Your Sensitivity Labels Are A Lie: The Collaborative AI Silo Crisis
This episode argues that sensitivity labels are widely misunderstood and often give organizations a false sense of security. While they appear to enforce governance, in reality they are static, incomplete, and poorly maintained—making them ineffective in dynamic, AI-driven environments.
The core issue is not the labeling technology itself, but the way organizations structure and manage their data. Most environments suffer from fragmented information spread across Teams, SharePoint, and other systems, creating silos that block both collaboration and effective AI usage. As a result, AI tools like Copilot cannot access the right data and are forced to generate outputs based on incomplete or outdated information.
This leads to what the episode describes as an “AI rework loop”: AI produces confident but incorrect results, and employees must spend significant time validating and fixing them. In many cases, a large portion of AI-generated work requires correction, eroding the expected productivity gains.
The episode emphasizes that AI does not fix these problems—it exposes them. Poor data architecture, inconsistent permissions, and low labeling coverage become more visible and more impactful when AI is introduced.
The traditional approach to governance—relying on manual labeling and static controls—is no longer sufficient. Instead, organizations must shift toward dynamic, context-aware access models that evaluate who should access what data in real time. This includes reducing permission sprawl, automating classification, and redesigning how data flows across the organization.
The key takeaway is that successful AI adoption depends on fixing the underlying data foundation. Sensitivity labels alone cannot provide meaningful protection or enable intelligent AI. Real value comes from moving away from siloed, static governance toward connected, context-driven data access that allows AI to operate accurately and safely.
You may wonder if Microsoft Copilot leaks your data. The answer is clear: many organizations have faced real risks. Recent reports show that 68% of companies experienced data leakage when employees used AI tools like Microsoft Copilot. This makes the need for strong governance urgent. You can boost your return on investment by adopting a team-based approach to Microsoft 365 Copilot. You must move beyond static controls and focus on dynamic, context-aware governance to protect your information.
Key Takeaways
- Data leakage is a serious risk with Microsoft Copilot, affecting 68% of organizations using AI tools.
- Over-permissioning is a major cause of data leaks; limit access to only what is necessary for users and apps.
- Prompt injection can trick Copilot into revealing sensitive information; train your team to write secure prompts.
- Regularly review and audit permissions in Microsoft 365 to prevent unauthorized access and data exposure.
- Implement dynamic access controls that adjust based on user behavior and risk to enhance data protection.
- Use Microsoft Purview to map and classify sensitive data, ensuring proper controls are in place.
- Continuous monitoring and real-time alerts can help detect and prevent data leaks before they escalate.
- Stay updated on compliance requirements and regularly train your team to maintain a strong security posture.
Is Microsoft Copilot Leaking Data?
What Data Leakage Means
You may hear the term "data leakage" and wonder what it really means for your organization. Data leakage happens when sensitive or confidential information leaves its intended environment. This can occur through accidental sharing, misconfigured permissions, or even targeted attacks. With tools like Microsoft Copilot, the risk of data exposure increases because the system can access large amounts of information across your Microsoft 365 environment.
Data leakage risks often come from over-permissioning, where users or apps have more access than they need. For example, recent security reports show that over 3% of business-sensitive data was shared organization-wide without proper review. This means confidential files, emails, or documents could reach people who should not see them.
Here are some common types of data leaks you should watch for:
- Over-permissioning and excessive data exposure: When users or apps have broad access, confidential data can be exposed.
- Prompt injection and jailbreak attacks: Attackers may trick Copilot into revealing restricted information.
- Data exfiltration via connected apps and APIs: Weak controls can allow sensitive data to move to external services.
- Integration vulnerabilities across the Microsoft 365 ecosystem: Misconfigurations can create new paths for external leaks.
- Compliance gaps in regulated environments: Poorly set policies can lead to compliance violations and even a breach.
A recent incident highlighted these risks. In January 2026, Microsoft Copilot processed confidential emails, including legal memos and protected health information, due to a code error. Microsoft responded quickly, but the fix took time to reach all users. The U.K.’s National Health Service flagged this incident, raising concerns about compliance and data handling.
Common Misunderstandings
Many users misunderstand how Microsoft Copilot handles data. Some believe that Copilot can access all data in Microsoft 365, but it only follows existing permission settings. Others think that Copilot stores prompts and responses in new databases, but it does not create extra data repositories.
Here are some myths and facts:
- Myth: Copilot shares your data with Microsoft for training.
Fact: Customer data is not used to train models or accessed by unauthorized personnel. - Myth: Copilot compromises data privacy.
Fact: Microsoft uses industry-standard encryption and follows privacy laws like GDPR. - Myth: Access controls always prevent data exposure.
Fact: Over-permissioning and misconfigurations can still lead to external leaks. - Myth: Copilot trains on your organizational data.
Fact: Copilot does not retain individual tenant information.
You should also know that Copilot can sometimes surface the wrong data, which may lead to harmful decisions. Users often overestimate how well access controls protect against a breach. Understanding these points helps you manage data leakage risks and use Microsoft 365 Copilot safely.
Microsoft Copilot Security Risks
When you use Microsoft Copilot, you must understand the security risks that come with it. These risks can lead to sensitive information exposure, internal oversharing, and even a breach. You need to know how over-permissioning, prompt injection, and recent bugs can put your data at risk.
Over-Permissioning Issues
Over-permissioning happens when users or apps have more access than they need. If you give broad permissions, you increase the chance of confidential documents being seen by the wrong people. Many organizations do not review access controls often enough. This can result in internal oversharing, where sensitive data spreads across teams or departments without proper checks. You should always limit permissions to only what is necessary. This step helps reduce the risk of exposure and supports better risk management.
Prompt Injection and Data Oversharing
Prompt injection is a growing concern with AI tools like Copilot. Attackers or even regular users can craft prompts that trick the system into revealing information that should stay private. Poorly structured prompts may cause Copilot to share financial records, HR files, or other confidential data. This can lead to compliance issues and possible violations of laws such as GDPR or HIPAA.
- Prompt injection can lead to unintended actions that expose sensitive information.
- Poorly structured prompts may inadvertently reveal sensitive data such as financial records or HR information.
- Compliance risks arise as prompt injection can bypass security measures, leading to potential violations of regulations like GDPR and HIPAA.
You must train your team to write clear and secure prompts. You should also monitor how Copilot responds to requests, especially when handling confidential documents.
Real-World Incidents
Several incidents have shown how Microsoft Copilot can expose confidential information. These cases highlight the importance of strong security and risk management practices.
| Incident Description | Source |
|---|---|
| Microsoft Copilot accessed confidential emails due to a bug in Microsoft 365, bypassing data-protection policies. | TechCrunch |
| AI system ingested and summarized privileged communications that were marked as off-limits, raising concerns for attorneys. | TimeNetLaw |
| The bug was logged on the NHS’s internal IT support dashboard, leading to the European Parliament disabling AI features on staff devices. | Tom's Guide |
These incidents show that even with strong security tools, you must stay alert. Regular reviews and updates to your security policies help prevent future exposure. By understanding these risks, you can use Microsoft Copilot safely and protect your organization’s data.
Data Governance Challenges
Collaborative AI Silo Crisis
You face new challenges when you use collaborative AI tools like Microsoft Copilot. These tools can break down old barriers, but they also create new risks. Employees may ask Copilot questions that reveal confidential strategies. Sometimes, the AI can access sensitive information even if you did not open those files yourself. The AI can also connect pieces of data in ways that go beyond your normal permissions.
Here are some common challenges you might see:
- Poor data classification and governance can lead to confusion about who should access what.
- Unmanaged agent proliferation means you may have too many AI agents acting without clear rules.
- Lack of incident response playbooks leaves you unprepared for unexpected leaks.
- Access escalation through AI can give users more power than you intended.
- Compliance gaps can put your organization at risk.
- You need user training and change management to keep up with these changes.
In regulated industries like healthcare and finance, Copilot can surface sensitive data in summaries or suggestions. This makes it harder to track who saw what, which complicates audits and compliance checks.
Static Labels vs. Dynamic Controls
Many organizations rely on static sensitivity labels to protect their data. These labels mark files as "Confidential" or "Internal," but they do not always stop leaks. If you do not enforce these labels with real controls, they become just words on a file. They give you a false sense of security.
Dynamic access controls work differently. They adjust in real time based on who you are, what device you use, and what you try to do. This approach gives you stronger protection. For example, a label that says "Confidential" does nothing unless you also set up encryption or access restrictions. Without dynamic enforcement, static labels are only decorative and do not protect your sensitive information.
You need to move toward dynamic, context-aware controls. This shift helps you keep your data safe as you use AI tools like Copilot.
Why Traditional DLP Fails
Traditional data loss prevention tools were built for simple, static environments. They use fixed rules and look for patterns in emails or documents. These tools cannot keep up with the fast, complex world of AI-driven workflows. Once your data enters an AI workflow, traditional DLP often loses visibility and control.
Here are some reasons why these tools fall short:
- They were designed for static channels and binary policies, not for AI.
- The dynamic nature of AI makes it hard for DLP to see and control data movement.
- Regex-based rules and static inspection points cannot protect data in AI environments.
You need a zero-trust mindset. Always verify who accesses your data, and never assume that old tools will protect you. By updating your governance strategy, you can better manage risks and keep your organization safe.
Compliance and Microsoft Copilot
Regulatory Risks
You face unique regulatory risks when you use Microsoft Copilot in industries with strict compliance requirements. If Copilot accesses or exposes confidential data, you may violate laws like HIPAA or GLBA. This can lead to fines or even a breach notification obligation. You also need to keep clear audit trails. Without proper logging of AI interactions, you may struggle during audits for regulations such as SOX or FDA. Using Copilot in unapproved scenarios can increase your risk of regulatory exposure and penalties if an incident occurs.
Tip: Always review your organization's compliance requirements before deploying Copilot in sensitive environments.
Here are some common regulatory risks:
- Unauthorized data exposure can trigger legal action or regulatory investigations.
- Inadequate audit trails make it hard to prove compliance during reviews.
- Unrestricted use of Copilot may result in errors that violate industry rules.
Legal and Privacy Concerns
You must also consider data privacy laws when you use Copilot. Laws like GDPR and CCPA set strict rules for how you collect, store, and use data. You need to make sure your Copilot deployment follows these laws to avoid legal trouble.
| Privacy Law | Compliance Strategy | Description |
|---|---|---|
| GDPR | Data Minimization | Only collect data needed for Copilot to work, reducing unnecessary exposure. |
| User Consent | Give users clear consent forms and let them manage or withdraw consent easily. | |
| Right to Be Forgotten | Allow users to request deletion of their personal data. | |
| CCPA | Data Subject Access Requests | Provide tools for users to access, correct, or understand their data use. |
| Opt-Out Options | Let users opt out of data sales and respect their choices. |
Microsoft acts as a data processor under strict agreements. Prompts and responses stay within the Microsoft 365 cloud boundary. The system uses encryption, tenant isolation, and role-based access controls to protect your data. Copilot does not use your data to train foundation models, which helps reduce privacy risks.
Protecting Intellectual Property
You need to protect your intellectual property when you use Copilot. Set clear guidelines on ownership and usage rights for anything Copilot helps create. Use strong data protection rules to address privacy and security concerns. Security measures like code obfuscation, encryption, and secure storage help keep your confidential information safe.
Note: Regularly review your security policies to ensure your data and intellectual property remain protected.
Best practices for intellectual property protection include:
- Establishing clear rules for IP ownership and usage.
- Applying data protection rules for privacy and security.
- Using encryption and secure storage for sensitive data.
By understanding these compliance challenges and following best practices, you can reduce the risk of data exposure, meet regulatory compliance, and protect your organization from a breach.
Fixing and Preventing Data Leaks
Review Copilot Permissions
You must start by reviewing permissions for Microsoft Copilot across your organization. Many data leaks happen because users or apps have more access than they need. You can reduce this risk by following a clear process:
- Check permissions in SharePoint, OneDrive, and Teams. Make sure Copilot only responds with approved content.
- Schedule regular reviews of file permissions and group memberships. This helps you spot changes or unwanted access.
- Audit your Microsoft 365 environment for over-permissioning. Scan for broad or inherited permissions that could lead to oversharing.
- Educate your team about prompt and data handling risks. Training sessions help users understand how to avoid a breach.
- Monitor Microsoft Graph and API access patterns. Look for unusual behavior that could signal a security issue.
- Start with a limited pilot program. Test your security policies and see how Copilot works in real situations.
- Use a human-in-the-loop review for high-risk cases. This adds an extra layer of governance.
- Set clear go/no-go criteria for each deployment phase.
Tip: Implement monitoring systems that flag policy violations in real time. Maintain audit trails to support compliance and security investigations.
Tracking user activity helps you understand adoption and spot unusual behavior. You can also monitor data access to support privacy audits and meet compliance requirements.
Map and Classify Sensitive Data
Before you enable Microsoft Copilot, you need to know where your sensitive data lives. Mapping and classifying data helps you set the right controls and avoid accidental exposure of confidential information.
- Use Microsoft Purview to map and classify data based on sensitivity levels. You can use predefined or custom sensitivity labels.
- Apply sensitivity labels and data protection policies either manually or automatically. This ensures consistent labeling and protection across all storage and sharing platforms.
- Enable sensitivity labels and Data Loss Prevention (DLP) policies for prompts and responses. This step helps you control what Copilot can access and share.
- Review container-level labels in Teams and SharePoint. Make sure item-level coverage is in place for better security.
- Apply sensitivity labels early. This ensures that Copilot respects labeling and access boundaries from the start.
Microsoft Purview also helps you classify files, chats, and emails. This keeps your data secure and supports your compliance goals.
Implement Dynamic Access Controls
Static labels alone cannot protect your data in a fast-changing environment. You need dynamic access controls that adjust in real time based on user behavior, device health, and risk level.
- Set device-based restrictions. Limit access from unmanaged devices or high-risk locations.
- Use behavioral analytics to spot unusual usage patterns. This helps you detect attempts to access information outside normal parameters.
- Establish time-based restrictions for sensitive functions. For example, limit access to financial data during non-business hours.
- Configure strict user permissions and data access policies. Follow the least privilege principle to reduce risk.
- Conduct regular access reviews and permission audits. Update permissions as roles and needs change.
Note: Regularly review and update access permissions to prevent unauthorized data exposure. Train employees on safe Copilot usage to minimize accidental sharing.
Dynamic controls give you stronger governance and help you respond quickly to new threats. You can measure the effectiveness of your strategies using key metrics:
| Metric Type | Metrics |
|---|---|
| Security Metrics | DLP policy violations, suspicious access attempts, oversharing incidents |
| Compliance Metrics | DLP policy violations per 1,000 Copilot actions, percentage of users completing security training, time to remediate permission issues, audit-ready status for compliance requirements |
| Access Governance Metrics | Number of users with unnecessary admin rights, MFA coverage percentage, device compliance rate, conditional access policy coverage |
| Risk Reduction Metrics | Percentage of sites with proper access governance, number of files with sensitivity labels, reduction in overshared sensitive files, time to detect and respond to access anomalies |
By following these steps, you can secure Microsoft 365 Copilot, protect confidential data, and reduce the chance of a breach. You also support ongoing compliance and strengthen your overall security posture.
Continuous Monitoring and Alerts
You need to watch your Microsoft Copilot environment at all times. Continuous monitoring helps you spot risks before they turn into real problems. You can use advanced tools that track how Copilot interacts with your data. These tools give you a clear view of who accesses what and when. They also help you find weak spots in your security.
Here is a table that shows important features you should look for in monitoring tools:
| Feature | Description |
|---|---|
| Unified Visibility | Shows data access across Microsoft 365, Azure, and Copilot. Helps you find exposure points. |
| Continuous Monitoring | Tracks Copilot’s actions with sensitive data in SharePoint and Teams. |
| Automated Detection | Finds misconfigured permissions and too much access. Stops leaks before they happen. |
| Policy Enforcement | Lets you set rules for Copilot’s data access. Keeps you in line with company policies. |
| Context-Aware Alerts | Sends real-time alerts for strange Copilot activity. Helps you act fast if something goes wrong. |
| AI Usage Control | Watches Copilot in real time. Spots risky patterns and high-risk use. |
| Risk Assessment | Finds unauthorized apps and hidden connections. Flags new security risks. |
You should set up alerts for unusual activity. For example, if Copilot tries to access files it should not see, you get a warning right away. This quick response can stop a data leak before it spreads. You also need to review logs often. Logs show you patterns and help you understand how users interact with Copilot. By using these tools, you keep your data safe and stay ahead of threats.
Update Policies and Training
You must keep your security policies up to date. Risks change fast, and old rules may not protect you anymore. Regular updates make sure your policies match the latest threats. You also need to train your team often. Training helps everyone understand how to use Copilot safely and what to do if something goes wrong.
- Update your security policies on a regular schedule. This keeps your defenses strong against new risks.
- Train your team often. Frequent training helps users spot dangers and avoid mistakes.
- Give clear guidance to everyone who uses Copilot. Simple instructions lead to safer results.
- Review your policies after every major update or incident. This helps you fix gaps and improve your response.
- Make training part of your onboarding process. New users learn safe habits from the start.
Tip: Regular updates and training help you stay ahead of threats. Clear rules and guidance make Copilot safer for everyone.
By keeping your policies fresh and your team informed, you build a strong defense against data leaks. You also create a culture of security that protects your organization as technology evolves.
Proactive AI Data Governance
You need to take a forward-thinking approach to managing AI in your organization. Proactive AI data governance helps you stay ahead of risks and make the most of Microsoft Copilot. You can build a strong foundation by using advanced tools and regular improvement cycles.
Ongoing Oversight
You should keep a close watch on how users interact with Copilot. Ongoing oversight gives you real-time coaching and support. When someone tries to share sensitive information, Copilot can give a warning right away. This helps users learn safe habits as they work. You can also use built-in tutorials and learning pathways. These tools teach your team how to handle information without slowing down their work.
- Real-time compliance coaching helps users follow rules.
- Interactive tutorials give hands-on practice.
- Instant feedback from Copilot reinforces good data handling.
Tip: Make oversight part of your daily routine. This keeps everyone alert and reduces mistakes.
Regular Audits
You need to check your systems often. Regular audits help you find weak spots and fix them before they cause problems. You can use Microsoft Purview to review sensitivity labels, DLP rules, and audit logs. This keeps your organization ready for any compliance review.
Here are the key components of a proactive strategy:
| Component | Description |
|---|---|
| Security Architecture | Uses Zero Trust, tenant isolation, and encryption to protect information. |
| Compliance Tools | Microsoft Purview supports governance with labels, DLP rules, and audit logs. |
| Deployment Phases | Five steps: readiness assessment, license mapping, pilot deployment, policy enforcement, improvement. |
You should schedule audits on a regular basis. This helps you spot trends and adjust your policies as needed.
Adapting to AI Evolution
AI changes quickly. You must adapt your governance framework to keep up. Start by using dynamic access controls that change based on user behavior and risk. Set up review cycles to check how people use Copilot. Define clear use cases so everyone knows what is allowed.
- Monitor policy violations and near-misses.
- Analyze usage patterns and new risks.
- Gather feedback from users to improve your policies.
Note: Continuous improvement keeps your governance strong as AI evolves.
By following these steps, you can protect your data, meet compliance needs, and get the most value from Microsoft Copilot.
You face urgent risks if you ignore data leaks from Microsoft Copilot. To prevent a breach, you should take these steps:
- Patch all Microsoft 365 systems and monitor for new vulnerabilities.
- Audit permissions and restrict Copilot access to only what is needed.
- Train users on AI risks and review Copilot activity logs.
Prioritize dynamic, context-aware governance. Stay alert as Microsoft and Copilot evolve, and make data protection a top priority.
FAQ
What is the main cause of data leaks with Microsoft Copilot?
You often see data leaks when users or apps have too many permissions. Copilot follows your existing access rules. If you do not review these rules, sensitive data may become exposed.
Can Copilot access all my files in Microsoft 365?
Copilot only accesses files you or your team can already see. It does not bypass your organization’s permission settings. You control what Copilot can reach by managing user access.
How do I know if Copilot leaked sensitive data?
Check your audit logs in Microsoft 365. Look for unusual access or sharing events. Set up alerts for suspicious Copilot activity to catch leaks early.
Does Copilot use my data to train its AI models?
No, Copilot does not use your organizational data to train its models. Your data stays within your Microsoft 365 environment and remains protected by Microsoft’s privacy standards.
What steps should I take before enabling Copilot?
- Map and classify your sensitive data.
- Review and update user permissions.
- Set up dynamic access controls.
- Train your team on safe Copilot use.
How can I stop Copilot from sharing confidential information?
You can use sensitivity labels, DLP policies, and dynamic access controls. These tools help you limit what Copilot can access and share. Regular reviews keep your controls strong.
Is Copilot safe for regulated industries like healthcare or finance?
Copilot supports compliance features, but you must configure them correctly. Always review your industry’s rules and set up strict access controls before using Copilot in sensitive environments.
What should I do if I suspect a data leak?
Act fast. Disable Copilot access for affected users, review audit logs, and update your security policies. Notify your compliance team and follow your incident response plan.
🚀 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 --> 00:00:02,640
You deploy co-pilot and wait for the productivity wave to hit,
2
00:00:02,640 --> 00:00:03,840
but instead of a breakthrough,
3
00:00:03,840 --> 00:00:06,960
you see a 300% spike in data loss prevention events.
4
00:00:06,960 --> 00:00:09,360
The assumption was that the AI would find our brilliance,
5
00:00:09,360 --> 00:00:12,360
but in reality, it is mostly finding our permission debt.
6
00:00:12,360 --> 00:00:14,320
The everyone group in your SharePoint environment
7
00:00:14,320 --> 00:00:17,320
is currently the largest security hole in your organization,
8
00:00:17,320 --> 00:00:19,800
and it acts like an open door that nobody bothered to close
9
00:00:19,800 --> 00:00:21,760
because the data was buried in the basement.
10
00:00:21,760 --> 00:00:23,360
Now that data is conversational.
11
00:00:23,360 --> 00:00:26,480
Your current sensitivity labeling strategy is not a shield.
12
00:00:26,480 --> 00:00:28,880
It is a data graveyard that hides information
13
00:00:28,880 --> 00:00:30,440
from the people who actually needed,
14
00:00:30,440 --> 00:00:32,600
while doing nothing to stop the AI from surfacing
15
00:00:32,600 --> 00:00:34,400
the wrong things to the wrong people.
16
00:00:34,400 --> 00:00:36,800
This is the collaborative AI-style or crisis.
17
00:00:36,800 --> 00:00:39,000
Today we are moving from containment to context,
18
00:00:39,000 --> 00:00:40,840
and we are going to save your AI investment
19
00:00:40,840 --> 00:00:42,760
by fixing the structural lie at the heart
20
00:00:42,760 --> 00:00:45,920
of your data governance model, the inheritance paradox.
21
00:00:45,920 --> 00:00:47,920
The fundamental flaw in most AI rollouts
22
00:00:47,920 --> 00:00:50,080
is a misunderstanding of how the machine thinks,
23
00:00:50,080 --> 00:00:52,440
and while you might assume that co-pilot makes decisions
24
00:00:52,440 --> 00:00:54,680
about who sees what, it actually does not.
25
00:00:54,680 --> 00:00:57,280
Co-pilot is a mirror that inherits your existing organizational
26
00:00:57,280 --> 00:00:59,040
entropy, so it does not create new access.
27
00:00:59,040 --> 00:01:02,520
It simply operationizes the access you already gave away years ago.
28
00:01:02,520 --> 00:01:04,280
Think of it as the spotlight effect.
29
00:01:04,280 --> 00:01:06,520
Before AI oversharing was latent,
30
00:01:06,520 --> 00:01:09,680
like a messy filing cabinet in a basement nobody visited,
31
00:01:09,680 --> 00:01:13,240
but AI makes that latent oversharing visible at machine speed.
32
00:01:13,240 --> 00:01:16,480
If a junior analyst asks for a summary of recent strategy shifts,
33
00:01:16,480 --> 00:01:18,560
and your M&A folder is set to internal,
34
00:01:18,560 --> 00:01:21,400
the AI will summarize the merger you have not announced yet.
35
00:01:21,400 --> 00:01:24,440
This is not a breach, it is a feature of your own architecture.
36
00:01:24,440 --> 00:01:26,240
The reality is that manual labeling coverage
37
00:01:26,240 --> 00:01:28,920
rarely exceeds 10% in the average enterprise,
38
00:01:28,920 --> 00:01:31,320
because we have spent a decade telling users
39
00:01:31,320 --> 00:01:34,000
to tag their files and they simply have not done it.
40
00:01:34,000 --> 00:01:37,440
We relied on a set it and forget it mentality for sensitivity tags,
41
00:01:37,440 --> 00:01:39,760
and assumed that once a file was marked public,
42
00:01:39,760 --> 00:01:41,600
it stayed that way forever.
43
00:01:41,600 --> 00:01:44,560
A data is fluid, which means a document that was public yesterday
44
00:01:44,560 --> 00:01:47,440
might contain a sensitive prompt response today.
45
00:01:47,440 --> 00:01:49,080
When we rely on static labels,
46
00:01:49,080 --> 00:01:52,480
we create invisible barriers to cross-departmental intelligence,
47
00:01:52,480 --> 00:01:54,960
and a highly confidential label on a project plan
48
00:01:54,960 --> 00:01:57,360
might stop the marketing team from seeing a launch date
49
00:01:57,360 --> 00:01:59,520
they actually need to do their jobs.
50
00:01:59,520 --> 00:02:01,400
It creates a silo, and these silos
51
00:02:01,400 --> 00:02:04,000
are the reason your AI feels less like an assistant
52
00:02:04,000 --> 00:02:05,640
and more like a restricted librarian.
53
00:02:05,640 --> 00:02:07,280
We have built a system based on the assumption
54
00:02:07,280 --> 00:02:09,520
that we can classify the world once and be done with it,
55
00:02:09,520 --> 00:02:11,920
but in a collaborative environment, that model breaks.
56
00:02:11,920 --> 00:02:14,080
If the labels are missing the AI sees too much,
57
00:02:14,080 --> 00:02:16,200
and if the labels are there, they are often too rigid,
58
00:02:16,200 --> 00:02:19,280
preventing the AI from connecting the dots between departments.
59
00:02:19,280 --> 00:02:21,360
We are currently seeing a massive gap between
60
00:02:21,360 --> 00:02:23,360
what the C-suite expects from AI,
61
00:02:23,360 --> 00:02:25,280
and what the data architecture allows,
62
00:02:25,280 --> 00:02:27,560
and you cannot have a high performing AI agent
63
00:02:27,560 --> 00:02:30,960
if your data is locked in 1990s style folder hierarchies.
64
00:02:30,960 --> 00:02:34,960
The inheritance paradox means the AI is only as secure
65
00:02:34,960 --> 00:02:37,960
as your worst SharePoint admins mistake from 2014.
66
00:02:37,960 --> 00:02:40,400
We have to admit that the current approach to classification
67
00:02:40,400 --> 00:02:44,360
is failing because it is static, manual, and treats data
68
00:02:44,360 --> 00:02:48,640
as an object to be locked away, rather than an asset to be used.
69
00:02:48,640 --> 00:02:51,080
Most organizations are currently in pilot purgatory
70
00:02:51,080 --> 00:02:54,480
because they realized six weeks in that their permissions are a mess.
71
00:02:54,480 --> 00:02:56,680
They try to fix it by adding more labels,
72
00:02:56,680 --> 00:02:58,560
but adding more labels to a broken system
73
00:02:58,560 --> 00:03:00,600
just creates more silos and does not solve
74
00:03:00,600 --> 00:03:02,640
the underlying problem of visibility.
75
00:03:02,640 --> 00:03:04,320
The shift we need is not about better tags,
76
00:03:04,320 --> 00:03:05,400
it is about a better model.
77
00:03:05,400 --> 00:03:08,560
We have been focusing on the what, which is the file itself,
78
00:03:08,560 --> 00:03:12,080
but we need to start focusing on the who, the where, and the why.
79
00:03:12,080 --> 00:03:14,200
The problem is not just that labels are missing,
80
00:03:14,200 --> 00:03:16,000
it is that the ones we have are static
81
00:03:16,000 --> 00:03:18,880
and do not account for the reality of how work happens today.
82
00:03:18,880 --> 00:03:21,800
Work is dynamic, so our security must be dynamic too.
83
00:03:21,800 --> 00:03:24,760
If we do not fix this, the AI will continue to hallucinate
84
00:03:24,760 --> 00:03:27,000
because it is looking at the wrong data or worse,
85
00:03:27,000 --> 00:03:30,000
it will tell the truth about data that should have been hidden.
86
00:03:30,000 --> 00:03:31,240
To fix the rework loop,
87
00:03:31,240 --> 00:03:34,480
we have to look at the underlying model of how we grant access,
88
00:03:34,480 --> 00:03:35,880
the high cost of rework.
89
00:03:35,880 --> 00:03:37,840
We talk about the productivity gains of AI
90
00:03:37,840 --> 00:03:39,280
as if they are a gross number
91
00:03:39,280 --> 00:03:41,600
and we see the nine hours saved per month
92
00:03:41,600 --> 00:03:43,200
and assume the job is done.
93
00:03:43,200 --> 00:03:45,520
But the reality is much messier because currently,
94
00:03:45,520 --> 00:03:48,400
about 40% of the time your employees saved by using AI
95
00:03:48,400 --> 00:03:51,040
is immediately lost to fixing what the AI got wrong.
96
00:03:51,040 --> 00:03:52,840
This is the hidden friction of the silo.
97
00:03:52,840 --> 00:03:55,400
When co-pilot doesn't have access to the good data,
98
00:03:55,400 --> 00:03:57,560
the finalized contract, the actual budget
99
00:03:57,560 --> 00:04:00,000
or the verified project plan, it doesn't stop.
100
00:04:00,000 --> 00:04:00,800
It guesses.
101
00:04:00,800 --> 00:04:02,440
It looks at the dark data it can see,
102
00:04:02,440 --> 00:04:04,160
like an old draft from a shared folder
103
00:04:04,160 --> 00:04:05,720
or speculative chat message
104
00:04:05,720 --> 00:04:07,720
and it builds a response based on that.
105
00:04:07,720 --> 00:04:11,280
The result is a rework loop that is quietly draining your ROI.
106
00:04:11,280 --> 00:04:13,320
We can actually quantify this digital debt.
107
00:04:13,320 --> 00:04:15,560
For your top performers, the power users
108
00:04:15,560 --> 00:04:17,600
who are leaning into these tools the most,
109
00:04:17,600 --> 00:04:20,280
this rework accounts for about one and a half weeks
110
00:04:20,280 --> 00:04:21,720
of lost time per year.
111
00:04:21,720 --> 00:04:22,560
Think about that.
112
00:04:22,560 --> 00:04:24,920
You are paying for a premium license to save time,
113
00:04:24,920 --> 00:04:26,520
but because your data is siloed,
114
00:04:26,520 --> 00:04:28,800
your best people are spending nearly two full work weeks
115
00:04:28,800 --> 00:04:31,080
just auditing machine generated errors.
116
00:04:31,080 --> 00:04:33,120
They are essentially acting as high paid proofreaders
117
00:04:33,120 --> 00:04:35,400
for a system that was supposed to be their co-pilot.
118
00:04:35,400 --> 00:04:38,400
This happens because rigid classification creates a vacuum.
119
00:04:38,400 --> 00:04:40,520
If the AI is barred from the system of record
120
00:04:40,520 --> 00:04:42,400
by an outdated sensitivity label,
121
00:04:42,400 --> 00:04:43,480
it will fill that vacuum
122
00:04:43,480 --> 00:04:46,280
with whatever system of convenience it can find.
123
00:04:46,280 --> 00:04:48,040
In departments like Finance and HR,
124
00:04:48,040 --> 00:04:51,080
this leads to outputs that are confidently wrong.
125
00:04:51,080 --> 00:04:54,000
A finance manager might ask for a summary of quarterly spend.
126
00:04:54,000 --> 00:04:56,320
If the actual ERP export is locked down,
127
00:04:56,320 --> 00:04:58,240
but a messy, unclassified working draft
128
00:04:58,240 --> 00:04:59,800
is sitting in a public teams channel,
129
00:04:59,800 --> 00:05:01,880
the AI will summarize the draft.
130
00:05:01,880 --> 00:05:04,960
The manager then spends 20 minutes finding the discrepancies.
131
00:05:04,960 --> 00:05:07,920
This is the friction between gross efficiency and net value.
132
00:05:07,920 --> 00:05:10,200
If it takes you two minutes to generate a report,
133
00:05:10,200 --> 00:05:12,080
but 20 minutes to verify the numbers,
134
00:05:12,080 --> 00:05:14,560
because the AI couldn't see the source of truth,
135
00:05:14,560 --> 00:05:16,240
you haven't actually gained anything.
136
00:05:16,240 --> 00:05:18,880
You've just shifted the labor from creation to correction.
137
00:05:18,880 --> 00:05:20,400
This rework isn't just a nuisance,
138
00:05:20,400 --> 00:05:22,440
it's a symptom of a structural failure.
139
00:05:22,440 --> 00:05:25,800
We've spent years building data silos to protect information,
140
00:05:25,800 --> 00:05:27,560
but in the age of generative AI,
141
00:05:27,560 --> 00:05:29,240
those silos act as blindfolds.
142
00:05:29,240 --> 00:05:31,400
When the AI is blindfolded, it hallucinates.
143
00:05:31,400 --> 00:05:33,320
It tries to please the user by connecting dots
144
00:05:33,320 --> 00:05:34,520
that shouldn't be connected.
145
00:05:34,520 --> 00:05:37,400
We are seeing organizations where the rework rate is so high
146
00:05:37,400 --> 00:05:40,040
that employees are starting to lose trust in the tool.
147
00:05:40,040 --> 00:05:42,240
They stop asking the AI for complex analysis
148
00:05:42,240 --> 00:05:44,800
and revert to using it for basic email drafting.
149
00:05:44,800 --> 00:05:47,520
The moment that happens, your AI strategy has failed.
150
00:05:47,520 --> 00:05:50,400
You are paying for an engine, but only using the headlights.
151
00:05:50,400 --> 00:05:52,920
To fix this rework loop, we have to stop looking at the AI
152
00:05:52,920 --> 00:05:54,360
and start looking at the plumbing.
153
00:05:54,360 --> 00:05:55,880
The good data exists.
154
00:05:55,880 --> 00:05:57,480
The brilliant insights are there.
155
00:05:57,480 --> 00:05:59,840
But they are trapped behind a wall of legacy permissions
156
00:05:59,840 --> 00:06:01,720
and static labels that were designed for a world
157
00:06:01,720 --> 00:06:03,280
where humans did all the searching.
158
00:06:03,280 --> 00:06:05,840
We need to move toward a model where the AI can see
159
00:06:05,840 --> 00:06:08,440
what it needs to see, exactly when it needs to see it,
160
00:06:08,440 --> 00:06:09,720
without compromising security.
161
00:06:09,720 --> 00:06:12,960
We have to look at the underlying model of how we grant access.
162
00:06:12,960 --> 00:06:15,680
Because as long as the AI is working with partial context,
163
00:06:15,680 --> 00:06:18,680
your employees will be working overtime to fix the results.
164
00:06:18,680 --> 00:06:21,680
That is a cost no organization can afford to ignore.
165
00:06:21,680 --> 00:06:24,240
From containment to context, we need to stop thinking
166
00:06:24,240 --> 00:06:25,760
about security as a wall.
167
00:06:25,760 --> 00:06:29,200
For decades, the industry relied on a hardened perimeter model.
168
00:06:29,200 --> 00:06:31,360
You were either inside the network or you were out,
169
00:06:31,360 --> 00:06:32,520
but that world is gone.
170
00:06:32,520 --> 00:06:34,200
Today, the perimeter is porous.
171
00:06:34,200 --> 00:06:35,160
People work from home.
172
00:06:35,160 --> 00:06:37,280
They use unmanaged devices and they collaborate
173
00:06:37,280 --> 00:06:38,720
with external vendors.
174
00:06:38,720 --> 00:06:41,240
In this environment, the old model of containment,
175
00:06:41,240 --> 00:06:43,760
where you lock data in a folder and assume it's safe
176
00:06:43,760 --> 00:06:46,760
because the folder is private is a dangerous illusion.
177
00:06:46,760 --> 00:06:49,680
It's dangerous because it assumes the threat is only outside.
178
00:06:49,680 --> 00:06:51,560
But the real threat in the age of co-pilot
179
00:06:51,560 --> 00:06:54,000
is the internal permission creep that has been accumulating
180
00:06:54,000 --> 00:06:54,680
for years.
181
00:06:54,680 --> 00:06:56,560
We need to move from a strategy of containment
182
00:06:56,560 --> 00:06:58,120
to a strategy of context.
183
00:06:58,120 --> 00:06:59,480
The old model was built for structure.
184
00:06:59,480 --> 00:07:02,840
It assumed that a file sensitivity was a fixed attribute.
185
00:07:02,840 --> 00:07:04,680
You tagged the document as confidential,
186
00:07:04,680 --> 00:07:06,160
and that was the end of the story.
187
00:07:06,160 --> 00:07:08,000
But in a collaborative AI environment,
188
00:07:08,000 --> 00:07:09,040
that tag is too blunt.
189
00:07:09,040 --> 00:07:10,080
It doesn't tell us enough.
190
00:07:10,080 --> 00:07:11,800
It doesn't account for who is asking,
191
00:07:11,800 --> 00:07:14,040
what device they are using, or where they are located.
192
00:07:14,040 --> 00:07:15,560
This is why we are seeing a massive shift
193
00:07:15,560 --> 00:07:18,560
toward dynamic access control or DAC.
194
00:07:18,560 --> 00:07:20,200
This isn't just a technical upgrade.
195
00:07:20,200 --> 00:07:21,440
It's a philosophical shift.
196
00:07:21,440 --> 00:07:23,680
It moves the decision-making process from the moment
197
00:07:23,680 --> 00:07:26,600
the file is created to the millisecond the data is requested.
198
00:07:26,600 --> 00:07:29,920
It's the difference between a static lock and a smart sensor.
199
00:07:29,920 --> 00:07:32,400
By 2026, the industry is mandating a shift
200
00:07:32,400 --> 00:07:35,360
toward attribute-based access control or ABAC.
201
00:07:35,360 --> 00:07:38,000
This is the foundation of the modern identity fabric.
202
00:07:38,000 --> 00:07:39,920
Instead of relying on a single static label,
203
00:07:39,920 --> 00:07:42,800
ABAC evaluates a dozen different risk signals in real time.
204
00:07:42,800 --> 00:07:44,480
It looks at the user's role, sure,
205
00:07:44,480 --> 00:07:46,320
but it also looks at the health of their laptop.
206
00:07:46,320 --> 00:07:48,040
It looks at their physical location.
207
00:07:48,040 --> 00:07:49,440
It even looks at their behavior.
208
00:07:49,440 --> 00:07:52,360
If a user who typically only accesses marketing files suddenly
209
00:07:52,360 --> 00:07:55,440
asks co-pilot to summarize the entire payroll database
210
00:07:55,440 --> 00:07:57,520
from a coffee shop in a different country,
211
00:07:57,520 --> 00:08:00,080
the system shouldn't just look at the sensitivity label.
212
00:08:00,080 --> 00:08:01,840
It should look at the context and say, no.
213
00:08:01,840 --> 00:08:04,160
This is how we move from, is this file labeled?
214
00:08:04,160 --> 00:08:06,360
Should this user see this data now?
215
00:08:06,360 --> 00:08:08,840
It's a subtle shift with massive implications.
216
00:08:08,840 --> 00:08:10,960
It allows us to be more permissive when the risk is low
217
00:08:10,960 --> 00:08:12,880
and more restrictive when the risk is high.
218
00:08:12,880 --> 00:08:16,520
It breaks the silos because it allows for just-in-time access.
219
00:08:16,520 --> 00:08:20,000
If the marketing team needs to see that highly confidential launch date,
220
00:08:20,000 --> 00:08:22,000
the system can grant them temporary,
221
00:08:22,000 --> 00:08:25,360
read-only access based on the context of their current project
222
00:08:25,360 --> 00:08:28,680
without permanently breaking the security of the file.
223
00:08:28,680 --> 00:08:31,720
This is the only way to enable the kind of cross-departmental
224
00:08:31,720 --> 00:08:34,760
intelligence that AI promises without turning your tenant
225
00:08:34,760 --> 00:08:37,040
into a data free for all.
226
00:08:37,040 --> 00:08:40,080
The ROI of these dynamic systems is already becoming clear.
227
00:08:40,080 --> 00:08:42,120
Organizations that move away from static labels
228
00:08:42,120 --> 00:08:45,680
to water context-aware model are seeing 15 to 30% reductions
229
00:08:45,680 --> 00:08:47,240
in their cyber insurance premiums.
230
00:08:47,240 --> 00:08:47,760
Why?
231
00:08:47,760 --> 00:08:50,960
Because insurers know that static labels are easy to bypass or ignore.
232
00:08:50,960 --> 00:08:53,960
Dynamic systems, on the other hand, are much harder to exploit.
233
00:08:53,960 --> 00:08:56,160
They provide a level of continuous verification
234
00:08:56,160 --> 00:08:58,120
that static models simply can't match.
235
00:08:58,120 --> 00:09:00,400
They reduce the blast radius of a compromised account
236
00:09:00,400 --> 00:09:03,760
because the attacker can't just inherit years of sloppy permissions.
237
00:09:03,760 --> 00:09:05,360
Every request is a new evaluation.
238
00:09:05,360 --> 00:09:08,520
Every interaction is a new opportunity to verify trust.
239
00:09:08,520 --> 00:09:11,200
This isn't just a theoretical shift for IT architects.
240
00:09:11,200 --> 00:09:14,520
It is a survival requirement for the next wave of a genetic AI.
241
00:09:14,520 --> 00:09:17,440
As we move toward agents that can take actions on our behalf,
242
00:09:17,440 --> 00:09:20,280
scheduling meetings, moving money, or updating records,
243
00:09:20,280 --> 00:09:23,080
we cannot rely on a "set it and forget it" security model.
244
00:09:23,080 --> 00:09:25,560
We need a system that understands the intent behind the action.
245
00:09:25,560 --> 00:09:28,840
We need security that is as smart as the AI it is protecting.
246
00:09:28,840 --> 00:09:30,800
This leads us to the technical roadmap.
247
00:09:30,800 --> 00:09:32,320
Because while the philosophy is changing,
248
00:09:32,320 --> 00:09:36,080
the tools to implement it are finally arriving in the purview ecosystem.
249
00:09:36,080 --> 00:09:40,240
The 2026 purview roadmap, the shift to what context-aware intelligence
250
00:09:40,240 --> 00:09:44,440
is finally showing up in the actual architecture of the Microsoft ecosystem.
251
00:09:44,440 --> 00:09:48,400
We are moving past the era where purview was just a tool for compliance reporting
252
00:09:48,400 --> 00:09:52,240
and by 2026 it has evolved into a real-time diagnostic engine.
253
00:09:52,240 --> 00:09:57,320
One of the most significant changes is the move to always on diagnostics for endpoint DLP.
254
00:09:57,320 --> 00:10:01,800
In the past, troubleshooting a policy failure meant a week of back-and-forth between IT and the user
255
00:10:01,800 --> 00:10:04,200
while they tried to reproduce a glitch that happened once.
256
00:10:04,200 --> 00:10:07,400
Now those diagnostic traces are stored locally on the device
257
00:10:07,400 --> 00:10:10,520
in a secure, compressed format for 90 days.
258
00:10:10,520 --> 00:10:13,560
This allows the system to analyze exactly why a specific file was blocked
259
00:10:13,560 --> 00:10:17,800
or why a label failed to apply without ever sending the actual sensitive content to the cloud.
260
00:10:17,800 --> 00:10:20,680
It is local visibility paired with global enforcement.
261
00:10:20,680 --> 00:10:24,760
This diagnostic power is being amplified by AI-powered explanation tools.
262
00:10:24,760 --> 00:10:28,440
We have reached a level of complexity where human admins can no longer manually track the difference
263
00:10:28,440 --> 00:10:30,600
between a dozen different DLP policies.
264
00:10:30,600 --> 00:10:34,520
When a policy changes, the system now generates a natural language summary
265
00:10:34,520 --> 00:10:38,280
to explain exactly what was modified and which user groups are affected.
266
00:10:38,280 --> 00:10:41,720
It also shows what the potential impact on collaborative workflows will be
267
00:10:41,720 --> 00:10:43,400
which reduces the risk of policy drift.
268
00:10:43,400 --> 00:10:47,160
This is when security rules become so tangled that they start blocking legitimate work.
269
00:10:47,160 --> 00:10:50,840
These tools allow the security team to act as enablers rather than gatekeepers
270
00:10:50,840 --> 00:10:52,120
by providing clarity.
271
00:10:52,120 --> 00:10:55,240
On the rules of the road in a language, the business can actually understand.
272
00:10:55,240 --> 00:10:59,880
A critical piece of this roadmap is the role of restricted SharePoint search or RSS.
273
00:10:59,880 --> 00:11:02,360
This is the emergency break for your AI deployment.
274
00:11:02,360 --> 00:11:05,480
It allows you to explicitly exclude high-risk sites like your board papers,
275
00:11:05,480 --> 00:11:09,960
your payroll data, or your legal strategy from being used as grounding for co-pilot.
276
00:11:09,960 --> 00:11:14,200
Before you even touch a sensitivity label, you can use RSS to ensure that the AI
277
00:11:14,200 --> 00:11:17,000
simply cannot see the most dangerous corners of your tenant.
278
00:11:17,000 --> 00:11:21,800
It is a foundational layer of protection that recognizes that not all data is created equal.
279
00:11:22,120 --> 00:11:24,440
This gives you the breathing room to fix your permission debt.
280
00:11:24,440 --> 00:11:27,320
Without having to shut down the entire AI experiment,
281
00:11:27,320 --> 00:11:30,680
we are also seeing a massive expansion in auto labeling capabilities.
282
00:11:30,680 --> 00:11:33,960
The goal for 2026 is to be secured by default.
283
00:11:33,960 --> 00:11:38,120
Your SharePoint libraries are no longer passive buckets because they are becoming active participants
284
00:11:38,120 --> 00:11:39,480
in your security posture.
285
00:11:39,480 --> 00:11:41,800
When a file is uploaded to a specific library,
286
00:11:41,800 --> 00:11:45,320
the default label of that library is applied at rest automatically.
287
00:11:45,320 --> 00:11:47,720
If that file contains sensitive information types,
288
00:11:47,720 --> 00:11:51,080
the system can now override lower priority manual labels.
289
00:11:51,080 --> 00:11:54,440
This closes the coverage gap that has plagued manual systems for years.
290
00:11:54,440 --> 00:11:58,840
It ensures that even if a user forgets to tag a document, the platform has their back.
291
00:11:58,840 --> 00:12:02,760
It is the move from a voluntary system to a mandatory automated infrastructure.
292
00:12:02,760 --> 00:12:04,040
To manage this at scale,
293
00:12:04,040 --> 00:12:08,200
we are increasingly relying on the Graph API for custom usage analytics.
294
00:12:08,200 --> 00:12:11,800
We can now look under the hood to see exactly which departments are siloed.
295
00:12:11,800 --> 00:12:16,760
If the R&D team has a 90% rework rate because they are blocked from seeing the engineering specs,
296
00:12:16,760 --> 00:12:17,960
the data will show it.
297
00:12:17,960 --> 00:12:21,800
We can identify the bottlenecks where rigid classification is killing productivity
298
00:12:21,800 --> 00:12:23,400
and adjust the policies in real time.
299
00:12:23,400 --> 00:12:27,720
This is how we move from a reactive security posture to a proactive governance strategy.
300
00:12:27,720 --> 00:12:31,160
We are using the data to tell us where the system is breaking rather than waiting for
301
00:12:31,160 --> 00:12:33,160
a user to complain or a breach to occur.
302
00:12:33,160 --> 00:12:34,680
But even with the best tools,
303
00:12:34,680 --> 00:12:37,160
the strategy fails without executive alignment.
304
00:12:37,160 --> 00:12:39,800
You can have the most advanced purview setup in the world,
305
00:12:39,800 --> 00:12:42,840
but if your leadership still views data as something to be hoarded,
306
00:12:42,840 --> 00:12:44,840
you will never realize the full value of AI.
307
00:12:44,840 --> 00:12:47,720
This brings us to a new kind of risk that is emerging from the ground up.
308
00:12:47,880 --> 00:12:50,840
It is a challenge that is not coming from hackers or external threats,
309
00:12:50,840 --> 00:12:53,320
but from your own employees trying to be productive.
310
00:12:53,320 --> 00:12:55,400
We need to talk about the Citizen Developer Delimmer.
311
00:12:55,400 --> 00:12:57,960
The Citizen Developer Delimmer,
312
00:12:57,960 --> 00:13:01,560
we are currently witnessing the largest explosion of unregulated creation
313
00:13:01,560 --> 00:13:03,400
in the history of enterprise computing.
314
00:13:03,400 --> 00:13:07,960
There are now over 1 million low-code assets living inside the Microsoft ecosystem.
315
00:13:07,960 --> 00:13:10,600
We have empowered employees to build their own apps,
316
00:13:10,600 --> 00:13:13,080
their own flows, and now their own AI agents.
317
00:13:13,080 --> 00:13:15,480
But this empowerment has a dark side.
318
00:13:15,480 --> 00:13:17,720
We have moved from Shadow IT to Shadow AI,
319
00:13:17,720 --> 00:13:22,200
where non-technical users are building complex agents on top of ungoverned messy data.
320
00:13:22,200 --> 00:13:23,720
This is the risk of agents sprawl.
321
00:13:23,720 --> 00:13:28,200
It is what happens when a marketing coordinator builds a custom GPT and co-pilot studio
322
00:13:28,200 --> 00:13:29,640
to help with vendor queries,
323
00:13:29,640 --> 00:13:32,680
but they inadvertently ground that agent on a sharepoint side
324
00:13:32,680 --> 00:13:35,880
containing every contract the company has signed since 2012.
325
00:13:35,880 --> 00:13:38,280
The dilemma is that we cannot simply turn it off.
326
00:13:38,280 --> 00:13:40,840
If you kill the ability for people to solve their own problems,
327
00:13:40,840 --> 00:13:43,880
you kill the very agility that AI is supposed to provide.
328
00:13:43,880 --> 00:13:46,440
Instead, we have to move toward a tiered governance model.
329
00:13:46,440 --> 00:13:47,880
Not all agents are created equal.
330
00:13:47,880 --> 00:13:50,680
A low-code tool that helps an individual summarise their own emails
331
00:13:50,680 --> 00:13:53,400
using standard Teams connectors is a low-risk asset.
332
00:13:53,400 --> 00:13:56,120
It should bypass the heavy duty review process.
333
00:13:56,120 --> 00:13:59,400
But the moment an agent touches a system of record-like dataverse
334
00:13:59,400 --> 00:14:02,360
or a sensitive SQL database, it needs a gatekeeper.
335
00:14:02,360 --> 00:14:05,720
We have to stop treating citizen development as a single category
336
00:14:05,720 --> 00:14:07,800
and start treating it as a spectrum of risk.
337
00:14:07,800 --> 00:14:11,080
The industry is pivoting from the old concept of citizen development
338
00:14:11,080 --> 00:14:14,120
to what we now call AI-augmented personal building.
339
00:14:14,120 --> 00:14:15,720
The difference is subtle but crucial.
340
00:14:15,720 --> 00:14:17,880
In the old world, the user was a developer,
341
00:14:17,880 --> 00:14:20,440
but in the new world, the user is an orchestrator.
342
00:14:20,440 --> 00:14:23,240
They are directing the AI to build the solution for them.
343
00:14:23,240 --> 00:14:27,000
This makes the guardrails in co-pilot studio more important than the code itself.
344
00:14:27,000 --> 00:14:28,920
We need to implement content moderation layers
345
00:14:28,920 --> 00:14:32,040
that prevent these agents from hallucinating or leaking data
346
00:14:32,040 --> 00:14:33,480
through prompt injection attacks
347
00:14:33,480 --> 00:14:35,480
that the user did not even know where possible.
348
00:14:35,480 --> 00:14:38,280
You would not let a summer intern write your firewall rules
349
00:14:38,280 --> 00:14:42,280
so we should not let them build agents that have read access to the entire corporate wiki.
350
00:14:42,280 --> 00:14:43,320
The fix is structural.
351
00:14:43,320 --> 00:14:46,520
We need to automate the enforcement of data loss prevention policies
352
00:14:46,520 --> 00:14:48,040
within the low-code environment.
353
00:14:48,040 --> 00:14:52,680
If a user tries to connect an AI agent to an external unmanaged dropbox account,
354
00:14:52,680 --> 00:14:54,440
the system should block it instantly.
355
00:14:54,440 --> 00:14:57,160
If they try to build a bot that queries a library marked
356
00:14:57,160 --> 00:14:59,320
with a highly confidential sensitivity label,
357
00:14:59,320 --> 00:15:03,080
the deployment should trigger an automatic manual review by the ITOPS team.
358
00:15:03,080 --> 00:15:05,560
This creates a secure-by-design pathway for innovation.
359
00:15:05,560 --> 00:15:09,560
It allows the builders to build while ensuring that the most sensitive assets of the organization
360
00:15:09,560 --> 00:15:12,440
remain behind the context-aware walls we have been discussing.
361
00:15:12,440 --> 00:15:17,080
We are moving into an era where every employee is a potential architect of the company's intelligence.
362
00:15:17,080 --> 00:15:19,800
That is a massive competitive advantage if managed correctly,
363
00:15:19,800 --> 00:15:22,280
but it is a catastrophic liability if left to chance.
364
00:15:22,280 --> 00:15:25,560
The goal is not to stop the sprawl, but to govern the flow.
365
00:15:25,560 --> 00:15:29,320
We need to provide the citizen builders with pre-approved templates and safe zones
366
00:15:29,320 --> 00:15:32,280
where they can experiment without risking the crown jewels.
367
00:15:32,280 --> 00:15:36,360
This ensures that the agents they create are an extension of our governance strategy
368
00:15:36,360 --> 00:15:37,960
rather than an exception to it.
369
00:15:37,960 --> 00:15:42,520
This leads us to the final hurdle, which is getting the C-suite to treat access as a strategic asset.
370
00:15:42,520 --> 00:15:44,680
You cannot govern a million agents from the basement.
371
00:15:44,680 --> 00:15:46,200
You need executive alignment.
372
00:15:46,200 --> 00:15:49,080
It is time to move the conversation from the server room to the boardroom.
373
00:15:49,080 --> 00:15:52,040
The executive buy-in strategy.
374
00:15:52,040 --> 00:15:55,240
The technical battle for context-aware security is being won in the code,
375
00:15:55,240 --> 00:15:58,040
but the strategic battle is still being lost in the boardroom.
376
00:15:58,040 --> 00:16:01,640
For years, identity and access management was treated like a plumbing issue.
377
00:16:01,640 --> 00:16:05,720
It was something for the basement teams to handle while the business focused on growth.
378
00:16:05,720 --> 00:16:08,840
In 2026, that mindset is a liability.
379
00:16:08,840 --> 00:16:13,480
We have to reframe the entire conversation because identity is no longer just a login.
380
00:16:13,480 --> 00:16:16,040
It is the core control plane for digital trust.
381
00:16:16,040 --> 00:16:20,120
If your leadership doesn't understand that, your AI deployment will remain a series of expensive,
382
00:16:20,120 --> 00:16:21,400
disconnected experiments.
383
00:16:21,400 --> 00:16:24,440
We need to stop pitching security as a cost center.
384
00:16:24,440 --> 00:16:27,400
Instead, we have to frame it as a decision-villacity enabler.
385
00:16:27,400 --> 00:16:30,360
Most executives view security as a series of no moments.
386
00:16:30,360 --> 00:16:31,560
No, you can't use that tool.
387
00:16:31,560 --> 00:16:33,080
No, you can't share that file.
388
00:16:33,080 --> 00:16:35,240
But a modern identity fabric does the opposite.
389
00:16:35,240 --> 00:16:38,520
It provides the yes that allows the business to move faster.
390
00:16:38,520 --> 00:16:41,160
When you have continuous context-aware verification,
391
00:16:41,160 --> 00:16:44,520
you can onboard a new partner in hours instead of weeks.
392
00:16:44,520 --> 00:16:48,360
You can give a consultant access to exactly what they need for a three-day project,
393
00:16:48,360 --> 00:16:51,880
and then you can watch that access vanish automatically the moment they finished.
394
00:16:51,880 --> 00:16:53,000
This isn't about restriction.
395
00:16:53,000 --> 00:16:54,120
It's about agility.
396
00:16:54,120 --> 00:16:57,960
It's about creating an environment where the right people can act on the right information
397
00:16:57,960 --> 00:16:59,800
without the friction of manual approvals.
398
00:16:59,800 --> 00:17:02,680
We can see the cost of ignoring this in the M&A market.
399
00:17:02,680 --> 00:17:05,560
There is a massive maturity gap emerging between organizations
400
00:17:05,560 --> 00:17:10,200
that have mastered identity governance and those that are still relying on static legacy models.
401
00:17:10,200 --> 00:17:12,600
When a high maturity firm acquires a competitor,
402
00:17:12,600 --> 00:17:16,360
they can integrate the new workforce into their AI ecosystem in days.
403
00:17:16,360 --> 00:17:18,360
But for firms with poor identity hygiene,
404
00:17:18,360 --> 00:17:20,040
that integration drags on for months,
405
00:17:20,040 --> 00:17:24,600
and this delay bleeds value while stalling the very synergies the merger was supposed to create.
406
00:17:24,600 --> 00:17:27,880
Poor identity governance is a direct hit to post-merger value.
407
00:17:27,880 --> 00:17:32,200
It is a drag on the balance sheet that most CEOs haven't even quantified yet.
408
00:17:32,200 --> 00:17:35,880
This is why the pitch for a 5-to-20% budget increase for identity fabric
409
00:17:35,880 --> 00:17:37,880
isn't a request for more securities bend.
410
00:17:37,880 --> 00:17:39,640
It's an investment in growth infrastructure.
411
00:17:39,640 --> 00:17:43,640
It's the digital equivalent of building a high-speed rail network between your data silos.
412
00:17:43,640 --> 00:17:46,760
You are moving from a hardened perimeter that everyone tries to bypass
413
00:17:46,760 --> 00:17:50,840
to a continuous verification model that actually supports the way people work.
414
00:17:50,840 --> 00:17:56,200
You are telling the C-suite that if they want the five times productivity boost that AI promises,
415
00:17:56,200 --> 00:17:58,520
they have to pay for the foundation that makes it safe.
416
00:17:58,520 --> 00:18:03,400
You have to move the goalposts from protecting the data to managing the flow of intelligence.
417
00:18:03,400 --> 00:18:07,080
The shift to a context-aware model is the only way to build a resilient enterprise
418
00:18:07,080 --> 00:18:08,680
in an age of a genetic AI.
419
00:18:08,680 --> 00:18:12,680
It allows the organization to scale its intelligence without scaling its risk.
420
00:18:12,680 --> 00:18:15,560
When the C-suite treats access as a strategic asset,
421
00:18:15,560 --> 00:18:17,720
the entire culture of the company shifts.
422
00:18:17,720 --> 00:18:19,560
Security becomes a shared responsibility,
423
00:18:19,560 --> 00:18:22,520
and data becomes a liquid resource rather than a hoarded secret.
424
00:18:22,520 --> 00:18:24,840
It's time to stop looking at the cost of the tools
425
00:18:24,840 --> 00:18:26,840
and start looking at the cost of the friction.
426
00:18:26,840 --> 00:18:29,880
Because in the AI era, the fastest company wins,
427
00:18:29,880 --> 00:18:32,680
and you can't be fast if your data is trapped in a lie.
428
00:18:32,680 --> 00:18:34,920
Your 2026 mandate is clear.
429
00:18:34,920 --> 00:18:38,520
You must audit your permission debt before you scale your AI licenses.
430
00:18:38,520 --> 00:18:41,400
The technology is ready, but your architecture likely isn't.
431
00:18:41,400 --> 00:18:45,240
The challenge for the next quarter is to implement mandatory auto labeling
432
00:18:45,240 --> 00:18:49,000
and execute a 90-day site cleanup to close the most obvious gaps.
433
00:18:49,000 --> 00:18:50,600
Stop the hooding, start the flow.
434
00:18:50,600 --> 00:18:55,400
If you want to share your silo-breaking progress or discuss how your team is handling the shift
435
00:18:55,400 --> 00:18:58,840
to context-aware permissions, connect with Mirko Peters on LinkedIn.
436
00:18:58,840 --> 00:19:01,160
Let's build an intelligence model that actually works.
437
00:19:01,160 --> 00:19:03,640
Subscribe for more deep dives into the future of work.
438
00:19:03,640 --> 00:19:05,160
Stay secure, stay productive.
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.
