Is Copilot Studio Replacing Low-Code Developers: The Future of Managed Business Logic

1
00:00:00,000 --> 00:00:01,920
Most developers in the Microsoft ecosystem

2
00:00:01,920 --> 00:00:04,080
are still spending their days building screens.

3
00:00:04,080 --> 00:00:06,680
They are focused on canvas apps, navigation panels,

4
00:00:06,680 --> 00:00:09,040
and button controls while tying power fx formulas

5
00:00:09,040 --> 00:00:10,360
to various gallery components.

6
00:00:10,360 --> 00:00:12,000
That work is real, it ships to users,

7
00:00:12,000 --> 00:00:13,600
and it definitely solves problems.

8
00:00:13,600 --> 00:00:15,160
But the platform underneath those screens

9
00:00:15,160 --> 00:00:17,880
has been rebuilt, and most people simply haven't noticed it yet.

10
00:00:17,880 --> 00:00:19,360
The assumption that low-code always means

11
00:00:19,360 --> 00:00:21,080
building a canvas app is broken.

12
00:00:21,080 --> 00:00:22,840
Copilot Studio now owns the logic layer

13
00:00:22,840 --> 00:00:25,640
because Microsoft moved their entire center of investment.

14
00:00:25,640 --> 00:00:27,880
Developers who don't understand what that shift means

15
00:00:27,880 --> 00:00:30,560
will soon find themselves building on the wrong foundation.

16
00:00:30,560 --> 00:00:32,400
This episode is about that structural shift

17
00:00:32,400 --> 00:00:35,120
rather than the marketing version you see in slide decks.

18
00:00:35,120 --> 00:00:36,880
We are looking at what actually changed,

19
00:00:36,880 --> 00:00:39,040
why the old model has a flaw baked into it,

20
00:00:39,040 --> 00:00:40,760
and what the new architecture looks like

21
00:00:40,760 --> 00:00:42,600
at the layer most people are ignoring.

22
00:00:42,600 --> 00:00:45,200
If you want to stay ahead of where this platform is actually going

23
00:00:45,200 --> 00:00:47,440
instead of just following what was announced last month,

24
00:00:47,440 --> 00:00:48,200
hit subscribe.

25
00:00:48,200 --> 00:00:50,240
This is the M365FM podcast.

26
00:00:50,240 --> 00:00:52,120
I'm Mirko Peters, let's get into it.

27
00:00:52,120 --> 00:00:53,200
The old model.

28
00:00:53,200 --> 00:00:55,400
What low-code development actually was?

29
00:00:55,400 --> 00:00:57,080
To understand what is breaking right now,

30
00:00:57,080 --> 00:00:58,560
you have to understand what was built

31
00:00:58,560 --> 00:01:00,280
and why it made sense at the time.

32
00:01:00,280 --> 00:01:02,680
From roughly 2018 through 2024,

33
00:01:02,680 --> 00:01:05,840
the M365 developer story had one dominant answer

34
00:01:05,840 --> 00:01:07,040
for business automation.

35
00:01:07,040 --> 00:01:08,840
That answer was Power Apps Canvas.

36
00:01:08,840 --> 00:01:11,280
You had a process that lived in spreadsheets or email,

37
00:01:11,280 --> 00:01:13,320
a business user who needed something faster,

38
00:01:13,320 --> 00:01:15,600
and a developer who could turn that around in days.

39
00:01:15,600 --> 00:01:17,000
That was the value proposition.

40
00:01:17,000 --> 00:01:19,680
It was about speed and proximity to the business

41
00:01:19,680 --> 00:01:21,600
rather than architectural elegance.

42
00:01:21,600 --> 00:01:22,560
And it worked.

43
00:01:22,560 --> 00:01:24,800
The alternative back then was a full development project

44
00:01:24,800 --> 00:01:27,520
with a team of engineers in a six-month backlog.

45
00:01:27,520 --> 00:01:30,440
Requirements processes often outlasted the original problem,

46
00:01:30,440 --> 00:01:32,480
so Canvas Apps filled a real gap.

47
00:01:32,480 --> 00:01:34,880
They let organizations build real solutions quickly

48
00:01:34,880 --> 00:01:36,320
with people who understood the business

49
00:01:36,320 --> 00:01:39,280
instead of people who only understood distributed systems.

50
00:01:39,280 --> 00:01:41,680
The assumption baked into every Canvas app was simple.

51
00:01:41,680 --> 00:01:43,080
Users know what they are looking for.

52
00:01:43,080 --> 00:01:44,680
Navigation is the interface.

53
00:01:44,680 --> 00:01:45,760
Screens are the solution.

54
00:01:45,760 --> 00:01:49,040
You build a form for data entry, a gallery for data display,

55
00:01:49,040 --> 00:01:50,560
and a button to trigger a flow.

56
00:01:50,560 --> 00:01:53,400
Then you wire it all together with PowerFX.

57
00:01:53,400 --> 00:01:55,200
That model is intuitive because it mirrors

58
00:01:55,200 --> 00:01:57,760
how people think about software as a collection of pages

59
00:01:57,760 --> 00:01:58,760
you move between.

60
00:01:58,760 --> 00:02:01,120
Local developers were valued for exactly that.

61
00:02:01,120 --> 00:02:03,960
They offered speed of delivery and closeness to business users.

62
00:02:03,960 --> 00:02:05,920
They had the ability to turn a whiteboard sketch

63
00:02:05,920 --> 00:02:09,240
into a working app before the meeting notes were even written up.

64
00:02:09,240 --> 00:02:11,160
Architectural depth was never the point.

65
00:02:11,160 --> 00:02:13,600
Getting the job done was the only thing that mattered.

66
00:02:13,600 --> 00:02:15,560
But there was a structural flaw in that model

67
00:02:15,560 --> 00:02:18,240
that nobody talked about much because the apps were small enough

68
00:02:18,240 --> 00:02:19,440
that it didn't matter yet.

69
00:02:19,440 --> 00:02:21,800
The flaw was this, logic lived on screens.

70
00:02:21,800 --> 00:02:23,880
Business rules weren't kept in a logic layer.

71
00:02:23,880 --> 00:02:25,320
They were scattered across controls.

72
00:02:25,320 --> 00:02:27,880
You had visibility formulas on buttons and validation logic

73
00:02:27,880 --> 00:02:29,080
in on-change events.

74
00:02:29,080 --> 00:02:30,960
Conditional formatting was tied to variables

75
00:02:30,960 --> 00:02:32,600
that were set three screens back.

76
00:02:32,600 --> 00:02:34,280
If you needed to know what the app actually

77
00:02:34,280 --> 00:02:37,440
did under specific conditions, you didn't read a specification.

78
00:02:37,440 --> 00:02:39,040
You traced formulas across controls

79
00:02:39,040 --> 00:02:40,800
and hoped you hadn't missed one.

80
00:02:40,800 --> 00:02:42,600
Every new requirement meant another screen.

81
00:02:42,600 --> 00:02:44,360
And that usually meant another set of formulas

82
00:02:44,360 --> 00:02:46,800
that partially duplicated what already existed.

83
00:02:46,800 --> 00:02:48,480
There was no single source of truth

84
00:02:48,480 --> 00:02:50,040
for what the business rules were.

85
00:02:50,040 --> 00:02:51,840
The app itself was the documentation.

86
00:02:51,840 --> 00:02:53,520
That worked when apps were small.

87
00:02:53,520 --> 00:02:55,920
When a department needed one app for one process,

88
00:02:55,920 --> 00:02:58,520
a canvas solution with 15 screens was manageable.

89
00:02:58,520 --> 00:03:01,160
A competent developer could hold the whole thing in their head.

90
00:03:01,160 --> 00:03:02,880
But organizations grew their appestates

91
00:03:02,880 --> 00:03:04,600
and processes got more complex.

92
00:03:04,600 --> 00:03:07,400
The same logic needed to work across multiple apps, channels,

93
00:03:07,400 --> 00:03:08,200
and devices.

94
00:03:08,200 --> 00:03:10,760
Suddenly, the canvas model didn't just feel stretched.

95
00:03:10,760 --> 00:03:13,800
It was actively working against maintainability.

96
00:03:13,800 --> 00:03:15,640
Apps with more than 20 or 30 screens

97
00:03:15,640 --> 00:03:18,160
became effectively un-maintainable by anyone

98
00:03:18,160 --> 00:03:19,640
except the person who built them.

99
00:03:19,640 --> 00:03:21,960
Sometimes even the original creator couldn't figure it out

100
00:03:21,960 --> 00:03:23,120
six months later.

101
00:03:23,120 --> 00:03:26,040
The delegation problem made performance unpredictable at scale,

102
00:03:26,040 --> 00:03:28,480
non-delegable queries ran fine in development,

103
00:03:28,480 --> 00:03:32,080
but fell apart in production with real data volumes.

104
00:03:32,080 --> 00:03:35,440
Logic embedded in UI controls couldn't be shared, reused,

105
00:03:35,440 --> 00:03:37,120
or tested in isolation.

106
00:03:37,120 --> 00:03:40,520
That model wasn't wrong for what it was solving in 2019.

107
00:03:40,520 --> 00:03:42,960
But the platform underneath it has been rebuilt.

108
00:03:42,960 --> 00:03:44,600
The rebuild didn't just add features.

109
00:03:44,600 --> 00:03:47,360
It changed where logic is supposed to live.

110
00:03:47,360 --> 00:03:48,720
The platform shifted.

111
00:03:48,720 --> 00:03:50,520
What Microsoft actually changed.

112
00:03:50,520 --> 00:03:52,360
Here is what Microsoft actually did.

113
00:03:52,360 --> 00:03:54,880
And why it matters more than most people realize.

114
00:03:54,880 --> 00:03:56,560
The roadmap pivot isn't speculation.

115
00:03:56,560 --> 00:03:58,920
It is documented in release wave announcements,

116
00:03:58,920 --> 00:04:01,680
developer tooling decisions, and where the product teams

117
00:04:01,680 --> 00:04:03,880
are concentrating engineering resources.

118
00:04:03,880 --> 00:04:07,280
Around 2025 and into 2026, the center of innovation

119
00:04:07,280 --> 00:04:08,440
at Microsoft moved.

120
00:04:08,440 --> 00:04:10,240
It didn't move away from the power platform entirely,

121
00:04:10,240 --> 00:04:12,560
but it shifted decisively away from Canvas apps

122
00:04:12,560 --> 00:04:13,960
as the primary way to build.

123
00:04:13,960 --> 00:04:16,760
The new focus is co-pilot studio and declarative agents

124
00:04:16,760 --> 00:04:19,880
as the first class development story for AI first scenarios.

125
00:04:19,880 --> 00:04:21,400
Canvas apps are not deprecated.

126
00:04:21,400 --> 00:04:24,040
That needs to be said clearly because the nuance matters.

127
00:04:24,040 --> 00:04:26,360
Microsoft is still shipping features for power apps,

128
00:04:26,360 --> 00:04:28,760
connectors are improving, and the AI assisted build

129
00:04:28,760 --> 00:04:31,160
experience inside the designer is getting better.

130
00:04:31,160 --> 00:04:32,520
Canvas is not going away.

131
00:04:32,520 --> 00:04:35,840
But still supported and where new investment is concentrated

132
00:04:35,840 --> 00:04:37,360
are two different things.

133
00:04:37,360 --> 00:04:39,120
Confusing them is how you end up optimizing

134
00:04:39,120 --> 00:04:41,240
for a layer the platform is moving past.

135
00:04:41,240 --> 00:04:42,480
The signal is in the tooling.

136
00:04:42,480 --> 00:04:45,560
Microsoft released the M365 agent toolkit,

137
00:04:45,560 --> 00:04:47,800
which is a visual studio code extension purpose

138
00:04:47,800 --> 00:04:50,520
built for creating, deploying, and publishing agents

139
00:04:50,520 --> 00:04:54,000
directly into Microsoft 365 co-pilot.

140
00:04:54,000 --> 00:04:56,600
That toolkit covers agent creation, packaging, provisioning,

141
00:04:56,600 --> 00:04:58,680
testing, debugging, and store publication.

142
00:04:58,680 --> 00:05:00,640
It is a full development life cycle,

143
00:05:00,640 --> 00:05:02,800
and it is built around agents instead of apps.

144
00:05:02,800 --> 00:05:04,720
When a platform ships a dedicated toolkit

145
00:05:04,720 --> 00:05:07,640
for a new paradigm that is not just a feature announcement,

146
00:05:07,640 --> 00:05:10,280
it is a statement about where the developer story is heading.

147
00:05:10,280 --> 00:05:12,360
The declarative agent model is worth pausing on

148
00:05:12,360 --> 00:05:14,280
because it is architecturally different

149
00:05:14,280 --> 00:05:15,960
from everything that came before it.

150
00:05:15,960 --> 00:05:17,880
When you build a declarative agent,

151
00:05:17,880 --> 00:05:20,120
you are not writing imperative code.

152
00:05:20,120 --> 00:05:21,240
You are writing instructions.

153
00:05:21,240 --> 00:05:23,760
You define who the agent is, what it knows, what it can do,

154
00:05:23,760 --> 00:05:26,600
and how it should behave in JSON, manifests, and instruction

155
00:05:26,600 --> 00:05:27,240
files.

156
00:05:27,240 --> 00:05:29,760
Co-pilot handles the reasoning, the orchestration,

157
00:05:29,760 --> 00:05:31,400
the memory, and the scaling.

158
00:05:31,400 --> 00:05:34,400
You are describing outcomes rather than programming steps.

159
00:05:34,400 --> 00:05:36,960
In the March of 2026, Microsoft 365

160
00:05:36,960 --> 00:05:38,880
and Power Platform Community call,

161
00:05:38,880 --> 00:05:40,240
Microsoft described it directly.

162
00:05:40,240 --> 00:05:42,840
They said you do not write code when building declarative agents.

163
00:05:42,840 --> 00:05:44,160
You write instructions.

164
00:05:44,160 --> 00:05:46,040
That is not a simplification for beginners.

165
00:05:46,040 --> 00:05:47,520
That is the architectural model.

166
00:05:47,520 --> 00:05:49,000
The logic layer is no longer something

167
00:05:49,000 --> 00:05:50,640
you wire together manually.

168
00:05:50,640 --> 00:05:53,480
It is something you configure, constrain, and govern.

169
00:05:53,480 --> 00:05:56,240
What that means structurally is that the app as a visual container

170
00:05:56,240 --> 00:05:58,280
is losing its role as the primary front door

171
00:05:58,280 --> 00:05:59,920
for enterprise interactions.

172
00:05:59,920 --> 00:06:02,800
Microsoft is positioning co-pilot as the universal client.

173
00:06:02,800 --> 00:06:05,440
Users start with intent rather than navigation.

174
00:06:05,440 --> 00:06:07,800
They do not open an app and look for the right screen,

175
00:06:07,800 --> 00:06:09,280
but instead they express a goal

176
00:06:09,280 --> 00:06:11,280
and the system figures out how to fulfill it.

177
00:06:11,280 --> 00:06:13,320
The canvas model assumed users would navigate

178
00:06:13,320 --> 00:06:14,320
to what they needed.

179
00:06:14,320 --> 00:06:17,440
The agentic model assumes users will describe what they need

180
00:06:17,440 --> 00:06:19,480
and the system will decide what to do next.

181
00:06:19,480 --> 00:06:20,800
This is not a cosmetic change.

182
00:06:20,800 --> 00:06:22,800
It is a change in where decision-making authority

183
00:06:22,800 --> 00:06:24,240
lives in the architecture.

184
00:06:24,240 --> 00:06:25,920
In the canvas model, the developer made

185
00:06:25,920 --> 00:06:27,840
the decisions about what screens exist,

186
00:06:27,840 --> 00:06:31,320
what paths users can take, and what happens when a condition is met.

187
00:06:31,320 --> 00:06:33,440
In the agentic model, the agent makes those decisions

188
00:06:33,440 --> 00:06:35,720
at runtime constrained by the instructions

189
00:06:35,720 --> 00:06:37,920
and governance rules the architect defined.

190
00:06:37,920 --> 00:06:40,680
The developer's job shifts from building the decision tree

191
00:06:40,680 --> 00:06:43,320
to designing the boundaries within which the agent operates.

192
00:06:43,320 --> 00:06:45,760
That shift has implications for what skills matter,

193
00:06:45,760 --> 00:06:48,880
what roles get hired and what kind of work survives automation.

194
00:06:48,880 --> 00:06:51,400
But before we get there, we need to understand exactly

195
00:06:51,400 --> 00:06:52,880
what co-pilot studio is.

196
00:06:52,880 --> 00:06:54,880
There is a widespread misconception about it

197
00:06:54,880 --> 00:06:57,960
that is causing teams to under-invest in the right places.

198
00:06:57,960 --> 00:07:00,000
Most people think of it as a chatbot builder.

199
00:07:00,000 --> 00:07:01,400
That is the wrong mental model

200
00:07:01,400 --> 00:07:04,600
and it is leading to the wrong architecture decisions.

201
00:07:04,600 --> 00:07:07,640
Co-pilot studio is not a chatbot, the reasoning layer explained.

202
00:07:07,640 --> 00:07:09,240
Let's deal with the misconception directly

203
00:07:09,240 --> 00:07:11,600
because it is the most expensive mistake teams make

204
00:07:11,600 --> 00:07:14,040
when they first encounter co-pilot studio.

205
00:07:14,040 --> 00:07:17,160
The assumption most people bring to it is that it is a fancier bot builder.

206
00:07:17,160 --> 00:07:20,080
They see it as a low-code wrapper around a language model.

207
00:07:20,080 --> 00:07:22,240
It might be a step up from power virtual agents,

208
00:07:22,240 --> 00:07:24,920
but they think it is fundamentally the same category of thing.

209
00:07:24,920 --> 00:07:26,760
They view it as a conversational interface.

210
00:07:26,760 --> 00:07:28,240
You bolt onto an existing process

211
00:07:28,240 --> 00:07:30,560
so users can ask questions instead of clicking buttons.

212
00:07:30,560 --> 00:07:31,680
That framing is wrong.

213
00:07:31,680 --> 00:07:35,000
Building on it produces architecture that will fail at scale.

214
00:07:35,000 --> 00:07:37,800
Co-pilot studio is a gold-driven orchestration engine.

215
00:07:37,800 --> 00:07:39,400
The distinction is not cosmetic.

216
00:07:39,400 --> 00:07:42,040
A chatbot receives a message and returns a response.

217
00:07:42,040 --> 00:07:44,120
Co-pilot studio receives a goal,

218
00:07:44,120 --> 00:07:47,000
expressed as natural language, a structured input,

219
00:07:47,000 --> 00:07:48,160
or an event trigger,

220
00:07:48,160 --> 00:07:50,440
and then it decides what needs to happen to fulfill it.

221
00:07:50,440 --> 00:07:51,520
It interprets intent.

222
00:07:51,520 --> 00:07:54,240
It selects the right tools from the set available to it.

223
00:07:54,240 --> 00:07:55,720
It executes across systems.

224
00:07:55,720 --> 00:07:58,240
It evaluates whether the result actually addressed the goal.

225
00:07:58,240 --> 00:08:01,240
If something changes or fails mid-execution, it adapts.

226
00:08:01,240 --> 00:08:03,560
That is a fundamentally different thing from a bot.

227
00:08:03,560 --> 00:08:05,040
The clearest way to see the difference

228
00:08:05,040 --> 00:08:07,440
is to compare it directly with power automate.

229
00:08:07,440 --> 00:08:09,480
When you build a flow in power automate,

230
00:08:09,480 --> 00:08:11,360
you are the one making every decision.

231
00:08:11,360 --> 00:08:13,560
You decide the trigger you define every branch

232
00:08:13,560 --> 00:08:15,520
and you specify the exact sequence of actions.

233
00:08:15,520 --> 00:08:17,680
The flow executes exactly what you built

234
00:08:17,680 --> 00:08:19,800
in the order you defined at every single time.

235
00:08:19,800 --> 00:08:20,720
It does not interpret.

236
00:08:20,720 --> 00:08:21,560
It does not evaluate.

237
00:08:21,560 --> 00:08:22,960
It runs the steps you wrote.

238
00:08:22,960 --> 00:08:25,040
Co-pilot studio agents do not run your steps.

239
00:08:25,040 --> 00:08:26,480
They decide which steps to run.

240
00:08:26,480 --> 00:08:28,520
You provide instructions that describe the goal,

241
00:08:28,520 --> 00:08:30,440
the constraints, the available tools,

242
00:08:30,440 --> 00:08:32,080
and the acceptable behaviors.

243
00:08:32,080 --> 00:08:34,840
The agent reasons over those instructions at runtime

244
00:08:34,840 --> 00:08:37,600
and determines what to do based on the context it is operating

245
00:08:37,600 --> 00:08:39,680
in two requests that look similar on the surface

246
00:08:39,680 --> 00:08:41,880
might result in completely different action sequences

247
00:08:41,880 --> 00:08:43,160
because the context is different.

248
00:08:43,160 --> 00:08:45,360
This is not a subtle difference in implementation.

249
00:08:45,360 --> 00:08:46,920
It is a different execution model.

250
00:08:46,920 --> 00:08:49,360
The implication for business logic is significant.

251
00:08:49,360 --> 00:08:51,280
In the Canvas app model, business logic

252
00:08:51,280 --> 00:08:53,320
was encoded in formulas and flows,

253
00:08:53,320 --> 00:08:55,440
which was explicit, imperative, and manual.

254
00:08:55,440 --> 00:08:56,680
You wrote what happened.

255
00:08:56,680 --> 00:08:58,560
In the agentic model, business logic

256
00:08:58,560 --> 00:09:01,080
is encoded in instructions, tool configurations,

257
00:09:01,080 --> 00:09:02,400
and orchestration rules.

258
00:09:02,400 --> 00:09:05,440
The agent interprets those rules and applies them dynamically.

259
00:09:05,440 --> 00:09:07,480
Logic does not live in a specific formula

260
00:09:07,480 --> 00:09:08,760
on a specific screen.

261
00:09:08,760 --> 00:09:10,640
It lives in the system-level definition

262
00:09:10,640 --> 00:09:13,120
of what the agent is allowed to do, what it knows,

263
00:09:13,120 --> 00:09:15,160
and what outcomes it is trying to achieve.

264
00:09:15,160 --> 00:09:17,640
That is a harder thing to build correctly.

265
00:09:17,640 --> 00:09:19,760
It requires a different kind of thinking.

266
00:09:19,760 --> 00:09:21,920
You aren't asking what steps happen in what order,

267
00:09:21,920 --> 00:09:23,480
but rather what constraints and goals

268
00:09:23,480 --> 00:09:26,360
define correct behavior across all possible inputs.

269
00:09:26,360 --> 00:09:28,920
That shift from procedural to declarative logic design

270
00:09:28,920 --> 00:09:31,280
is exactly why the skills needed to work effectively

271
00:09:31,280 --> 00:09:33,640
with Copilot Studio are different from the skills

272
00:09:33,640 --> 00:09:35,920
that made someone effective in Power Apps.

273
00:09:35,920 --> 00:09:37,720
One more thing worth being precise about

274
00:09:37,720 --> 00:09:40,160
because it comes up constantly in team discussions

275
00:09:40,160 --> 00:09:43,240
is that Copilot Studio is not a wrapper for a public language

276
00:09:43,240 --> 00:09:44,000
model.

277
00:09:44,000 --> 00:09:45,560
The agents you build in Copilot Studio

278
00:09:45,560 --> 00:09:48,440
run against enterprise data under enterprise permissions

279
00:09:48,440 --> 00:09:51,320
inside the Microsoft 365 security boundary.

280
00:09:51,320 --> 00:09:53,400
They can be grounded on SharePoint content,

281
00:09:53,400 --> 00:09:55,680
Dataverse Records, external APIs, and knowledge

282
00:09:55,680 --> 00:09:56,920
bases that you define.

283
00:09:56,920 --> 00:09:59,240
The reasoning happens inside a governed environment,

284
00:09:59,240 --> 00:10:00,680
not on top of a public API.

285
00:10:00,680 --> 00:10:02,960
That distinction matters enormously for compliance,

286
00:10:02,960 --> 00:10:05,840
for data handling, and for what kinds of use cases

287
00:10:05,840 --> 00:10:08,120
are actually viable in regulated industries.

288
00:10:08,120 --> 00:10:09,360
So when someone in your organization

289
00:10:09,360 --> 00:10:12,400
says we should build a Copilot Studio agent for this,

290
00:10:12,400 --> 00:10:15,000
the question to ask is not what the bot should say.

291
00:10:15,000 --> 00:10:17,680
The question is, what goal we are giving the agent,

292
00:10:17,680 --> 00:10:20,120
what tools we wanted to use, and what constraints

293
00:10:20,120 --> 00:10:21,880
we need to put around its decision making.

294
00:10:21,880 --> 00:10:23,200
That is an architecture question.

295
00:10:23,200 --> 00:10:24,720
It is not a UX question.

296
00:10:24,720 --> 00:10:26,120
And that brings us to the mental model

297
00:10:26,120 --> 00:10:29,160
that once you have it, reframes almost every decision

298
00:10:29,160 --> 00:10:30,880
you will make on this platform.

299
00:10:30,880 --> 00:10:32,240
Judgment versus logic.

300
00:10:32,240 --> 00:10:34,520
The architectural split, that changes everything.

301
00:10:34,520 --> 00:10:36,160
Think of this as your new mental model.

302
00:10:36,160 --> 00:10:38,200
You need to hold onto this one because it reframes

303
00:10:38,200 --> 00:10:39,600
every single architecture decision

304
00:10:39,600 --> 00:10:41,760
you will make on this platform from here and out.

305
00:10:41,760 --> 00:10:43,680
In any automated system, you are dealing

306
00:10:43,680 --> 00:10:45,720
with two fundamentally different types of work.

307
00:10:45,720 --> 00:10:47,840
There is judgment, which is the act of deciding

308
00:10:47,840 --> 00:10:49,560
what should happen, and then there is logic,

309
00:10:49,560 --> 00:10:51,160
which is the act of making it happen.

310
00:10:51,160 --> 00:10:53,400
These two forces have always existed inside every piece

311
00:10:53,400 --> 00:10:55,520
of software ever built, but something has changed.

312
00:10:55,520 --> 00:10:57,400
We finally have a platform where these two things

313
00:10:57,400 --> 00:10:59,560
can be handled by completely different layers,

314
00:10:59,560 --> 00:11:02,840
and each layer is optimized for its own specific job.

315
00:11:02,840 --> 00:11:04,240
Power automate owns the logic.

316
00:11:04,240 --> 00:11:05,760
This isn't a limitation of the tool.

317
00:11:05,760 --> 00:11:07,760
It is a core design principle.

318
00:11:07,760 --> 00:11:10,720
When your input is structured, your rules are set in stone,

319
00:11:10,720 --> 00:11:11,920
and your volume is high.

320
00:11:11,920 --> 00:11:13,440
Power automate is the right choice.

321
00:11:13,440 --> 00:11:15,560
You use it when the output needs to be auditable

322
00:11:15,560 --> 00:11:17,760
and reproducible every single time.

323
00:11:17,760 --> 00:11:20,040
It runs the exact same steps in the exact same way,

324
00:11:20,040 --> 00:11:21,240
which is not a weakness.

325
00:11:21,240 --> 00:11:23,080
In fact, that is exactly what you want

326
00:11:23,080 --> 00:11:25,920
when you are provisioning accounts, assigning licenses,

327
00:11:25,920 --> 00:11:27,280
or routing approvals.

328
00:11:27,280 --> 00:11:30,200
Deterministic processes require deterministic engines

329
00:11:30,200 --> 00:11:31,280
to stay reliable.

330
00:11:31,280 --> 00:11:33,840
Copilot Studio, on the other hand, owns the judgment.

331
00:11:33,840 --> 00:11:35,360
This is the territory of the agent.

332
00:11:35,360 --> 00:11:38,120
You use it when the input is messy, or the right action

333
00:11:38,120 --> 00:11:41,600
depends on context that you can't reduce to a simple if-then rule.

334
00:11:41,600 --> 00:11:44,200
It is built to reason across multiple data sources

335
00:11:44,200 --> 00:11:46,120
and adapt based on what it finds.

336
00:11:46,120 --> 00:11:48,560
Think about exception handling, dynamic routing,

337
00:11:48,560 --> 00:11:50,520
or requests that arrive in natural language

338
00:11:50,520 --> 00:11:52,040
with half the information missing.

339
00:11:52,040 --> 00:11:53,920
That kind of work breaks the power automate flow

340
00:11:53,920 --> 00:11:55,920
because flows weren't designed to think.

341
00:11:55,920 --> 00:11:57,880
They were designed to execute decisions

342
00:11:57,880 --> 00:12:00,360
that a human already made, whereas the agent

343
00:12:00,360 --> 00:12:02,400
is the thing actually making the decision.

344
00:12:02,400 --> 00:12:04,680
The mistake most teams make is using power automate

345
00:12:04,680 --> 00:12:07,480
for judgment calls, and it ends up being a very expensive error.

346
00:12:07,480 --> 00:12:08,720
You have probably seen this yourself.

347
00:12:08,720 --> 00:12:11,680
It starts with a flow that has 15 nested conditions

348
00:12:11,680 --> 00:12:13,280
branching across every edge case

349
00:12:13,280 --> 00:12:14,600
the developer could imagine.

350
00:12:14,600 --> 00:12:16,440
Then, after the flow goes live,

351
00:12:16,440 --> 00:12:18,280
a dozen more cases get bolted on.

352
00:12:18,280 --> 00:12:20,360
Every new exception adds a new branch,

353
00:12:20,360 --> 00:12:23,320
and every branch adds more surface area to maintain.

354
00:12:23,320 --> 00:12:24,800
What started as clean automation

355
00:12:24,800 --> 00:12:26,360
turns into a fragile decision tree

356
00:12:26,360 --> 00:12:28,360
that snaps the moment reality doesn't match

357
00:12:28,360 --> 00:12:29,840
a condition written six months ago.

358
00:12:29,840 --> 00:12:31,720
That is what happens when you force judgment work

359
00:12:31,720 --> 00:12:32,840
through a logic engine.

360
00:12:32,840 --> 00:12:34,040
The other mistake is less common,

361
00:12:34,040 --> 00:12:35,280
but it is actually more dangerous.

362
00:12:35,280 --> 00:12:38,040
This is when people use co-pilot studio for work

363
00:12:38,040 --> 00:12:41,360
that requires exact 100% reproducibility.

364
00:12:41,360 --> 00:12:44,080
If you are running a process where the output must be identical

365
00:12:44,080 --> 00:12:47,440
every time, like a financial calculation or a compliance check,

366
00:12:47,440 --> 00:12:49,480
you do not want an agent making that call.

367
00:12:49,480 --> 00:12:51,000
Probabilistic reasoning has no business

368
00:12:51,000 --> 00:12:53,240
being inside a deterministic process.

369
00:12:53,240 --> 00:12:56,240
The risk of a hallucination isn't just a theory in these cases.

370
00:12:56,240 --> 00:12:58,960
It is a design flaw that is waiting to surface

371
00:12:58,960 --> 00:13:00,520
in your production environment.

372
00:13:00,520 --> 00:13:03,600
A healthy architecture keeps these layers clean and explicit.

373
00:13:03,600 --> 00:13:05,120
The agent handles what should happen

374
00:13:05,120 --> 00:13:07,080
while the flows handle how it happens.

375
00:13:07,080 --> 00:13:08,280
The agent receives a goal,

376
00:13:08,280 --> 00:13:10,840
reasons through the data, determines the right move,

377
00:13:10,840 --> 00:13:12,680
and then calls the flow to execute it.

378
00:13:12,680 --> 00:13:14,240
The flow doesn't decide anything.

379
00:13:14,240 --> 00:13:17,240
It simply does what it is told, reliably and at scale.

380
00:13:17,240 --> 00:13:19,480
The agent doesn't execute the individual steps.

381
00:13:19,480 --> 00:13:21,400
It orchestrates which steps get executed

382
00:13:21,400 --> 00:13:22,440
and when they need to happen.

383
00:13:22,440 --> 00:13:23,600
This isn't just a theory.

384
00:13:23,600 --> 00:13:26,400
This is the exact pattern that Microsoft and enterprise architects

385
00:13:26,400 --> 00:13:27,840
are moving toward right now.

386
00:13:27,840 --> 00:13:30,400
If you look at the technical guidance coming from Microsoft,

387
00:13:30,400 --> 00:13:32,560
or the patterns used by the partner community,

388
00:13:32,560 --> 00:13:34,840
or the architecture reviews of companies running real

389
00:13:34,840 --> 00:13:37,880
agente workflows, they all land in the same place.

390
00:13:37,880 --> 00:13:39,800
The agent acts as the reasoning layer

391
00:13:39,800 --> 00:13:42,320
and the flow acts as the execution layer.

392
00:13:42,320 --> 00:13:44,440
There is a clean, hard line between the two.

393
00:13:44,440 --> 00:13:47,440
Once you see the split, the flows of the old model become obvious.

394
00:13:47,440 --> 00:13:49,320
The old canvas app model collapse judgment

395
00:13:49,320 --> 00:13:51,800
and logic into the same place, which was the screen.

396
00:13:51,800 --> 00:13:54,280
Power FX formulas were trying to do both at once.

397
00:13:54,280 --> 00:13:55,520
They were deciding what should happen

398
00:13:55,520 --> 00:13:57,840
and making it happen inside a single UI control.

399
00:13:57,840 --> 00:14:01,160
That is why canvas apps became impossible to maintain at scale.

400
00:14:01,160 --> 00:14:03,400
It wasn't because PowerFX was a bad language.

401
00:14:03,400 --> 00:14:05,720
It was because putting judgment and logic in the same place

402
00:14:05,720 --> 00:14:07,880
means you can't change one without breaking the other.

403
00:14:07,880 --> 00:14:10,880
You can't govern the decisions independently from the execution

404
00:14:10,880 --> 00:14:13,840
and you certainly can't reuse those layers in other systems.

405
00:14:13,840 --> 00:14:16,720
The agente architecture does more than just change your tools.

406
00:14:16,720 --> 00:14:18,480
It finally fixes a structural problem

407
00:14:18,480 --> 00:14:20,720
that has been there from the beginning.

408
00:14:20,720 --> 00:14:22,840
Where canvas apps were actually breaking.

409
00:14:22,840 --> 00:14:25,600
We need to get specific about how these systems fail

410
00:14:25,600 --> 00:14:29,720
because saying canvas apps don't scale is too vague to be helpful.

411
00:14:29,720 --> 00:14:31,160
The failure isn't a random event.

412
00:14:31,160 --> 00:14:32,880
It follows a very predictable pattern

413
00:14:32,880 --> 00:14:35,240
that anyone working on a mature power platform estate

414
00:14:35,240 --> 00:14:36,720
has seen a hundred times.

415
00:14:36,720 --> 00:14:38,960
It usually starts with the duplication problem.

416
00:14:38,960 --> 00:14:40,920
When your business logic lives inside a button

417
00:14:40,920 --> 00:14:43,600
or a gallery on a screen, it doesn't get reused.

418
00:14:43,600 --> 00:14:44,680
It gets copied.

419
00:14:44,680 --> 00:14:47,440
A validation rule on screen 4 gets rebuilt on screen 11

420
00:14:47,440 --> 00:14:49,640
because the developer didn't know the first one existed

421
00:14:49,640 --> 00:14:51,320
or they couldn't get to it easily.

422
00:14:51,320 --> 00:14:53,120
Maybe they just needed a slightly different version

423
00:14:53,120 --> 00:14:54,920
and it felt faster to write it from scratch

424
00:14:54,920 --> 00:14:56,640
than to try and abstract the original.

425
00:14:56,640 --> 00:14:58,320
When you multiply that across in a state

426
00:14:58,320 --> 00:15:00,120
with dozens of apps built by different people

427
00:15:00,120 --> 00:15:02,480
over several years, you don't have a cohesive system.

428
00:15:02,480 --> 00:15:05,320
You have a pile of independent copies of business rules

429
00:15:05,320 --> 00:15:07,520
that are all slowly diverging from each other.

430
00:15:07,520 --> 00:15:09,200
When a business rule eventually changes,

431
00:15:09,200 --> 00:15:11,200
every single one of those copies has to be found

432
00:15:11,200 --> 00:15:12,400
and updated manually.

433
00:15:12,400 --> 00:15:13,920
Some of them always get missed.

434
00:15:13,920 --> 00:15:15,880
The ones that are forgotten create inconsistencies

435
00:15:15,880 --> 00:15:18,840
that stay hidden until they turn into errors in production.

436
00:15:18,840 --> 00:15:21,640
Usually you find them during an audit or a process handoff

437
00:15:21,640 --> 00:15:23,400
where one app produces an output

438
00:15:23,400 --> 00:15:25,160
that the next app doesn't recognize

439
00:15:25,160 --> 00:15:26,600
and then you hit the performance cliff.

440
00:15:26,600 --> 00:15:28,560
Canvas apps look great in a dev environment

441
00:15:28,560 --> 00:15:30,040
with a few dozen test records,

442
00:15:30,040 --> 00:15:33,080
but they degrade fast when they hit real world data volumes.

443
00:15:33,080 --> 00:15:34,400
This is the delegation problem.

444
00:15:34,400 --> 00:15:36,280
When you use a non-delegable function,

445
00:15:36,280 --> 00:15:38,640
the app has to pull records down to the user's device

446
00:15:38,640 --> 00:15:40,760
and filter them there instead of letting the database

447
00:15:40,760 --> 00:15:41,680
do the work.

448
00:15:41,680 --> 00:15:43,600
At a small scale, you won't even notice the lag.

449
00:15:43,600 --> 00:15:45,240
At an enterprise scale where you are dealing

450
00:15:45,240 --> 00:15:47,240
with SharePoint lists or dataverse tables

451
00:15:47,240 --> 00:15:49,120
with hundreds of thousands of rows,

452
00:15:49,120 --> 00:15:53,000
you end up loading thousands of records just to show 10.

453
00:15:53,000 --> 00:15:55,000
The performance tanks, the users complain

454
00:15:55,000 --> 00:15:56,880
and developers start adding hacks

455
00:15:56,880 --> 00:15:58,760
that make the logic even harder to read.

456
00:15:58,760 --> 00:16:01,280
The governance problem is even deeper than the technical issues.

457
00:16:01,280 --> 00:16:02,880
When your logic lives inside the UI,

458
00:16:02,880 --> 00:16:06,040
the only thing you can really govern is who has access to that UI.

459
00:16:06,040 --> 00:16:07,840
You can control who opens the app,

460
00:16:07,840 --> 00:16:09,920
but you can't easily govern the business decisions

461
00:16:09,920 --> 00:16:11,680
the app makes once a user is inside.

462
00:16:11,680 --> 00:16:14,000
There is no separation between who can use the tool

463
00:16:14,000 --> 00:16:15,760
and what the tool actually does.

464
00:16:15,760 --> 00:16:17,440
The decision logic and the access control

465
00:16:17,440 --> 00:16:19,000
are fused into the same file.

466
00:16:19,000 --> 00:16:21,080
That might be fine for a simple calculator,

467
00:16:21,080 --> 00:16:22,920
but it is a massive compliance nightmare

468
00:16:22,920 --> 00:16:26,000
for anything involving regulated data or audit requirements.

469
00:16:26,000 --> 00:16:28,400
Many organizations learned this lesson the hard way

470
00:16:28,400 --> 00:16:30,720
when they tried to move their apps to new channels.

471
00:16:30,720 --> 00:16:33,280
They wanted a mobile version of a desktop app

472
00:16:33,280 --> 00:16:36,120
or a team's integration or a portal for their customers.

473
00:16:36,120 --> 00:16:37,520
Every time they expanded,

474
00:16:37,520 --> 00:16:39,400
they had to rebuild the logic layer from scratch

475
00:16:39,400 --> 00:16:41,000
because that logic wasn't portable.

476
00:16:41,000 --> 00:16:43,760
It was stuck inside the controls of a specific screen layout.

477
00:16:43,760 --> 00:16:45,600
You couldn't just lift the business rules out

478
00:16:45,600 --> 00:16:47,640
and put them somewhere else because they were inseparable

479
00:16:47,640 --> 00:16:48,440
from the UI.

480
00:16:48,440 --> 00:16:51,440
This is the honest truth about what the Canvas era gave us.

481
00:16:51,440 --> 00:16:52,920
It solved the delivery problem.

482
00:16:52,920 --> 00:16:56,080
Companies that used to wait years for IT to build a basic tool

483
00:16:56,080 --> 00:16:58,560
were suddenly getting those tools in a matter of weeks

484
00:16:58,560 --> 00:17:01,160
that provided real value and we shouldn't ignore that success.

485
00:17:01,160 --> 00:17:04,080
But the same choice that made Canvas apps fast to build,

486
00:17:04,080 --> 00:17:06,560
putting logic directly into the presentation layer,

487
00:17:06,560 --> 00:17:08,880
is exactly what made them expensive to maintain.

488
00:17:08,880 --> 00:17:11,400
It made them impossible to govern at the rule level

489
00:17:11,400 --> 00:17:13,960
and kept them from being reused in different contexts.

490
00:17:13,960 --> 00:17:16,240
That technical debt didn't appear on day one.

491
00:17:16,240 --> 00:17:17,560
It grew quietly for years

492
00:17:17,560 --> 00:17:19,920
and now many organizations are stuck with Apple States

493
00:17:19,920 --> 00:17:23,120
where the business logic is scattered across thousands of formulas.

494
00:17:23,120 --> 00:17:26,120
It is trapped inside UI controls with no way to get it out

495
00:17:26,120 --> 00:17:28,200
other than rebuilding the whole thing from the ground up.

496
00:17:28,200 --> 00:17:30,240
The agentic model is more than just a better way

497
00:17:30,240 --> 00:17:31,400
to build a front end.

498
00:17:31,400 --> 00:17:33,080
It provides a completely different answer

499
00:17:33,080 --> 00:17:34,960
to the question of where logic should live

500
00:17:34,960 --> 00:17:36,480
and that answer changes everything

501
00:17:36,480 --> 00:17:38,040
about how we build for the future.

502
00:17:38,040 --> 00:17:39,800
How agentic logic works?

503
00:17:39,800 --> 00:17:41,600
Inside the orchestration engine.

504
00:17:41,600 --> 00:17:44,600
So what actually happens when a co-pilot studio agent receives a goal?

505
00:17:44,600 --> 00:17:46,160
I'm not talking about the marketing version.

506
00:17:46,160 --> 00:17:47,920
I am talking about the mechanical version.

507
00:17:47,920 --> 00:17:49,520
Once you understand the sequence,

508
00:17:49,520 --> 00:17:52,920
every architectural decision you make downstream starts to make obvious sense.

509
00:17:52,920 --> 00:17:54,760
It starts with intent interpretation.

510
00:17:54,760 --> 00:17:56,080
The agent receives an input

511
00:17:56,080 --> 00:17:58,600
which might be natural language from a user,

512
00:17:58,600 --> 00:18:00,600
a structured payload from an event

513
00:18:00,600 --> 00:18:02,800
or an email passed by a connector.

514
00:18:02,800 --> 00:18:04,240
But that input is rarely clean.

515
00:18:04,240 --> 00:18:05,960
It is often incomplete or ambiguous.

516
00:18:05,960 --> 00:18:09,480
If a user says, help me onboard the new contractor starting Monday

517
00:18:09,480 --> 00:18:11,480
that is not a structured API call.

518
00:18:11,480 --> 00:18:14,280
It is a goal filled with missing fields and implied dependencies.

519
00:18:14,280 --> 00:18:16,480
The agent has to figure out what is actually being asked,

520
00:18:16,480 --> 00:18:17,760
what information is missing

521
00:18:17,760 --> 00:18:19,920
and what it already knows from the context.

522
00:18:19,920 --> 00:18:21,720
Once the intent is clear enough to act on,

523
00:18:21,720 --> 00:18:23,160
the agent selects its tools.

524
00:18:23,160 --> 00:18:24,720
This is where it diverges sharply

525
00:18:24,720 --> 00:18:26,760
from the old world of canvas apps.

526
00:18:26,760 --> 00:18:28,560
The agent does not follow a predefined path.

527
00:18:28,560 --> 00:18:30,240
It looks at everything available to it,

528
00:18:30,240 --> 00:18:33,000
power automate flows, connectors, knowledge sources,

529
00:18:33,000 --> 00:18:36,320
and other agents, and then it decides which ones are relevant.

530
00:18:36,320 --> 00:18:38,000
It chooses the order and the parameters

531
00:18:38,000 --> 00:18:40,360
that run time based on the specific situation.

532
00:18:40,360 --> 00:18:42,200
If you give the same agent a similar request

533
00:18:42,200 --> 00:18:43,440
but change the context,

534
00:18:43,440 --> 00:18:45,760
it might take a completely different sequence of actions

535
00:18:45,760 --> 00:18:47,120
and still be correct.

536
00:18:47,120 --> 00:18:48,480
Execution follows selection,

537
00:18:48,480 --> 00:18:50,360
but execution is not the end of the loop.

538
00:18:50,360 --> 00:18:52,280
After the agent calls a tool and gets a result,

539
00:18:52,280 --> 00:18:54,800
it evaluates whether that result actually moves the needle.

540
00:18:54,800 --> 00:18:56,120
If a flow returns an error,

541
00:18:56,120 --> 00:18:58,280
the agent does not just fail silently.

542
00:18:58,280 --> 00:19:00,080
It can retry with different parameters,

543
00:19:00,080 --> 00:19:01,280
escalate to a human,

544
00:19:01,280 --> 00:19:02,960
or surface the failure with enough detail

545
00:19:02,960 --> 00:19:04,200
for a person to fix it.

546
00:19:04,200 --> 00:19:05,920
This adaptive loop is what makes agents

547
00:19:05,920 --> 00:19:08,080
genuinely different from standard flows.

548
00:19:08,080 --> 00:19:10,760
A flow fails at the step where the error occurred and stops,

549
00:19:10,760 --> 00:19:13,320
but an agent fails and asks what to do next.

550
00:19:13,320 --> 00:19:15,480
Multi-step orchestration is where this gets interesting

551
00:19:15,480 --> 00:19:16,480
for the enterprise.

552
00:19:16,480 --> 00:19:18,160
A single goal might trigger a sequence

553
00:19:18,160 --> 00:19:19,840
that spans five different systems,

554
00:19:19,840 --> 00:19:21,320
involves waiting for an approval,

555
00:19:21,320 --> 00:19:23,680
and then resumes once that approval is granted.

556
00:19:23,680 --> 00:19:25,800
The agent holds the state of the overall goal

557
00:19:25,800 --> 00:19:27,520
across every one of those steps.

558
00:19:27,520 --> 00:19:28,920
It knows what has been finished,

559
00:19:28,920 --> 00:19:30,120
what is still pending,

560
00:19:30,120 --> 00:19:31,920
and what the next move should be.

561
00:19:31,920 --> 00:19:33,440
Building that kind of stateful orchestration

562
00:19:33,440 --> 00:19:35,480
used to require massive custom development,

563
00:19:35,480 --> 00:19:36,680
but in the agentic model,

564
00:19:36,680 --> 00:19:39,560
it is just a design choice in how you configure behavior.

565
00:19:39,560 --> 00:19:41,600
Now, I want to be precise about a distinction

566
00:19:41,600 --> 00:19:43,280
that causes a lot of confusion.

567
00:19:43,280 --> 00:19:45,440
There are agent flows and there are cloud flows,

568
00:19:45,440 --> 00:19:46,480
and they are not the same thing,

569
00:19:46,480 --> 00:19:48,200
even though they share the same engine.

570
00:19:48,200 --> 00:19:50,880
A cloud flow in Power Automate is a general purpose tool.

571
00:19:50,880 --> 00:19:52,880
It runs under Power Automate licensing.

572
00:19:52,880 --> 00:19:55,040
It is accessible across the whole platform,

573
00:19:55,040 --> 00:19:58,280
and it is designed for predictable, stand-alone processes.

574
00:19:58,280 --> 00:19:59,600
An agent flow is different.

575
00:19:59,600 --> 00:20:01,040
It is a flow that has been converted

576
00:20:01,040 --> 00:20:03,040
to run inside co-pilot studio capacity,

577
00:20:03,040 --> 00:20:05,200
specifically for agent-driven execution.

578
00:20:05,200 --> 00:20:07,400
It runs under the co-pilot studio licensing model,

579
00:20:07,400 --> 00:20:09,720
which means it consumes message capacity

580
00:20:09,720 --> 00:20:11,440
rather than flow entitlements.

581
00:20:11,440 --> 00:20:12,560
And here is the kicker.

582
00:20:12,560 --> 00:20:14,760
Once you convert a flow to an agent flow,

583
00:20:14,760 --> 00:20:16,200
you cannot convert it back.

584
00:20:16,200 --> 00:20:17,800
That is not just a technicality.

585
00:20:17,800 --> 00:20:19,280
It is an architectural commitment.

586
00:20:19,280 --> 00:20:21,400
When you turn something into an agent flow,

587
00:20:21,400 --> 00:20:23,320
you are stating that this logic exists

588
00:20:23,320 --> 00:20:25,680
to serve the agent not to run on its own.

589
00:20:25,680 --> 00:20:26,960
It is a tool the agent calls,

590
00:20:26,960 --> 00:20:28,920
not a process that stands alone.

591
00:20:28,920 --> 00:20:30,520
Keeping this distinction clear,

592
00:20:30,520 --> 00:20:32,160
prevents a whole category of problems

593
00:20:32,160 --> 00:20:34,560
around licensing and governance that teams run into

594
00:20:34,560 --> 00:20:36,720
when they start mixing the two without a plan.

595
00:20:36,720 --> 00:20:38,440
The agent owns the what and the when,

596
00:20:38,440 --> 00:20:39,720
the flow owns the how.

597
00:20:39,720 --> 00:20:41,880
This separation is what makes the system maintainable

598
00:20:41,880 --> 00:20:42,680
as it grows.

599
00:20:42,680 --> 00:20:45,280
It is the exact opposite of how we used to build canvas apps,

600
00:20:45,280 --> 00:20:47,200
where everything lived in the same messy layer

601
00:20:47,200 --> 00:20:48,680
with no separation at all.

602
00:20:48,680 --> 00:20:50,600
But this architecture has a constraint

603
00:20:50,600 --> 00:20:52,440
that most teams do not discover

604
00:20:52,440 --> 00:20:54,120
until they hit it in production.

605
00:20:54,120 --> 00:20:56,360
And when they hit it, it is frustrating,

606
00:20:56,360 --> 00:20:59,400
because nothing in the documentation prepares you for it.

607
00:20:59,400 --> 00:21:01,120
The safety summarization problem,

608
00:21:01,120 --> 00:21:02,960
what agente AI hides from you.

609
00:21:02,960 --> 00:21:04,080
Here is the constraint.

610
00:21:04,080 --> 00:21:05,000
And it is a real one.

611
00:21:05,000 --> 00:21:07,320
When you build a multi agent architecture,

612
00:21:07,320 --> 00:21:10,120
a master agent coordinating specialized sub agents,

613
00:21:10,120 --> 00:21:12,840
you run into a wall that isn't mentioned in the brochures.

614
00:21:12,840 --> 00:21:15,280
The master agent does not receive the full raw response

615
00:21:15,280 --> 00:21:16,120
from a sub agent.

616
00:21:16,120 --> 00:21:18,800
It receives a summarized version, a safe version.

617
00:21:18,800 --> 00:21:21,960
Citations get stripped out and document links often disappear.

618
00:21:21,960 --> 00:21:23,320
The raw payload from the sub agent

619
00:21:23,320 --> 00:21:25,080
gets processed through an orchestration layer

620
00:21:25,080 --> 00:21:26,880
before it ever reaches the master.

621
00:21:26,880 --> 00:21:29,000
And that processing is very opinionated.

622
00:21:29,000 --> 00:21:30,400
You might write an instruction that says,

623
00:21:30,400 --> 00:21:32,400
"Return the exact response from the sub agent,

624
00:21:32,400 --> 00:21:34,000
"including all source links."

625
00:21:34,000 --> 00:21:36,360
But that instruction will frequently be ignored.

626
00:21:36,360 --> 00:21:38,040
This isn't because the agent is broken.

627
00:21:38,040 --> 00:21:40,520
It is because the orchestration layer has its own policies

628
00:21:40,520 --> 00:21:42,240
about what it passes upstream.

629
00:21:42,240 --> 00:21:43,760
And those policies take precedence

630
00:21:43,760 --> 00:21:45,400
over whatever you put in your prompt.

631
00:21:45,400 --> 00:21:46,760
To be clear, this is not a bug.

632
00:21:46,760 --> 00:21:49,440
This is a deliberate, safety, and compliance behavior.

633
00:21:49,440 --> 00:21:51,640
The orchestration layer sanitizes responses

634
00:21:51,640 --> 00:21:53,320
to prevent things like prompt injection

635
00:21:53,320 --> 00:21:54,960
from moving through an agent chain.

636
00:21:54,960 --> 00:21:57,200
It stops a sub agent's output from becoming a vector

637
00:21:57,200 --> 00:21:58,640
to manipulate the master agent.

638
00:21:58,640 --> 00:22:01,480
Microsoft is working on improving this fidelity over time.

639
00:22:01,480 --> 00:22:02,840
But as of today,

640
00:22:02,840 --> 00:22:04,320
this is the reality of the platform.

641
00:22:04,320 --> 00:22:06,200
The architectural implication here is massive.

642
00:22:06,200 --> 00:22:08,160
You cannot rely on agent to agent orchestration

643
00:22:08,160 --> 00:22:10,200
if data fidelity is a hard requirement.

644
00:22:10,200 --> 00:22:12,640
If your use case needs the user to see specific document

645
00:22:12,640 --> 00:22:16,600
citations, exact URLs, or raw API payloads,

646
00:22:16,600 --> 00:22:19,000
the direct orchestration path will let you down.

647
00:22:19,000 --> 00:22:20,440
If you build on that assumption,

648
00:22:20,440 --> 00:22:22,920
you will discover the problem during a high stakes demo

649
00:22:22,920 --> 00:22:24,400
rather than a design review.

650
00:22:24,400 --> 00:22:26,360
The work around exists, and once you know it,

651
00:22:26,360 --> 00:22:28,040
it becomes a standard pattern.

652
00:22:28,040 --> 00:22:29,680
Instead of calling the sub agent directly

653
00:22:29,680 --> 00:22:31,200
and hoping the response survives,

654
00:22:31,200 --> 00:22:32,800
you expose that logic as an API.

655
00:22:32,800 --> 00:22:35,400
You build a power automate flow with an HTTP trigger

656
00:22:35,400 --> 00:22:37,480
that calls your system on knowledge source.

657
00:22:37,480 --> 00:22:39,400
That flow aggregates the raw response,

658
00:22:39,400 --> 00:22:41,480
keeps all the links and citations intact,

659
00:22:41,480 --> 00:22:43,160
and returns it in a structured format

660
00:22:43,160 --> 00:22:44,600
that master agent can use.

661
00:22:44,600 --> 00:22:47,560
The flow acts as a bridge that preserves the data fidelity

662
00:22:47,560 --> 00:22:50,400
that the agent to agent channel would otherwise scrub away.

663
00:22:50,400 --> 00:22:52,480
This is a more complex way to build.

664
00:22:52,480 --> 00:22:53,920
It requires you to treat the flow

665
00:22:53,920 --> 00:22:55,840
as an explicit data contract,

666
00:22:55,840 --> 00:22:57,520
rather than just a background task.

667
00:22:57,520 --> 00:23:00,120
But it is also more governable and much easier to test.

668
00:23:00,120 --> 00:23:01,840
You can see exactly what the flow returned.

669
00:23:01,840 --> 00:23:03,160
You conversion the schema,

670
00:23:03,160 --> 00:23:05,320
and you can monitor the call independently

671
00:23:05,320 --> 00:23:07,240
from the agent's reasoning loop.

672
00:23:07,240 --> 00:23:09,960
This is the kind of constraint that separates an architect

673
00:23:09,960 --> 00:23:11,760
who has actually built these systems

674
00:23:11,760 --> 00:23:13,800
from someone who has only read the manual.

675
00:23:13,800 --> 00:23:16,040
The documentation tells you what the platform can do,

676
00:23:16,040 --> 00:23:17,920
but experiences you where the gaps are

677
00:23:17,920 --> 00:23:19,400
and how to navigate around them.

678
00:23:19,400 --> 00:23:21,680
Every platform has these kinds of hidden rules.

679
00:23:21,680 --> 00:23:23,280
The question is not whether they exist

680
00:23:23,280 --> 00:23:24,120
because they always do.

681
00:23:24,120 --> 00:23:26,160
The real question is whether someone on your team

682
00:23:26,160 --> 00:23:27,400
knows where the landmines are

683
00:23:27,400 --> 00:23:29,600
before you build a production system on top of them.

684
00:23:29,600 --> 00:23:31,800
That person, the one who maps the constraints

685
00:23:31,800 --> 00:23:33,560
and designs around them is exactly who

686
00:23:33,560 --> 00:23:35,840
the next shift in technology is looking for.

687
00:23:35,840 --> 00:23:39,560
The role that is actually growing from maker to logic architect.

688
00:23:39,560 --> 00:23:40,960
The job title is shifting,

689
00:23:40,960 --> 00:23:43,280
but the demand underneath it is accelerating.

690
00:23:43,280 --> 00:23:46,360
And the direction it's moving is worth paying close attention to

691
00:23:46,360 --> 00:23:48,320
because it tells you exactly where the authority

692
00:23:48,320 --> 00:23:50,280
is consolidating in this ecosystem.

693
00:23:50,280 --> 00:23:52,680
Organizations aren't looking for more screen builders.

694
00:23:52,680 --> 00:23:55,320
They're looking for people who can design systems.

695
00:23:55,320 --> 00:23:57,680
They want people who can look at a business problem

696
00:23:57,680 --> 00:23:59,480
and produce a solution architecture.

697
00:23:59,480 --> 00:24:02,360
Not just an app, not just a flow, but a layered design

698
00:24:02,360 --> 00:24:05,040
that specifies which component owns which decision,

699
00:24:05,040 --> 00:24:08,080
how data moves between layers, where governance applies,

700
00:24:08,080 --> 00:24:09,720
and what happens when something fails.

701
00:24:09,720 --> 00:24:12,400
That's a different scope of work than delivering a canvas app.

702
00:24:12,400 --> 00:24:13,800
And it's the scope that's commanding

703
00:24:13,800 --> 00:24:16,080
the highest attention in hiring right now.

704
00:24:16,080 --> 00:24:20,400
The 2026 M365 architect job description makes this explicit.

705
00:24:20,400 --> 00:24:23,080
Copilot governance is listed as a required competency

706
00:24:23,080 --> 00:24:26,040
in postings that two years ago wouldn't have mentioned AI at all.

707
00:24:26,040 --> 00:24:27,240
AI Foundry Integration,

708
00:24:27,240 --> 00:24:30,080
Agentec Workflow Design, and Responsible AI Frameworks

709
00:24:30,080 --> 00:24:32,280
are showing up alongside the traditional exchange

710
00:24:32,280 --> 00:24:34,520
online and SharePoint architecture work.

711
00:24:34,520 --> 00:24:37,640
One posting from a major university lists AB900.

712
00:24:37,640 --> 00:24:41,360
Microsoft 365 Copilot and Agent Administration Fundamentals.

713
00:24:41,360 --> 00:24:43,600
As a preferred certification, that certification

714
00:24:43,600 --> 00:24:45,200
didn't exist a few years ago.

715
00:24:45,200 --> 00:24:47,440
And the fact that it's now appearing in job requirements

716
00:24:47,440 --> 00:24:49,680
is a signal about which layer of the stack employers

717
00:24:49,680 --> 00:24:50,880
are prioritizing.

718
00:24:50,880 --> 00:24:53,320
The skills that matter now are architectural in nature,

719
00:24:53,320 --> 00:24:54,360
API thinking.

720
00:24:54,360 --> 00:24:55,840
Not just knowing how to call a connector,

721
00:24:55,840 --> 00:24:57,720
but understanding what an API contract means,

722
00:24:57,720 --> 00:25:00,040
what happens when it changes and how to design systems

723
00:25:00,040 --> 00:25:03,440
that don't break when the underlying service evolves.

724
00:25:03,440 --> 00:25:04,720
State management.

725
00:25:04,720 --> 00:25:07,440
Understanding how to design a system that maintains context

726
00:25:07,440 --> 00:25:09,280
across a long-running multi-step process

727
00:25:09,280 --> 00:25:12,760
without that context getting corrupted or lost.

728
00:25:12,760 --> 00:25:13,880
Data governance.

729
00:25:13,880 --> 00:25:17,040
Not at the who has access to this app level,

730
00:25:17,040 --> 00:25:20,560
but at the level of data quality, lineage, classification,

731
00:25:20,560 --> 00:25:23,040
and what it means when an agent retrieves a document

732
00:25:23,040 --> 00:25:25,000
with a sensitivity label attached.

733
00:25:25,000 --> 00:25:27,400
Separation of concerns, which we've been discussing throughout

734
00:25:27,400 --> 00:25:29,880
this episode in the form of judgment versus logic

735
00:25:29,880 --> 00:25:31,440
and human in the loop design.

736
00:25:31,440 --> 00:25:34,240
Building systems where AI outputs feed into human decision

737
00:25:34,240 --> 00:25:37,360
points that are explicit, auditable, and recoverable.

738
00:25:37,360 --> 00:25:39,000
The skills that are declining in value

739
00:25:39,000 --> 00:25:41,440
are the ones that were central to the canvas era.

740
00:25:41,440 --> 00:25:43,600
PixelPerfect UI design matters less when

741
00:25:43,600 --> 00:25:45,880
Copilot is generating the interface.

742
00:25:45,880 --> 00:25:48,200
Deep PowerFX Formula Expertise matters less

743
00:25:48,200 --> 00:25:50,080
when the logic layer is moving out of the screen

744
00:25:50,080 --> 00:25:51,600
and into agent instructions.

745
00:25:51,600 --> 00:25:53,520
Canvas screen navigation logic.

746
00:25:53,520 --> 00:25:55,080
Knowing how to wire gallery selections

747
00:25:55,080 --> 00:25:57,480
into context variables across 15 screens

748
00:25:57,480 --> 00:25:59,960
is valuable in a world where screens are the primary interaction

749
00:25:59,960 --> 00:26:00,480
model.

750
00:26:00,480 --> 00:26:02,160
It's less valuable in a world where users

751
00:26:02,160 --> 00:26:04,080
are expressing goals in natural language

752
00:26:04,080 --> 00:26:06,240
and the agent is deciding what to surface.

753
00:26:06,240 --> 00:26:08,040
That's not a dismissal of those skills.

754
00:26:08,040 --> 00:26:10,200
It's a recognition that their relative weight in the value

755
00:26:10,200 --> 00:26:11,440
stack is shifting.

756
00:26:11,440 --> 00:26:13,360
Someone who only knows how to build canvas apps

757
00:26:13,360 --> 00:26:15,640
is in a narrower position than they were three years ago.

758
00:26:15,640 --> 00:26:17,240
Someone who understands both.

759
00:26:17,240 --> 00:26:20,560
Who can build a clean power apps UI when that's the right answer

760
00:26:20,560 --> 00:26:22,440
and design the orchestration layer when that's

761
00:26:22,440 --> 00:26:23,480
the right answer?

762
00:26:23,480 --> 00:26:26,280
Is in a stronger position than either specialist alone?

763
00:26:26,280 --> 00:26:28,680
That's the t-shaped profile the market is rewarding.

764
00:26:28,680 --> 00:26:30,840
Deep expertise in one layer with enough fluency

765
00:26:30,840 --> 00:26:32,640
in the adjacent layers to collaborate,

766
00:26:32,640 --> 00:26:36,120
make architectural trade-offs, and communicate across disciplines.

767
00:26:36,120 --> 00:26:38,360
A logic architect who understands enough about UI

768
00:26:38,360 --> 00:26:40,800
to know when to hand off to a canvas component.

769
00:26:40,800 --> 00:26:43,640
A UI developer who understands enough about orchestration

770
00:26:43,640 --> 00:26:45,680
to know where the boundaries of their responsibility

771
00:26:45,680 --> 00:26:46,480
should end.

772
00:26:46,480 --> 00:26:48,480
The salary data reflects this directly.

773
00:26:48,480 --> 00:26:50,160
AI architects in mature markets

774
00:26:50,160 --> 00:26:53,880
command 150,000 to $330,000 and above.

775
00:26:53,880 --> 00:26:56,440
Pure low-code specialists even experienced ones

776
00:26:56,440 --> 00:26:58,680
cluster significantly below that range.

777
00:26:58,680 --> 00:27:00,200
The gap isn't about coding ability.

778
00:27:00,200 --> 00:27:01,560
It's about scope.

779
00:27:01,560 --> 00:27:03,360
The architect is responsible for the behavior

780
00:27:03,360 --> 00:27:05,200
of a system across all conditions.

781
00:27:05,200 --> 00:27:07,240
The screen builder is responsible for the behavior

782
00:27:07,240 --> 00:27:09,120
of a UI in the happy path.

783
00:27:09,120 --> 00:27:11,000
The transition from one to the other

784
00:27:11,000 --> 00:27:12,800
isn't about learning harder things.

785
00:27:12,800 --> 00:27:14,160
It's about expanding what you consider

786
00:27:14,160 --> 00:27:16,840
to be within your remit, from delivering an application

787
00:27:16,840 --> 00:27:19,720
to governing the logic that runs beneath it.

788
00:27:19,720 --> 00:27:21,160
The skill gap is real.

789
00:27:21,160 --> 00:27:23,400
What low-code developers are actually missing.

790
00:27:23,400 --> 00:27:26,000
So if the role is expanding towards system-level design,

791
00:27:26,000 --> 00:27:28,960
the natural next question is, what specifically is missing?

792
00:27:28,960 --> 00:27:29,960
Not in the abstract.

793
00:27:29,960 --> 00:27:34,000
In practice, what does a capable, experienced low-code developer

794
00:27:34,000 --> 00:27:36,560
not know that an AI architect needs to know?

795
00:27:36,560 --> 00:27:38,320
The gap is not visual development skill.

796
00:27:38,320 --> 00:27:40,040
That's not what's holding anyone back.

797
00:27:40,040 --> 00:27:43,080
The gap is architectural depth across five specific areas

798
00:27:43,080 --> 00:27:45,400
and being precise about where the gaps actually are matters.

799
00:27:45,400 --> 00:27:47,400
Because the upskilling path looks very different,

800
00:27:47,400 --> 00:27:49,120
depending on which ones you're addressing.

801
00:27:49,120 --> 00:27:51,480
The first gap is data governance.

802
00:27:51,480 --> 00:27:54,960
Low-code developers work with data as an input to a workflow.

803
00:27:54,960 --> 00:27:57,240
Data comes in, something happens, data goes out.

804
00:27:57,240 --> 00:27:59,920
That mental model works fine when the data is well-structured,

805
00:27:59,920 --> 00:28:02,560
clean, and governed upstream by someone else.

806
00:28:02,560 --> 00:28:04,720
But AI systems don't just consume data.

807
00:28:04,720 --> 00:28:05,680
They reason about it.

808
00:28:05,680 --> 00:28:07,480
They retrieve it from unstructured sources

809
00:28:07,480 --> 00:28:09,240
that they synthesize it across documents

810
00:28:09,240 --> 00:28:10,520
with different provenance.

811
00:28:10,520 --> 00:28:13,480
They produce outputs that inherit the quality of the inputs.

812
00:28:13,480 --> 00:28:15,720
An AI architect has to think about data quality

813
00:28:15,720 --> 00:28:18,160
before it enters the system, about lineage,

814
00:28:18,160 --> 00:28:20,080
where the data came from and whether that origin

815
00:28:20,080 --> 00:28:21,960
matters for how it should be used.

816
00:28:21,960 --> 00:28:24,040
And about drift, which is what happens when the data,

817
00:28:24,040 --> 00:28:26,400
the system was designed around, stops reflecting the data

818
00:28:26,400 --> 00:28:27,640
it's actually operating on.

819
00:28:27,640 --> 00:28:29,160
That's a different relationship to data

820
00:28:29,160 --> 00:28:31,360
than connector-based integration requires.

821
00:28:31,360 --> 00:28:33,440
The second gap is model evaluation.

822
00:28:33,440 --> 00:28:35,760
A low-code developer knows whether a process works.

823
00:28:35,760 --> 00:28:37,640
Does the flow run, does the form submit,

824
00:28:37,640 --> 00:28:39,160
does the output land in the right place?

825
00:28:39,160 --> 00:28:40,960
That's functional correctness, and it's

826
00:28:40,960 --> 00:28:43,160
the right standard for deterministic systems.

827
00:28:43,160 --> 00:28:45,480
But AI systems aren't deterministic.

828
00:28:45,480 --> 00:28:47,400
The question isn't just, did it run?

829
00:28:47,400 --> 00:28:49,720
Is it accurate and is it consistently accurate

830
00:28:49,720 --> 00:28:52,280
across the range of inputs it will encounter in production?

831
00:28:52,280 --> 00:28:53,960
Is it producing safe outputs?

832
00:28:53,960 --> 00:28:55,520
Is it stable when conditions shift?

833
00:28:55,520 --> 00:28:57,920
Evaluating those things requires a different frame

834
00:28:57,920 --> 00:28:58,960
than functional testing.

835
00:28:58,960 --> 00:29:00,440
It requires defining what good looks like

836
00:29:00,440 --> 00:29:02,760
probabilistically running structured evaluations

837
00:29:02,760 --> 00:29:05,520
against that definition and monitoring for degradation

838
00:29:05,520 --> 00:29:06,280
over time.

839
00:29:06,280 --> 00:29:07,920
Most low-code developers have never

840
00:29:07,920 --> 00:29:10,680
had to think in those terms, because the systems they built

841
00:29:10,680 --> 00:29:12,440
didn't require it.

842
00:29:12,440 --> 00:29:14,560
The third gap is human in the loop design.

843
00:29:14,560 --> 00:29:16,880
When you build a deterministic system, failure modes

844
00:29:16,880 --> 00:29:18,920
are predictable enough that you can handle most of them

845
00:29:18,920 --> 00:29:20,080
in the flow itself.

846
00:29:20,080 --> 00:29:21,720
Retry branch fail with a message.

847
00:29:21,720 --> 00:29:24,120
AI systems fail in less predictable ways.

848
00:29:24,120 --> 00:29:26,040
The output might be technically complete,

849
00:29:26,040 --> 00:29:27,320
but wrong in context.

850
00:29:27,320 --> 00:29:29,720
It might be accurate 95% of the time,

851
00:29:29,720 --> 00:29:31,760
and badly wrong 5% of the time.

852
00:29:31,760 --> 00:29:33,960
And both cases look identical to the system.

853
00:29:33,960 --> 00:29:36,200
Designing the escalation paths, the approval checkpoints,

854
00:29:36,200 --> 00:29:38,280
and the fallback behaviors that make an AI system

855
00:29:38,280 --> 00:29:39,880
safe to operate at scale.

856
00:29:39,880 --> 00:29:42,880
Those need to be explicit design decisions, not afterthoughts.

857
00:29:42,880 --> 00:29:44,240
Low-code developers haven't had

858
00:29:44,240 --> 00:29:46,400
to design for probabilistic failure before.

859
00:29:46,400 --> 00:29:47,320
Now they do.

860
00:29:47,320 --> 00:29:49,000
The fourth gap is integration architecture

861
00:29:49,000 --> 00:29:51,920
depth, calling a connector and designing an integration

862
00:29:51,920 --> 00:29:53,160
are not the same thing.

863
00:29:53,160 --> 00:29:56,040
Rack pipelines vector databases identity-aware retrieval,

864
00:29:56,040 --> 00:29:58,400
observability tooling, permission-filtered search.

865
00:29:58,400 --> 00:29:59,680
These require understanding what's

866
00:29:59,680 --> 00:30:01,360
happening under the abstraction layer,

867
00:30:01,360 --> 00:30:03,160
not just how to configure the surface.

868
00:30:03,160 --> 00:30:05,240
When something breaks in a connector-based integration,

869
00:30:05,240 --> 00:30:06,400
you fix the connector.

870
00:30:06,400 --> 00:30:08,280
When something breaks in a Rack pipeline,

871
00:30:08,280 --> 00:30:10,200
you need to know whether the problem is in the chunking

872
00:30:10,200 --> 00:30:12,000
strategy, the retrieval configuration,

873
00:30:12,000 --> 00:30:14,240
the permission mapping, or the prompt.

874
00:30:14,240 --> 00:30:16,520
That diagnostic depth requires technical fluency

875
00:30:16,520 --> 00:30:19,280
that most connector-based development simply doesn't build.

876
00:30:19,280 --> 00:30:21,760
The fifth gap is AI risk management, understanding

877
00:30:21,760 --> 00:30:23,360
what can go wrong with an AI system

878
00:30:23,360 --> 00:30:25,960
in ways that have no equivalent in traditional automation,

879
00:30:25,960 --> 00:30:28,520
prompt injection, retrieval poisoning, agents

880
00:30:28,520 --> 00:30:30,880
acquiring permissions that weren't intentionally granted

881
00:30:30,880 --> 00:30:32,360
because of a misconfigured scope.

882
00:30:32,360 --> 00:30:33,560
These aren't hypothetical.

883
00:30:33,560 --> 00:30:35,760
They're documented risks in production deployments.

884
00:30:35,760 --> 00:30:37,040
And here's the honest framing, though.

885
00:30:37,040 --> 00:30:38,840
Low-code developers already have the two things

886
00:30:38,840 --> 00:30:40,360
that take the longest to develop.

887
00:30:40,360 --> 00:30:42,280
Business process, fluency, and delivery speed,

888
00:30:42,280 --> 00:30:44,120
the gaps are real, but they're edible.

889
00:30:44,120 --> 00:30:45,880
The foundation is already there.

890
00:30:45,880 --> 00:30:49,240
The upskilling path from screen builder to AI orchestrator.

891
00:30:49,240 --> 00:30:51,360
Knowing the gaps exist doesn't close them.

892
00:30:51,360 --> 00:30:53,480
We need to talk about the actual path forward

893
00:30:53,480 --> 00:30:56,160
because in this transition, the sequence of what you learn

894
00:30:56,160 --> 00:30:58,480
matters just as much as the content itself.

895
00:30:58,480 --> 00:31:00,280
If you jump straight into governance frameworks

896
00:31:00,280 --> 00:31:03,400
before you understand why AI systems behave the way they do,

897
00:31:03,400 --> 00:31:05,320
you'll run into a familiar wall.

898
00:31:05,320 --> 00:31:07,080
It is the same mistake as building an app

899
00:31:07,080 --> 00:31:09,560
before you understand the business process.

900
00:31:09,560 --> 00:31:12,480
You end up assembling a system whose failure modes you simply

901
00:31:12,480 --> 00:31:13,440
cannot predict.

902
00:31:13,440 --> 00:31:16,240
Stage one focuses on AI and systems foundations.

903
00:31:16,240 --> 00:31:18,720
This isn't about model training or deep data science.

904
00:31:18,720 --> 00:31:22,000
Your goal here is to move from simply calling an AI API

905
00:31:22,000 --> 00:31:25,480
to truly understanding how these systems behave when they fail.

906
00:31:25,480 --> 00:31:27,280
It is a narrower target than it sounds,

907
00:31:27,280 --> 00:31:29,760
and you can reach it without a background in machine learning.

908
00:31:29,760 --> 00:31:31,680
You need to internalize how language models

909
00:31:31,680 --> 00:31:33,240
work at the operational level.

910
00:31:33,240 --> 00:31:35,720
You need to know what happens when you exceed a context window

911
00:31:35,720 --> 00:31:38,400
or why retrieval behavior changes when the index shifts.

912
00:31:38,400 --> 00:31:39,760
You don't need to master the math, but you

913
00:31:39,760 --> 00:31:41,640
do need to understand the behavior well enough

914
00:31:41,640 --> 00:31:43,280
to design around the constraints.

915
00:31:43,280 --> 00:31:45,840
This takes a few months of deliberate experimentation,

916
00:31:45,840 --> 00:31:48,280
and it serves as the foundation for everything else.

917
00:31:48,280 --> 00:31:50,840
Stage two runs in parallel with the first stage,

918
00:31:50,840 --> 00:31:53,520
focusing on the basics of data and cloud architecture.

919
00:31:53,520 --> 00:31:56,440
The objective is to understand what lives under your low-code apps

920
00:31:56,440 --> 00:31:58,000
instead of just what connects to them.

921
00:31:58,000 --> 00:32:00,800
You should know where data verse actually stores information,

922
00:32:00,800 --> 00:32:04,120
and what a storage tier decision does to your retrieval speed.

923
00:32:04,120 --> 00:32:05,960
Understanding the difference between structured

924
00:32:05,960 --> 00:32:08,000
and unstructured data at the architecture level

925
00:32:08,000 --> 00:32:11,480
is vital when you are designing what a rag pipeline retrieves.

926
00:32:11,480 --> 00:32:13,000
You aren't trying to become a data engineer,

927
00:32:13,000 --> 00:32:15,800
but you need enough fluency to have an informed conversation

928
00:32:15,800 --> 00:32:16,480
with one.

929
00:32:16,480 --> 00:32:19,120
This helps you recognize when a problem in an agent's behavior

930
00:32:19,120 --> 00:32:21,800
is actually a data issue dressed up as a prompt problem.

931
00:32:21,800 --> 00:32:24,000
Stage three is where the identity shift happens

932
00:32:24,000 --> 00:32:26,360
as you move from app builder to solution architect.

933
00:32:26,360 --> 00:32:28,960
This is the transition from designing inside a single platform

934
00:32:28,960 --> 00:32:30,880
to designing across an entire ecosystem.

935
00:32:30,880 --> 00:32:33,440
You start asking what the end-to-end solution looks like

936
00:32:33,440 --> 00:32:35,520
and which layer owns which specific decision.

937
00:32:35,520 --> 00:32:37,840
You have to decide when data verse makes more sense

938
00:32:37,840 --> 00:32:40,240
than SharePoint, or when a flow should be replaced

939
00:32:40,240 --> 00:32:41,360
by a function app.

940
00:32:41,360 --> 00:32:44,160
At this level, you start producing architecture artifacts

941
00:32:44,160 --> 00:32:46,320
like sequence diagrams and solution blueprints

942
00:32:46,320 --> 00:32:48,800
that someone else could implement without you in the room.

943
00:32:48,800 --> 00:32:51,080
Your portfolio shouldn't be a collection of apps you build,

944
00:32:51,080 --> 00:32:53,000
but a collection of architectures you designed.

945
00:32:53,000 --> 00:32:55,040
This proves you can think at the system level

946
00:32:55,040 --> 00:32:56,880
rather than just the screen level.

947
00:32:56,880 --> 00:33:00,200
Stage four is about learning just enough code, specifically

948
00:33:00,200 --> 00:33:03,240
Python, because that is what the AI tooling ecosystem runs on.

949
00:33:03,240 --> 00:33:05,440
You aren't trying to become a full stack developer.

950
00:33:05,440 --> 00:33:07,880
You are learning to call APIs, orchestrate workflows,

951
00:33:07,880 --> 00:33:09,680
and write data transformations.

952
00:33:09,680 --> 00:33:12,120
The point is to remove the ceiling on what you can understand

953
00:33:12,120 --> 00:33:12,800
and design.

954
00:33:12,800 --> 00:33:16,360
When you can read a Python script that calls the Azure Open I API,

955
00:33:16,360 --> 00:33:17,960
you can finally have a real conversation

956
00:33:17,960 --> 00:33:19,600
with the engineers building the backend.

957
00:33:19,600 --> 00:33:20,840
If you can write a basic script,

958
00:33:20,840 --> 00:33:22,400
you can prototype an integration pattern

959
00:33:22,400 --> 00:33:23,880
before committing to a full build.

960
00:33:23,880 --> 00:33:26,320
That design credibility is worth far more than the lines

961
00:33:26,320 --> 00:33:27,520
of code themselves.

962
00:33:27,520 --> 00:33:29,560
Stage five is where real authority consolidates

963
00:33:29,560 --> 00:33:31,280
in AI governance and leadership.

964
00:33:31,280 --> 00:33:33,000
This isn't just about knowing the frameworks.

965
00:33:33,000 --> 00:33:35,400
It is about being the person who can run a design review

966
00:33:35,400 --> 00:33:38,560
and communicate technical risk to stakeholders who don't code.

967
00:33:38,560 --> 00:33:40,280
You become the one who defines the policies

968
00:33:40,280 --> 00:33:42,600
that constrain agent behavior and takes accountability

969
00:33:42,600 --> 00:33:44,040
for system level outcomes.

970
00:33:44,040 --> 00:33:46,200
This is a leadership position that you build over time

971
00:33:46,200 --> 00:33:49,120
through experience rather than acquiring it from a single course.

972
00:33:49,120 --> 00:33:51,280
One thing I want to say directly is that your portfolio matters

973
00:33:51,280 --> 00:33:52,880
more than any certification.

974
00:33:52,880 --> 00:33:54,920
A reference architecture showing how an agent,

975
00:33:54,920 --> 00:33:57,360
a flow, and a canvas app work together

976
00:33:57,360 --> 00:34:00,360
is worth more in a job interview than a digital badge.

977
00:34:00,360 --> 00:34:01,880
Build the artifact and write the record

978
00:34:01,880 --> 00:34:04,320
explaining why you made those specific choices.

979
00:34:04,320 --> 00:34:07,080
That document is the proof of architectural thinking

980
00:34:07,080 --> 00:34:10,040
that a standard resume can never convey.

981
00:34:10,040 --> 00:34:13,400
What governance actually means now, the new core skill.

982
00:34:13,400 --> 00:34:15,800
In most organizations, governance is treated

983
00:34:15,800 --> 00:34:19,080
like a compliance function or a boring checklist you run through

984
00:34:19,080 --> 00:34:21,000
right before a project goes live.

985
00:34:21,000 --> 00:34:23,280
It is usually a set of policies maintained by IT

986
00:34:23,280 --> 00:34:26,400
in a document that nobody actually reads until something breaks.

987
00:34:26,400 --> 00:34:29,280
That framing made sense when systems were deterministic

988
00:34:29,280 --> 00:34:32,360
and followed the exact steps a developer encoded into them.

989
00:34:32,360 --> 00:34:35,080
But that model doesn't work for agentic systems.

990
00:34:35,080 --> 00:34:37,000
The architects who understand why it fails

991
00:34:37,000 --> 00:34:39,560
are the ones building real authority right now.

992
00:34:39,560 --> 00:34:41,760
Governance for agents is a design discipline,

993
00:34:41,760 --> 00:34:44,360
not an audit function or a post-built review.

994
00:34:44,360 --> 00:34:46,000
You have to embed it into the architecture

995
00:34:46,000 --> 00:34:49,480
from the very first conversation about what the agent is supposed to do.

996
00:34:49,480 --> 00:34:52,920
The decisions you make early on regarding what an agent can access

997
00:34:52,920 --> 00:34:55,480
and how it escalates will define the risk surface

998
00:34:55,480 --> 00:34:56,720
for the life of the system.

999
00:34:56,720 --> 00:34:58,480
You cannot simply retrofit governance

1000
00:34:58,480 --> 00:35:00,640
onto an agent that was built without it.

1001
00:35:00,640 --> 00:35:02,400
If you try, you usually end up having

1002
00:35:02,400 --> 00:35:04,160
to redesign the whole thing from scratch.

1003
00:35:04,160 --> 00:35:06,920
Microsoft breaks this down into a five-phase life cycle

1004
00:35:06,920 --> 00:35:08,880
that helps an architect focus their attention.

1005
00:35:08,880 --> 00:35:11,360
Discovery and planning is where you define the use cases

1006
00:35:11,360 --> 00:35:14,520
and risk posture before anyone writes a single line of code.

1007
00:35:14,520 --> 00:35:16,400
Architecture and design is where you decide

1008
00:35:16,400 --> 00:35:19,200
on the environment strategy and the identity model.

1009
00:35:19,200 --> 00:35:20,720
During build and integration,

1010
00:35:20,720 --> 00:35:23,400
you configure the connectors and authentication patterns

1011
00:35:23,400 --> 00:35:25,400
using the principle of least privilege.

1012
00:35:25,400 --> 00:35:28,920
Testing and deployment is where you validate security and compliance

1013
00:35:28,920 --> 00:35:30,880
before anything reaches production.

1014
00:35:30,880 --> 00:35:32,880
Finally, monitoring and optimization

1015
00:35:32,880 --> 00:35:34,880
involves the ongoing work of catching drift

1016
00:35:34,880 --> 00:35:37,400
and retiring agents that no longer serve a purpose.

1017
00:35:37,400 --> 00:35:39,480
That last phase is the one most team skip,

1018
00:35:39,480 --> 00:35:42,240
which leads to the specific problem of ownerless agents.

1019
00:35:42,240 --> 00:35:44,720
When the person who built an agent leaves the company,

1020
00:35:44,720 --> 00:35:46,640
the agent stays behind in the tenant.

1021
00:35:46,640 --> 00:35:49,560
It remains connected to data sources and authorized to act,

1022
00:35:49,560 --> 00:35:51,400
but no one is accountable for what it does.

1023
00:35:51,400 --> 00:35:53,280
At a small scale, this is just a nuisance,

1024
00:35:53,280 --> 00:35:54,760
but at an enterprise scale,

1025
00:35:54,760 --> 00:35:57,440
it is a massive security incident waiting to happen.

1026
00:35:57,440 --> 00:35:59,840
Ownership assignment and decommissioning policies

1027
00:35:59,840 --> 00:36:02,200
are not optional features for a mature model.

1028
00:36:02,200 --> 00:36:03,880
They are the minimum structure required

1029
00:36:03,880 --> 00:36:05,840
for a fleet of agents that will likely outlive

1030
00:36:05,840 --> 00:36:07,160
the people who created them.

1031
00:36:07,160 --> 00:36:09,400
A risk tiered model makes this governance actually

1032
00:36:09,400 --> 00:36:11,240
workable because not every agent needs

1033
00:36:11,240 --> 00:36:12,640
the same level of scrutiny.

1034
00:36:12,640 --> 00:36:14,600
A personal assistant agent that only looks

1035
00:36:14,600 --> 00:36:16,800
at a single sharepoint site for one user

1036
00:36:16,800 --> 00:36:18,080
carries very little risk.

1037
00:36:18,080 --> 00:36:20,360
A customer facing agent that can update financial records

1038
00:36:20,360 --> 00:36:23,200
or access HR data is a completely different category

1039
00:36:23,200 --> 00:36:23,920
of system.

1040
00:36:23,920 --> 00:36:25,800
Using a green, yellow, and red zone framing

1041
00:36:25,800 --> 00:36:28,320
allows you to map controls to the actual risk level.

1042
00:36:28,320 --> 00:36:30,080
You can have lighter approval requirements

1043
00:36:30,080 --> 00:36:33,000
for low stakes experiments while keeping strict change control

1044
00:36:33,000 --> 00:36:35,000
for enterprise critical deployments.

1045
00:36:35,000 --> 00:36:37,320
This proportionality ensures that governance acts

1046
00:36:37,320 --> 00:36:39,560
as an enabler instead of a bottleneck.

1047
00:36:39,560 --> 00:36:41,600
The expanded data loss prevention surface

1048
00:36:41,600 --> 00:36:43,720
is something most teams don't account for

1049
00:36:43,720 --> 00:36:45,880
until they are already deep into the build.

1050
00:36:45,880 --> 00:36:48,480
In the past, DLP policies in the power platform

1051
00:36:48,480 --> 00:36:51,080
focused mostly on which services could exchange data

1052
00:36:51,080 --> 00:36:51,920
with each other.

1053
00:36:51,920 --> 00:36:54,440
In an agentic model, that surface grows much larger

1054
00:36:54,440 --> 00:36:56,760
because DLP now applies to agent connectors

1055
00:36:56,760 --> 00:36:57,720
and knowledge sources.

1056
00:36:57,720 --> 00:36:59,800
If an agent is grounded on a sharepoint site

1057
00:36:59,800 --> 00:37:03,000
or calls an external API, it is subject to these policies.

1058
00:37:03,000 --> 00:37:04,920
The control surface has expanded,

1059
00:37:04,920 --> 00:37:06,960
and your governance model has to expand with it

1060
00:37:06,960 --> 00:37:08,080
to stay coherent.

1061
00:37:08,080 --> 00:37:10,520
By the year 2026, a center of excellence

1062
00:37:10,520 --> 00:37:12,600
will look very different than it does today.

1063
00:37:12,600 --> 00:37:14,920
It won't just be an inventory of apps and flows.

1064
00:37:14,920 --> 00:37:17,800
It will include agent inventories, evaluation frameworks

1065
00:37:17,800 --> 00:37:20,680
for safety, and prompt libraries with approved patterns.

1066
00:37:20,680 --> 00:37:23,240
You will see lifecycle pipelines that move agents into production

1067
00:37:23,240 --> 00:37:25,680
with the same rigor used for professional software.

1068
00:37:25,680 --> 00:37:27,120
The architect who owns this model

1069
00:37:27,120 --> 00:37:28,640
isn't just doing compliance work.

1070
00:37:28,640 --> 00:37:30,560
They are designing the specific conditions

1071
00:37:30,560 --> 00:37:33,200
that allow an organization to scale these systems

1072
00:37:33,200 --> 00:37:34,880
without losing control.

1073
00:37:34,880 --> 00:37:36,560
That is infrastructure level authority.

1074
00:37:36,560 --> 00:37:38,520
When something goes wrong at 2am,

1075
00:37:38,520 --> 00:37:40,880
the company doesn't call the person who built the agent.

1076
00:37:40,880 --> 00:37:43,640
They call the person who designed the system that governs it,

1077
00:37:43,640 --> 00:37:45,400
scaling agentic workflows.

1078
00:37:45,400 --> 00:37:47,640
What enterprise adoption actually looks like?

1079
00:37:47,640 --> 00:37:49,280
The shift from theoretical architecture

1080
00:37:49,280 --> 00:37:52,280
to production deployment is where the real picture emerges.

1081
00:37:52,280 --> 00:37:53,840
We aren't talking about the polished version

1082
00:37:53,840 --> 00:37:55,280
you hear in a conference talk,

1083
00:37:55,280 --> 00:37:57,640
but the version where organizations actually commit budget

1084
00:37:57,640 --> 00:37:58,760
and assign ownership,

1085
00:37:58,760 --> 00:38:00,920
this is where teams find out exactly what breaks

1086
00:38:00,920 --> 00:38:03,280
when agent-driven systems meet real data

1087
00:38:03,280 --> 00:38:04,680
and real users at scale.

1088
00:38:04,680 --> 00:38:07,800
By 2026, about 25% of large enterprises

1089
00:38:07,800 --> 00:38:09,680
are actively deploying agentic AI

1090
00:38:09,680 --> 00:38:11,280
in their production workflows.

1091
00:38:11,280 --> 00:38:14,400
This puts us squarely in the early majority phase of adoption,

1092
00:38:14,400 --> 00:38:16,680
meaning we are no longer on the bleeding edge.

1093
00:38:16,680 --> 00:38:18,360
There are real case studies to study

1094
00:38:18,360 --> 00:38:20,120
and real failure modes to learn from,

1095
00:38:20,120 --> 00:38:23,160
even if most organizations don't have a mature playbook yet.

1096
00:38:23,160 --> 00:38:24,320
The companies moving right now

1097
00:38:24,320 --> 00:38:26,320
are building the patterns that everyone else will follow

1098
00:38:26,320 --> 00:38:28,040
in 18 to 24 months.

1099
00:38:28,040 --> 00:38:31,320
And that gap is closing much faster than most teams realize.

1100
00:38:31,320 --> 00:38:33,440
The pattern of adoption stays remarkably consistent

1101
00:38:33,440 --> 00:38:35,520
across different industries and geographies.

1102
00:38:35,520 --> 00:38:37,880
Organizations don't start with their most complex

1103
00:38:37,880 --> 00:38:40,680
or highest stakes processes because the risk is too high.

1104
00:38:40,680 --> 00:38:44,080
Instead, they focus on high volume, semi-structured work,

1105
00:38:44,080 --> 00:38:45,960
where the cost of an error is recoverable,

1106
00:38:45,960 --> 00:38:47,560
and the time savings are visible enough

1107
00:38:47,560 --> 00:38:49,640
to justify the governance investment.

1108
00:38:49,640 --> 00:38:52,480
IT service triage is the most common first deployment,

1109
00:38:52,480 --> 00:38:54,160
where agents receive incident reports

1110
00:38:54,160 --> 00:38:56,440
and gather diagnostic contacts before routing the ticket

1111
00:38:56,440 --> 00:39:00,760
to the right team without a human ever touching the first three steps.

1112
00:39:00,760 --> 00:39:03,200
HR Policy Q&A is another popular choice

1113
00:39:03,200 --> 00:39:05,400
as grounding an agent on actual policy documents

1114
00:39:05,400 --> 00:39:07,400
allows employees to get accurate answers

1115
00:39:07,400 --> 00:39:11,320
without burdening the HR team with repetitive questions.

1116
00:39:11,320 --> 00:39:13,840
Customer support routing and procurement request intake

1117
00:39:13,840 --> 00:39:16,400
follow this same pattern of high volume and clear business rules

1118
00:39:16,400 --> 00:39:18,880
to drive meaningful cost reduction.

1119
00:39:18,880 --> 00:39:20,680
What these use cases share is a strategy

1120
00:39:20,680 --> 00:39:22,400
that starts with reading and classifying

1121
00:39:22,400 --> 00:39:24,080
before moving to writing and acting.

1122
00:39:24,080 --> 00:39:25,760
The sequencing isn't accidental,

1123
00:39:25,760 --> 00:39:27,400
but reflects the scaling principle

1124
00:39:27,400 --> 00:39:30,400
that experienced architects apply deliberately to every project.

1125
00:39:30,400 --> 00:39:32,240
You earn the right to take actions

1126
00:39:32,240 --> 00:39:35,840
by first demonstrating that your agents understand the domain correctly.

1127
00:39:35,840 --> 00:39:38,200
An agent that reads invoices and flags exceptions

1128
00:39:38,200 --> 00:39:40,240
can be validated against human judgment

1129
00:39:40,240 --> 00:39:41,960
before it ever receives the authority

1130
00:39:41,960 --> 00:39:43,800
to approve or reject a payment.

1131
00:39:43,800 --> 00:39:45,480
Starting with a read-only scope

1132
00:39:45,480 --> 00:39:48,000
lets you build confidence in the reasoning of the system

1133
00:39:48,000 --> 00:39:49,680
before you extend its power.

1134
00:39:49,680 --> 00:39:51,080
The multi-agent pattern is emerging

1135
00:39:51,080 --> 00:39:52,840
as the natural architecture for anything

1136
00:39:52,840 --> 00:39:54,400
beyond simple triage.

1137
00:39:54,400 --> 00:39:56,880
A master agent coordinates specialized subagents

1138
00:39:56,880 --> 00:39:58,720
with one focused on policy retrieval

1139
00:39:58,720 --> 00:40:01,160
and another on system actions or approval routing.

1140
00:40:01,160 --> 00:40:03,320
Each sub agent operates with a defined scope

1141
00:40:03,320 --> 00:40:04,760
and a governed connector set

1142
00:40:04,760 --> 00:40:07,560
while the master agent handles the intent and orchestration.

1143
00:40:07,560 --> 00:40:09,240
This separation keeps individual agents

1144
00:40:09,240 --> 00:40:11,680
small enough to evaluate and govern effectively,

1145
00:40:11,680 --> 00:40:14,200
which allows the overall system to handle complexity

1146
00:40:14,200 --> 00:40:15,800
that a single monolithic agent simply

1147
00:40:15,800 --> 00:40:17,240
couldn't manage reliably.

1148
00:40:17,240 --> 00:40:19,440
The biggest obstacles to scale aren't technical

1149
00:40:19,440 --> 00:40:21,040
because the technology is already capable

1150
00:40:21,040 --> 00:40:23,920
of handling the use cases organizations are deploying.

1151
00:40:23,920 --> 00:40:25,600
The real hurdles are data readiness,

1152
00:40:25,600 --> 00:40:27,680
change management and governance maturity,

1153
00:40:27,680 --> 00:40:30,000
and they usually appear in that specific order.

1154
00:40:30,000 --> 00:40:32,960
Agents are only as useful as the quality of the data they can reach

1155
00:40:32,960 --> 00:40:35,680
and organizations that build messy sharepoint environments

1156
00:40:35,680 --> 00:40:37,640
over the years are discovering that grounding

1157
00:40:37,640 --> 00:40:39,840
an agent on poorly organized content

1158
00:40:39,840 --> 00:40:41,960
produces poorly organized answers.

1159
00:40:41,960 --> 00:40:44,160
The governance work that should have happened five years ago

1160
00:40:44,160 --> 00:40:46,080
has become a prerequisite for adoption

1161
00:40:46,080 --> 00:40:47,960
rather than a background task.

1162
00:40:47,960 --> 00:40:49,440
Change management is the friction point

1163
00:40:49,440 --> 00:40:52,720
that most technology focus teams underestimate during a rollout.

1164
00:40:52,720 --> 00:40:55,760
Users who have spent years learning to navigate specific apps

1165
00:40:55,760 --> 00:40:58,880
often resist the shift to expressing intent conversationally.

1166
00:40:58,880 --> 00:41:01,160
This isn't because the agent interfaces worse,

1167
00:41:01,160 --> 00:41:03,280
but because unfamiliar tools feel slower

1168
00:41:03,280 --> 00:41:05,160
until the user builds a new habit.

1169
00:41:05,160 --> 00:41:07,320
The organization's moving fastest on adoption

1170
00:41:07,320 --> 00:41:10,080
are the ones investing in enablement alongside architecture

1171
00:41:10,080 --> 00:41:12,600
rather than treating the human side of the equation

1172
00:41:12,600 --> 00:41:16,200
as something that happens after the technical build is finished.

1173
00:41:16,200 --> 00:41:19,240
Organizations that already standardized on M365 and Azure

1174
00:41:19,240 --> 00:41:21,320
are consolidating their agentic infrastructure

1175
00:41:21,320 --> 00:41:24,000
on co-pilot studio as the central orchestration layer.

1176
00:41:24,000 --> 00:41:25,760
The integrations that would take weeks of work

1177
00:41:25,760 --> 00:41:28,080
for anyone else are already present in the stack.

1178
00:41:28,080 --> 00:41:30,080
The identity model, the security framework

1179
00:41:30,080 --> 00:41:33,000
and the data verse substrate are all there and ready to use.

1180
00:41:33,000 --> 00:41:34,400
This means the architecture decisions

1181
00:41:34,400 --> 00:41:36,240
being made inside those organizations right now

1182
00:41:36,240 --> 00:41:38,080
are establishing the foundation for everything

1183
00:41:38,080 --> 00:41:39,960
that scales from here.

1184
00:41:39,960 --> 00:41:43,320
Power Apps is not dead, where UI still wins.

1185
00:41:43,320 --> 00:41:45,600
Before this episode creates the wrong impression,

1186
00:41:45,600 --> 00:41:47,600
we need to be direct about one thing.

1187
00:41:47,600 --> 00:41:49,480
The argument here is not that Canvas apps

1188
00:41:49,480 --> 00:41:51,400
should be replaced with agents everywhere.

1189
00:41:51,400 --> 00:41:53,360
That would be a massive over correction

1190
00:41:53,360 --> 00:41:55,160
and it is just as wrong as pretending

1191
00:41:55,160 --> 00:41:57,040
that nothing in the industry has changed.

1192
00:41:57,040 --> 00:42:01,680
A 2026 community survey of about 380 power platform practitioners

1193
00:42:01,680 --> 00:42:04,200
found that 88% plan to build the same number

1194
00:42:04,200 --> 00:42:07,720
or more Canvas apps that year compared to 2025.

1195
00:42:07,720 --> 00:42:10,280
For organizations with over 25,000 employees,

1196
00:42:10,280 --> 00:42:13,360
64% expected their Canvas app usage to increase.

1197
00:42:13,360 --> 00:42:15,440
Those aren't the numbers of a dying technology,

1198
00:42:15,440 --> 00:42:18,600
but the numbers of a mature one finding its right level of use

1199
00:42:18,600 --> 00:42:19,800
in a changing ecosystem.

1200
00:42:19,800 --> 00:42:21,400
So where does a traditional UI still win?

1201
00:42:21,400 --> 00:42:23,440
We can look at a few concrete examples.

1202
00:42:23,440 --> 00:42:26,280
Complex multi-step forms with strict validation logic

1203
00:42:26,280 --> 00:42:28,040
are still better suited for apps.

1204
00:42:28,040 --> 00:42:31,000
When a user needs to enter data across a structured process,

1205
00:42:31,000 --> 00:42:33,200
like a procurement request with conditional fields

1206
00:42:33,200 --> 00:42:36,240
or a safety inspection with mandatory photo capture,

1207
00:42:36,240 --> 00:42:39,760
a Canvas app handles the job better than a conversational interface.

1208
00:42:39,760 --> 00:42:42,600
The agent asks questions, but the form enforces rules.

1209
00:42:42,600 --> 00:42:43,760
Those are two different jobs.

1210
00:42:43,760 --> 00:42:46,320
And when enforcement matters more than flexibility,

1211
00:42:46,320 --> 00:42:47,600
the form wins every time.

1212
00:42:47,600 --> 00:42:50,440
Offline scenarios are another area where agents struggle.

1213
00:42:50,440 --> 00:42:52,000
Agents require a constant connection

1214
00:42:52,000 --> 00:42:53,800
to reason and process information,

1215
00:42:53,800 --> 00:42:56,840
but power apps can operate against locally cashed data

1216
00:42:56,840 --> 00:42:58,840
and sync when a connection is restored.

1217
00:42:58,840 --> 00:43:01,600
Field workers, warehouse staff and technicians

1218
00:43:01,600 --> 00:43:03,680
in environments with unreliable signals

1219
00:43:03,680 --> 00:43:05,480
need an interface that works regardless

1220
00:43:05,480 --> 00:43:06,840
of what the network is doing.

1221
00:43:06,840 --> 00:43:09,800
No amount of clever conversational design solves that problem,

1222
00:43:09,800 --> 00:43:12,000
making offline capability a canvas strength

1223
00:43:12,000 --> 00:43:13,520
that agents simply don't have.

1224
00:43:13,520 --> 00:43:16,560
Specialized device integration remains a major factor as well.

1225
00:43:16,560 --> 00:43:19,440
Camera access for document capture, barcode scanning,

1226
00:43:19,440 --> 00:43:21,960
signature collection, and GPS for location tagging

1227
00:43:21,960 --> 00:43:23,840
are all built into the Canvas app model.

1228
00:43:23,840 --> 00:43:26,600
These integrations work reliably across mobile platforms.

1229
00:43:26,600 --> 00:43:29,320
And while agents can receive the output of these interactions,

1230
00:43:29,320 --> 00:43:31,640
they cannot replace the interaction itself.

1231
00:43:31,640 --> 00:43:33,480
Dense data grids with complex navigation

1232
00:43:33,480 --> 00:43:35,440
also favor a visual interface.

1233
00:43:35,440 --> 00:43:38,320
When a user needs to review 200 records, apply filters,

1234
00:43:38,320 --> 00:43:39,680
and act on them in a batch.

1235
00:43:39,680 --> 00:43:41,560
A Canvas app or model-driven interface

1236
00:43:41,560 --> 00:43:45,440
handles that workflow far more efficiently than a conversation.

1237
00:43:45,440 --> 00:43:48,120
Some work is inherently visual and spatial in nature,

1238
00:43:48,120 --> 00:43:50,440
trying to express that work as a dialogue creates friction

1239
00:43:50,440 --> 00:43:52,360
that doesn't exist on a well-designed screen.

1240
00:43:52,360 --> 00:43:54,600
The hybrid pattern emerging in mature deployments

1241
00:43:54,600 --> 00:43:56,400
reflects this reality honestly.

1242
00:43:56,400 --> 00:43:59,280
Copilot Studio handles the discovery, triage,

1243
00:43:59,280 --> 00:44:01,160
and routing at the front of the interaction

1244
00:44:01,160 --> 00:44:02,560
where intent is ambiguous.

1245
00:44:02,560 --> 00:44:04,320
Power apps handles the structured completion

1246
00:44:04,320 --> 00:44:05,520
at the back of the interaction

1247
00:44:05,520 --> 00:44:07,840
where precision and user control matter most.

1248
00:44:07,840 --> 00:44:09,520
The agent gets the user to the right place

1249
00:44:09,520 --> 00:44:11,240
with the right context preloaded,

1250
00:44:11,240 --> 00:44:14,120
and then the Canvas app closes the transaction correctly.

1251
00:44:14,120 --> 00:44:16,720
Model-driven apps are finding a similar positioning shift

1252
00:44:16,720 --> 00:44:19,560
as they become the operator console for agent systems.

1253
00:44:19,560 --> 00:44:20,920
This is the interface where someone

1254
00:44:20,920 --> 00:44:23,480
configures agent behavior, reviews what an agent

1255
00:44:23,480 --> 00:44:26,760
decided, and manages the data layer that agents read from.

1256
00:44:26,760 --> 00:44:28,920
It isn't the front door for end users anymore,

1257
00:44:28,920 --> 00:44:31,520
but the control surface for the people responsible

1258
00:44:31,520 --> 00:44:33,200
for what the system actually does.

1259
00:44:33,200 --> 00:44:35,640
The mistake to avoid is the reflexive replacement

1260
00:44:35,640 --> 00:44:37,680
of every Canvas app with a chatbot just

1261
00:44:37,680 --> 00:44:40,080
because conversational interfaces feel more modern,

1262
00:44:40,080 --> 00:44:42,880
conversational interfaces are not inherently better

1263
00:44:42,880 --> 00:44:45,880
as they are great for some things and much worse for others.

1264
00:44:45,880 --> 00:44:48,120
An interface that makes a knowledge retrieval task

1265
00:44:48,120 --> 00:44:51,680
effortless can make a dense data entry task feel exhausting.

1266
00:44:51,680 --> 00:44:52,840
The right question has never been

1267
00:44:52,840 --> 00:44:54,600
whether to use a Canvas app or an agent.

1268
00:44:54,600 --> 00:44:55,960
It has always been the same question

1269
00:44:55,960 --> 00:44:58,320
we've been building toward throughout this episode.

1270
00:44:58,320 --> 00:45:01,160
We have to decide which layer should own the decision

1271
00:45:01,160 --> 00:45:04,440
and which layer should execute the steps.

1272
00:45:04,440 --> 00:45:06,160
The hybrid architecture and practice,

1273
00:45:06,160 --> 00:45:07,880
how the layers actually work together.

1274
00:45:07,880 --> 00:45:08,960
Let's make this concrete.

1275
00:45:08,960 --> 00:45:10,960
Instead of looking at a whiteboard diagram,

1276
00:45:10,960 --> 00:45:12,720
we should trace an actual architecture

1277
00:45:12,720 --> 00:45:15,320
from the first click to the final audit log.

1278
00:45:15,320 --> 00:45:17,000
HR onboarding is the perfect example

1279
00:45:17,000 --> 00:45:19,400
because it touches every single layer we have discussed.

1280
00:45:19,400 --> 00:45:22,560
You have conversational intake, structured data, approval

1281
00:45:22,560 --> 00:45:25,640
routing, and provisioning actions all happening at once.

1282
00:45:25,640 --> 00:45:27,720
On top of that, you need governance controls

1283
00:45:27,720 --> 00:45:30,280
that manage everything simultaneously.

1284
00:45:30,280 --> 00:45:32,080
Imagine this scenario, a hiring manager

1285
00:45:32,080 --> 00:45:34,720
needs to onboard a new contractor who starts in two weeks,

1286
00:45:34,720 --> 00:45:36,440
so they send that request through teams,

1287
00:45:36,440 --> 00:45:37,960
the agent picks it up immediately.

1288
00:45:37,960 --> 00:45:39,840
The first thing the agent does is figure out

1289
00:45:39,840 --> 00:45:41,480
what the manager actually needs.

1290
00:45:41,480 --> 00:45:43,640
A new contractor isn't a structured record yet.

1291
00:45:43,640 --> 00:45:46,240
It could mean a full-time person on a short contract,

1292
00:45:46,240 --> 00:45:48,840
a vendor with limited access or an agency worker

1293
00:45:48,840 --> 00:45:50,480
who follows a specific template.

1294
00:45:50,480 --> 00:45:52,320
To find the answer, the agent looks at the company

1295
00:45:52,320 --> 00:45:54,280
HR policy stored in SharePoint.

1296
00:45:54,280 --> 00:45:56,240
It uses a permission-aware search index

1297
00:45:56,240 --> 00:45:57,480
to find the right documents

1298
00:45:57,480 --> 00:45:59,560
and asks the manager clarifying questions

1299
00:45:59,560 --> 00:46:01,040
to pick the right category.

1300
00:46:01,040 --> 00:46:03,400
This whole conversation takes about 30 seconds.

1301
00:46:03,400 --> 00:46:05,800
In the old model, this required a long email chain

1302
00:46:05,800 --> 00:46:08,160
or a phone call, which created a massive delay

1303
00:46:08,160 --> 00:46:10,080
before the real work even started.

1304
00:46:10,080 --> 00:46:12,040
Once the intent is clear, the agent layer

1305
00:46:12,040 --> 00:46:13,440
makes the routing decision.

1306
00:46:13,440 --> 00:46:15,560
It figures out which onboarding workflow applies

1307
00:46:15,560 --> 00:46:18,000
and which approvals are needed based on the contractor type.

1308
00:46:18,000 --> 00:46:19,400
The agent handles these decisions

1309
00:46:19,400 --> 00:46:22,280
by looking up policies and using context-aware reasoning.

1310
00:46:22,280 --> 00:46:25,160
It doesn't ask a flow to figure out which branch to take.

1311
00:46:25,160 --> 00:46:26,440
Instead, it makes a judgment call

1312
00:46:26,440 --> 00:46:28,120
and hands the flow a structured instruction

1313
00:46:28,120 --> 00:46:29,680
to provision a specific account type

1314
00:46:29,680 --> 00:46:31,800
with set permissions for a fixed duration.

1315
00:46:31,800 --> 00:46:33,520
The power-automate flows that kick in now

1316
00:46:33,520 --> 00:46:35,480
are intentionally narrow and specific.

1317
00:46:35,480 --> 00:46:37,920
They handle account provisioning, license assignments

1318
00:46:37,920 --> 00:46:40,320
from an approved list, and system notifications

1319
00:46:40,320 --> 00:46:42,000
to IT and the hiring manager.

1320
00:46:42,000 --> 00:46:44,280
They also root approvals through the right chain

1321
00:46:44,280 --> 00:46:46,560
based on the access level of the contractor.

1322
00:46:46,560 --> 00:46:48,600
These are deterministic steps with known inputs

1323
00:46:48,600 --> 00:46:50,080
and very clear outputs.

1324
00:46:50,080 --> 00:46:51,360
They don't need to think or reason.

1325
00:46:51,360 --> 00:46:53,440
They just need to execute the task reliably

1326
00:46:53,440 --> 00:46:55,680
and log every single action so the compliance team

1327
00:46:55,680 --> 00:46:56,840
can check them later.

1328
00:46:56,840 --> 00:46:59,680
That is exactly what a well-designed cloud flow is built to do.

1329
00:46:59,680 --> 00:47:01,160
The power app shows up when it is time

1330
00:47:01,160 --> 00:47:02,960
for the structured completion step.

1331
00:47:02,960 --> 00:47:04,600
The hiring manager gets a notification

1332
00:47:04,600 --> 00:47:06,400
that the agent has gathered all the context

1333
00:47:06,400 --> 00:47:08,320
and pre-filled a role-specific form.

1334
00:47:08,320 --> 00:47:10,680
When the manager opens that form, it isn't a blank screen.

1335
00:47:10,680 --> 00:47:13,120
It is already filled out with everything the agent collected,

1336
00:47:13,120 --> 00:47:15,720
showing only the specific fields that need a human to confirm

1337
00:47:15,720 --> 00:47:16,440
or adjust.

1338
00:47:16,440 --> 00:47:18,800
The manager reviews the details, makes a quick change

1339
00:47:18,800 --> 00:47:20,840
of necessary, and hits confirm.

1340
00:47:20,840 --> 00:47:23,560
That one click triggers the next phase of the flow sequence.

1341
00:47:23,560 --> 00:47:25,960
Dataverse sits underneath the entire process

1342
00:47:25,960 --> 00:47:27,480
as the shared source of truth.

1343
00:47:27,480 --> 00:47:30,120
Both the agent and the power app read from and write

1344
00:47:30,120 --> 00:47:31,840
to this exact same data model.

1345
00:47:31,840 --> 00:47:33,560
The agent doesn't keep its own private notes

1346
00:47:33,560 --> 00:47:34,760
that the app can't see.

1347
00:47:34,760 --> 00:47:36,480
And the app doesn't create records in a silo

1348
00:47:36,480 --> 00:47:37,880
that the agent can't understand.

1349
00:47:37,880 --> 00:47:40,200
Because they share a data model, they share context.

1350
00:47:40,200 --> 00:47:42,760
This means the system stays consistent no matter which interface

1351
00:47:42,760 --> 00:47:44,920
the user chooses to use at any given moment.

1352
00:47:44,920 --> 00:47:47,800
The governance layer is the thread that ties everything together.

1353
00:47:47,800 --> 00:47:50,240
DLP policies apply to the agent's connectors and knowledge

1354
00:47:50,240 --> 00:47:52,240
sources just like they do for flows.

1355
00:47:52,240 --> 00:47:54,800
The same framework that controls what a flow can do

1356
00:47:54,800 --> 00:47:57,520
also limits what the agent can find and act on.

1357
00:47:57,520 --> 00:48:00,480
Since agent flows run under co-pilot studio capacity,

1358
00:48:00,480 --> 00:48:03,720
the organization tracks licensing and usage at the agent level

1359
00:48:03,720 --> 00:48:06,440
rather than hunting through individual flow entitlements.

1360
00:48:06,440 --> 00:48:08,360
Every interaction, execution, and submission

1361
00:48:08,360 --> 00:48:09,680
feeds into purview.

1362
00:48:09,680 --> 00:48:11,520
This gives the compliance team one single view

1363
00:48:11,520 --> 00:48:13,480
of what happened and who authorized it.

1364
00:48:13,480 --> 00:48:14,800
But here is the real insight.

1365
00:48:14,800 --> 00:48:17,480
The complexity of the process didn't actually go away.

1366
00:48:17,480 --> 00:48:19,840
Onboarding still has the same business rules it always had,

1367
00:48:19,840 --> 00:48:22,080
but the big shift is where that complexity lives.

1368
00:48:22,080 --> 00:48:23,840
In the old way of doing things, complexity

1369
00:48:23,840 --> 00:48:25,680
was scattered across different screens,

1370
00:48:25,680 --> 00:48:27,880
messy formulas, and manual handoffs.

1371
00:48:27,880 --> 00:48:29,960
Now it lives in the orchestration layer.

1372
00:48:29,960 --> 00:48:32,120
You have agent instructions that are easy to read and test,

1373
00:48:32,120 --> 00:48:33,840
modular flows that are easy to fix,

1374
00:48:33,840 --> 00:48:35,600
and a data model that everyone shares.

1375
00:48:35,600 --> 00:48:38,800
Moving that complexity into governed, auditable pieces

1376
00:48:38,800 --> 00:48:42,040
is what makes the systems stay manageable as the company grows.

1377
00:48:42,040 --> 00:48:44,080
An architect designed this entire flow,

1378
00:48:44,080 --> 00:48:45,480
someone had to decide which layer

1379
00:48:45,480 --> 00:48:47,200
owned which specific responsibility.

1380
00:48:47,200 --> 00:48:48,800
That person's logic is the only reason

1381
00:48:48,800 --> 00:48:50,720
the whole thing actually functions.

1382
00:48:50,720 --> 00:48:53,120
The security implications nobody is talking about.

1383
00:48:53,120 --> 00:48:54,960
The architecture we just walked through is great

1384
00:48:54,960 --> 00:48:56,520
when it is designed correctly.

1385
00:48:56,520 --> 00:48:58,600
But there is a side to this shift that most teams

1386
00:48:58,600 --> 00:49:00,720
are building on Copilot Studio completely ignore

1387
00:49:00,720 --> 00:49:02,760
until a problem forces them to look at it.

1388
00:49:02,760 --> 00:49:04,640
By the time they realized there is a gap,

1389
00:49:04,640 --> 00:49:06,720
the design decisions that created the risk

1390
00:49:06,720 --> 00:49:08,280
are already running in production.

1391
00:49:08,280 --> 00:49:10,520
Agente workflows open up the attack surface

1392
00:49:10,520 --> 00:49:12,720
in a way that old canvas apps never did.

1393
00:49:12,720 --> 00:49:14,840
This isn't because agents are less secure by nature.

1394
00:49:14,840 --> 00:49:17,280
In fact, they can be much safer than the tools they replace

1395
00:49:17,280 --> 00:49:18,920
if you build them the right way.

1396
00:49:18,920 --> 00:49:21,160
The problem is that the threat model has changed,

1397
00:49:21,160 --> 00:49:23,280
and most people working on the power platform

1398
00:49:23,280 --> 00:49:24,680
were trained on a security model

1399
00:49:24,680 --> 00:49:26,480
that just doesn't apply here anymore.

1400
00:49:26,480 --> 00:49:28,640
It starts with the identity of the agent.

1401
00:49:28,640 --> 00:49:31,080
Every Copilot Studio agent is a machine identity

1402
00:49:31,080 --> 00:49:33,680
in Microsoft, EntraID, with its own credentials

1403
00:49:33,680 --> 00:49:34,600
and access policies.

1404
00:49:34,600 --> 00:49:37,160
It isn't a human user, and it isn't a flow running

1405
00:49:37,160 --> 00:49:38,800
under someone's delegated permissions.

1406
00:49:38,800 --> 00:49:39,920
It is a service principle.

1407
00:49:39,920 --> 00:49:42,360
You have to treat it with the same discipline you would use

1408
00:49:42,360 --> 00:49:44,520
for a service principle in an Azure workload.

1409
00:49:44,520 --> 00:49:46,760
That means you need least privilege scoping,

1410
00:49:46,760 --> 00:49:49,200
credential rotation and conditional access rules.

1411
00:49:49,200 --> 00:49:51,480
If you deploy an agent without these controls,

1412
00:49:51,480 --> 00:49:53,440
you are running a powerful identity

1413
00:49:53,440 --> 00:49:56,160
with whatever default permissions the platform gave it.

1414
00:49:56,160 --> 00:49:58,320
And those defaults are almost never restricted enough.

1415
00:49:58,320 --> 00:50:00,840
The problem of over-pimissioning is a very real risk.

1416
00:50:00,840 --> 00:50:03,000
We aren't talking about a hypothetical situation

1417
00:50:03,000 --> 00:50:04,600
when we see misconfigured agents

1418
00:50:04,600 --> 00:50:07,040
with read access to an entire SharePoint tenant.

1419
00:50:07,040 --> 00:50:08,720
This happens when people build too fast

1420
00:50:08,720 --> 00:50:11,800
and forget to think about scope during the design phase.

1421
00:50:11,800 --> 00:50:13,880
An agent that can read every document in your company

1422
00:50:13,880 --> 00:50:16,720
can find information that the user shouldn't be able

1423
00:50:16,720 --> 00:50:18,120
to see through the normal UI.

1424
00:50:18,120 --> 00:50:20,120
The interface used to enforce those permissions,

1425
00:50:20,120 --> 00:50:21,320
but the agent bypassed them

1426
00:50:21,320 --> 00:50:24,000
because the retrieval layer wasn't filtered correctly.

1427
00:50:24,000 --> 00:50:26,400
That retrieval layer is where the rag exposure lives.

1428
00:50:26,400 --> 00:50:28,200
Agents use search to find documents

1429
00:50:28,200 --> 00:50:30,040
before they generate an answer for a user.

1430
00:50:30,040 --> 00:50:31,960
If that search doesn't use the same permission trimming

1431
00:50:31,960 --> 00:50:33,360
that SharePoint usually does,

1432
00:50:33,360 --> 00:50:35,160
the agent might show content from a file

1433
00:50:35,160 --> 00:50:36,720
the user has no business seeing.

1434
00:50:36,720 --> 00:50:39,120
The UI that used to protect that data is gone.

1435
00:50:39,120 --> 00:50:42,280
Now, the agent's retrieval layer is the new security boundary

1436
00:50:42,280 --> 00:50:45,320
and you have to configure it to respect underlying permissions

1437
00:50:45,320 --> 00:50:47,760
because it doesn't always inherit them automatically.

1438
00:50:47,760 --> 00:50:51,040
Prompt injection is the third big risk you need to watch out for.

1439
00:50:51,040 --> 00:50:53,200
An agent that just answers questions is a target

1440
00:50:53,200 --> 00:50:55,720
but an agent that can take actions is a much bigger price.

1441
00:50:55,720 --> 00:50:58,240
If an agent can create records, trigger approvals

1442
00:50:58,240 --> 00:50:59,800
or change system settings,

1443
00:50:59,800 --> 00:51:03,560
it becomes a major target for anyone looking to abuse those powers.

1444
00:51:03,560 --> 00:51:05,640
Prompt injection happens when malicious instructions

1445
00:51:05,640 --> 00:51:07,920
are hidden in the content and agent processes.

1446
00:51:07,920 --> 00:51:09,560
Causing it to do things it wasn't supposed to do,

1447
00:51:09,560 --> 00:51:12,480
there is no equivalent for this in traditional app security.

1448
00:51:12,480 --> 00:51:14,240
You can't SQL inject a simple gallery,

1449
00:51:14,240 --> 00:51:16,280
but you can definitely prompt inject an agent

1450
00:51:16,280 --> 00:51:17,840
if you don't have the right at guard rails.

1451
00:51:17,840 --> 00:51:19,080
Fixing these governance issues

1452
00:51:19,080 --> 00:51:21,200
isn't a separate project from the architecture work,

1453
00:51:21,200 --> 00:51:24,240
it is the same work just with security as a main requirement.

1454
00:51:24,240 --> 00:51:25,760
You need agent IDs managed in entry

1455
00:51:25,760 --> 00:51:27,480
with the smallest possible access levels.

1456
00:51:27,480 --> 00:51:29,000
You need conditional access policies

1457
00:51:29,000 --> 00:51:31,520
that look at the context of what the agent is doing.

1458
00:51:31,520 --> 00:51:33,840
You also need DLP policies for every connector.

1459
00:51:33,840 --> 00:51:36,800
Finally, you need to pipe all your audit logs into Sentinel.

1460
00:51:36,800 --> 00:51:39,000
This way, if an agent starts touching data

1461
00:51:39,000 --> 00:51:41,960
it is never used before or sends a massive spike of actions,

1462
00:51:41,960 --> 00:51:43,640
you see it immediately instead of finding out

1463
00:51:43,640 --> 00:51:44,840
after the damage is done.

1464
00:51:44,840 --> 00:51:47,160
If an architect thinks security is a separate task

1465
00:51:47,160 --> 00:51:48,760
for another team to handle later,

1466
00:51:48,760 --> 00:51:50,200
they aren't really an architect.

1467
00:51:50,200 --> 00:51:51,640
They are a liability to the company.

1468
00:51:51,640 --> 00:51:54,400
Security isn't a gate you pass through at the end of the build.

1469
00:51:54,400 --> 00:51:56,120
It has to be part of every single layer

1470
00:51:56,120 --> 00:51:58,920
and every design decision we have talked about today.

1471
00:51:58,920 --> 00:52:00,960
What junior roles actually look like now?

1472
00:52:00,960 --> 00:52:02,880
Stack Overflow recently released data

1473
00:52:02,880 --> 00:52:04,960
that confirms what many of us already felt.

1474
00:52:04,960 --> 00:52:07,760
Entry level tech hiring dropped 25% in 24.

1475
00:52:07,760 --> 00:52:09,960
That isn't just a small dip or a rounding error.

1476
00:52:09,960 --> 00:52:13,520
It's a structural shift in how companies value early career talent

1477
00:52:13,520 --> 00:52:15,600
and if you look closely at why this is happening.

1478
00:52:15,600 --> 00:52:18,800
It tells you exactly where the entire ecosystem is headed.

1479
00:52:18,800 --> 00:52:20,480
The roles under the most pressure right now

1480
00:52:20,480 --> 00:52:23,280
are the ones that used to define the first two years of a career.

1481
00:52:23,280 --> 00:52:27,160
Scaffolding, boilerplate, basic thread work, routine documentation.

1482
00:52:27,160 --> 00:52:28,760
We used to call these the easy assignments

1483
00:52:28,760 --> 00:52:30,120
not because they were mindless,

1484
00:52:30,120 --> 00:52:31,760
but because the skills were easy to learn

1485
00:52:31,760 --> 00:52:33,560
and the cost of a mistake was low,

1486
00:52:33,560 --> 00:52:35,320
but those tasks weren't just busy work.

1487
00:52:35,320 --> 00:52:36,680
They were how you built judgment.

1488
00:52:36,680 --> 00:52:38,720
You learned how systems actually behaved

1489
00:52:38,720 --> 00:52:41,040
by building the small pieces that connected them

1490
00:52:41,040 --> 00:52:42,360
and you understood data models

1491
00:52:42,360 --> 00:52:44,520
by writing the queries that pulled from them.

1492
00:52:44,520 --> 00:52:46,320
The repetitive work was never the point,

1493
00:52:46,320 --> 00:52:48,760
the deep understanding you gained while doing it was.

1494
00:52:48,760 --> 00:52:51,640
Now, a gentigai is handling a massive chunk of that volume.

1495
00:52:51,640 --> 00:52:54,360
It isn't doing all of it and it certainly isn't doing it perfectly,

1496
00:52:54,360 --> 00:52:57,600
but it's doing enough that the supply of hands-on repetitive tasks

1497
00:52:57,600 --> 00:52:59,200
for new hires is drying up.

1498
00:52:59,200 --> 00:53:01,640
The older apprenticeship model relied on a steady stream

1499
00:53:01,640 --> 00:53:04,440
of small concrete problems where mistakes were survivable.

1500
00:53:04,440 --> 00:53:06,440
That supply is shrinking and what's replacing it

1501
00:53:06,440 --> 00:53:08,000
is a different kind of work that requires

1502
00:53:08,000 --> 00:53:09,920
a completely different starting posture.

1503
00:53:09,920 --> 00:53:12,000
The tasks that matter now at the entry level

1504
00:53:12,000 --> 00:53:14,400
are about judgment rather than production.

1505
00:53:14,400 --> 00:53:16,840
It's about writing precise specifications

1506
00:53:16,840 --> 00:53:19,400
that an agent can execute without getting confused

1507
00:53:19,400 --> 00:53:22,480
and it's about reviewing AI-generated logic for correctness

1508
00:53:22,480 --> 00:53:24,760
instead of just checking if the code runs.

1509
00:53:24,760 --> 00:53:27,240
Juniors are now being asked to inspect security boundaries

1510
00:53:27,240 --> 00:53:28,800
and understand system architecture

1511
00:53:28,800 --> 00:53:31,280
much earlier than previous generations ever had to

1512
00:53:31,280 --> 00:53:33,480
and this is the part that makes people uncomfortable.

1513
00:53:33,480 --> 00:53:35,080
The traditional career ladder worked

1514
00:53:35,080 --> 00:53:36,480
because you could be productive

1515
00:53:36,480 --> 00:53:38,680
while knowing very little about the full system,

1516
00:53:38,680 --> 00:53:40,040
your work was scoped so narrowly

1517
00:53:40,040 --> 00:53:42,560
that local knowledge was enough to get the job done.

1518
00:53:42,560 --> 00:53:44,400
But this shift is compressing the learning curve

1519
00:53:44,400 --> 00:53:46,000
whether we like it or not.

1520
00:53:46,000 --> 00:53:48,240
A junior developer, reviewing agent output

1521
00:53:48,240 --> 00:53:50,480
needs to know what the agent was supposed to do,

1522
00:53:50,480 --> 00:53:53,680
what it actually did and why those two things might be different

1523
00:53:53,680 --> 00:53:55,840
that requires a level of architectural context

1524
00:53:55,840 --> 00:53:57,600
that used to take years to build.

1525
00:53:57,600 --> 00:54:00,400
The reality is that career ladders built for the old world

1526
00:54:00,400 --> 00:54:02,240
are now producing the wrong outcomes.

1527
00:54:02,240 --> 00:54:05,240
If the entry point used to be build these specific things

1528
00:54:05,240 --> 00:54:06,840
and that task is now automated,

1529
00:54:06,840 --> 00:54:07,920
the entry point has to change.

1530
00:54:07,920 --> 00:54:10,120
Organizations that don't redesign their onboarding

1531
00:54:10,120 --> 00:54:11,800
are going to struggle to develop talent

1532
00:54:11,800 --> 00:54:14,040
that can actually survive in this environment.

1533
00:54:14,040 --> 00:54:16,560
The easy starter tasks are gone and they aren't coming back.

1534
00:54:16,560 --> 00:54:18,480
But the judgment those tasks were meant to build

1535
00:54:18,480 --> 00:54:20,080
still has to come from somewhere.

1536
00:54:20,080 --> 00:54:22,240
This isn't a story about people being replaced.

1537
00:54:22,240 --> 00:54:25,040
A generation of developers isn't being eliminated.

1538
00:54:25,040 --> 00:54:26,360
They're being redefined.

1539
00:54:26,360 --> 00:54:28,280
They're being asked to contribute earlier

1540
00:54:28,280 --> 00:54:30,440
and in ways the industry hasn't quite figured out

1541
00:54:30,440 --> 00:54:31,640
how to support yet.

1542
00:54:31,640 --> 00:54:34,960
Is co-pilot studio actually replacing local developers?

1543
00:54:34,960 --> 00:54:36,280
So let's get straight to the question

1544
00:54:36,280 --> 00:54:38,680
that the title of this video has been holding over us.

1545
00:54:38,680 --> 00:54:41,480
Is co-pilot studio replacing local developers?

1546
00:54:41,480 --> 00:54:42,360
The honest answer is,

1547
00:54:42,360 --> 00:54:44,400
it depends on which part of the job you're talking about.

1548
00:54:44,400 --> 00:54:46,800
What's actually being displaced is a very specific type

1549
00:54:46,800 --> 00:54:47,880
of practitioner.

1550
00:54:47,880 --> 00:54:50,160
I'm talking about the manual screen builder.

1551
00:54:50,160 --> 00:54:52,720
The person whose primary value was building canvas apps

1552
00:54:52,720 --> 00:54:53,720
as fast as possible.

1553
00:54:53,720 --> 00:54:56,040
This was the power FX specialist who knew exactly

1554
00:54:56,040 --> 00:54:58,040
how to wire a gallery selection into a variable

1555
00:54:58,040 --> 00:55:00,240
or how to structure a complex patch function.

1556
00:55:00,240 --> 00:55:02,640
That role existed because there was a massive gap

1557
00:55:02,640 --> 00:55:04,680
between what a business user could do

1558
00:55:04,680 --> 00:55:07,720
and what a professional dev team could deliver in six months.

1559
00:55:07,720 --> 00:55:09,680
Canvas apps filled that gap perfectly

1560
00:55:09,680 --> 00:55:12,320
and the people who could build them were incredibly valuable.

1561
00:55:12,320 --> 00:55:13,400
The gap is still there.

1562
00:55:13,400 --> 00:55:14,520
But it has moved.

1563
00:55:14,520 --> 00:55:16,520
The delivery problem is increasingly being solved

1564
00:55:16,520 --> 00:55:18,360
by AI-assisted generation.

1565
00:55:18,360 --> 00:55:20,760
Co-pilot in Power Apps can now scaffold a canvas app

1566
00:55:20,760 --> 00:55:22,040
from a simple conversation

1567
00:55:22,040 --> 00:55:24,920
and the M365 agent toolkit can provision a working agent

1568
00:55:24,920 --> 00:55:26,640
from a single instruction file.

1569
00:55:26,640 --> 00:55:28,360
The speed advantage that screen builders had

1570
00:55:28,360 --> 00:55:30,440
over traditional developers is disappearing

1571
00:55:30,440 --> 00:55:33,040
because of the same platform shift hitting everyone else

1572
00:55:33,040 --> 00:55:35,280
when the tool builds the first version for you.

1573
00:55:35,280 --> 00:55:37,680
The skill that matters isn't how fast you can click.

1574
00:55:37,680 --> 00:55:40,480
It's how well you can design what the system should actually do.

1575
00:55:40,480 --> 00:55:42,400
That is the part that isn't being replaced.

1576
00:55:42,400 --> 00:55:44,440
The person who understands the business process,

1577
00:55:44,440 --> 00:55:45,840
who can sit with a department head

1578
00:55:45,840 --> 00:55:48,480
and find the real problem hidden under a request.

1579
00:55:48,480 --> 00:55:50,400
That person is more valuable now than ever.

1580
00:55:50,400 --> 00:55:52,400
The platform can generate the artifacts,

1581
00:55:52,400 --> 00:55:54,120
but it cannot replace judgment.

1582
00:55:54,120 --> 00:55:56,520
It can't decide which layer should own a specific decision

1583
00:55:56,520 --> 00:55:57,480
and it doesn't know what to do

1584
00:55:57,480 --> 00:56:00,240
when reality doesn't match the original design.

1585
00:56:00,240 --> 00:56:01,960
The data shows both sides of this story

1586
00:56:01,960 --> 00:56:03,240
happening at the same time.

1587
00:56:03,240 --> 00:56:05,560
About 78% of Power Platform teams

1588
00:56:05,560 --> 00:56:07,840
are experimenting with CoPilot Studio.

1589
00:56:07,840 --> 00:56:10,560
Yet 88% are still building Canvas apps.

1590
00:56:10,560 --> 00:56:12,080
Those numbers aren't fighting each other.

1591
00:56:12,080 --> 00:56:13,520
They're describing two different parts

1592
00:56:13,520 --> 00:56:15,000
of the same modern architecture.

1593
00:56:15,000 --> 00:56:17,080
CoPilot Studio is becoming the orchestration layer

1594
00:56:17,080 --> 00:56:18,840
that sits on top of Power Apps.

1595
00:56:18,840 --> 00:56:20,480
It isn't a replacement for it.

1596
00:56:20,480 --> 00:56:22,240
The organization is moving the fastest

1597
00:56:22,240 --> 00:56:24,440
on choosing one tool over the other.

1598
00:56:24,440 --> 00:56:27,520
They're designing systems where each layer does what it does best

1599
00:56:27,520 --> 00:56:30,720
and they need people who understand where those boundaries sit.

1600
00:56:30,720 --> 00:56:33,520
The real shift isn't CoPilot Studio instead of Power Apps.

1601
00:56:33,520 --> 00:56:35,680
It's CoPilot Studio acting as a reasoning layer

1602
00:56:35,680 --> 00:56:37,840
that makes Power Apps better by taking away the work

1603
00:56:37,840 --> 00:56:39,520
that screens were never meant to handle.

1604
00:56:39,520 --> 00:56:41,960
If a developer only knows how to build screens,

1605
00:56:41,960 --> 00:56:43,280
they are being displaced,

1606
00:56:43,280 --> 00:56:45,720
but they aren't being replaced by a specific tool.

1607
00:56:45,720 --> 00:56:47,920
They are being replaced by the architectural shift

1608
00:56:47,920 --> 00:56:49,520
that the tool represents.

1609
00:56:49,520 --> 00:56:51,040
The platform rebuilt the foundation

1610
00:56:51,040 --> 00:56:53,080
and if you only built on the old surface,

1611
00:56:53,080 --> 00:56:56,200
you're going to find that surface is no longer where the value is.

1612
00:56:56,200 --> 00:56:57,640
This isn't a personal failure.

1613
00:56:57,640 --> 00:56:58,720
It's just a platform cycle.

1614
00:56:58,720 --> 00:57:00,680
Every cycle produces the same result.

1615
00:57:00,680 --> 00:57:03,040
The people tied most tightly to the old way of doing things

1616
00:57:03,040 --> 00:57:04,840
that face the most pressure to change.

1617
00:57:04,840 --> 00:57:08,360
On the other hand, the developer who understands logic and governance

1618
00:57:08,360 --> 00:57:10,440
is in a stronger position than they've ever been.

1619
00:57:10,440 --> 00:57:12,880
The demand for that kind of system level thinking

1620
00:57:12,880 --> 00:57:15,960
is growing much faster than the supply of people who can do it.

1621
00:57:15,960 --> 00:57:17,280
We're seeing roles pop up today

1622
00:57:17,280 --> 00:57:19,360
that didn't even exist two years ago,

1623
00:57:19,360 --> 00:57:22,400
like AI product owners and logic architecture leads.

1624
00:57:22,400 --> 00:57:24,360
These aren't just rebranded screen builder jobs.

1625
00:57:24,360 --> 00:57:26,320
They are high-level positions being filled by people

1626
00:57:26,320 --> 00:57:28,480
who saw the shift coming and expanded their scope

1627
00:57:28,480 --> 00:57:29,680
before they were forced to.

1628
00:57:29,680 --> 00:57:33,000
The question was never really about CoPilot Studio replacing developers.

1629
00:57:33,000 --> 00:57:36,640
The real question is, which part of what you do is actually irreplaceable.

1630
00:57:36,640 --> 00:57:38,200
The answer is the same as it's always been.

1631
00:57:38,200 --> 00:57:40,200
It's your judgment, your context,

1632
00:57:40,200 --> 00:57:43,720
and your ability to design a system that works when reality gets messy.

1633
00:57:43,720 --> 00:57:47,760
That part hasn't changed even if everything else around it has.

1634
00:57:47,760 --> 00:57:49,960
The three moves that matter right now.

1635
00:57:49,960 --> 00:57:52,840
Everything we have covered leads to one practical question.

1636
00:57:52,840 --> 00:57:55,120
Given where the platform is and where it is heading,

1637
00:57:55,120 --> 00:57:57,560
what do you actually do differently starting this week?

1638
00:57:57,560 --> 00:57:58,400
You need three moves.

1639
00:57:58,400 --> 00:58:01,400
This is not a six-month training plan or a long certification roadmap.

1640
00:58:01,400 --> 00:58:05,320
These are three decisions about how you approach the work already sitting on your desk.

1641
00:58:05,320 --> 00:58:09,000
Move one is to stop designing in screens and start designing in layers.

1642
00:58:09,000 --> 00:58:11,120
Every solution you work on from this point forward

1643
00:58:11,120 --> 00:58:14,000
needs an explicit answer to one question before you build anything.

1644
00:58:14,000 --> 00:58:16,120
You have to decide which layer owns the decision.

1645
00:58:16,120 --> 00:58:17,920
It is not about what screens you need

1646
00:58:17,920 --> 00:58:20,000
or how many flows the project requires.

1647
00:58:20,000 --> 00:58:22,480
You must determine if the agent, the flow or the data model

1648
00:58:22,480 --> 00:58:24,640
is responsible for the reasoning that drives each step.

1649
00:58:24,640 --> 00:58:26,680
If you cannot answer that question cleanly,

1650
00:58:26,680 --> 00:58:28,320
your design is not finished yet.

1651
00:58:28,320 --> 00:58:30,760
This is not about adding more documentation to your day.

1652
00:58:30,760 --> 00:58:32,920
It is about catching a structural mistake

1653
00:58:32,920 --> 00:58:34,480
before it gets baked into the system.

1654
00:58:34,480 --> 00:58:37,600
A structural mistake in an agentic architecture is much harder to fix

1655
00:58:37,600 --> 00:58:39,840
than a messy power FX formula ever was.

1656
00:58:39,840 --> 00:58:42,880
Move two is to learn the governance model before you actually need it.

1657
00:58:42,880 --> 00:58:45,400
The five-phase life cycle and the risk-tiered zoning

1658
00:58:45,400 --> 00:58:47,920
are where authority will live in 2026.

1659
00:58:47,920 --> 00:58:49,520
The same applies to the DLP surface

1660
00:58:49,520 --> 00:58:52,040
that now covers agent connectors and knowledge sources.

1661
00:58:52,040 --> 00:58:54,400
Authority will not come from how fast you deliver

1662
00:58:54,400 --> 00:58:56,160
or how many features you pack in.

1663
00:58:56,160 --> 00:58:58,320
It will come from your ability to design systems

1664
00:58:58,320 --> 00:59:00,720
that the organization can trust with real data

1665
00:59:00,720 --> 00:59:02,600
and real business processes at scale.

1666
00:59:02,600 --> 00:59:04,800
The company's figuring out agent governance right now

1667
00:59:04,800 --> 00:59:07,680
are writing the playbooks that everyone else will have to follow later.

1668
00:59:07,680 --> 00:59:09,680
Architects who join those conversations today

1669
00:59:09,680 --> 00:59:11,520
have a two-year head start on everyone

1670
00:59:11,520 --> 00:59:13,200
who waits for a crisis to happen.

1671
00:59:13,200 --> 00:59:16,360
Getting fluent in this model before an incident forces the conversation

1672
00:59:16,360 --> 00:59:17,720
is how you position yourself.

1673
00:59:17,720 --> 00:59:20,240
You want to be the person who designed the guardrails,

1674
00:59:20,240 --> 00:59:22,720
not the person called in to explain why they were missing.

1675
00:59:22,720 --> 00:59:25,480
Move three is to build one reference architecture.

1676
00:59:25,480 --> 00:59:27,240
I'm not talking about a tutorial project

1677
00:59:27,240 --> 00:59:29,880
or a proof of concept that just shows off a feature.

1678
00:59:29,880 --> 00:59:32,120
You need a real design artifact with diagrams

1679
00:59:32,120 --> 00:59:34,120
and architecture decision records.

1680
00:59:34,120 --> 00:59:35,520
It should include a written rationale

1681
00:59:35,520 --> 00:59:38,240
for why every component sits in its specific layer.

1682
00:59:38,240 --> 00:59:40,680
This artifact shows an agent, a flow,

1683
00:59:40,680 --> 00:59:42,720
and a canvas app working together

1684
00:59:42,720 --> 00:59:44,560
under one shared governance model.

1685
00:59:44,560 --> 00:59:46,560
A resume cannot carry this kind of signal,

1686
00:59:46,560 --> 00:59:48,040
but a portfolio can.

1687
00:59:48,040 --> 00:59:49,840
It proves that you think at the system level

1688
00:59:49,840 --> 00:59:51,320
instead of just the screen level.

1689
00:59:51,320 --> 00:59:54,280
This is the document you bring into a design review to show your worth,

1690
00:59:54,280 --> 00:59:56,960
building it once against a realistic use case

1691
00:59:56,960 --> 00:59:59,680
forces you to make every trade off we have discussed.

1692
00:59:59,680 --> 01:00:02,480
It turns theoretical ideas into actual decisions.

1693
01:00:02,480 --> 01:00:04,480
The career signals show why this is urgent.

1694
01:00:04,480 --> 01:00:07,200
Organizations are actively posting for AI product owners

1695
01:00:07,200 --> 01:00:08,840
and agent COE leads.

1696
01:00:08,840 --> 01:00:11,400
These logic architecture titles did not even exist

1697
01:00:11,400 --> 01:00:13,480
in job listings 24 months ago.

1698
01:00:13,480 --> 01:00:17,040
The AB 900 certification for Microsoft 365 co-pilot

1699
01:00:17,040 --> 01:00:19,920
and agent administration is now a preferred qualification

1700
01:00:19,920 --> 01:00:21,480
for many architect roles.

1701
01:00:21,480 --> 01:00:23,440
The platform is telling you exactly what it values

1702
01:00:23,440 --> 01:00:24,520
through the job market.

1703
01:00:24,520 --> 01:00:26,240
The window to be early to this transition

1704
01:00:26,240 --> 01:00:28,400
is measured in months, not years.

1705
01:00:28,400 --> 01:00:30,240
This is not about learning a new tool.

1706
01:00:30,240 --> 01:00:31,640
Co-pilot studio is just a tool,

1707
01:00:31,640 --> 01:00:34,320
but the shift happening underneath it is architectural.

1708
01:00:34,320 --> 01:00:37,080
The organization's building fluency in that architecture now

1709
01:00:37,080 --> 01:00:38,920
are creating a foundation that will grow

1710
01:00:38,920 --> 01:00:41,200
when the next wave of capability arrives.

1711
01:00:41,200 --> 01:00:43,560
The people who wait will simply inherit those decisions

1712
01:00:43,560 --> 01:00:44,880
rather than making them.

1713
01:00:44,880 --> 01:00:47,800
What you build in the next six months determines your path.

1714
01:00:47,800 --> 01:00:50,280
You will either be the person who designed this layer

1715
01:00:50,280 --> 01:00:52,720
or the person who was handed a design by someone else.

1716
01:00:52,720 --> 01:00:54,360
Both paths are still open right now.

1717
01:00:54,360 --> 01:00:56,360
The one you take is a choice you make

1718
01:00:56,360 --> 01:00:58,120
in the work you are doing today.

1719
01:00:58,120 --> 01:01:00,560
The value in local development has moved from the screen

1720
01:01:00,560 --> 01:01:01,400
to the system.

1721
01:01:01,400 --> 01:01:03,560
We are moving away from building interfaces

1722
01:01:03,560 --> 01:01:06,160
and toward governing the logic that runs beneath them.

1723
01:01:06,160 --> 01:01:09,120
That is the shift everything else we talked about in this episode

1724
01:01:09,120 --> 01:01:11,160
is just a consequence of that change.

1725
01:01:11,160 --> 01:01:12,760
Here is your challenge for this week.

1726
01:01:12,760 --> 01:01:14,440
In your next project, find one place

1727
01:01:14,440 --> 01:01:16,920
where business logic currently lives on a screen.

1728
01:01:16,920 --> 01:01:20,040
Ask yourself if that logic belongs in an agent instruction,

1729
01:01:20,040 --> 01:01:22,440
a flow or a data layer rule instead.

1730
01:01:22,440 --> 01:01:23,040
Just find one.

1731
01:01:23,040 --> 01:01:25,600
That single question is how you start the transition.

1732
01:01:25,600 --> 01:01:27,560
If this changed how you think about your career

1733
01:01:27,560 --> 01:01:29,920
in the Microsoft ecosystem, please leave a review.

1734
01:01:29,920 --> 01:01:32,240
It helps more people find this conversation.

1735
01:01:32,240 --> 01:01:34,280
The people who need this information most

1736
01:01:34,280 --> 01:01:36,600
are often the ones who have not found it yet.

1737
01:01:36,600 --> 01:01:38,840
Follow me, Mirko Peters on LinkedIn for more analysis

1738
01:01:38,840 --> 01:01:40,640
on where this platform is actually going.

1739
01:01:40,640 --> 01:01:42,600
I focus on the logic, not just the announcements.

1740
01:01:42,600 --> 01:01:44,280
I will see you in the next one.