need identity.

2023
01:11:32,000 --> 01:11:33,320
So agents are accountable.

2024
01:11:33,320 --> 01:11:34,320
You need ontologies.

2025
01:11:34,320 --> 01:11:35,520
You need reason correctly.

2026
01:11:35,520 --> 01:11:37,960
You need governance hubs, so policies stay consistent.

2027
01:11:37,960 --> 01:11:41,160
You need landing zones, so the system scales without breaking.

2028
01:11:41,160 --> 01:11:44,880
The organizations that build this foundation first will treat AI as a managed, auditable

2029
01:11:44,880 --> 01:11:45,880
capability.

2030
01:11:45,880 --> 01:11:48,560
The ones that don't, they're just building governance debt.

2031
01:11:48,560 --> 01:11:50,960
They'll spend the next decade trying to pay it off.

2032
01:11:50,960 --> 01:11:54,360
The future belongs to the companies that treat AI infrastructure like any other critical

2033
01:11:54,360 --> 01:11:56,360
system designed on purpose.

2034
01:11:56,360 --> 01:11:57,360
Built methodically.

2035
01:11:57,360 --> 01:11:58,360
Gov.

2036
01:11:58,360 --> 01:11:59,360
Every single day.

2037
01:11:59,360 --> 01:12:00,360
That's what the agent measures.

2038
01:12:00,360 --> 01:12:01,360
It isn't a technology choice.

2039
01:12:01,360 --> 01:12:05,240
the structural decision about how you're going to operate at scale.

Mirko Peters

Founder of m365.fm, m365.show and m365con.net

Mirko Peters is a Microsoft 365 expert, content creator, and founder of m365.fm, a platform dedicated to sharing practical insights on modern workplace technologies. His work focuses on Microsoft 365 governance, security, collaboration, and real-world implementation strategies.

Through his podcast and written content, Mirko provides hands-on guidance for IT professionals, architects, and business leaders navigating the complexities of Microsoft 365. He is known for translating complex topics into clear, actionable advice, often highlighting common mistakes and overlooked risks in real-world environments.

With a strong emphasis on community contribution and knowledge sharing, Mirko is actively building a platform that connects experts, shares experiences, and helps organizations get the most out of their Microsoft 365 investments.