The Agentic Advantage: Scaling Intelligence Without Chaos

1
00:00:00,000 --> 00:00:04,280
Most organizations hear more agents and assume more productivity.

2
00:00:04,280 --> 00:00:06,560
That assumption is comfortable, it's also wrong.

3
00:00:06,560 --> 00:00:08,840
More agents usually means more unmanaged authority,

4
00:00:08,840 --> 00:00:10,120
more automated side effects,

5
00:00:10,120 --> 00:00:13,240
and more places where accountability quietly disappears.

6
00:00:13,240 --> 00:00:14,760
This isn't an episode about features,

7
00:00:14,760 --> 00:00:17,040
it's about scale, risk, and an operating model

8
00:00:17,040 --> 00:00:19,240
that survives contact with reality.

9
00:00:19,240 --> 00:00:20,760
You'll leave with three failure modes

10
00:00:20,760 --> 00:00:21,880
that kill agent programs,

11
00:00:21,880 --> 00:00:24,040
the four layer control plane that prevents drift

12
00:00:24,040 --> 00:00:26,800
and the questions executives should demand answers to.

13
00:00:26,800 --> 00:00:28,280
Now, start with the misunderstanding

14
00:00:28,280 --> 00:00:29,600
that causes the chaos.

15
00:00:29,600 --> 00:00:33,240
The foundational misunderstanding agents aren't assistants

16
00:00:33,240 --> 00:00:34,960
and assistant generates answers,

17
00:00:34,960 --> 00:00:36,200
then agent executes work.

18
00:00:36,200 --> 00:00:38,840
That distinction matters because execution creates state

19
00:00:38,840 --> 00:00:40,520
and state creates consequences.

20
00:00:40,520 --> 00:00:43,000
A copilot style assistant can write a summary.

21
00:00:43,000 --> 00:00:45,520
An agent can open a ticket, change a permission,

22
00:00:45,520 --> 00:00:48,480
update a record, trigger a flow, notify a manager,

23
00:00:48,480 --> 00:00:49,960
and then do it again tomorrow

24
00:00:49,960 --> 00:00:51,680
because you told it that's the process.

25
00:00:51,680 --> 00:00:54,600
The system didn't help, it acted.

26
00:00:54,600 --> 00:00:56,680
Most enterprises keep talking about agents

27
00:00:56,680 --> 00:00:58,800
like their chatbots with better prompts.

28
00:00:58,800 --> 00:00:59,640
They're not.

29
00:00:59,640 --> 00:01:01,920
The agentic model is tools plus memory plus loops.

30
00:01:01,920 --> 00:01:03,520
It takes a goal, it calls something,

31
00:01:03,520 --> 00:01:05,840
it evaluates the result, it calls something else,

32
00:01:05,840 --> 00:01:08,520
and it keeps going until it decides it's done.

33
00:01:08,520 --> 00:01:10,440
That loop is the real architectural shift.

34
00:01:10,440 --> 00:01:11,920
Here's the uncomfortable truth.

35
00:01:11,920 --> 00:01:15,360
Once you cross from answer generation to tool execution,

36
00:01:15,360 --> 00:01:17,480
you are no longer governing a conversation.

37
00:01:17,480 --> 00:01:19,440
You are governing a distributed decision engine

38
00:01:19,440 --> 00:01:20,960
that can mutate systems.

39
00:01:20,960 --> 00:01:23,760
Think about the things that were previously hard to do at scale.

40
00:01:23,760 --> 00:01:25,200
Not because they were technically hard,

41
00:01:25,200 --> 00:01:26,960
but because they required human friction.

42
00:01:26,960 --> 00:01:28,720
Humans hesitate, humans ask questions,

43
00:01:28,720 --> 00:01:31,400
humans get tired, humans escalate, agents don't,

44
00:01:31,400 --> 00:01:33,840
agents are obedient, they don't resist,

45
00:01:33,840 --> 00:01:36,880
they don't get that gut feeling that something feels off.

46
00:01:36,880 --> 00:01:38,520
They just run the workflow you gave them,

47
00:01:38,520 --> 00:01:40,960
with whatever data and permissions you accidentally handed them

48
00:01:40,960 --> 00:01:44,080
along the way, and that's why the network of agents' narrative

49
00:01:44,080 --> 00:01:45,680
should make you nervous.

50
00:01:45,680 --> 00:01:48,840
The marketing version is coordination, agents collaborating,

51
00:01:48,840 --> 00:01:51,440
handing off tasks, accelerating outcomes.

52
00:01:51,440 --> 00:01:53,240
The operational version is chaining.

53
00:01:53,240 --> 00:01:57,000
Agent A calls tool X, which triggers system Y, which wakes agent B,

54
00:01:57,000 --> 00:01:59,760
which requests approval from a person who clicks approve

55
00:01:59,760 --> 00:02:01,600
because the card looks legitimate,

56
00:02:01,600 --> 00:02:04,280
which then causes agency to run a right operation

57
00:02:04,280 --> 00:02:06,400
that nobody can easily unwind.

58
00:02:06,400 --> 00:02:09,320
Actions, triggers, approval loops, chaining.

59
00:02:09,320 --> 00:02:10,800
This is not a linear flow chart.

60
00:02:10,800 --> 00:02:13,320
It's an authorization graph with automation attached.

61
00:02:13,320 --> 00:02:14,360
Now add scale.

62
00:02:14,360 --> 00:02:15,760
In the research you provided,

63
00:02:15,760 --> 00:02:17,640
the billions of agents' prediction shows up.

64
00:02:17,640 --> 00:02:19,560
Fine, treat that as hype if you want.

65
00:02:19,560 --> 00:02:22,280
Enterprise reality doesn't need billions to fail.

66
00:02:22,280 --> 00:02:23,240
It needs dozens.

67
00:02:23,240 --> 00:02:24,920
It needs one department shipping,

68
00:02:24,920 --> 00:02:27,440
helpful agents faster than anyone can catalog.

69
00:02:27,440 --> 00:02:29,560
It needs one vendor adding a pre-built agent

70
00:02:29,560 --> 00:02:31,880
that quietly brings its own connector strategy.

71
00:02:31,880 --> 00:02:34,440
It needs one maker discovering they can share something

72
00:02:34,440 --> 00:02:35,480
just to my team.

73
00:02:35,480 --> 00:02:37,360
And then it spreads because it's useful.

74
00:02:37,360 --> 00:02:38,920
Entropy doesn't require malice.

75
00:02:38,920 --> 00:02:41,080
It requires convenience.

76
00:02:41,080 --> 00:02:42,560
The foundational mistake is assuming

77
00:02:42,560 --> 00:02:44,400
that an agent is a feature you deploy.

78
00:02:44,400 --> 00:02:46,720
In reality, an agent is a product you operate.

79
00:02:46,720 --> 00:02:49,040
It has an identity surface, a permission surface,

80
00:02:49,040 --> 00:02:52,280
a data surface, a tool surface, and a behavioral surface.

81
00:02:52,280 --> 00:02:54,400
And if you don't define those surfaces deliberately,

82
00:02:54,400 --> 00:02:56,320
the platform will define them accidentally.

83
00:02:56,320 --> 00:02:59,040
Let's break down the assistant versus agent difference

84
00:02:59,040 --> 00:03:01,160
in a way that makes governance obvious.

85
00:03:01,160 --> 00:03:02,960
Assistance mostly create text.

86
00:03:02,960 --> 00:03:05,560
The worst case is misinformation, tone issues,

87
00:03:05,560 --> 00:03:06,760
or bad summary.

88
00:03:06,760 --> 00:03:08,880
That's real damage, but it's usually reversible.

89
00:03:08,880 --> 00:03:10,760
You correct the output, you move on.

90
00:03:10,760 --> 00:03:12,320
Agents create side effects.

91
00:03:12,320 --> 00:03:13,640
Side effects are the point.

92
00:03:13,640 --> 00:03:15,160
They exist to change something.

93
00:03:15,160 --> 00:03:18,160
A ticket, a document state, a sharepoint permission,

94
00:03:18,160 --> 00:03:20,320
a mailbox rule, a team's membership,

95
00:03:20,320 --> 00:03:22,960
and as your resource, a power platform environment,

96
00:03:22,960 --> 00:03:24,320
a purchase request.

97
00:03:24,320 --> 00:03:26,720
Side effects are also what auditors and incident responders

98
00:03:26,720 --> 00:03:29,200
care about because side effects are evidence of authority.

99
00:03:29,200 --> 00:03:31,760
So when leadership asks, our agent's accurate,

100
00:03:31,760 --> 00:03:34,560
then the right answer is accuracy is not the primary risk.

101
00:03:34,560 --> 00:03:37,600
Authority is, nobody gets fired because a bot wrote an awkward

102
00:03:37,600 --> 00:03:38,120
paragraph.

103
00:03:38,120 --> 00:03:40,360
People get fired because the bot did something

104
00:03:40,360 --> 00:03:42,800
and nobody can prove who authorized it.

105
00:03:42,800 --> 00:03:44,840
And this is where the open loop matters.

106
00:03:44,840 --> 00:03:46,640
Everyone wants the agentic decade.

107
00:03:46,640 --> 00:03:48,160
They want a fleet of digital workers.

108
00:03:48,160 --> 00:03:49,160
They want autonomy.

109
00:03:49,160 --> 00:03:50,080
They want scale.

110
00:03:50,080 --> 00:03:52,960
But autonomy without governance doesn't scale intelligence.

111
00:03:52,960 --> 00:03:54,160
It scales ambiguity.

112
00:03:54,160 --> 00:03:56,640
Over time, policies drift away from intent.

113
00:03:56,640 --> 00:03:57,880
Exceptions accumulate.

114
00:03:57,880 --> 00:04:00,320
Ownership changes, service accounts, multiply,

115
00:04:00,320 --> 00:04:01,760
connectors proliferate.

116
00:04:01,760 --> 00:04:04,240
And every single one of those is an entropy generator.

117
00:04:04,240 --> 00:04:06,880
The program keeps moving, but the control plane erodes

118
00:04:06,880 --> 00:04:07,680
underneath it.

119
00:04:07,680 --> 00:04:09,800
So the first job in an agentic strategy

120
00:04:09,800 --> 00:04:11,160
is in building smarter agents.

121
00:04:11,160 --> 00:04:13,800
It's deciding what the enterprise will treat as enforceable

122
00:04:13,800 --> 00:04:14,600
truth.

123
00:04:14,600 --> 00:04:16,320
If something can act, it must have

124
00:04:16,320 --> 00:04:17,000
an identity.

125
00:04:17,000 --> 00:04:20,160
If it can call a tool, the tool must be governed as infrastructure.

126
00:04:20,160 --> 00:04:23,280
If it can see data, the boundary must be defined and monitored.

127
00:04:23,280 --> 00:04:26,760
If it behaves strangely, containment must be fast and attributable.

128
00:04:26,760 --> 00:04:28,120
That's the mental model shift.

129
00:04:28,120 --> 00:04:29,280
Agents aren't assistants.

130
00:04:29,280 --> 00:04:32,040
They are automated actors inside your control plane.

131
00:04:32,040 --> 00:04:35,400
Next, define what agent sprawl actually means without the marketing

132
00:04:35,400 --> 00:04:38,720
fog and why it always shows up first as identity drift.

133
00:04:38,720 --> 00:04:41,760
Define agent sprawl without the marketing fog.

134
00:04:41,760 --> 00:04:44,160
Agent sprawl isn't, we have a lot of agents.

135
00:04:44,160 --> 00:04:45,640
That's the poster version.

136
00:04:45,640 --> 00:04:47,520
The operational version is uglier.

137
00:04:47,520 --> 00:04:50,960
Uncontrolled growth across six things that are hard to inventory

138
00:04:50,960 --> 00:04:52,320
and even harder to reverse.

139
00:04:52,320 --> 00:04:55,400
Identities, tools, prompts, permissions, owners, versions.

140
00:04:55,400 --> 00:04:57,840
If you can't name those six for every agent,

141
00:04:57,840 --> 00:04:59,280
you don't have an ecosystem.

142
00:04:59,280 --> 00:05:00,000
You have a rumor.

143
00:05:00,000 --> 00:05:01,160
Start with identity.

144
00:05:01,160 --> 00:05:04,360
Every agent that can act needs some form of authentication path.

145
00:05:04,360 --> 00:05:07,720
If the enterprise can't point to a stable, non-human identity

146
00:05:07,720 --> 00:05:11,520
for that agent, then every action it takes becomes a debate later.

147
00:05:11,520 --> 00:05:13,880
Identity sprawl is the first crack in the foundation,

148
00:05:13,880 --> 00:05:17,680
because everything else, tool calls, audits, containment,

149
00:05:17,680 --> 00:05:19,960
hangs off it, then tools.

150
00:05:19,960 --> 00:05:22,000
Agents don't exist as pure language.

151
00:05:22,000 --> 00:05:24,480
They exist as a bundle of tool invocations,

152
00:05:24,480 --> 00:05:27,080
graph calls, connector actions, custom APIs,

153
00:05:27,080 --> 00:05:30,160
MCP tools, flows, web retrieval, code interpreter,

154
00:05:30,160 --> 00:05:32,240
whatever the platform exposes this week.

155
00:05:32,240 --> 00:05:33,880
Tools sprawl means you stop knowing

156
00:05:33,880 --> 00:05:36,240
what capabilities are available, where they're reused,

157
00:05:36,240 --> 00:05:38,160
and which ones became the unofficial standard

158
00:05:38,160 --> 00:05:39,840
because someone shipped them first.

159
00:05:39,840 --> 00:05:41,240
Then prompts and instructions.

160
00:05:41,240 --> 00:05:43,600
People treat prompts like documentation they can.

161
00:05:43,600 --> 00:05:44,440
They aren't.

162
00:05:44,440 --> 00:05:45,760
Prompts are logic.

163
00:05:45,760 --> 00:05:47,880
They are policy expressed as text,

164
00:05:47,880 --> 00:05:50,000
usually without change control.

165
00:05:50,000 --> 00:05:52,160
Prompts sprawl means the enterprise ends up

166
00:05:52,160 --> 00:05:54,120
with dozens of slightly different interpretations

167
00:05:54,120 --> 00:05:55,040
of the same rule.

168
00:05:55,040 --> 00:05:57,840
What counts as sensitive, what counts as approval worthy,

169
00:05:57,840 --> 00:06:01,160
what counts as safe to share, what counts as authoritative.

170
00:06:01,160 --> 00:06:02,360
Drift isn't a bug here.

171
00:06:02,360 --> 00:06:03,720
Drift is the default behavior.

172
00:06:03,720 --> 00:06:04,560
Then permissions.

173
00:06:04,560 --> 00:06:06,480
This is where sprawl becomes expensive.

174
00:06:06,480 --> 00:06:09,040
Agents get access through the same pathways humans do.

175
00:06:09,040 --> 00:06:11,920
App permissions, delegated access, connector credentials,

176
00:06:11,920 --> 00:06:14,680
service principles, environment roles, sharepoint permissions,

177
00:06:14,680 --> 00:06:17,080
and the quiet, just give it contributor for now,

178
00:06:17,080 --> 00:06:19,200
moments that never get revisited.

179
00:06:19,200 --> 00:06:21,840
Permission sprawl is how a deterministic security model

180
00:06:21,840 --> 00:06:22,960
becomes probabilistic.

181
00:06:22,960 --> 00:06:24,040
One exception doesn't hurt.

182
00:06:24,040 --> 00:06:26,320
A thousand exceptions becomes policy theater.

183
00:06:26,320 --> 00:06:27,160
Then owners.

184
00:06:27,160 --> 00:06:28,360
Someone has to be accountable.

185
00:06:28,360 --> 00:06:29,560
Not the maker who built it,

186
00:06:29,560 --> 00:06:31,960
but the sponsor who owns the business outcome,

187
00:06:31,960 --> 00:06:33,880
the risk and the life cycle.

188
00:06:33,880 --> 00:06:35,520
Owners sprawl looks like this.

189
00:06:35,520 --> 00:06:37,400
The agent is shared, it becomes popular,

190
00:06:37,400 --> 00:06:38,800
the original maker changes roles,

191
00:06:38,800 --> 00:06:41,200
and suddenly nobody can answer a basic question like,

192
00:06:41,200 --> 00:06:43,000
who approves changes to this thing?

193
00:06:43,000 --> 00:06:44,400
The agent stays live anyway.

194
00:06:44,400 --> 00:06:45,080
Of course it does.

195
00:06:45,080 --> 00:06:46,080
And finally versions.

196
00:06:46,080 --> 00:06:47,240
Every agent evolves.

197
00:06:47,240 --> 00:06:49,440
New instructions, new tools, new knowledge sources,

198
00:06:49,440 --> 00:06:51,360
new connectors, new models, new guardrails.

199
00:06:51,360 --> 00:06:54,800
Version sprawl is what turns troubleshooting into archaeology.

200
00:06:54,800 --> 00:06:56,440
You stop debugging behavior,

201
00:06:56,440 --> 00:06:59,760
you start guessing which variant of the agent a user even hit.

202
00:06:59,760 --> 00:07:02,280
Now how does this sprawl actually enter the tenant?

203
00:07:02,280 --> 00:07:04,560
It usually arrives through three vectors,

204
00:07:04,560 --> 00:07:06,280
and all three feel legitimate.

205
00:07:06,280 --> 00:07:08,920
Vector1 is maker led that I built a helpful agent

206
00:07:08,920 --> 00:07:09,840
for my team story.

207
00:07:09,840 --> 00:07:11,640
It starts personal, then departmental,

208
00:07:11,640 --> 00:07:12,880
then someone forwards the link,

209
00:07:12,880 --> 00:07:14,200
then it becomes semi-official.

210
00:07:14,200 --> 00:07:15,880
And suddenly it's in a critical workflow

211
00:07:15,880 --> 00:07:17,160
without a review gate.

212
00:07:17,160 --> 00:07:19,880
Maker led sprawl is fast because the platform is designed

213
00:07:19,880 --> 00:07:20,600
to be easy.

214
00:07:20,600 --> 00:07:22,160
Vector2 is vendor-provided.

215
00:07:22,160 --> 00:07:23,840
Microsoft ships agents.

216
00:07:23,840 --> 00:07:25,480
Third parties ship agents.

217
00:07:25,480 --> 00:07:27,480
Internal platform teams ship starter agents.

218
00:07:27,480 --> 00:07:29,320
They're pre-built, they demo well,

219
00:07:29,320 --> 00:07:31,680
and they often arrive with assumptions baked in.

220
00:07:31,680 --> 00:07:34,640
Permissions, connectors, data paths, and tool contracts.

221
00:07:34,640 --> 00:07:37,320
Vender led sprawl is dangerous because it feels sanctioned.

222
00:07:37,320 --> 00:07:38,880
People stop asking what it can do

223
00:07:38,880 --> 00:07:40,800
and start asking why they can't have it.

224
00:07:40,800 --> 00:07:43,800
Vector3 is marketplace and third party integrations.

225
00:07:43,800 --> 00:07:45,560
MCP makes tool on boarding easier.

226
00:07:45,560 --> 00:07:48,240
That's the point, but the other point nobody wants to admit

227
00:07:48,240 --> 00:07:50,680
is that easier on boarding also means easier propagation

228
00:07:50,680 --> 00:07:51,440
of mistakes.

229
00:07:51,440 --> 00:07:53,240
One bad tool contract doesn't stay local.

230
00:07:53,240 --> 00:07:54,440
It becomes reusable.

231
00:07:54,440 --> 00:07:55,960
It spreads.

232
00:07:55,960 --> 00:07:59,160
And here's the hidden multiplier most organizations miss.

233
00:07:59,160 --> 00:08:02,120
Every agent adds surface area across Microsoft 365,

234
00:08:02,120 --> 00:08:04,640
Power Platform, and Azure at the same time.

235
00:08:04,640 --> 00:08:06,280
This isn't three separate platforms.

236
00:08:06,280 --> 00:08:09,600
It's one connected control plane with different admin centers

237
00:08:09,600 --> 00:08:11,480
pretending to be separate worlds.

238
00:08:11,480 --> 00:08:14,200
So agent sprawl doesn't just grow your AI footprint.

239
00:08:14,200 --> 00:08:16,600
It grows your identity footprint, your connector footprint,

240
00:08:16,600 --> 00:08:18,920
your audit footprint, and your incident footprint.

241
00:08:18,920 --> 00:08:20,520
You'll know your in sprawl early

242
00:08:20,520 --> 00:08:22,800
when a new question shows up in meetings.

243
00:08:22,800 --> 00:08:24,720
Which agent should I use?

244
00:08:24,720 --> 00:08:25,600
That sounds harmless.

245
00:08:25,600 --> 00:08:26,440
It isn't.

246
00:08:26,440 --> 00:08:27,440
It's a productivity tax.

247
00:08:27,440 --> 00:08:29,920
It means discoverability is failing, duplication is happening,

248
00:08:29,920 --> 00:08:31,080
and trust is fragmenting.

249
00:08:31,080 --> 00:08:33,400
People stop using the ecosystem because it's noisy,

250
00:08:33,400 --> 00:08:35,280
inconsistent, and unpredictable.

251
00:08:35,280 --> 00:08:37,600
And once that happens, the next failure mode

252
00:08:37,600 --> 00:08:39,880
arrives on schedule, identity drift.

253
00:08:39,880 --> 00:08:41,640
Not because you meant to lose accountability,

254
00:08:41,640 --> 00:08:43,520
but because you scaled agents faster

255
00:08:43,520 --> 00:08:44,920
than you scaled attribution.

256
00:08:44,920 --> 00:08:46,280
That distinction matters.

257
00:08:46,280 --> 00:08:49,840
Risk one, identity drift, when accountability collapses.

258
00:08:49,840 --> 00:08:52,760
Identity drift is what happens when an agent can take action,

259
00:08:52,760 --> 00:08:54,760
but the enterprise can't reliably prove

260
00:08:54,760 --> 00:08:57,400
which non-human identity perform that action

261
00:08:57,400 --> 00:08:59,560
under what authority and with whose approval

262
00:08:59,560 --> 00:09:01,560
and the collapse is silent at first.

263
00:09:01,560 --> 00:09:03,200
Because everything still works.

264
00:09:03,200 --> 00:09:04,640
Tickets still get created.

265
00:09:04,640 --> 00:09:05,720
Files still get moved.

266
00:09:05,720 --> 00:09:06,960
Permissions still get granted.

267
00:09:06,960 --> 00:09:08,120
People still get answers.

268
00:09:08,120 --> 00:09:10,960
The system stays productive right up until the moment someone

269
00:09:10,960 --> 00:09:13,640
asks the one question you can't afford to fumble.

270
00:09:13,640 --> 00:09:14,280
Who did this?

271
00:09:14,280 --> 00:09:15,800
Not which user asked the question

272
00:09:15,800 --> 00:09:18,200
and not which department owns the workflow.

273
00:09:18,200 --> 00:09:20,600
Auditors don't care about your intent story.

274
00:09:20,600 --> 00:09:23,240
They care about attribution, a stable identity,

275
00:09:23,240 --> 00:09:25,600
a permission chain, and an evidence trail.

276
00:09:25,600 --> 00:09:28,160
Identity drift shows up when actions get executed

277
00:09:28,160 --> 00:09:29,880
through anonymous pathways.

278
00:09:29,880 --> 00:09:31,120
Shared credentials.

279
00:09:31,120 --> 00:09:34,040
Maker connections, inherited delegated permissions,

280
00:09:34,040 --> 00:09:37,640
or tool calls that log as some generic application context

281
00:09:37,640 --> 00:09:41,120
that nobody can map back to an accountable sponsor.

282
00:09:41,120 --> 00:09:43,920
Over time, the organization stops seeing agent actions

283
00:09:43,920 --> 00:09:44,440
as actions.

284
00:09:44,440 --> 00:09:46,640
They start seeing them as automation.

285
00:09:46,640 --> 00:09:48,600
And automation historically has been treated

286
00:09:48,600 --> 00:09:49,760
like background noise.

287
00:09:49,760 --> 00:09:50,720
That's a mistake.

288
00:09:50,720 --> 00:09:53,280
Because the risk isn't that the agent makes a wrong decision.

289
00:09:53,280 --> 00:09:55,080
The risk is that the enterprise can't prove

290
00:09:55,080 --> 00:09:56,440
it made any decision at all.

291
00:09:56,440 --> 00:09:59,520
Here's what incident response looks like under identity drift.

292
00:09:59,520 --> 00:10:01,640
Something changes in SharePoint permissions.

293
00:10:01,640 --> 00:10:02,640
A mailbox rule appears.

294
00:10:02,640 --> 00:10:03,920
A team's membership changes.

295
00:10:03,920 --> 00:10:05,040
A record gets updated.

296
00:10:05,040 --> 00:10:06,880
The user says, I didn't do that.

297
00:10:06,880 --> 00:10:08,520
The admin says, we have logs.

298
00:10:08,520 --> 00:10:10,800
The security team pulls the event, and it points

299
00:10:10,800 --> 00:10:12,640
to a service account, or a connector,

300
00:10:12,640 --> 00:10:15,760
or a vague app identity used by five different agents,

301
00:10:15,760 --> 00:10:19,560
owned by nobody, sponsored by nobody, and documented nowhere.

302
00:10:19,560 --> 00:10:22,720
Now, the organization is no longer investigating a technical event.

303
00:10:22,720 --> 00:10:25,920
It's negotiating a narrative, and narratives don't survive audits.

304
00:10:25,920 --> 00:10:27,600
Auditors ask three basic questions.

305
00:10:27,600 --> 00:10:29,960
And identity drift makes all three painful.

306
00:10:29,960 --> 00:10:31,320
First, who perform the action?

307
00:10:31,320 --> 00:10:34,040
Second, under what authority did they have permission?

308
00:10:34,040 --> 00:10:36,440
Third, who approved that authority and when?

309
00:10:36,440 --> 00:10:39,560
If your answer sounds like, well, it was probably the service desk agent,

310
00:10:39,560 --> 00:10:40,480
you don't have evidence.

311
00:10:40,480 --> 00:10:41,280
You have suspicion.

312
00:10:41,280 --> 00:10:44,480
And the moment the business hears, probably, the program freezes.

313
00:10:44,480 --> 00:10:47,280
Because that's the real consequence of identity drift.

314
00:10:47,280 --> 00:10:51,120
One visible incident can pause the entire agent roll out across the company.

315
00:10:51,120 --> 00:10:53,160
The patents are depressingly consistent.

316
00:10:53,160 --> 00:10:55,960
Share credentials are the classic entropy generator.

317
00:10:55,960 --> 00:10:57,720
Someone creates a bot account.

318
00:10:57,720 --> 00:11:00,080
Shares it with a few makers and calls it progress.

319
00:11:00,080 --> 00:11:01,960
It isn't its identity debt.

320
00:11:01,960 --> 00:11:03,400
Maker connections are the next one.

321
00:11:03,400 --> 00:11:06,080
An agent gets built using a person's delegated access,

322
00:11:06,080 --> 00:11:08,200
then shared, then used by others.

323
00:11:08,200 --> 00:11:10,520
The agent now acts with the makers shadow authority

324
00:11:10,520 --> 00:11:14,480
until the day that maker leaves, changes roles, or gets their access removed.

325
00:11:14,480 --> 00:11:18,680
Then the agent breaks, or worse, it keeps working in partial, unpredictable ways.

326
00:11:18,680 --> 00:11:21,040
Then there's the service principle everywhere, anti-pattern.

327
00:11:21,040 --> 00:11:23,760
A single app registration ends up powering multiple agents,

328
00:11:23,760 --> 00:11:25,480
multiple tools, multiple environments.

329
00:11:25,480 --> 00:11:26,800
It looks neat in a spreadsheet.

330
00:11:26,800 --> 00:11:29,280
In reality, it's a blast radius amplifier.

331
00:11:29,280 --> 00:11:32,680
One compromise, one mis-scoped permission, one forgotten secret rotation,

332
00:11:32,680 --> 00:11:35,320
and you've attached a privileged identity to a fleet.

333
00:11:35,320 --> 00:11:37,560
And the weirdest one is anonymous tool calls,

334
00:11:37,560 --> 00:11:40,760
not anonymous in the literal sense, everything logs somewhere,

335
00:11:40,760 --> 00:11:42,720
but anonymous in the governance sense.

336
00:11:42,720 --> 00:11:46,920
The logs don't map cleanly to an owned, life cycle managed agent identity.

337
00:11:46,920 --> 00:11:49,680
You can't answer basic questions like which agent invoked,

338
00:11:49,680 --> 00:11:53,200
which tool on behalf of what workflow, with what approval state.

339
00:11:53,200 --> 00:11:55,760
So the operational consequence isn't theoretical.

340
00:11:55,760 --> 00:11:57,640
It's immediate. You can't do containment.

341
00:11:57,640 --> 00:12:00,680
Because you can't decide what to disable without collateral damage.

342
00:12:00,680 --> 00:12:04,880
If ten agents share one identity, blocking the identity blocks ten business workflows,

343
00:12:04,880 --> 00:12:06,320
you won't do it. You'll hesitate.

344
00:12:06,320 --> 00:12:09,440
And that hesitation is how incidents become outages.

345
00:12:09,440 --> 00:12:12,920
This is why identity drift is the first catastrophic failure mode.

346
00:12:12,920 --> 00:12:14,680
Not because identity is important,

347
00:12:14,680 --> 00:12:18,240
because identity is the anchor that makes the rest of governance possible.

348
00:12:18,240 --> 00:12:21,440
No identity means no least privilege enforcement that holds over time.

349
00:12:21,440 --> 00:12:24,440
No identity means no reliable audit trail.

350
00:12:24,440 --> 00:12:28,520
No identity means defender detections that can't be tied to a specific actor.

351
00:12:28,520 --> 00:12:32,920
No identity means purview exposure reports that can't be attributed to responsible owner.

352
00:12:32,920 --> 00:12:36,520
In other words, you don't lose accountability in a dramatic moment.

353
00:12:36,520 --> 00:12:38,880
You lose it one convenient shortcut at a time.

354
00:12:38,880 --> 00:12:40,800
And when the shortcut becomes normal,

355
00:12:40,800 --> 00:12:43,120
the organization crosses an invisible line.

356
00:12:43,120 --> 00:12:47,040
It moves from a deterministic security model to a probabilistic one.

357
00:12:47,040 --> 00:12:48,520
That distinction matters.

358
00:12:48,520 --> 00:12:52,440
Next, map identity drift to Microsoft's enforcement mechanism,

359
00:12:52,440 --> 00:12:58,400
Entraagent ID, and why Entra needs to be treated as a distributed decision engine, not a directory.

360
00:12:58,400 --> 00:13:02,000
Control plane layer one, Entraagent ID as the audit anchor.

361
00:13:02,000 --> 00:13:04,320
Entraagent ID exists for one reason,

362
00:13:04,320 --> 00:13:08,280
to stop identity drift from becoming your default operating model.

363
00:13:08,280 --> 00:13:10,440
The rule is simple and it's non-negotiable at scale.

364
00:13:10,440 --> 00:13:14,200
If it can act, it must have an identity, not a shared user, not whoever built it,

365
00:13:14,200 --> 00:13:15,800
not the connector account.

366
00:13:15,800 --> 00:13:20,840
An actual non-human identity that the enterprise can point to scope, monitor and retire,

367
00:13:20,840 --> 00:13:24,760
because the moment an agent can call tools, it is no longer a chat experience.

368
00:13:24,760 --> 00:13:26,400
It is an actor in your environment.

369
00:13:26,400 --> 00:13:30,400
Actors need identities, identities need ownership, ownership needs life cycle.

370
00:13:30,400 --> 00:13:32,320
That distinction matters.

371
00:13:32,320 --> 00:13:34,760
Most organizations treat Entra like a directory,

372
00:13:34,760 --> 00:13:38,800
uses groups, apps, and the occasional service principle nobody wants to own.

373
00:13:38,800 --> 00:13:40,120
That's the comfortable framing.

374
00:13:40,120 --> 00:13:41,960
It's also incomplete in architectural terms.

375
00:13:41,960 --> 00:13:43,720
Entra is a distributed decision engine.

376
00:13:43,720 --> 00:13:47,240
It sits in the middle of every authorization path continuously,

377
00:13:47,240 --> 00:13:50,520
and it enforces whatever you actually configured, not what you intended.

378
00:13:50,520 --> 00:13:54,000
And this is why Entra Agent ID is not another object in Entra.

379
00:13:54,000 --> 00:13:58,600
It is the ordered anchor that lets you build an evidence chain around agent behavior.

380
00:13:58,600 --> 00:14:00,640
When an agent has a stable identity,

381
00:14:00,640 --> 00:14:04,560
the enterprise can answer the questions that stop programs from getting shut down,

382
00:14:04,560 --> 00:14:07,360
which agent did the action, which permissions enabled it,

383
00:14:07,360 --> 00:14:13,200
which tool invocation executed it, and which human approved it, if approval was required.

384
00:14:13,200 --> 00:14:16,720
Without that identity, every downstream control becomes hand-wavy.

385
00:14:16,720 --> 00:14:20,080
So what changes when you make Entra Agent ID a first principle?

386
00:14:20,080 --> 00:14:22,440
First, least privilege stops being optional.

387
00:14:22,440 --> 00:14:24,960
At small scale, people tolerate sloppy permissions

388
00:14:24,960 --> 00:14:27,360
because the blast radius feels contained.

389
00:14:27,360 --> 00:14:32,040
At agent scale, sloppy permissions turn into repeated automation of your worst assumptions.

390
00:14:32,040 --> 00:14:34,560
The agent doesn't occasionally overreach.

391
00:14:34,560 --> 00:14:37,560
It overreaches consistently.

392
00:14:37,560 --> 00:14:41,680
Entra Agent ID forces you to define permission boundaries explicitly.

393
00:14:41,680 --> 00:14:45,040
What this agent can read, what it can write, and what it can never touch.

394
00:14:45,040 --> 00:14:48,040
And when teams ask for exceptions, you can finally see what they're doing.

395
00:14:48,040 --> 00:14:49,640
They're not requesting convenience.

396
00:14:49,640 --> 00:14:51,720
They're manufacturing an entropy generator.

397
00:14:51,720 --> 00:14:56,480
Second, accountability becomes a configured property, not a social agreement.

398
00:14:56,480 --> 00:14:58,920
Every agent identity needs an owner and a sponsor.

399
00:14:58,920 --> 00:15:00,960
The owner handles the build and maintenance.

400
00:15:00,960 --> 00:15:04,520
The sponsor owns the business outcome, risk acceptance and funding.

401
00:15:04,520 --> 00:15:07,800
Those are different roles, and enterprises keep pretending they aren't.

402
00:15:07,800 --> 00:15:09,600
This is where life cycle workflows matter.

403
00:15:09,600 --> 00:15:13,240
If the sponsor leaves, the agent can't be allowed to drift into orphan status.

404
00:15:13,240 --> 00:15:16,560
Often agents don't retire, they rot, they keep their permissions.

405
00:15:16,560 --> 00:15:18,920
They keep being used, they keep producing incidents

406
00:15:18,920 --> 00:15:20,160
that nobody can explain.

407
00:15:20,160 --> 00:15:24,480
So the identity object needs life cycle controls, creation gate, periodic review,

408
00:15:24,480 --> 00:15:26,400
reattestation and deprecation.

409
00:15:26,400 --> 00:15:29,760
If you don't enforce that, agent products become zombie services.

410
00:15:29,760 --> 00:15:32,520
Third, incident response becomes containment, not politics.

411
00:15:32,520 --> 00:15:35,280
With Entra Agent ID, you can disable the agent identity

412
00:15:35,280 --> 00:15:37,240
without disabling the entire program.

413
00:15:37,240 --> 00:15:38,320
That sounds small.

414
00:15:38,320 --> 00:15:39,160
It isn't.

415
00:15:39,160 --> 00:15:42,200
That is the difference between, we contain the incident,

416
00:15:42,200 --> 00:15:45,160
and we paused all agents until further notice.

417
00:15:45,160 --> 00:15:48,480
Executives don't fund programs that can't be surgically contained.

418
00:15:48,480 --> 00:15:51,240
And now the subtle but critical part, Entra Agent ID

419
00:15:51,240 --> 00:15:52,760
doesn't magically secure agents.

420
00:15:52,760 --> 00:15:54,920
It makes security enforceable.

421
00:15:54,920 --> 00:15:59,440
It's the anchor that lets purview exposure map to a specific actor.

422
00:15:59,440 --> 00:16:03,600
It's the anchor that lets defender detections turn into an action you can take.

423
00:16:03,600 --> 00:16:06,400
It's the anchor that makes tool usage attributable.

424
00:16:06,400 --> 00:16:08,360
Identity is not the whole control plane.

425
00:16:08,360 --> 00:16:10,640
Identity is the thing the other layers attach to.

426
00:16:10,640 --> 00:16:11,600
Here's the micro story.

427
00:16:11,600 --> 00:16:15,280
Every audit team recognizes a service desk agent updates a ticket queue.

428
00:16:15,280 --> 00:16:18,440
A week later, someone sees a set of tickets reassigned incorrectly.

429
00:16:18,440 --> 00:16:21,520
Under Identity Drift, you get, we think the bot did it.

430
00:16:21,520 --> 00:16:23,400
Under Entra Agent ID, you get,

431
00:16:23,400 --> 00:16:25,840
Agent Identity X, Tool Action Y,

432
00:16:25,840 --> 00:16:29,560
Executed at time Z, approved by Person A, using permission set B.

433
00:16:29,560 --> 00:16:30,640
That's not a better story.

434
00:16:30,640 --> 00:16:31,400
That's evidence.

435
00:16:31,400 --> 00:16:33,240
And evidence is what keeps the program alive.

436
00:16:33,240 --> 00:16:36,440
So, treat Entra Agent ID as the starting line, not the finish line.

437
00:16:36,440 --> 00:16:37,880
It gives you attribution.

438
00:16:37,880 --> 00:16:38,800
It gives you boundaries.

439
00:16:38,800 --> 00:16:40,280
It gives you life cycle hooks.

440
00:16:40,280 --> 00:16:44,200
It gives you the ability to disable one actor without burning the whole village down.

441
00:16:44,200 --> 00:16:46,480
But Identity alone doesn't prevent damage.

442
00:16:46,480 --> 00:16:48,080
Identity tells you who acted.

443
00:16:48,080 --> 00:16:49,760
It doesn't limit what happens when they act.

444
00:16:49,760 --> 00:16:51,040
That's the next layer.

445
00:16:51,040 --> 00:16:53,480
And it's where most enterprises lose control.

446
00:16:53,480 --> 00:16:56,520
Tools define blast radius.

447
00:16:56,520 --> 00:17:00,760
Next, MCP as the tool contract that turns integrations into infrastructure

448
00:17:00,760 --> 00:17:04,960
and why standardization can either save you or amplify mistakes at scale.

449
00:17:04,960 --> 00:17:08,280
Risk 2, data leakage via grounding plus tools.

450
00:17:08,280 --> 00:17:12,680
Data leakage in agent systems rarely looks like an attacker X-filterating secrets.

451
00:17:12,680 --> 00:17:16,560
It looks like a helpful workflow doing exactly what it was configured to do

452
00:17:16,560 --> 00:17:20,720
with exactly the access it was granted using exactly the data it was allowed to retrieve.

453
00:17:20,720 --> 00:17:22,640
Agents don't leak maliciously.

454
00:17:22,640 --> 00:17:24,240
They leak obediently.

455
00:17:24,240 --> 00:17:28,120
That distinction matters because it changes the post-incident conversation.

456
00:17:28,120 --> 00:17:31,440
If the organization believes leakage is an intent problem,

457
00:17:31,440 --> 00:17:36,080
it will keep buying more training, more awareness and more policies written in PowerPoint.

458
00:17:36,080 --> 00:17:37,040
The system won't care.

459
00:17:37,040 --> 00:17:39,760
The system will keep following the configured pathways.

460
00:17:39,760 --> 00:17:43,280
Leakage happens at the boundary between grounding and tools.

461
00:17:43,280 --> 00:17:47,960
Grounding is where the agent pulls context. Files, emails, chats, SharePoint pages,

462
00:17:47,960 --> 00:17:52,240
confluence, tickets, CRM records, the knowledge base, someone's wall was clean.

463
00:17:52,240 --> 00:17:56,560
Tools are where the agent takes action, exporting, summarizing, sending, posting,

464
00:17:56,560 --> 00:17:58,160
creating, updating, attaching.

465
00:17:58,160 --> 00:18:01,720
If you overscope either one, the agent becomes a high-speed compliance incident.

466
00:18:01,720 --> 00:18:05,160
Here are the common leakage pathways and none of them require someone to be evil.

467
00:18:05,160 --> 00:18:07,600
First, overbroad retrieval.

468
00:18:07,600 --> 00:18:11,240
Someone grounds an agent on the whole department SharePoint because it's convenient

469
00:18:11,240 --> 00:18:14,360
or all of team's messages because the agent feels smarter.

470
00:18:14,360 --> 00:18:16,920
Or they use a connector that returns more than the task needs

471
00:18:16,920 --> 00:18:19,440
because nobody enforced a minimal data set contract.

472
00:18:19,440 --> 00:18:24,240
The agent now has access to a blended pool of data with different sensitivity levels,

473
00:18:24,240 --> 00:18:26,720
different retention rules and different audiences.

474
00:18:26,720 --> 00:18:28,400
Then a user asks an innocent question.

475
00:18:28,400 --> 00:18:29,880
The agent answers correctly.

476
00:18:29,880 --> 00:18:33,680
And the correct answer contains something that was never meant to be surfaced in that context.

477
00:18:33,680 --> 00:18:35,080
The agent didn't hack anything.

478
00:18:35,080 --> 00:18:36,960
It just retrieved what you let it retrieve.

479
00:18:36,960 --> 00:18:39,480
Second, cross-domain context reuse.

480
00:18:39,480 --> 00:18:40,720
This is the one people miss.

481
00:18:40,720 --> 00:18:43,840
Agents operate in loops and they maintain state across steps.

482
00:18:43,840 --> 00:18:47,520
That means content retrieved for one part of a workflow can bleed into another part,

483
00:18:47,520 --> 00:18:50,400
especially when you chain agents or reuse tool outputs.

484
00:18:50,400 --> 00:18:52,720
A procurement agent retrieves a contract clause.

485
00:18:52,720 --> 00:18:56,400
A separate right-the-email agent gets handed that clause as context.

486
00:18:56,400 --> 00:19:00,040
The email agent posts it into a team's channel that includes external guests

487
00:19:00,040 --> 00:19:02,200
because that's where requests are handled.

488
00:19:02,200 --> 00:19:04,040
Everyone claims they followed policy.

489
00:19:04,040 --> 00:19:05,240
The system still leaked.

490
00:19:05,240 --> 00:19:06,320
The leak wasn't the model.

491
00:19:06,320 --> 00:19:08,000
The leak was the choreography.

492
00:19:08,000 --> 00:19:11,360
Third, tool outputs get treated as safe artifacts.

493
00:19:11,360 --> 00:19:13,000
Tools produce outputs.

494
00:19:13,000 --> 00:19:17,600
Summaries, exports, spreadsheets, tickets, cards, notifications, file attachments.

495
00:19:17,600 --> 00:19:20,520
Those outputs then get reused as inputs elsewhere.

496
00:19:20,520 --> 00:19:22,600
If you don't carry provenance and sensitivity forward,

497
00:19:22,600 --> 00:19:26,120
the system quietly launders restricted information into unrestricted channels.

498
00:19:26,120 --> 00:19:29,360
This is where a helpful agent design turns into helpful breach design

499
00:19:29,360 --> 00:19:32,040
because the agent can do the dangerous thing with a polite tone.

500
00:19:32,040 --> 00:19:35,320
Now the executive level trap in all of this is what happens after the incident.

501
00:19:35,320 --> 00:19:37,200
Leadership asks, why did the agent do that?

502
00:19:37,200 --> 00:19:39,840
And the honest answer is because it worked as designed.

503
00:19:39,840 --> 00:19:41,320
That's when trust collapses.

504
00:19:41,320 --> 00:19:42,840
Not because the incident happened.

505
00:19:42,840 --> 00:19:45,400
Incidents always happen, but because the enterprise realizes

506
00:19:45,400 --> 00:19:47,920
it didn't actually design enforceable boundaries.

507
00:19:47,920 --> 00:19:49,280
It designed aspirations.

508
00:19:49,280 --> 00:19:51,080
And aspirations don't survive scale.

509
00:19:51,080 --> 00:19:53,920
The most uncomfortable version of this is when the agent answers

510
00:19:53,920 --> 00:19:55,400
from the wrong policy source.

511
00:19:55,400 --> 00:19:56,200
Not a wrong answer.

512
00:19:56,200 --> 00:19:58,560
A correct answer from a non-authoritative source.

513
00:19:58,560 --> 00:20:02,440
This happens constantly in enterprises because available and authoritative

514
00:20:02,440 --> 00:20:03,680
are not the same thing.

515
00:20:03,680 --> 00:20:07,000
The agent retrieves the first relevant looking document it can access.

516
00:20:07,000 --> 00:20:10,640
That document might be outdated, it might be a draft, it might be a local variant,

517
00:20:10,640 --> 00:20:13,520
it might be a wiki page, someone updated during a fire drill,

518
00:20:13,520 --> 00:20:17,160
it might be a screenshot of a policy pasted into a team's message.

519
00:20:17,160 --> 00:20:19,320
The agent then gives the answer with confidence.

520
00:20:19,320 --> 00:20:22,560
The user acts.compliance inherits the blast radius.

521
00:20:22,560 --> 00:20:26,360
And when you investigate, you discover the agent never violated access controls.

522
00:20:26,360 --> 00:20:28,760
It had access, it retrieved, it responded.

523
00:20:28,760 --> 00:20:30,000
The failure was governance.

524
00:20:30,000 --> 00:20:33,640
You never told the system which sources are allowed to be truth.

525
00:20:33,640 --> 00:20:36,520
So when executives ask, how do we stop data leakage?

526
00:20:36,520 --> 00:20:39,080
The correct answer is not make the model safer.

527
00:20:39,080 --> 00:20:43,000
The correct answer is enforce the data boundary before retrieval

528
00:20:43,000 --> 00:20:47,080
and enforce the tool boundary before action and keep provenance attached to outputs.

529
00:20:47,080 --> 00:20:49,960
If you don't, you're just accelerating the wrong behavior.

530
00:20:49,960 --> 00:20:51,960
And this is why a tool contract matters.

531
00:20:51,960 --> 00:20:56,160
Because bespoke connectors and one off integrations are where boundary enforcement dies.

532
00:20:56,160 --> 00:20:59,120
Every custom integration expresses its own assumptions.

533
00:20:59,120 --> 00:21:00,840
Every shortcut creates a new pathway.

534
00:21:00,840 --> 00:21:02,920
Every pathway becomes a leak candidate.

535
00:21:02,920 --> 00:21:05,560
A standardized tool contract doesn't eliminate mistakes.

536
00:21:05,560 --> 00:21:09,320
It makes mistakes visible, reviewable, and reusable in a controlled way.

537
00:21:09,320 --> 00:21:10,320
That's the transition.

538
00:21:10,320 --> 00:21:12,320
Identity tells you who acted.

539
00:21:12,320 --> 00:21:14,080
But tools determine what they can break.

540
00:21:14,080 --> 00:21:16,960
Next, MCP, reframed as a tool contract.

541
00:21:16,960 --> 00:21:20,560
Integration becomes infrastructure and infrastructure is the only thing that scales

542
00:21:20,560 --> 00:21:22,520
without becoming conditional chaos.

543
00:21:22,520 --> 00:21:24,280
Control plane layer 2.

544
00:21:24,280 --> 00:21:26,240
MCP as the tool contract.

545
00:21:26,240 --> 00:21:28,440
MCP is not a new connector type.

546
00:21:28,440 --> 00:21:29,760
That's the comfortable framing.

547
00:21:29,760 --> 00:21:32,880
In architectural terms, MCP is a tool contract,

548
00:21:32,880 --> 00:21:38,000
a standardized way for agents to discover capabilities, call them, and receive structured results.

549
00:21:38,000 --> 00:21:41,440
And that matters because tools' brawl is where governance goes to die.

550
00:21:41,440 --> 00:21:44,160
Every bespoke integration is a new snowflake.

551
00:21:44,160 --> 00:21:45,160
Different auth.

552
00:21:45,160 --> 00:21:46,160
Different logging.

553
00:21:46,160 --> 00:21:47,480
Different error behavior.

554
00:21:47,480 --> 00:21:49,600
Different assumptions about what safe means.

555
00:21:49,600 --> 00:21:51,440
At 10 agents, you can survive that.

556
00:21:51,440 --> 00:21:52,440
At 100, you can't.

557
00:21:52,440 --> 00:21:55,160
At 1000, you're not operating an ecosystem anymore.

558
00:21:55,160 --> 00:21:56,360
You're operating an accident.

559
00:21:56,360 --> 00:21:58,960
So MCP turns integration into infrastructure.

560
00:21:58,960 --> 00:22:02,600
It makes tools something the platform team can treat like a shared service instead of

561
00:22:02,600 --> 00:22:04,160
a per agent science project.

562
00:22:04,160 --> 00:22:05,480
That is the actual value.

563
00:22:05,480 --> 00:22:06,800
Here's what most people miss.

564
00:22:06,800 --> 00:22:08,920
Standard tool discovery isn't about developer convenience.

565
00:22:08,920 --> 00:22:10,080
It's about predictability.

566
00:22:10,080 --> 00:22:14,600
The system can't enforce intent if every agent calls a different implementation of create

567
00:22:14,600 --> 00:22:16,560
ticket or update record.

568
00:22:16,560 --> 00:22:17,800
You don't just get duplication.

569
00:22:17,800 --> 00:22:21,960
You get divergent behavior and divergent behavior becomes inconsistent outcomes.

570
00:22:21,960 --> 00:22:23,320
In other words, decision debt.

571
00:22:23,320 --> 00:22:28,160
With MCP, the enterprise can standardize the verbs, create, read, update, approve, notify,

572
00:22:28,160 --> 00:22:31,720
export, and then govern those verbs like they are part of the control plane.

573
00:22:31,720 --> 00:22:33,160
Because they are.

574
00:22:33,160 --> 00:22:34,440
Now there's a tension here.

575
00:22:34,440 --> 00:22:37,040
And pretending it doesn't exist is how programs get hurt.

576
00:22:37,040 --> 00:22:38,680
MCP reduces bespoke chaos.

577
00:22:38,680 --> 00:22:40,920
It also makes mistakes propagate faster.

578
00:22:40,920 --> 00:22:44,320
When you build one custom connector wrong, you usually hurt one agent.

579
00:22:44,320 --> 00:22:48,480
When you publish one MCP tool contract wrong, you potentially hurt every agent that discovers

580
00:22:48,480 --> 00:22:49,480
and reuses it.

581
00:22:49,480 --> 00:22:51,040
That's not a reason to avoid MCP.

582
00:22:51,040 --> 00:22:55,840
It's a reason to treat MCP like production infrastructure, versioning, review, and blast

583
00:22:55,840 --> 00:22:56,840
radius thinking.

584
00:22:56,840 --> 00:22:59,200
So what does contract thinking actually mean?

585
00:22:59,200 --> 00:23:02,960
It means tools are enforceable interfaces, not developer shortcuts.

586
00:23:02,960 --> 00:23:05,280
An enforceable interface has three properties.

587
00:23:05,280 --> 00:23:07,200
One, it's explicit about what it does.

588
00:23:07,200 --> 00:23:09,320
Not in marketing language, in operational language.

589
00:23:09,320 --> 00:23:10,320
Does it read?

590
00:23:10,320 --> 00:23:11,320
Does it write?

591
00:23:11,320 --> 00:23:12,320
Does it delete?

592
00:23:12,320 --> 00:23:14,320
Does it create side effects that can't be undone?

593
00:23:14,320 --> 00:23:18,720
Two, it's explicit about what inputs it accepts and what outputs it produces.

594
00:23:18,720 --> 00:23:21,480
If the output is free text, you just created a laundering pathway.

595
00:23:21,480 --> 00:23:24,280
The next agent can't reliably know what it received.

596
00:23:24,280 --> 00:23:28,840
If the output is structured, fields, identifiers, source metadata, you can carry provenance

597
00:23:28,840 --> 00:23:31,080
forward and you can audit behavior later.

598
00:23:31,080 --> 00:23:33,120
Three, it's explicit about authority.

599
00:23:33,120 --> 00:23:37,280
Which identity calls it, with which permission scope against which system boundary.

600
00:23:37,280 --> 00:23:41,520
If the tool can do privileged actions, then privileged actions require an approval pattern.

601
00:23:41,520 --> 00:23:43,600
Not a policy document, a real gate.

602
00:23:43,600 --> 00:23:47,000
This is why MCP matters even when you already have connectors.

603
00:23:47,000 --> 00:23:48,560
Connectors get you, it works.

604
00:23:48,560 --> 00:23:52,280
MCP gets you, it works the same way everywhere, which is the only thing that scales.

605
00:23:52,280 --> 00:23:56,400
And this is also why MCP belongs in the control plane, not buried inside app teams.

606
00:23:56,400 --> 00:24:01,520
Two contracts need central review because they define the blast radius of autonomous behavior.

607
00:24:01,520 --> 00:24:05,640
If you let every team publish MCP tools at Hawk, you didn't decentralize innovation.

608
00:24:05,640 --> 00:24:07,240
You decentralized authority.

609
00:24:07,240 --> 00:24:09,200
That's the same mistake, just with nicer Jason.

610
00:24:09,200 --> 00:24:11,360
So what should executives demand here?

611
00:24:11,360 --> 00:24:13,520
They should demand a curated tool catalog.

612
00:24:13,520 --> 00:24:17,800
Which MCP servers are allowed, which tools are exposed, and which actions are deliberately

613
00:24:17,800 --> 00:24:18,800
disabled.

614
00:24:18,800 --> 00:24:21,280
We enabled MCP is not a control, it's a headline.

615
00:24:21,280 --> 00:24:24,000
They should demand version control and deprecation rules.

616
00:24:24,000 --> 00:24:27,960
Actors change, APIs change, workflows change, if the contract changes silently, agent behavior

617
00:24:27,960 --> 00:24:29,040
changes silently.

618
00:24:29,040 --> 00:24:33,480
And that's how it worked last week turns into, we can't explain what happened.

619
00:24:33,480 --> 00:24:36,680
And they should demand that every tool output includes provenance.

620
00:24:36,680 --> 00:24:40,600
If a tool returns policy guidance, it must return the policy source, identify a version

621
00:24:40,600 --> 00:24:41,600
and timestamp.

622
00:24:41,600 --> 00:24:45,680
If it returns a ticket action, it must return the ticket ID and the actor identity that

623
00:24:45,680 --> 00:24:46,680
executed it.

624
00:24:46,680 --> 00:24:49,840
Otherwise, you're building an ecosystem that can't prove its own actions.

625
00:24:49,840 --> 00:24:52,920
Now the boundary problem doesn't go away just because tools are standardized.

626
00:24:52,920 --> 00:24:57,360
A well-governed tool can still retrieve the wrong data if the data boundary is weak.

627
00:24:57,360 --> 00:25:00,200
And a well-governed tool can still leak if the agent can read too much.

628
00:25:00,200 --> 00:25:02,000
So MCP reduces bespoke chaos.

629
00:25:02,000 --> 00:25:03,720
It does not prevent obedient leakage.

630
00:25:03,720 --> 00:25:05,520
That's why the next layer exists.

631
00:25:05,520 --> 00:25:10,120
Perview DSPM for AI, not as a compliance dashboard, but as the data boundary that decides

632
00:25:10,120 --> 00:25:14,000
what agents may see and what they may never bring back.

633
00:25:14,000 --> 00:25:17,800
Control plane layer three, perview DSPM for AI as the data boundary.

634
00:25:17,800 --> 00:25:21,800
Perview DSPM for AI exists for a reason most programs try to skip.

635
00:25:21,800 --> 00:25:24,280
Governance law, you can't govern what you can't see.

636
00:25:24,280 --> 00:25:25,920
Identity tells you who acted.

637
00:25:25,920 --> 00:25:28,120
MCP standardizes what they can do.

638
00:25:28,120 --> 00:25:32,160
But neither one answers the question that actually determines whether your agents become

639
00:25:32,160 --> 00:25:33,320
a leak factory.

640
00:25:33,320 --> 00:25:34,480
What data did they touch?

641
00:25:34,480 --> 00:25:35,560
What did they retrieve?

642
00:25:35,560 --> 00:25:37,080
And what did they return?

643
00:25:37,080 --> 00:25:40,760
At enterprise scale, we think it only used approved data is not a control.

644
00:25:40,760 --> 00:25:42,400
It's wishful thinking with a dashboard.

645
00:25:42,400 --> 00:25:45,880
Perview is the layer that makes data boundaries observable.

646
00:25:45,880 --> 00:25:50,600
Not just in the classic DLP sense of, don't share credit cards, but in the agentic sense.

647
00:25:50,600 --> 00:25:56,160
There is sensitive data exposed to AI systems, which agents are accessing it and which interactions

648
00:25:56,160 --> 00:26:00,560
create risk signals you can act on before the incident becomes a headline.

649
00:26:00,560 --> 00:26:04,400
This is the foundational misunderstanding with data governance in the agent era.

650
00:26:04,400 --> 00:26:06,120
Most teams try to govern the response.

651
00:26:06,120 --> 00:26:09,440
They let the agent read whatever it can read, then they try to prevent it from saying the

652
00:26:09,440 --> 00:26:10,440
wrong thing.

653
00:26:10,440 --> 00:26:11,440
That's backwards.

654
00:26:11,440 --> 00:26:14,000
If the agent is allowed to read it, the agent can leak it.

655
00:26:14,000 --> 00:26:17,440
Even if you block the output, you've still created an access pathway in evidence trail

656
00:26:17,440 --> 00:26:18,920
and an attack surface.

657
00:26:18,920 --> 00:26:20,600
The breach might not be in the response.

658
00:26:20,600 --> 00:26:24,840
It might be in the retrieval logs, the tool outputs, the intermediate state or the downstream

659
00:26:24,840 --> 00:26:26,880
workflow that reuses the data.

660
00:26:26,880 --> 00:26:29,040
Blocking the final sentence doesn't undo that.

661
00:26:29,040 --> 00:26:34,080
So the boundary mindset is, prevent reading when you must, not allow reading and hope you

662
00:26:34,080 --> 00:26:35,800
can sanitize later.

663
00:26:35,800 --> 00:26:38,200
This is where Perview becomes more than compliance theatre.

664
00:26:38,200 --> 00:26:40,720
It becomes the system that helps you define three things.

665
00:26:40,720 --> 00:26:42,920
The business never defines cleanly on its own.

666
00:26:42,920 --> 00:26:46,920
First, what data is sensitive, where it lives and who is allowed to access it in the

667
00:26:46,920 --> 00:26:47,920
first place.

668
00:26:47,920 --> 00:26:50,720
The old problem is still unsolved in most tenants.

669
00:26:50,720 --> 00:26:54,120
Second, which data sources are allowed to be used as grounding for agents.

670
00:26:54,120 --> 00:26:55,760
This is the new problem.

671
00:26:55,760 --> 00:26:59,000
Allowed for humans is not the same as allowed for agents.

672
00:26:59,000 --> 00:27:00,720
Agents scale retrieval.

673
00:27:00,720 --> 00:27:01,720
Humans don't.

674
00:27:01,720 --> 00:27:04,360
A human opening a document is a single event.

675
00:27:04,360 --> 00:27:10,720
An agent can retrieve a hundred related documents in a loop because the goal required context.

676
00:27:10,720 --> 00:27:14,000
Same permissions, totally different behavior.

677
00:27:14,000 --> 00:27:17,040
Third, what constitutes authoritative versus available.

678
00:27:17,040 --> 00:27:19,040
And this is the part that collapses trust.

679
00:27:19,040 --> 00:27:23,080
If your agent pulls policy from the most accessible location instead of the authoritative

680
00:27:23,080 --> 00:27:27,200
repository, you've just automated misinformation with a professional tone.

681
00:27:27,200 --> 00:27:31,200
Perview helps by giving you visibility and exposure tracking across the content plane

682
00:27:31,200 --> 00:27:34,800
and by surfacing risk signals tied to AI usage patterns.

683
00:27:34,800 --> 00:27:37,240
It doesn't magically fix your information architecture.

684
00:27:37,240 --> 00:27:39,840
It just stops you from pretending the architecture doesn't matter.

685
00:27:39,840 --> 00:27:45,400
Now, executives usually want one thing from Perview, reporting that maps to business reality.

686
00:27:45,400 --> 00:27:47,360
They don't want a heat map of labels.

687
00:27:47,360 --> 00:27:51,480
They want answers to questions like which agents access sensitive data, which users rely

688
00:27:51,480 --> 00:27:55,800
on those agents and where policy violations are accumulating over time.

689
00:27:55,800 --> 00:27:58,960
Because that's what determines whether the program scales or gets paused.

690
00:27:58,960 --> 00:28:02,960
So the practical framing is DSPM for AI as an executive dashboard for adoption, exposure

691
00:28:02,960 --> 00:28:03,960
and violations.

692
00:28:03,960 --> 00:28:07,360
Not as a monthly compliance report, as an operational control plane signal.

693
00:28:07,360 --> 00:28:10,160
And here's the important distinction to keep straight.

694
00:28:10,160 --> 00:28:11,480
Visibility is not enforcement.

695
00:28:11,480 --> 00:28:13,200
Perview can show you exposure.

696
00:28:13,200 --> 00:28:14,200
It can show you risk.

697
00:28:14,200 --> 00:28:16,200
It can help you classify and define boundaries.

698
00:28:16,200 --> 00:28:19,640
But if you stop there, you've built a better review mirror, not a safer vehicle.

699
00:28:19,640 --> 00:28:22,320
This is why Perview has to connect back to the other layers.

700
00:28:22,320 --> 00:28:24,680
Entra agent ID gives you attribution.

701
00:28:24,680 --> 00:28:27,640
Which agent identity performed the access?

702
00:28:27,640 --> 00:28:29,680
MCP gives you the tool boundary.

703
00:28:29,680 --> 00:28:32,560
Which tool path retrieved or exported the data?

704
00:28:32,560 --> 00:28:34,360
Perview gives you the data boundary.

705
00:28:34,360 --> 00:28:37,240
Which content was sensitive, which sources were used, and

706
00:28:37,240 --> 00:28:39,560
whether the interaction crossed a policy line.

707
00:28:39,560 --> 00:28:42,680
And then you still need a layer that detects behavior drift over time.

708
00:28:42,680 --> 00:28:45,320
Because agents don't fail only through bad design.

709
00:28:45,320 --> 00:28:47,120
They fail through evolving usage.

710
00:28:47,120 --> 00:28:47,960
People probe them.

711
00:28:47,960 --> 00:28:48,720
People push them.

712
00:28:48,720 --> 00:28:50,840
People accidentally feed them the wrong context.

713
00:28:50,840 --> 00:28:52,200
Attackers deliberately do it.

714
00:28:52,200 --> 00:28:56,560
And the normal behavior of an agent changes as the organization changes around it.

715
00:28:56,560 --> 00:28:58,440
That's the part compliance teams hate.

716
00:28:58,440 --> 00:29:01,080
Because it means the boundary is not a one-time design.

717
00:29:01,080 --> 00:29:02,400
It's ongoing entropy management.

718
00:29:02,400 --> 00:29:03,920
So the rule for this layer is simple.

719
00:29:03,920 --> 00:29:07,680
If you want agents at scale, you must treat data governance as a runtime boundary,

720
00:29:07,680 --> 00:29:09,480
not a documentation exercise.

721
00:29:09,480 --> 00:29:12,200
If you can't see what agents are touching, you can't defend it.

722
00:29:12,200 --> 00:29:14,960
And if you only try to govern what they say, you're governing the wrong surface.

723
00:29:14,960 --> 00:29:18,520
Perview gives you the visibility and the boundary signals to stop guessing.

724
00:29:18,520 --> 00:29:20,240
But visibility isn't defense.

725
00:29:20,240 --> 00:29:22,280
Behavior still drifts under pressure.

726
00:29:22,280 --> 00:29:23,880
Next layer, defender for AI.

727
00:29:23,880 --> 00:29:27,680
Because at scale, security becomes behavior based, whether you like it or not.

728
00:29:27,680 --> 00:29:29,480
Control plane layer four.

729
00:29:29,480 --> 00:29:32,200
Defender for AI as behavioral detection.

730
00:29:32,200 --> 00:29:37,800
Defender for AI is the layer that forces the enterprise to accept a basic reality.

731
00:29:37,800 --> 00:29:40,520
Agent security can't be intent-based.

732
00:29:40,520 --> 00:29:43,280
Intent is what humans claim after something goes wrong.

733
00:29:43,280 --> 00:29:45,240
Behavior is what systems actually do.

734
00:29:45,240 --> 00:29:50,400
And once agents act through tools across data sources in loops, you're no longer defending

735
00:29:50,400 --> 00:29:51,560
a chat interface.

736
00:29:51,560 --> 00:29:54,960
You're defending an execution surface that changes shape every day.

737
00:29:54,960 --> 00:29:57,080
Because the organization keeps changing around it.

738
00:29:57,080 --> 00:29:59,600
That distinction matters.

739
00:29:59,600 --> 00:30:02,680
Most security programs are still built on a comfortable assumption.

740
00:30:02,680 --> 00:30:06,200
If you lock down identities and you label data, you're covered.

741
00:30:06,200 --> 00:30:08,160
That was already optimistic before agents.

742
00:30:08,160 --> 00:30:10,080
In the agent era, it's inadequate.

743
00:30:10,080 --> 00:30:12,560
Because the failure mode isn't only unauthorized access.

744
00:30:12,560 --> 00:30:15,360
It's authorized access used in unauthorized patterns.

745
00:30:15,360 --> 00:30:17,880
The same identity can be valid and still dangerous.

746
00:30:17,880 --> 00:30:20,440
The same tool can be approved and still abused.

747
00:30:20,440 --> 00:30:24,240
The same data source can be allowed and still produce a leak when combined with the wrong

748
00:30:24,240 --> 00:30:25,480
prompt at the wrong time.

749
00:30:25,480 --> 00:30:29,480
So defenders job in this control plane is not to judge whether an agent meant well.

750
00:30:29,480 --> 00:30:33,400
It's to detect when an agent's behavior stops matching the enterprise's assumptions.

751
00:30:33,400 --> 00:30:35,960
That's what behavioral detection actually means in practice.

752
00:30:35,960 --> 00:30:41,200
That means looking for patterns that indicate drift, coercion, or automation gone feral.

753
00:30:41,200 --> 00:30:45,960
Prompt injection attempts, tool abuse, anomalous access sequences, and why is this agent suddenly

754
00:30:45,960 --> 00:30:47,280
doing that?

755
00:30:47,280 --> 00:30:51,280
Events that don't show up in your design documents start with prompt injection.

756
00:30:51,280 --> 00:30:54,800
Because it's the attack everyone would pretend is edge case until it's in their incident

757
00:30:54,800 --> 00:30:55,800
report.

758
00:30:55,800 --> 00:30:57,200
Prompt injection isn't magic.

759
00:30:57,200 --> 00:31:02,400
It's just an attacker or a careless user smuggling instructions into the context stream.

760
00:31:02,400 --> 00:31:06,560
The agent reads it as part of the problem then follows it as part of the solution.

761
00:31:06,560 --> 00:31:11,320
If the agent can call tools the injection isn't just about bad answers, it's about bad actions.

762
00:31:11,320 --> 00:31:15,320
Defenders role is to detect those attempts and raise signals you can act on.

763
00:31:15,320 --> 00:31:19,720
Suspicious instruction patterns unexpected content shaping and the characteristic ignore

764
00:31:19,720 --> 00:31:23,560
previous instructions style coercion that shows up in real attacks.

765
00:31:23,560 --> 00:31:24,800
Then tool abuse.

766
00:31:24,800 --> 00:31:29,440
Tool abuse looks like an agent calling a high impact tool at an unusual time, at an unusual

767
00:31:29,440 --> 00:31:31,840
volume or in an unusual sequence.

768
00:31:31,840 --> 00:31:35,880
The service desk agent that normally creates tickets suddenly starts updating existing

769
00:31:35,880 --> 00:31:36,880
tickets in bulk.

770
00:31:36,880 --> 00:31:40,880
A policy agent that normally retrieves one document starts pulling hundreds.

771
00:31:40,880 --> 00:31:44,400
An approvals agent that normally waits for humans suddenly chains actions without the

772
00:31:44,400 --> 00:31:45,720
expected checkpoints.

773
00:31:45,720 --> 00:31:47,120
The system isn't broken.

774
00:31:47,120 --> 00:31:48,600
Your assumptions are.

775
00:31:48,600 --> 00:31:51,600
And those assumptions erode faster than your governance committees meet.

776
00:31:51,600 --> 00:31:54,640
This is why defender for AI becomes mandatory at scale.

777
00:31:54,640 --> 00:31:56,160
Not because Microsoft says so.

778
00:31:56,160 --> 00:32:00,040
Because autonomous behavior plus human pressure creates misunderstood design omissions

779
00:32:00,040 --> 00:32:01,200
on schedule.

780
00:32:01,200 --> 00:32:03,840
And will ask agents to do things they shouldn't.

781
00:32:03,840 --> 00:32:06,160
Makers will add tools they didn't fully review.

782
00:32:06,160 --> 00:32:07,560
Vendors will ship updates.

783
00:32:07,560 --> 00:32:10,080
Somebody will paste the wrong thing into the wrong place.

784
00:32:10,080 --> 00:32:14,160
And the platform will keep operating security that doesn't announce itself it accumulates.

785
00:32:14,160 --> 00:32:16,680
Defender gives you a way to see the accumulation as it happens.

786
00:32:16,680 --> 00:32:21,440
The practical outcome executives should care about is containment not containment of AI,

787
00:32:21,440 --> 00:32:25,360
containment of one agent identity, one tool pathway, one behavior pattern without shutting

788
00:32:25,360 --> 00:32:26,360
down the whole program.

789
00:32:26,360 --> 00:32:27,880
That's what keeps adoption alive.

790
00:32:27,880 --> 00:32:33,000
If the only response you have to suspicious behavior is disabled co-pilot or pause all agents,

791
00:32:33,000 --> 00:32:35,600
you've built a program that can't survive scrutiny.

792
00:32:35,600 --> 00:32:38,880
You've also trained the business to treat AI as fragile.

793
00:32:38,880 --> 00:32:42,040
They will stop investing the moment it causes inconvenience.

794
00:32:42,040 --> 00:32:43,400
Defender changes the playbook.

795
00:32:43,400 --> 00:32:45,880
With identity in Entra, you can target the actor.

796
00:32:45,880 --> 00:32:48,720
With MCP tool contracts, you can target the capability.

797
00:32:48,720 --> 00:32:51,320
With Pervue signals, you can target the exposure.

798
00:32:51,320 --> 00:32:53,800
With Defender, you can target the behavior in motion.

799
00:32:53,800 --> 00:32:56,840
So when something goes wrong, you don't need a cross-functional war room to debate what

800
00:32:56,840 --> 00:32:57,840
happened for three days.

801
00:32:57,840 --> 00:33:02,160
You isolate the agent, you quarantine the tool, you roll back the permission boundary,

802
00:33:02,160 --> 00:33:06,080
you preserve the evidence trail, and you keep the rest of the ecosystem running.

803
00:33:06,080 --> 00:33:07,080
That's the architectural win.

804
00:33:07,080 --> 00:33:09,800
You stop treating incidents as existential.

805
00:33:09,800 --> 00:33:10,960
Now there's a trap here too.

806
00:33:10,960 --> 00:33:14,160
If you deploy Defender without attribution, you get alerts you can't act on.

807
00:33:14,160 --> 00:33:18,320
You'll see anomalous behavior involving agents, but you won't have a clean identity chain

808
00:33:18,320 --> 00:33:19,320
to disable.

809
00:33:19,320 --> 00:33:22,600
If you deploy Defender without data boundaries, you'll detect leaks after they already

810
00:33:22,600 --> 00:33:27,160
happened. If you deploy Defender without tool contracts, you'll detect tool abuse without

811
00:33:27,160 --> 00:33:29,880
knowing which tool implementation was called.

812
00:33:29,880 --> 00:33:33,280
Detection without enforceability is just anxiety, so Defender isn't the control plane

813
00:33:33,280 --> 00:33:34,280
by itself.

814
00:33:34,280 --> 00:33:36,680
It's the last layer that makes the chain complete.

815
00:33:36,680 --> 00:33:40,720
Behavior signals tied to identities, tied to tools, tied to data boundaries, tied to actions

816
00:33:40,720 --> 00:33:42,320
you can actually take.

817
00:33:42,320 --> 00:33:45,120
And the reason it has to be the fourth layer is simple.

818
00:33:45,120 --> 00:33:47,160
Visibility isn't Defends.

819
00:33:47,160 --> 00:33:48,400
Boundaries aren't enough.

820
00:33:48,400 --> 00:33:50,960
At scale, behavior drifts.

821
00:33:50,960 --> 00:33:55,280
Vierter is how you see the drift early and how you contain it precisely before the drift

822
00:33:55,280 --> 00:33:58,880
becomes the forcing function that freezes the entire program.

823
00:33:58,880 --> 00:34:04,040
Next, connect the four layers into a single minimum viable control plane, Entra, MCP,

824
00:34:04,040 --> 00:34:08,640
Perview and Defender and why missing any one of them turns enforceable intelligence into

825
00:34:08,640 --> 00:34:09,880
conditional chaos.

826
00:34:09,880 --> 00:34:13,960
The minimum viable agent control plane, how the four layers interlock.

827
00:34:13,960 --> 00:34:18,120
Most organizations try to govern agents by buying one product, turning on one switch and

828
00:34:18,120 --> 00:34:19,120
declaring victory.

829
00:34:19,120 --> 00:34:20,120
That is not how this works.

830
00:34:20,120 --> 00:34:21,680
The control plane is not a product.

831
00:34:21,680 --> 00:34:23,080
It's an enforcement chain.

832
00:34:23,080 --> 00:34:26,000
And chains fail at their weakest link, not their most expensive one.

833
00:34:26,000 --> 00:34:30,040
So the minimum viable agent control plane is four layers that interlock cleanly.

834
00:34:30,040 --> 00:34:31,800
With no gaps you have to explain later.

835
00:34:31,800 --> 00:34:36,840
Entra agent ID for who, MCP, for what, Perview, for what data, Defender, for how it behaves.

836
00:34:36,840 --> 00:34:37,840
That's it.

837
00:34:37,840 --> 00:34:40,920
Four layers, not because it's elegant, because anything less becomes probabilistic.

838
00:34:40,920 --> 00:34:43,240
Start with the interlock model as a mental map.

839
00:34:43,240 --> 00:34:44,880
Entra agent ID answers.

840
00:34:44,880 --> 00:34:45,880
Who is acting?

841
00:34:45,880 --> 00:34:47,040
Not who asked.

842
00:34:47,040 --> 00:34:48,160
Who executed?

843
00:34:48,160 --> 00:34:52,440
Which non-human identity performed the tool called, Touch the Data, and created the side

844
00:34:52,440 --> 00:34:53,440
effect?

845
00:34:53,440 --> 00:34:54,440
MCP answers.

846
00:34:54,440 --> 00:34:55,440
What is the agent allowed to do?

847
00:34:55,440 --> 00:34:57,560
Not in philosophical terms, in verbs.

848
00:34:57,560 --> 00:34:58,560
Which tools exist?

849
00:34:58,560 --> 00:34:59,800
Which actions are exposed?

850
00:34:59,800 --> 00:35:01,080
Which actions are disabled?

851
00:35:01,080 --> 00:35:02,240
Which outputs are structured?

852
00:35:02,240 --> 00:35:03,920
Which ones carry provenance?

853
00:35:03,920 --> 00:35:04,920
Perview answers.

854
00:35:04,920 --> 00:35:06,400
What data can the agent touch?

855
00:35:06,400 --> 00:35:07,880
And what data can it return?

856
00:35:07,880 --> 00:35:09,040
This is the boundary layer.

857
00:35:09,040 --> 00:35:12,360
It turns, we think it shouldn't see that into it can't read it.

858
00:35:12,360 --> 00:35:16,000
And it turns, we hope it didn't use sensitive content into exposure signals.

859
00:35:16,000 --> 00:35:17,640
You can actually review.

860
00:35:17,640 --> 00:35:18,640
Defender answers.

861
00:35:18,640 --> 00:35:20,200
How is it behaving over time?

862
00:35:20,200 --> 00:35:25,560
Normal behavior becomes baseline, drift becomes detectable, abuse becomes containable, and

863
00:35:25,560 --> 00:35:30,000
the signals tie back to an identity you can disable, and a tool contract you can restrict.

864
00:35:30,000 --> 00:35:31,680
Now here's what most people miss.

865
00:35:31,680 --> 00:35:33,040
These layers don't just stack.

866
00:35:33,040 --> 00:35:34,200
They depend on each other.

867
00:35:34,200 --> 00:35:38,080
If Entra is missing, you can detect and monitor all day, but you can't attribute.

868
00:35:38,080 --> 00:35:39,400
You can't prove who acted.

869
00:35:39,400 --> 00:35:42,560
You can't surgically disable one agent without collateral damage.

870
00:35:42,560 --> 00:35:45,320
You end up pausing the program because that's the only lever left.

871
00:35:45,320 --> 00:35:49,920
If MCP discipline is missing, you can have perfect identity and perfect logs, but your integrations

872
00:35:49,920 --> 00:35:51,480
remain bespoke chaos.

873
00:35:51,480 --> 00:35:55,000
Every agent calls different endpoints in different ways, outputs, different shapes, and

874
00:35:55,000 --> 00:35:56,000
leaks provenance.

875
00:35:56,000 --> 00:35:59,400
So your incident response becomes reverse engineering, not containment.

876
00:35:59,400 --> 00:36:02,960
If Perview is missing, you can know who acted and what tool they used, but you don't know

877
00:36:02,960 --> 00:36:03,960
what they touched.

878
00:36:03,960 --> 00:36:07,320
You'll discover exposure after the fact and your post-incident report will read like a

879
00:36:07,320 --> 00:36:08,320
guess.

880
00:36:08,320 --> 00:36:10,360
That's the moment executive stop funding scale.

881
00:36:10,360 --> 00:36:13,840
If Defender is missing, you can label data, scope identities, and standardize tools,

882
00:36:13,840 --> 00:36:18,080
and still get burned by behavior drift because the environment changes, the prompts change,

883
00:36:18,080 --> 00:36:21,360
the agents change in new ways and your controls lag reality.

884
00:36:21,360 --> 00:36:23,040
Missing layers don't create obvious gaps.

885
00:36:23,040 --> 00:36:26,240
They create ambiguity and ambiguity is where agent programs die.

886
00:36:26,240 --> 00:36:29,000
So what does minimum viable mean in operational terms?

887
00:36:29,000 --> 00:36:33,880
It means you can trace any agent action end-to-end as a single evidence chain.

888
00:36:33,880 --> 00:36:39,800
Identity, tool call, data access, behavior signal, outcome, not screenshots, not a deck.

889
00:36:39,800 --> 00:36:40,800
Evidence.

890
00:36:40,800 --> 00:36:45,160
If you're not executed, you can show which agent identity created it, which MCP tool executed

891
00:36:45,160 --> 00:36:50,000
it, which data sources were read to ground it, and whether the behavior matched baseline.

892
00:36:50,000 --> 00:36:53,960
If a file gets shared, you can show the identity, the tool path, the data classification,

893
00:36:53,960 --> 00:36:56,320
and the behavioral context that led to the share.

894
00:36:56,320 --> 00:36:59,320
If a policy answer gets generated, you can show provenance.

895
00:36:59,320 --> 00:37:03,000
Which authoritative source was used and which non-authoritative sources were intentionally

896
00:37:03,000 --> 00:37:04,000
excluded.

897
00:37:04,000 --> 00:37:07,640
That is the control plane doing its job, enforcing intent at runtime.

898
00:37:07,640 --> 00:37:11,640
Now translate that into executive demands because this is where teams love to hide behind partial

899
00:37:11,640 --> 00:37:12,640
compliance.

900
00:37:12,640 --> 00:37:17,440
Executives should demand proof of enforcement end-to-end, not per component checklists.

901
00:37:17,440 --> 00:37:20,040
Not we have entra, not we turned on purview.

902
00:37:20,040 --> 00:37:21,720
Not we built an MCP server.

903
00:37:21,720 --> 00:37:23,040
Not defender is enabled.

904
00:37:23,040 --> 00:37:27,840
The demand is show the chain for a real scenario, pick a high impact workflow, show the identity,

905
00:37:27,840 --> 00:37:31,720
show the tool contract, show the data boundary, show the detection and containment path.

906
00:37:31,720 --> 00:37:35,920
If any part of that is, we don't have that yet, then the program is not ready to scale.

907
00:37:35,920 --> 00:37:39,520
It can still experiment, it can still learn, but it can't industrialize.

908
00:37:39,520 --> 00:37:42,960
And one more uncomfortable truth, the control plane doesn't reduce innovation.

909
00:37:42,960 --> 00:37:46,040
It prevents innovation from turning into unbounded authority.

910
00:37:46,040 --> 00:37:50,840
Because the moment you scale agents, you are scaling decisions, not just answers, decisions

911
00:37:50,840 --> 00:37:52,480
that mutate systems.

912
00:37:52,480 --> 00:37:57,080
So the minimum viable control plane is how you keep intelligence enforceable as it grows.

913
00:37:57,080 --> 00:37:58,520
Next this stops being theory.

914
00:37:58,520 --> 00:38:02,840
Apply the control plane to real scenarios, spoken, concrete, and uncomfortable, starting

915
00:38:02,840 --> 00:38:05,840
with the one every enterprise thinks is harmless.

916
00:38:05,840 --> 00:38:11,280
And IT service desk agent that succeeds so quickly it triggers sprawl.

917
00:38:11,280 --> 00:38:15,360
Scenario one setup, IT service desk sprawl starts as a success.

918
00:38:15,360 --> 00:38:17,560
The first scenario always starts as a win.

919
00:38:17,560 --> 00:38:18,560
That's why it spreads.

920
00:38:18,560 --> 00:38:23,000
A service desk team builds one triage agent, it sits in copilot or teams or a portal, doesn't

921
00:38:23,000 --> 00:38:24,000
matter.

922
00:38:24,000 --> 00:38:29,400
Users type the same messy tickets they've always typed, VPN broken, can't access sharepoint,

923
00:38:29,400 --> 00:38:33,200
new laptop, password reset, outlook is haunted.

924
00:38:33,200 --> 00:38:37,480
The agent cleans it up, asks the two missing questions humans never asked consistently

925
00:38:37,480 --> 00:38:41,760
and roots the request to the right queue with the same category and a clear summary.

926
00:38:41,760 --> 00:38:42,760
Ticket noise drops.

927
00:38:42,760 --> 00:38:44,560
MTTR starts to move.

928
00:38:44,560 --> 00:38:48,280
Leadership gets a slide that says AI reduced backlog at everyone claps.

929
00:38:48,280 --> 00:38:50,640
The team that built it gets asked to do a road show.

930
00:38:50,640 --> 00:38:53,800
And that's where the entropy begins because the next thing that happens isn't a security

931
00:38:53,800 --> 00:38:55,720
incident, it's success.

932
00:38:55,720 --> 00:38:59,440
The agent works well enough that other teams decide they need their own version.

933
00:38:59,440 --> 00:39:03,160
The endpoint team wants a triage agent, optimized for devices.

934
00:39:03,160 --> 00:39:06,440
The identity team wants one optimized for access requests.

935
00:39:06,440 --> 00:39:09,880
The network team wants one optimized for Wi-Fi and VPN.

936
00:39:09,880 --> 00:39:13,600
The application team wants one optimized for our line of business apps.

937
00:39:13,600 --> 00:39:15,400
And each of those requests sounds reasonable.

938
00:39:15,400 --> 00:39:17,400
Each team optimizes locally.

939
00:39:17,400 --> 00:39:18,800
That's the core problem.

940
00:39:18,800 --> 00:39:21,440
Local optimization creates global inconsistency.

941
00:39:21,440 --> 00:39:25,760
So you end up with five triage agents that all claim to be the best way to open a ticket,

942
00:39:25,760 --> 00:39:30,480
each with a slightly different intake process, a slightly different definition of severity,

943
00:39:30,480 --> 00:39:34,800
a slightly different set of tool calls, and a slightly different idea of what resolved

944
00:39:34,800 --> 00:39:35,960
means.

945
00:39:35,960 --> 00:39:38,080
Now the user experience collapses quietly.

946
00:39:38,080 --> 00:39:40,440
Users start asking which agent should I use.

947
00:39:40,440 --> 00:39:42,280
That question sounds like curiosity.

948
00:39:42,280 --> 00:39:45,720
It's actually the first sign that your service model has fragmented because if an ecosystem

949
00:39:45,720 --> 00:39:47,760
is healthy, users don't choose agents.

950
00:39:47,760 --> 00:39:48,760
The system roots them.

951
00:39:48,760 --> 00:39:50,000
The system has intent.

952
00:39:50,000 --> 00:39:52,080
In sprawl, the user becomes the router.

953
00:39:52,080 --> 00:39:54,640
And users root based on convenience, not control.

954
00:39:54,640 --> 00:39:56,200
Now add the orchestration failure.

955
00:39:56,200 --> 00:39:58,600
This is the part nobody diagrams until it hurts.

956
00:39:58,600 --> 00:40:00,920
One agent roots a ticket to the network queue.

957
00:40:00,920 --> 00:40:03,040
Another agent seeing the same symptoms.

958
00:40:03,040 --> 00:40:04,960
Roots a similar ticket to identity.

959
00:40:04,960 --> 00:40:09,840
A third agent auto suggests a self-service fix that's correct for one environment but wrong

960
00:40:09,840 --> 00:40:10,840
for another.

961
00:40:10,840 --> 00:40:11,840
Tickets bounce.

962
00:40:11,840 --> 00:40:13,680
Duplicate tickets get created.

963
00:40:13,680 --> 00:40:15,280
Humans re-triage the triage.

964
00:40:15,280 --> 00:40:19,080
And the thing you build to reduce noise starts generating noise in a new shape.

965
00:40:19,080 --> 00:40:21,400
The organization calls this teething issues.

966
00:40:21,400 --> 00:40:22,400
It isn't.

967
00:40:22,400 --> 00:40:25,640
It's the system doing exactly what you designed.

968
00:40:25,640 --> 00:40:28,040
Decentralized authority without shared intent.

969
00:40:28,040 --> 00:40:31,440
An ownership collapses because ownership always collapses in sprawl.

970
00:40:31,440 --> 00:40:33,240
The original service desk agent had an owner.

971
00:40:33,240 --> 00:40:34,240
It had a team.

972
00:40:34,240 --> 00:40:35,600
It had someone who could answer for it.

973
00:40:35,600 --> 00:40:39,720
But the moment other teams clone it, fork it or build their own, ownership becomes ambiguous.

974
00:40:39,720 --> 00:40:42,760
Who owns the user outcome when the wrong queue gets the ticket?

975
00:40:42,760 --> 00:40:46,360
Who owns the inconsistent answers when two agents contradict each other?

976
00:40:46,360 --> 00:40:51,680
Who owns the escalation path when the agent says, resolved but the user still can't work?

977
00:40:51,680 --> 00:40:52,920
Nobody owns the outcome.

978
00:40:52,920 --> 00:40:54,320
Everyone owns their component.

979
00:40:54,320 --> 00:40:56,120
That's how enterprise systems fail.

980
00:40:56,120 --> 00:40:57,640
Now put a real failure moment on it.

981
00:40:57,640 --> 00:41:00,600
A user requests access to a restricted SharePoint site.

982
00:41:00,600 --> 00:41:04,960
The triage agent decides it's an access request calls the tool and creates the ticket.

983
00:41:04,960 --> 00:41:07,400
Another team's agent decides it can do better.

984
00:41:07,400 --> 00:41:12,440
It triggers an automated workflow that grants temporary access to unblock productivity,

985
00:41:12,440 --> 00:41:16,960
because someone added that capability during a sprint and never removed it.

986
00:41:16,960 --> 00:41:20,800
A manager gets an adaptive card approval in teams, hits a prove because it looks routine

987
00:41:20,800 --> 00:41:22,520
and the user gets access.

988
00:41:22,520 --> 00:41:25,400
Except the site contains sensitive content the user shouldn't have.

989
00:41:25,400 --> 00:41:27,880
Now you have an incident and it's the worst kind.

990
00:41:27,880 --> 00:41:32,800
An incident created by helpful automation backed by a legitimate approval executed through

991
00:41:32,800 --> 00:41:35,640
a toolpath nobody reviewed end to end.

992
00:41:35,640 --> 00:41:38,760
And when leadership asks what happened you get five different answers.

993
00:41:38,760 --> 00:41:40,960
Service desk says we didn't build that workflow.

994
00:41:40,960 --> 00:41:43,760
Identity says we didn't approve those permissions.

995
00:41:43,760 --> 00:41:47,560
The team that built the second agent says it was approved by a manager.

996
00:41:47,560 --> 00:41:51,120
Security says we can't tell which agent executed the tool call.

997
00:41:51,120 --> 00:41:54,240
And the business says so are we pausing all agents?

998
00:41:54,240 --> 00:41:55,520
This is the lesson.

999
00:41:55,520 --> 00:41:59,240
Scaling agents without shared intent creates contradictory intelligence.

1000
00:41:59,240 --> 00:42:04,040
Not because people are incompetent, because the platform makes it easy to build local success

1001
00:42:04,040 --> 00:42:05,920
that erodes global control.

1002
00:42:05,920 --> 00:42:10,320
And once the service desk becomes the proving ground, sprawl becomes normalized.

1003
00:42:10,320 --> 00:42:12,120
Every department learns the wrong rule.

1004
00:42:12,120 --> 00:42:16,160
If you need a capability built in agent, not if you need a capability, design it as a

1005
00:42:16,160 --> 00:42:17,160
govern service.

1006
00:42:17,160 --> 00:42:18,840
Next, the uncomfortable part.

1007
00:42:18,840 --> 00:42:23,240
What changes when you enforce identity and shared intent is non-negotiable.

1008
00:42:23,240 --> 00:42:27,560
There one remediation, deterministic rooting and accountable action.

1009
00:42:27,560 --> 00:42:31,560
Fixing the service desk sprawl problem doesn't start with better prompts.

1010
00:42:31,560 --> 00:42:34,640
It starts with making rooting and action deterministic again.

1011
00:42:34,640 --> 00:42:38,240
Not deterministic in the sense that every ticket looks the same.

1012
00:42:38,240 --> 00:42:41,720
Deterministic in the sense that the enterprise can prove this agent did this thing through

1013
00:42:41,720 --> 00:42:46,640
this tool under this approval model with this data boundary every time.

1014
00:42:46,640 --> 00:42:51,840
So remediation begins with the simplest enforcement rule the organization avoided in the rollout.

1015
00:42:51,840 --> 00:42:55,000
The triage capability gets its own Entra agent ID.

1016
00:42:55,000 --> 00:42:56,800
Not one service desk but account.

1017
00:42:56,800 --> 00:43:00,080
Not one shared service principle that powers five different agents.

1018
00:43:00,080 --> 00:43:01,840
One identity per agent product.

1019
00:43:01,840 --> 00:43:03,360
Scoped to what it actually does.

1020
00:43:03,360 --> 00:43:05,600
That's what turns incident response back into engineering.

1021
00:43:05,600 --> 00:43:09,680
If network triage agent starts doing something wrong, the containment action is obvious.

1022
00:43:09,680 --> 00:43:10,680
Disable that identity.

1023
00:43:10,680 --> 00:43:12,160
You don't disable all automation.

1024
00:43:12,160 --> 00:43:13,480
You don't pause co-pilot.

1025
00:43:13,480 --> 00:43:16,560
You remove one actor from the system and you keep operating.

1026
00:43:16,560 --> 00:43:20,200
And Entra agent ID also forces the ownership model to stop being theater.

1027
00:43:20,200 --> 00:43:24,480
Each triage agent has an owner who can change configuration and a sponsor who owns the business

1028
00:43:24,480 --> 00:43:25,480
outcome.

1029
00:43:25,480 --> 00:43:29,760
If the sponsor changes roles, lifecycle workflows, reassign sponsorship, if the agent becomes

1030
00:43:29,760 --> 00:43:31,680
orphaned, it doesn't stay live by accident.

1031
00:43:31,680 --> 00:43:33,080
It gets reviewed or retired.

1032
00:43:33,080 --> 00:43:36,640
That is what run agent products looks like in the service desk.

1033
00:43:36,640 --> 00:43:38,360
Next, tools.

1034
00:43:38,360 --> 00:43:42,200
The sprawl pattern happened because each team built its own toolpath.

1035
00:43:42,200 --> 00:43:46,480
Different connectors, different flows, different create ticket implementations, different

1036
00:43:46,480 --> 00:43:48,280
rooting logic.

1037
00:43:48,280 --> 00:43:51,080
The remediation requires MCP discipline.

1038
00:43:51,080 --> 00:43:54,480
Standard ticket actions published once, reused everywhere.

1039
00:43:54,480 --> 00:43:56,720
One MCP tool for create incident.

1040
00:43:56,720 --> 00:43:59,400
One MCP tool for create service request.

1041
00:43:59,400 --> 00:44:02,400
One MCP tool for update ticket status.

1042
00:44:02,400 --> 00:44:04,760
One MCP tool for assigned to queue.

1043
00:44:04,760 --> 00:44:06,800
And those tools return structured results.

1044
00:44:06,800 --> 00:44:10,520
Ticket ID, QID, urgency classification and the source that justified it.

1045
00:44:10,520 --> 00:44:11,520
Not free text.

1046
00:44:11,520 --> 00:44:12,520
Not vibes.

1047
00:44:12,520 --> 00:44:14,800
Outputs that can be audited and replayed.

1048
00:44:14,800 --> 00:44:16,720
This is the quiet architectural win.

1049
00:44:16,720 --> 00:44:20,040
The tool reuse is where cost and control become the same decision.

1050
00:44:20,040 --> 00:44:23,680
You reduce duplicated work and you reduce duplicated failure modes.

1051
00:44:23,680 --> 00:44:24,680
Now the routing problem.

1052
00:44:24,680 --> 00:44:29,680
The only scalable routing model is one intake, one intent model and explicit escalation

1053
00:44:29,680 --> 00:44:30,680
paths.

1054
00:44:30,680 --> 00:44:34,200
If you let every team invent its own intake, you guarantee contradictions.

1055
00:44:34,200 --> 00:44:37,040
So the enterprise picks an orchestrator pattern.

1056
00:44:37,040 --> 00:44:41,760
A single front door triage experience that decides which specialized triage agent to

1057
00:44:41,760 --> 00:44:42,760
call.

1058
00:44:42,760 --> 00:44:46,040
Not because the specialized agents are bad because the user can't be a router.

1059
00:44:46,040 --> 00:44:47,320
The system has to be.

1060
00:44:47,320 --> 00:44:50,080
This is where deterministic routing shows up in practice.

1061
00:44:50,080 --> 00:44:54,400
The front door agent collects a minimum payload, symptom, impacted service, user identity

1062
00:44:54,400 --> 00:44:57,840
and whether it's an access request, an outage or a how do I question.

1063
00:44:57,840 --> 00:45:00,480
Then it roots to exactly one downstream capability.

1064
00:45:00,480 --> 00:45:04,680
If the downstream capability can't resolve it, it escalates to a human queue with the payload

1065
00:45:04,680 --> 00:45:05,680
attached.

1066
00:45:05,680 --> 00:45:10,200
No ping pong, no duplicate tickets, no conflicting answers from parallel agents.

1067
00:45:10,200 --> 00:45:14,760
Now apply data boundaries because service desk agents love to help themselves into trouble.

1068
00:45:14,760 --> 00:45:18,920
Per view boundaries block access to sensitive queues and restricted knowledge sources.

1069
00:45:18,920 --> 00:45:24,200
The service desk triage agent doesn't need HRK's data, it doesn't need legal investigations,

1070
00:45:24,200 --> 00:45:26,360
it doesn't need executive mailbox metadata.

1071
00:45:26,360 --> 00:45:28,800
So it cannot retrieve them even if a user asks.

1072
00:45:28,800 --> 00:45:33,520
And this is where the organization has to stop lying to itself about least privilege.

1073
00:45:33,520 --> 00:45:36,360
Least privilege is not a security best practice at agent scale.

1074
00:45:36,360 --> 00:45:40,520
It's the only way to keep the blast radius finite, finally defender because even with identity

1075
00:45:40,520 --> 00:45:42,680
and tool contracts behavior drifts.

1076
00:45:42,680 --> 00:45:47,560
So defender for AI baselines, the agents normal operating pattern, volume of ticket creation,

1077
00:45:47,560 --> 00:45:53,240
types of ticket actions, typical queues, typical time of day, typical user populations.

1078
00:45:53,240 --> 00:45:57,760
When the agent starts manipulating tickets in bulk or changing assignments unusually or

1079
00:45:57,760 --> 00:46:01,040
pulling sensitive artifacts, defender raises a signal you can act on.

1080
00:46:01,040 --> 00:46:03,600
And crucially, you can act on it without destroying the program.

1081
00:46:03,600 --> 00:46:07,880
You block the agent identity, you quarantine the tool, you disable a single capability,

1082
00:46:07,880 --> 00:46:09,640
the rest of the ecosystem continues.

1083
00:46:09,640 --> 00:46:12,160
That's what containment looks like when the control plane is real.

1084
00:46:12,160 --> 00:46:16,000
So the remediation isn't one big service desk agent that does everything.

1085
00:46:16,000 --> 00:46:21,560
It's a small portfolio of agent products within forced boundaries, identities you can disable,

1086
00:46:21,560 --> 00:46:26,400
tools you can govern, data sources you can restrict, and behaviors you can detect.

1087
00:46:26,400 --> 00:46:29,520
And when leadership asks, how do we scale this safely?

1088
00:46:29,520 --> 00:46:31,320
The answer is no longer a promise.

1089
00:46:31,320 --> 00:46:32,800
It's a demonstration.

1090
00:46:32,800 --> 00:46:37,880
Show the evidence chain for one-rooted ticket, enter identity, MCP tool call, purview bound

1091
00:46:37,880 --> 00:46:42,440
grounding, defender behavioral baseline, and the human approval step if the action is

1092
00:46:42,440 --> 00:46:43,640
irreversible.

1093
00:46:43,640 --> 00:46:47,920
Because once that chain exists, the service desk stops being the origin of sprawl.

1094
00:46:47,920 --> 00:46:51,120
It becomes the template for enforceable automation.

1095
00:46:51,120 --> 00:46:56,480
Next the part that still breaks even well-governed service desk agents, accuracy without provenance.

1096
00:46:56,480 --> 00:47:01,400
That's where policy and op-s knowledge turns into compliance fallout, even when the agents

1097
00:47:01,400 --> 00:47:04,120
write scenario two set up.

1098
00:47:04,120 --> 00:47:08,200
MCP backed policy plus ops agent and the provenance problem.

1099
00:47:08,200 --> 00:47:12,520
The second scenario is the one that makes leaders hate agents because it looks like competence

1100
00:47:12,520 --> 00:47:15,080
right up until it becomes compliance fallout.

1101
00:47:15,080 --> 00:47:17,400
A team builds a policy and operations agent.

1102
00:47:17,400 --> 00:47:18,760
The goal is clean.

1103
00:47:18,760 --> 00:47:22,280
Stop wasting senior people's time answering the same questions.

1104
00:47:22,280 --> 00:47:23,760
What's the change window?

1105
00:47:23,760 --> 00:47:25,160
What's the approved exception path?

1106
00:47:25,160 --> 00:47:28,680
What's the retention requirement for this data set?

1107
00:47:28,680 --> 00:47:30,560
Which policy applies to contractors?

1108
00:47:30,560 --> 00:47:31,560
It's all reasonable.

1109
00:47:31,560 --> 00:47:35,000
Repetitive and it's all the kind of work that feels like it should be automatable.

1110
00:47:35,000 --> 00:47:36,600
So they do it properly.

1111
00:47:36,600 --> 00:47:38,600
They ground the agent on policy libraries.

1112
00:47:38,600 --> 00:47:43,720
They connect MCP tools to pull live operational context from systems of record, ticketing,

1113
00:47:43,720 --> 00:47:48,120
CMDB, maybe a configuration repo, maybe a risk register, they structure outputs, they add

1114
00:47:48,120 --> 00:47:52,280
guardrails, they roll it out to a pilot group, and the agent performs beautifully.

1115
00:47:52,280 --> 00:47:56,600
It answers quickly, it answers confidently, it sounds consistent, it even sites sources

1116
00:47:56,600 --> 00:47:58,680
because somebody knew that citations matter.

1117
00:47:58,680 --> 00:48:00,960
The organization relaxes, that's the mistake.

1118
00:48:00,960 --> 00:48:03,680
Because the failure mode here isn't that the agent makes things up.

1119
00:48:03,680 --> 00:48:06,600
The failure mode is that it retrieves the wrong truth from the right place.

1120
00:48:06,600 --> 00:48:07,600
Here's the failure moment.

1121
00:48:07,600 --> 00:48:10,000
A user asks a question that looks simple.

1122
00:48:10,000 --> 00:48:13,640
Are we allowed to apply this policy exception for this scenario?

1123
00:48:13,640 --> 00:48:18,880
The agent pulls the relevant text, summarizes it, and replies, yes, here's the condition,

1124
00:48:18,880 --> 00:48:22,080
here's the justification, here's the process, the user follows the guidance.

1125
00:48:22,080 --> 00:48:23,440
They implement the exception.

1126
00:48:23,440 --> 00:48:24,440
Work continues.

1127
00:48:24,440 --> 00:48:29,400
Two weeks later an audit or a compliance review happens, or an incident triggers discovery.

1128
00:48:29,400 --> 00:48:33,880
And someone notices the exception was granted under an outdated policy version, or under a

1129
00:48:33,880 --> 00:48:38,240
departmental policy that was never meant to be enterprise-wide, or under a draft that was

1130
00:48:38,240 --> 00:48:42,400
published to a sharepoint site for review and never marked as non-authoritative, or under

1131
00:48:42,400 --> 00:48:45,280
a policy that applies to internal employees, not contractors.

1132
00:48:45,280 --> 00:48:47,320
The agent's answer was internally consistent.

1133
00:48:47,320 --> 00:48:49,200
It was also wrong in the only way that matters.

1134
00:48:49,200 --> 00:48:50,360
It wasn't authoritative.

1135
00:48:50,360 --> 00:48:53,960
So now the organization has a governance problem that's hard to explain because nobody can

1136
00:48:53,960 --> 00:48:57,160
point to malice, and nobody can point to a bug.

1137
00:48:57,160 --> 00:48:59,220
The agent did what it was built to do.

1138
00:48:59,220 --> 00:49:02,760
Retrieve content was allowed to access and produce a high quality summary.

1139
00:49:02,760 --> 00:49:06,360
This is why provenance is not a nice to have in the agent era.

1140
00:49:06,360 --> 00:49:08,280
Accuracy without provenance is still risk.

1141
00:49:08,280 --> 00:49:13,760
And enterprises confuse those two constantly because humans tend to evaluate answers by confidence

1142
00:49:13,760 --> 00:49:14,760
and fluency.

1143
00:49:14,760 --> 00:49:18,280
They don't evaluate by source authority unless they're already suspicious.

1144
00:49:18,280 --> 00:49:20,800
Agents amplify that bias because they speak like they know.

1145
00:49:20,800 --> 00:49:25,920
Now add MCP to the mix because MCP makes the blast radius bigger in a very specific way.

1146
00:49:25,920 --> 00:49:27,680
The agent doesn't just answer from documents.

1147
00:49:27,680 --> 00:49:32,080
It can also call tools that return policy like data from operational systems.

1148
00:49:32,080 --> 00:49:33,960
What's configured in production?

1149
00:49:33,960 --> 00:49:35,800
What's the current exception state?

1150
00:49:35,800 --> 00:49:37,480
What's the last change record?

1151
00:49:37,480 --> 00:49:40,400
What's the active retention policy in system X?

1152
00:49:40,400 --> 00:49:42,840
Those outputs look authoritative because they're live.

1153
00:49:42,840 --> 00:49:44,040
But live doesn't mean govern.

1154
00:49:44,040 --> 00:49:45,360
A CMDB record can be wrong.

1155
00:49:45,360 --> 00:49:46,760
A ticket can be mislabeled.

1156
00:49:46,760 --> 00:49:49,320
A config repo can contain legacy settings.

1157
00:49:49,320 --> 00:49:50,800
And MCP doesn't validate meaning.

1158
00:49:50,800 --> 00:49:52,240
It just standardizes access.

1159
00:49:52,240 --> 00:49:56,600
So the agent becomes a synthesis layer that merges three categories of information that humans

1160
00:49:56,600 --> 00:49:58,280
normally keep separate.

1161
00:49:58,280 --> 00:50:03,320
Written policy, operational reality and tribal knowledge embedded in tool outputs.

1162
00:50:03,320 --> 00:50:04,920
That synthesis is useful.

1163
00:50:04,920 --> 00:50:08,960
It is also where compliance dies if you don't enforce which sources are allowed to drive

1164
00:50:08,960 --> 00:50:12,200
decisions because the user isn't asking for bibliography.

1165
00:50:12,200 --> 00:50:15,320
They're asking for permission and the agent can accidentally grant it.

1166
00:50:15,320 --> 00:50:19,960
Now zoom out to the executive implication because this is where we'll iterate, stops working.

1167
00:50:19,960 --> 00:50:22,000
Trust collapses faster than adoption grows.

1168
00:50:22,000 --> 00:50:26,200
You can spend six months building momentum, training users, rolling out capabilities and

1169
00:50:26,200 --> 00:50:27,480
getting teams comfortable.

1170
00:50:27,480 --> 00:50:30,280
One provenance failure resets the trust model overnight.

1171
00:50:30,280 --> 00:50:33,360
The business stops asking, how do we use this and starts asking yet?

1172
00:50:33,360 --> 00:50:34,360
Can we rely on it?

1173
00:50:34,360 --> 00:50:36,600
Legal starts asking can we defend it?

1174
00:50:36,600 --> 00:50:39,200
Security starts asking, can we contain it?

1175
00:50:39,200 --> 00:50:43,840
And leadership starts asking, do we need to pause expansion until we can prove control?

1176
00:50:43,840 --> 00:50:48,360
This is the pattern that kills agent programs, not technical failure, but credibility failure.

1177
00:50:48,360 --> 00:50:52,360
Accredibility failure happens when the system can't prove that its answers came from the right

1178
00:50:52,360 --> 00:50:53,440
authority.

1179
00:50:53,440 --> 00:50:56,200
So in this scenario, the agent's correctness is a trap.

1180
00:50:56,200 --> 00:50:59,000
It's optimized for helpfulness, speed and clarity.

1181
00:50:59,000 --> 00:51:02,280
But the enterprise needs a different optimization target.

1182
00:51:02,280 --> 00:51:06,560
Authoritative knowledge, explicit provenance and enforceable boundaries that prevent non-authoritative

1183
00:51:06,560 --> 00:51:09,520
sources from being treated as truth.

1184
00:51:09,520 --> 00:51:11,040
That's the open loop to keep in mind.

1185
00:51:11,040 --> 00:51:13,600
The remediation isn't make it more accurate.

1186
00:51:13,600 --> 00:51:16,080
It's make it reliably authoritative.

1187
00:51:16,080 --> 00:51:20,480
And that means the next rule has to become non-negotiable across the ecosystem.

1188
00:51:20,480 --> 00:51:24,360
The agent must prefer authoritative knowledge over maximum knowledge because maximum knowledge

1189
00:51:24,360 --> 00:51:26,400
is just everything it can read.

1190
00:51:26,400 --> 00:51:29,760
And everything it can read is where the compliance incident is hiding.

1191
00:51:29,760 --> 00:51:31,000
Scenario 2.

1192
00:51:31,000 --> 00:51:32,800
Remediation.

1193
00:51:32,800 --> 00:51:33,920
Authoritative knowledge.

1194
00:51:33,920 --> 00:51:35,600
Not maximum knowledge.

1195
00:51:35,600 --> 00:51:39,520
Fixing the provenance problem doesn't start with better prompting.

1196
00:51:39,520 --> 00:51:42,480
It starts with admitting what the agent is actually doing.

1197
00:51:42,480 --> 00:51:43,480
It's not searching.

1198
00:51:43,480 --> 00:51:45,440
It's compiling an answer from whatever it can reach.

1199
00:51:45,440 --> 00:51:47,560
So the remediation rule is blunt.

1200
00:51:47,560 --> 00:51:51,800
The agent must prefer authoritative knowledge over maximum knowledge even when maximum knowledge

1201
00:51:51,800 --> 00:51:55,200
feels more helpful because helpful is what gets you sued.

1202
00:51:55,200 --> 00:52:00,440
The first move is purview, first classification and not in the abstract we labeled some stuff

1203
00:52:00,440 --> 00:52:01,600
since.

1204
00:52:01,600 --> 00:52:06,040
The organization needs an explicit definition of what counts as authoritative for each domain.

1205
00:52:06,040 --> 00:52:11,680
HR policy, security policy, operational runbooks, change management, finance procedures, legal

1206
00:52:11,680 --> 00:52:12,680
templates.

1207
00:52:12,680 --> 00:52:17,280
Authoritative means, owned, versioned, reviewed and intended to be relied on.

1208
00:52:17,280 --> 00:52:20,200
If the content doesn't meet that bar, it can exist.

1209
00:52:20,200 --> 00:52:22,520
But it can't be treated as truth by agents.

1210
00:52:22,520 --> 00:52:26,680
That distinction matters because the enterprise already has massive amounts of policy-shaped

1211
00:52:26,680 --> 00:52:28,600
content that is not policy.

1212
00:52:28,600 --> 00:52:33,080
Drafts, local copies, screenshots, meeting notes, wiki pages and email threads.

1213
00:52:33,080 --> 00:52:36,560
Humans can usually detect that difference because they recognize the context.

1214
00:52:36,560 --> 00:52:37,560
Agents can't.

1215
00:52:37,560 --> 00:52:38,960
They detect relevance, not authority.

1216
00:52:38,960 --> 00:52:43,080
If you don't enforce authority, the agent will optimize for most relevant text, which

1217
00:52:43,080 --> 00:52:45,000
is how a draft becomes a decision.

1218
00:52:45,000 --> 00:52:47,800
So you classify authoritative repositories as authoritative.

1219
00:52:47,800 --> 00:52:51,000
You mark non-authoritative repositories as non-authoritative.

1220
00:52:51,000 --> 00:52:53,960
And you stop pretending that it's in SharePoint makes it official.

1221
00:52:53,960 --> 00:52:55,200
Next is boundary enforcement.

1222
00:52:55,200 --> 00:52:57,200
You don't just tag content and hope.

1223
00:52:57,200 --> 00:52:59,800
You restrict what the agent is allowed to use as grounding.

1224
00:52:59,800 --> 00:53:04,200
That means the agent's identity gets access to the authoritative sources and it gets denied

1225
00:53:04,200 --> 00:53:06,240
access to everything else by default.

1226
00:53:06,240 --> 00:53:09,920
And read everything and side carefully, denied by design.

1227
00:53:09,920 --> 00:53:13,960
Because if the agent can read non-authoritative content, it will occasionally use it.

1228
00:53:13,960 --> 00:53:15,880
Not because it's reckless, because it is obedient.

1229
00:53:15,880 --> 00:53:18,440
Now, the second move is MCP discipline.

1230
00:53:18,440 --> 00:53:21,800
And this is where most teams accidentally reintroduce the problem.

1231
00:53:21,800 --> 00:53:24,680
Tools must return structured results with source identifiers.

1232
00:53:24,680 --> 00:53:29,920
If the agent calls an MCP tool to retrieve a policy or an operational rule, the tool responds

1233
00:53:29,920 --> 00:53:33,200
can't be a paragraph of text that looks authoritative.

1234
00:53:33,200 --> 00:53:38,760
It needs a schema, source system, document ID, version, effective date and classification.

1235
00:53:38,760 --> 00:53:42,160
If the result doesn't contain those fields, the agent can't prove provenance and your

1236
00:53:42,160 --> 00:53:43,920
back to sounds right.

1237
00:53:43,920 --> 00:53:47,320
This is also where you make outputs carry their own chain of custody.

1238
00:53:47,320 --> 00:53:49,160
The policy answer isn't just an answer.

1239
00:53:49,160 --> 00:53:50,760
It's an answer plus a receipt.

1240
00:53:50,760 --> 00:53:52,680
The receipt is what makes it defensible.

1241
00:53:52,680 --> 00:53:55,960
And no, citations aren't enough if citations can point to anything.

1242
00:53:55,960 --> 00:54:00,800
A link to a SharePoint page that got moved, duplicated or edited without governance is

1243
00:54:00,800 --> 00:54:01,800
not provenance.

1244
00:54:01,800 --> 00:54:06,320
Maintenance means this came from the authoritative system of record at this version effective

1245
00:54:06,320 --> 00:54:07,560
on this date.

1246
00:54:07,560 --> 00:54:10,800
Then enter enforcement turns that into a permission model that doesn't drift.

1247
00:54:10,800 --> 00:54:14,520
The policy agent identity gets scoped to the policy repositories.

1248
00:54:14,520 --> 00:54:16,080
Not everything it can read.

1249
00:54:16,080 --> 00:54:20,000
If teams want broader access to make it smarter, they're requesting the ability to blend

1250
00:54:20,000 --> 00:54:21,360
authority with noise.

1251
00:54:21,360 --> 00:54:22,360
That's not intelligence.

1252
00:54:22,360 --> 00:54:25,000
That's a compliance incident with better grammar.

1253
00:54:25,000 --> 00:54:28,160
And this is where the sponsor role becomes non-optional.

1254
00:54:28,160 --> 00:54:30,400
Someone in the business has to accept the trade off.

1255
00:54:30,400 --> 00:54:32,480
To throw a knowledge, higher reliability.

1256
00:54:32,480 --> 00:54:36,720
If the sponsor wants maximum knowledge, the sponsor is accepting maximum risk.

1257
00:54:36,720 --> 00:54:37,800
That's the actual decision.

1258
00:54:37,800 --> 00:54:38,800
Make it explicit.

1259
00:54:38,800 --> 00:54:40,400
Now add Defender monitoring.

1260
00:54:40,400 --> 00:54:43,000
Because even with boundaries, people will still try to bend the system.

1261
00:54:43,000 --> 00:54:44,400
You watch for two patterns.

1262
00:54:44,400 --> 00:54:46,600
One, unusual policy queries.

1263
00:54:46,600 --> 00:54:51,400
If a policy agent suddenly starts pulling large volumes of content or probing unusual domains

1264
00:54:51,400 --> 00:54:55,560
or repeatedly querying exceptions and bypass terms, something changed.

1265
00:54:55,560 --> 00:54:59,400
Either a user is trying to game it or the agent got new toolpaths or someone feted poison

1266
00:54:59,400 --> 00:55:00,400
context.

1267
00:55:00,400 --> 00:55:01,920
Some doesn't need to know which one first.

1268
00:55:01,920 --> 00:55:05,640
It needs to alert and contain two, mass retrieval.

1269
00:55:05,640 --> 00:55:09,840
The fastest way to turn a policy agent into an exfiltration tool is to ask it for all policies

1270
00:55:09,840 --> 00:55:12,080
related to X repeatedly and then export.

1271
00:55:12,080 --> 00:55:15,920
If you can detect that behavior early, you can block the agent identity, lock down the tool

1272
00:55:15,920 --> 00:55:18,000
and keep the rest of the ecosystem running.

1273
00:55:18,000 --> 00:55:19,560
And that's the actual remediation story.

1274
00:55:19,560 --> 00:55:21,800
You don't fix hallucinations.

1275
00:55:21,800 --> 00:55:23,000
You fix authority.

1276
00:55:23,000 --> 00:55:27,200
You reduce the agent's reachable surface area to what you're willing to defend.

1277
00:55:27,200 --> 00:55:31,800
You force every tool to return provenance in a structured way and you bind the whole thing

1278
00:55:31,800 --> 00:55:35,920
to an identity that can be contained when behavior drifts.

1279
00:55:35,920 --> 00:55:39,160
What this actually means is the agent stops being a search engine for your tenant and

1280
00:55:39,160 --> 00:55:41,480
becomes a governed interface to your official truth.

1281
00:55:41,480 --> 00:55:45,160
And once you do that, trust grows because trust has a mechanical basis.

1282
00:55:45,160 --> 00:55:47,320
The system can prove where it got the answer.

1283
00:55:47,320 --> 00:55:51,720
Next, the scalable pattern that makes this survivable across thousands of decisions

1284
00:55:51,720 --> 00:55:53,120
isn't more restrictions.

1285
00:55:53,120 --> 00:55:57,800
It's human approvals at the edge where irreversible actions and sensitive domains get a final

1286
00:55:57,800 --> 00:56:01,040
checkpoint without slowing everything to a crawl.

1287
00:56:01,040 --> 00:56:02,040
Scenario 3.

1288
00:56:02,040 --> 00:56:03,040
Set up.

1289
00:56:03,040 --> 00:56:05,480
Teams plus adaptive cards for approvals at scale.

1290
00:56:05,480 --> 00:56:10,120
Now the third scenario, because eventually every enterprise discovers the same limit.

1291
00:56:10,120 --> 00:56:14,600
You can't scale agent autonomy into irreversible actions and pretend governance will catch up

1292
00:56:14,600 --> 00:56:15,600
later.

1293
00:56:15,600 --> 00:56:19,000
So the pattern that actually survives is boring, repeatable and enforceable.

1294
00:56:19,000 --> 00:56:21,560
Teams plus adaptive cards for approvals at scale.

1295
00:56:21,560 --> 00:56:25,600
Not because teams is magical, because it's where work already happens and because the approval

1296
00:56:25,600 --> 00:56:29,840
surface is the only place you can reliably insert human accountability without sending

1297
00:56:29,840 --> 00:56:32,040
people into a separate portal.

1298
00:56:32,040 --> 00:56:33,040
Nobody opens.

1299
00:56:33,040 --> 00:56:34,720
Here's the setup.

1300
00:56:34,720 --> 00:56:36,480
An agent doesn't decide.

1301
00:56:36,480 --> 00:56:37,920
It prepares a decision.

1302
00:56:37,920 --> 00:56:38,920
It gathers context.

1303
00:56:38,920 --> 00:56:40,160
It normalizes input.

1304
00:56:40,160 --> 00:56:41,560
It checks policy boundaries.

1305
00:56:41,560 --> 00:56:42,680
It proposes an action.

1306
00:56:42,680 --> 00:56:43,680
And then it stops.

1307
00:56:43,680 --> 00:56:47,160
That stop is the entire point because the enterprise doesn't need agents that can do anything.

1308
00:56:47,160 --> 00:56:51,200
It needs agents that can do bounded things quickly with humans approving outcomes when

1309
00:56:51,200 --> 00:56:55,280
the action is irreversible, sensitive or politically explosive.

1310
00:56:55,280 --> 00:56:56,640
So the workflow looks like this.

1311
00:56:56,640 --> 00:57:00,040
A user asks for something that has a real blast radius.

1312
00:57:00,040 --> 00:57:01,680
Access to a restricted site.

1313
00:57:01,680 --> 00:57:03,320
An exception to a standard control.

1314
00:57:03,320 --> 00:57:04,920
A new vendor on board.

1315
00:57:04,920 --> 00:57:06,160
A mailbox delegation.

1316
00:57:06,160 --> 00:57:07,160
A data export.

1317
00:57:07,160 --> 00:57:08,160
A policy waiver.

1318
00:57:08,160 --> 00:57:09,680
A routing override.

1319
00:57:09,680 --> 00:57:13,360
The agent collects the missing fields that humans always forget.

1320
00:57:13,360 --> 00:57:14,600
Business justification.

1321
00:57:14,600 --> 00:57:15,920
Duration.

1322
00:57:15,920 --> 00:57:17,560
Impacted systems.

1323
00:57:17,560 --> 00:57:21,640
A user identity.

1324
00:57:21,640 --> 00:57:24,400
Then the agent package is that into an adaptive card.

1325
00:57:24,400 --> 00:57:26,880
And the adaptive card is not just nice UI.

1326
00:57:26,880 --> 00:57:28,040
It's a constrained mechanism.

1327
00:57:28,040 --> 00:57:29,640
It forces structured input.

1328
00:57:29,640 --> 00:57:33,560
It prevents the sure-go ahead approval that arrives as free text with no evidence.

1329
00:57:33,560 --> 00:57:37,640
It standardizes the decision shape, what was asked, what was recommended, what policy it

1330
00:57:37,640 --> 00:57:39,560
maps to and what the consequences are.

1331
00:57:39,560 --> 00:57:42,400
And critically, it keeps the approval in the flow of work.

1332
00:57:42,400 --> 00:57:43,400
No context switch.

1333
00:57:43,400 --> 00:57:44,800
No hunting for a link.

1334
00:57:44,800 --> 00:57:47,920
No separate approval portal with its own identity problems.

1335
00:57:47,920 --> 00:57:49,920
It shows up where they already live.

1336
00:57:49,920 --> 00:57:51,160
Teams.

1337
00:57:51,160 --> 00:57:53,920
Now the failure mode this avoids is subtle.

1338
00:57:53,920 --> 00:57:54,920
Approval theater.

1339
00:57:54,920 --> 00:57:57,080
Most organizations claim they have approvals.

1340
00:57:57,080 --> 00:58:02,080
But the approvals exist as emails or as screenshots or as the manager said yes and chat.

1341
00:58:02,080 --> 00:58:03,400
None of that is enforceable.

1342
00:58:03,400 --> 00:58:05,120
None of that is reliably auditable.

1343
00:58:05,120 --> 00:58:09,040
And none of that scales because it can't be correlated back to an action chain.

1344
00:58:09,040 --> 00:58:13,480
Adaptive cards fix that by making approvals a first class event in the workflow, not

1345
00:58:13,480 --> 00:58:14,680
an afterthought.

1346
00:58:14,680 --> 00:58:18,960
The agent sends the card to the right approver based on a deterministic rule.

1347
00:58:18,960 --> 00:58:21,640
The approver sees a bounded set of options.

1348
00:58:21,640 --> 00:58:23,880
Approve, reject, request clarification.

1349
00:58:23,880 --> 00:58:27,960
If they request clarification, the agent collects the missing information and resubmits.

1350
00:58:27,960 --> 00:58:30,640
If they reject, the workflow ends with a reason captured.

1351
00:58:30,640 --> 00:58:35,200
If they approve, the agent executes only the bounded action it was designed for.

1352
00:58:35,200 --> 00:58:36,200
That's how it scales.

1353
00:58:36,200 --> 00:58:37,200
Humans approve outcomes.

1354
00:58:37,200 --> 00:58:39,040
Agents execute bounded steps.

1355
00:58:39,040 --> 00:58:43,240
And the reason this scales operationally is that it reduces context switching in the exact

1356
00:58:43,240 --> 00:58:45,640
places where enterprises bleed time.

1357
00:58:45,640 --> 00:58:49,160
Approvers don't want to read a paragraph, interpret it and then remember what system they

1358
00:58:49,160 --> 00:58:50,880
need to log into to do the thing.

1359
00:58:50,880 --> 00:58:55,360
They want a decision interface that is one screen, one minute and one accountable click.

1360
00:58:55,360 --> 00:58:57,800
Adaptive cards give you that micro experience.

1361
00:58:57,800 --> 00:59:03,440
Now the real payoff is the audit trail because this is where the control plane becomes visible.

1362
00:59:03,440 --> 00:59:07,920
With this approval pattern, an enterprise can produce a clean evidence chain per action.

1363
00:59:07,920 --> 00:59:12,260
The agent identity that prepared the request, the user identity that initiated it, the

1364
00:59:12,260 --> 00:59:16,700
approval identity that authorized it, the tool call that executed it and the timestamp

1365
00:59:16,700 --> 00:59:17,700
for each step.

1366
00:59:17,700 --> 00:59:22,340
Identity plus approval plus tool call plus timestamp, that becomes the default.

1367
00:59:22,340 --> 00:59:25,420
And once that is the default, you stop arguing about whether the agent should have done

1368
00:59:25,420 --> 00:59:26,420
it.

1369
00:59:26,420 --> 00:59:29,020
You have the record, you have the boundary, you have the accountability.

1370
00:59:29,020 --> 00:59:33,500
This is also where the governance layers interlock without being abstract.

1371
00:59:33,500 --> 00:59:34,500
Entra agent ID.

1372
00:59:34,500 --> 00:59:38,300
The agent has a stable identity that prepared the request and executed the tool call,

1373
00:59:38,300 --> 00:59:39,300
MCP.

1374
00:59:39,300 --> 00:59:43,280
The execution happens through standardized tools with structured outputs that record what

1375
00:59:43,280 --> 00:59:44,280
happened.

1376
00:59:44,280 --> 00:59:48,460
Per view, the agent can't include sensitive data in the approval context if the boundary

1377
00:59:48,460 --> 00:59:53,020
blocks it and it can't ground on non-authoritative sources without being allowed.

1378
00:59:53,020 --> 00:59:54,020
Defender.

1379
00:59:54,020 --> 00:59:58,580
If the agent starts generating approvals at a weird rate or targeting unusual approvals

1380
00:59:58,580 --> 01:00:02,820
or chaining actions outside baseline, you detect it and contain it, so the approval

1381
01:00:02,820 --> 01:00:04,220
surface is an extra.

1382
01:00:04,220 --> 01:00:06,540
It's the seam wear enforcement becomes real.

1383
01:00:06,540 --> 01:00:08,260
Now one more uncomfortable truth.

1384
01:00:08,260 --> 01:00:11,900
This pattern forces you to admit that humans remain the accountability layer.

1385
01:00:11,900 --> 01:00:15,700
The enterprise can't outsource responsibility to a probabilistic system and then blame the

1386
01:00:15,700 --> 01:00:17,380
system when it acts.

1387
01:00:17,380 --> 01:00:20,300
Adaptive cards make that responsibility explicit and rootable.

1388
01:00:20,300 --> 01:00:24,060
And the organizations that scale cleanly are the ones that embrace that operating model

1389
01:00:24,060 --> 01:00:25,460
shift early.

1390
01:00:25,460 --> 01:00:28,060
Humans lead, agents operate, systems learn.

1391
01:00:28,060 --> 01:00:31,980
Next, that operating model because the tooling doesn't save you if you still treat agents

1392
01:00:31,980 --> 01:00:34,580
like side projects instead of products.

1393
01:00:34,580 --> 01:00:38,500
The operating model from build agents to run agent products.

1394
01:00:38,500 --> 01:00:41,580
Most enterprises fail at agent scale for the most boring reason.

1395
01:00:41,580 --> 01:00:43,060
They treat agents like projects.

1396
01:00:43,060 --> 01:00:47,620
The project has a start date, a finish date, a hand off and a slide that says delivered.

1397
01:00:47,620 --> 01:00:51,180
Then everyone moves on and the system starts drifting immediately.

1398
01:00:51,180 --> 01:00:53,780
Agents don't work like that and agent is not a build artifact.

1399
01:00:53,780 --> 01:00:55,020
It is a running product.

1400
01:00:55,020 --> 01:00:59,180
It has users, behaviors, dependencies and failure modes that evolve under pressure.

1401
01:00:59,180 --> 01:01:02,660
That distinction matters because the moment the organization ships an agent and walks

1402
01:01:02,660 --> 01:01:05,100
away, entropy starts writing the roadmap.

1403
01:01:05,100 --> 01:01:09,060
So the operating model has to change from build agents to run agent products.

1404
01:01:09,060 --> 01:01:10,260
A product has ownership.

1405
01:01:10,260 --> 01:01:11,260
It has a sponsor.

1406
01:01:11,260 --> 01:01:12,260
It has a backlog.

1407
01:01:12,260 --> 01:01:13,260
It has a deprecation plan.

1408
01:01:13,260 --> 01:01:16,740
It has an evidence trail and it has an explicit answer to the question executives will

1409
01:01:16,740 --> 01:01:18,660
ask after the first incident.

1410
01:01:18,660 --> 01:01:20,820
Who is accountable for this system's behavior?

1411
01:01:20,820 --> 01:01:24,060
If you can't answer that in one sentence, you don't have a product.

1412
01:01:24,060 --> 01:01:25,540
You have a liability.

1413
01:01:25,540 --> 01:01:27,020
Start with the simplest rule.

1414
01:01:27,020 --> 01:01:29,900
Every enterprise grade agent needs an owner and a sponsor.

1415
01:01:29,900 --> 01:01:32,340
The owner is responsible for the technical shape.

1416
01:01:32,340 --> 01:01:37,100
Instructions, tool contracts, identity scope, monitoring and change control.

1417
01:01:37,100 --> 01:01:39,780
The sponsor is responsible for the business outcome.

1418
01:01:39,780 --> 01:01:41,380
Why the agent exists?

1419
01:01:41,380 --> 01:01:43,060
What decisions it influences?

1420
01:01:43,060 --> 01:01:46,180
What risk it is allowed to carry and what happens when it fails?

1421
01:01:46,180 --> 01:01:49,780
Without a sponsor, the agent will drift toward whatever users ask for because that's the

1422
01:01:49,780 --> 01:01:51,740
path of least resistance.

1423
01:01:51,740 --> 01:01:56,260
And whatever users ask for is how you end up with agents that accidentally become policy

1424
01:01:56,260 --> 01:01:59,820
authorities, data brokers and shadow administrators.

1425
01:01:59,820 --> 01:02:05,700
And define life cycle workflows that treat agents like employees with badges, not like scripts.

1426
01:02:05,700 --> 01:02:10,300
Provisioning, identity is created, tools are approved, data boundaries are validated, monitoring

1427
01:02:10,300 --> 01:02:15,380
is enabled, operation, usage is tracked, behaviors are baseline, incidents are contained, outputs

1428
01:02:15,380 --> 01:02:17,380
are reviewed for provenance failures.

1429
01:02:17,380 --> 01:02:22,180
Change, tool contracts version, prompts update, model behavior shifts, connectors evolve

1430
01:02:22,180 --> 01:02:25,660
and someone signs off that the blast radius remains acceptable.

1431
01:02:25,660 --> 01:02:26,660
Retirement.

1432
01:02:26,660 --> 01:02:31,340
The agent is deprecated, access is removed, identities are disabled and the registry records

1433
01:02:31,340 --> 01:02:32,660
wide was shut down.

1434
01:02:32,660 --> 01:02:35,260
If any of those steps are optional, they won't happen.

1435
01:02:35,260 --> 01:02:36,940
Optional controls become folklore.

1436
01:02:36,940 --> 01:02:38,100
Folklore doesn't survive scale.

1437
01:02:38,100 --> 01:02:42,260
Now the roles, this is where organizations love to overcomplicate and call it governance.

1438
01:02:42,260 --> 01:02:45,220
The required roles are minimal and non-negotiable.

1439
01:02:45,220 --> 01:02:48,300
Platform, security, compliance and a business sponsor.

1440
01:02:48,300 --> 01:02:52,780
Platform owns the shared services, the curated tool catalog, MCP contract discipline, publishing

1441
01:02:52,780 --> 01:02:54,460
gates and the registry.

1442
01:02:54,460 --> 01:02:57,540
The owns identity boundaries and behavioral detection.

1443
01:02:57,540 --> 01:03:02,300
The entry identity model, conditional access decisions, defender baselines and containment

1444
01:03:02,300 --> 01:03:03,820
playbooks.

1445
01:03:03,820 --> 01:03:08,540
Compliance owns the data boundary, purview classification, authoritative source designation

1446
01:03:08,540 --> 01:03:09,980
and exposure reporting.

1447
01:03:09,980 --> 01:03:14,780
The business sponsor owns intent which outcomes matter what's acceptable and what isn't.

1448
01:03:14,780 --> 01:03:18,020
If any one of these roles is missing, you will still ship agents.

1449
01:03:18,020 --> 01:03:20,020
You will also ship contradictory authority.

1450
01:03:20,020 --> 01:03:21,940
It will feel fast, it will also be fragile.

1451
01:03:21,940 --> 01:03:26,300
It now comes the piece most organizations avoid because it forces transparency, a registry.

1452
01:03:26,300 --> 01:03:29,220
A registry isn't a catalog that exists to look mature.

1453
01:03:29,220 --> 01:03:33,260
It is the system of record for why an agent exists, who owns it, what it can do, what

1454
01:03:33,260 --> 01:03:35,500
it can touch and which tools it can call.

1455
01:03:35,500 --> 01:03:38,660
It answers the user question, which agent should I use?

1456
01:03:38,660 --> 01:03:41,260
And before that question becomes a productivity tax.

1457
01:03:41,260 --> 01:03:46,460
It answers the auditor question, which identities exist and what are they authorized to do?

1458
01:03:46,460 --> 01:03:48,580
Before that question becomes a program freeze.

1459
01:03:48,580 --> 01:03:52,700
And it answers the architecture question, how many overlapping agents do we have that solve

1460
01:03:52,700 --> 01:03:54,700
the same problem differently?

1461
01:03:54,700 --> 01:03:56,420
Before decision that becomes permanent.

1462
01:03:56,420 --> 01:04:00,660
The registry mindset also forces the enterprise to define discoverability properly.

1463
01:04:00,660 --> 01:04:03,260
Agents shouldn't be discovered by rumor and team's links.

1464
01:04:03,260 --> 01:04:06,820
They should be discoverable by purpose, data domain and allowed actions.

1465
01:04:06,820 --> 01:04:11,340
And every agent entry should state one uncomfortable truth plainly, the agent's boundary, what

1466
01:04:11,340 --> 01:04:14,380
it will not do, what it cannot access, what requires human approval.

1467
01:04:14,380 --> 01:04:18,940
And comes the entropy law because this is where governance becomes real, exceptions accumulate.

1468
01:04:18,940 --> 01:04:21,020
A team asks for just one more connector.

1469
01:04:21,020 --> 01:04:23,900
A leader wants just this one sensitive site.

1470
01:04:23,900 --> 01:04:26,980
A pilot needs temporary access that never gets removed.

1471
01:04:26,980 --> 01:04:32,260
A tool needs right permissions because it failed once and somebody unblocked it to hit a deadline.

1472
01:04:32,260 --> 01:04:34,260
Each exception is an entropy generator.

1473
01:04:34,260 --> 01:04:38,420
Over time, the system drifts away from the original intent, not because anyone chose chaos

1474
01:04:38,420 --> 01:04:43,020
but because the platform allowed intent to be negotiated instead of enforced.

1475
01:04:43,020 --> 01:04:46,980
So the operating model has one job, enforce assumptions at scale, not by telling people

1476
01:04:46,980 --> 01:04:51,220
to behave, by designing the control plane and the life cycle gates so that unsafe drift

1477
01:04:51,220 --> 01:04:52,940
is harder than safe iteration.

1478
01:04:52,940 --> 01:04:56,540
And when the organization does that, something surprising happens, innovation doesn't die.

1479
01:04:56,540 --> 01:05:00,260
It gets rooted into patterns that are reusable, auditable and containable, which is the only

1480
01:05:00,260 --> 01:05:03,460
kind of innovation that survives the agent decade.

1481
01:05:03,460 --> 01:05:07,300
Risk three, cost and decision debt, the slow burn failure.

1482
01:05:07,300 --> 01:05:10,940
Cost is the risk nobody leads with because it sounds like procurement, but it's the

1483
01:05:10,940 --> 01:05:15,340
forcing function that kills programs after the first wave of enthusiasm.

1484
01:05:15,340 --> 01:05:18,540
Security incidents pause you fast, cost pauses you permanently.

1485
01:05:18,540 --> 01:05:22,580
And the ugly part is that agent costs don't show up where people expect.

1486
01:05:22,580 --> 01:05:28,060
They don't show up as AI spend, they show up as a distributed tax across tokens, tool calls,

1487
01:05:28,060 --> 01:05:32,420
premium connectors, additional environments, extra monitoring and the one cost nobody budgets

1488
01:05:32,420 --> 01:05:33,420
for.

1489
01:05:33,420 --> 01:05:35,340
Human escalation when the agent can't finish.

1490
01:05:35,340 --> 01:05:37,540
So the enterprise thinks it's buying productivity.

1491
01:05:37,540 --> 01:05:39,780
What it's actually buying is a new kind of runtime.

1492
01:05:39,780 --> 01:05:43,980
Every agent you add becomes a consumer of compute, data access and decision making bandwidth.

1493
01:05:43,980 --> 01:05:45,940
Not once, continuously.

1494
01:05:45,940 --> 01:05:50,500
And because agents encourage more usage because they make it easy, your program doesn't scale

1495
01:05:50,500 --> 01:05:51,500
linearly.

1496
01:05:51,500 --> 01:05:55,180
It scales with demand you didn't previously see because humans avoided the friction.

1497
01:05:55,180 --> 01:05:57,220
Remove the friction and the request volume spikes.

1498
01:05:57,220 --> 01:05:58,460
This is the token side.

1499
01:05:58,460 --> 01:06:00,700
Every conversation is not a chat.

1500
01:06:00,700 --> 01:06:04,460
It's a chain, grounding retrieval, tool calls, summarization, follow-ups, retreats and

1501
01:06:04,460 --> 01:06:06,260
sometimes multi agent chaining.

1502
01:06:06,260 --> 01:06:11,380
The user sees one answer, the platform sees a loop, then the tool side.

1503
01:06:11,380 --> 01:06:17,340
Every tool call has a cost profile, API quotas, transaction charges, connector licensing,

1504
01:06:17,340 --> 01:06:19,540
downstream system load and audit storage.

1505
01:06:19,540 --> 01:06:22,660
If you let agents call tools freely, they will call tools freely.

1506
01:06:22,660 --> 01:06:26,220
Not because they're wasteful, because they're optimized to complete tasks and completion

1507
01:06:26,220 --> 01:06:29,260
often means try again with more context.

1508
01:06:29,260 --> 01:06:31,580
Retries aren't free, they become the default.

1509
01:06:31,580 --> 01:06:35,260
Now add premium connectors and licensing boundaries because this is where finance starts paying

1510
01:06:35,260 --> 01:06:36,220
attention.

1511
01:06:36,220 --> 01:06:40,500
Sometimes build agents in co-pilot studio connect to a premium system and nobody remembers

1512
01:06:40,500 --> 01:06:42,580
that one more connector has a bill.

1513
01:06:42,580 --> 01:06:48,140
Or worse, it has a per flow, per run, per capacity shape that only becomes obvious once

1514
01:06:48,140 --> 01:06:49,140
adoption grows.

1515
01:06:49,140 --> 01:06:50,660
The platform didn't trick you.

1516
01:06:50,660 --> 01:06:54,740
You just didn't model the cost per completed task and that's the metric that matters.

1517
01:06:54,740 --> 01:06:55,940
Not cost per chat.

1518
01:06:55,940 --> 01:06:58,900
A chat that ends in escalation is not a completed task.

1519
01:06:58,900 --> 01:07:02,180
A chat that creates two duplicate tickets is not a completed task.

1520
01:07:02,180 --> 01:07:06,140
A chat that retrieves 10 documents and still answers from the wrong policy source is not

1521
01:07:06,140 --> 01:07:07,460
a completed task.

1522
01:07:07,460 --> 01:07:10,820
Counting chats makes dashboards look good, it does not keep the program funded.

1523
01:07:10,820 --> 01:07:13,820
Now the second part of this risk is harder to see.

1524
01:07:13,820 --> 01:07:15,260
And it's more corrosive.

1525
01:07:15,260 --> 01:07:16,260
Decision debt.

1526
01:07:16,260 --> 01:07:19,580
Decision debt is what happens when multiple agents solve the same problem differently at

1527
01:07:19,580 --> 01:07:22,620
the same time with different assumptions and different toolchains.

1528
01:07:22,620 --> 01:07:26,500
It's the agent version of shadow IT except it doesn't just create duplicate apps.

1529
01:07:26,500 --> 01:07:28,100
It creates duplicate answers.

1530
01:07:28,100 --> 01:07:29,820
And duplicate answers create rework.

1531
01:07:29,820 --> 01:07:32,140
A user asks one agent for the policy path.

1532
01:07:32,140 --> 01:07:33,700
Other agent gives a different path.

1533
01:07:33,700 --> 01:07:37,100
The user pings a human, the human picks one, then the agent gets blamed anyway.

1534
01:07:37,100 --> 01:07:38,900
That's not AI unreliability.

1535
01:07:38,900 --> 01:07:40,900
That's an inconsistent system design.

1536
01:07:40,900 --> 01:07:42,580
Decision debt compounds in three ways.

1537
01:07:42,580 --> 01:07:45,780
First, overlapping agents create inconsistent outcomes.

1538
01:07:45,780 --> 01:07:49,580
The organization spends time reconciling contradictions instead of competing work.

1539
01:07:49,580 --> 01:07:52,540
Second, duplicated tool contracts create duplicated maintenance.

1540
01:07:52,540 --> 01:07:56,020
Even when MCP exists, teams still rebuild because they can.

1541
01:07:56,020 --> 01:07:59,020
That doubles cost and triples failure surface.

1542
01:07:59,020 --> 01:08:03,340
Third, exception growth turns every special case into a custom path.

1543
01:08:03,340 --> 01:08:08,500
Each custom path requires more tokens, more tool calls, more approvals, more monitoring,

1544
01:08:08,500 --> 01:08:10,180
and more incident response.

1545
01:08:10,180 --> 01:08:11,900
Exceptions aren't operational flexibility.

1546
01:08:11,900 --> 01:08:13,820
They're recurring invoices.

1547
01:08:13,820 --> 01:08:17,260
This is why cost and governance aren't separate conversations.

1548
01:08:17,260 --> 01:08:19,220
Cost is governance failing slowly.

1549
01:08:19,220 --> 01:08:23,460
And cost becomes visible when finance asks the only question that doesn't care about

1550
01:08:23,460 --> 01:08:25,060
your architecture narrative.

1551
01:08:25,060 --> 01:08:28,260
What are we paying for and what are we getting back?

1552
01:08:28,260 --> 01:08:30,620
The predictable trajectory is always the same.

1553
01:08:30,620 --> 01:08:31,620
Phase one.

1554
01:08:31,620 --> 01:08:32,620
Invoices are small.

1555
01:08:32,620 --> 01:08:33,620
Adoption is exciting.

1556
01:08:33,620 --> 01:08:34,620
Leadership is optimistic.

1557
01:08:34,620 --> 01:08:35,620
Phase two.

1558
01:08:35,620 --> 01:08:36,620
Usage grows.

1559
01:08:36,620 --> 01:08:37,620
More agents appear.

1560
01:08:37,620 --> 01:08:38,620
Tool calls increase.

1561
01:08:38,620 --> 01:08:40,420
And the first cost spike hits.

1562
01:08:40,420 --> 01:08:41,420
Phase three.

1563
01:08:41,420 --> 01:08:42,420
Someone panics.

1564
01:08:42,420 --> 01:08:43,420
The program gets frozen.

1565
01:08:43,420 --> 01:08:48,180
Not because agents didn't work, but because nobody can prove which spend created which outcome.

1566
01:08:48,180 --> 01:08:50,460
And nobody can guarantee the next quarter won't be worse.

1567
01:08:50,460 --> 01:08:51,780
This is the slow burn failure.

1568
01:08:51,780 --> 01:08:53,340
It doesn't come from one bad agent.

1569
01:08:53,340 --> 01:08:55,060
It comes from unmanaged multiplication.

1570
01:08:55,060 --> 01:08:59,620
And if you want a practical warning sign before the CFO does it for you, it's this.

1571
01:08:59,620 --> 01:09:01,380
When teams can't answer.

1572
01:09:01,380 --> 01:09:05,740
In one sentence, which agent owns which outcome, you already have decision debt.

1573
01:09:05,740 --> 01:09:09,820
When you can't tie tool usage to completed tasks, you already have cost drift.

1574
01:09:09,820 --> 01:09:13,940
When you can't deprecate agents cleanly, you've turned your ecosystem into a permanent subscription

1575
01:09:13,940 --> 01:09:14,940
to entropy.

1576
01:09:14,940 --> 01:09:18,500
Next, the metrics that stop this from becoming a budget driven shutdown.

1577
01:09:18,500 --> 01:09:22,380
Four metrics executives actually fund, because they translate enforceable intelligence

1578
01:09:22,380 --> 01:09:23,900
into defendable outcomes.

1579
01:09:23,900 --> 01:09:26,500
The four metrics executives actually fund.

1580
01:09:26,500 --> 01:09:28,580
Executives don't fund agent platforms.

1581
01:09:28,580 --> 01:09:30,820
They fund outcomes that survive scrutiny.

1582
01:09:30,820 --> 01:09:34,420
So if you want the program to live past the first invoice spike, the first audit question

1583
01:09:34,420 --> 01:09:39,380
or the first incident review, you need four metrics that don't collapse into marketing.

1584
01:09:39,380 --> 01:09:43,300
Four metrics that force an evidence chain, not adoption, not number of agents, not hours

1585
01:09:43,300 --> 01:09:44,940
saved from a survey.

1586
01:09:44,940 --> 01:09:46,700
Metrics that map to operating reality.

1587
01:09:46,700 --> 01:09:49,020
First MTR reduction, mean time to resolution.

1588
01:09:49,020 --> 01:09:51,740
This is the easiest executive win because it's measurable.

1589
01:09:51,740 --> 01:09:55,180
It's already tracked and service operations already feel the pain.

1590
01:09:55,180 --> 01:09:59,980
But the mistake teams make is treating MTR as a tool problem instead of a routing problem.

1591
01:09:59,980 --> 01:10:03,940
Clean payloads reduce MTR, deterministic routing reduces MTR.

1592
01:10:03,940 --> 01:10:05,740
Standardized tool actions reduce MTR.

1593
01:10:05,740 --> 01:10:08,500
And the control plane is what makes those three things repeatable.

1594
01:10:08,500 --> 01:10:13,300
A typical mature agent rollout that enforces routing and tool reuse sees MTR drop in

1595
01:10:13,300 --> 01:10:16,380
the 20 to 35% range on targeted workflows.

1596
01:10:16,380 --> 01:10:17,380
That's not magic.

1597
01:10:17,380 --> 01:10:22,580
Moving retriage, removing duplication and preventing tickets from bouncing between cues.

1598
01:10:22,580 --> 01:10:27,220
And the executive story becomes simple, fewer handoffs, less noise, faster closure.

1599
01:10:27,220 --> 01:10:29,660
Second, request to decision time.

1600
01:10:29,660 --> 01:10:34,180
This is where agents matter outside IT because every enterprise has decision loops that waste

1601
01:10:34,180 --> 01:10:35,180
days.

1602
01:10:35,180 --> 01:10:39,380
Access approvals, exception handling, policy waivers, change requests, vendor onboarding,

1603
01:10:39,380 --> 01:10:40,580
procurement checks.

1604
01:10:40,580 --> 01:10:43,700
The work itself isn't hard, the waiting is.

1605
01:10:43,700 --> 01:10:45,500
Agents don't eliminate governance.

1606
01:10:45,500 --> 01:10:49,620
They compress the latency, they collect the missing fields, they package the context,

1607
01:10:49,620 --> 01:10:53,020
they route to the correct approver and they preserve the evidence.

1608
01:10:53,020 --> 01:10:55,020
So the measure isn't how many chats.

1609
01:10:55,020 --> 01:10:58,580
It's how long it takes to go from request created to decision recorded.

1610
01:10:58,580 --> 01:11:02,820
In well governed approval patterns, it's normal to see a 40 to 60% reduction.

1611
01:11:02,820 --> 01:11:06,660
Days to hours, sometimes hours to minutes because humans stop doing the administrative

1612
01:11:06,660 --> 01:11:07,900
choreography.

1613
01:11:07,900 --> 01:11:11,180
They just approve or reject with context already assembled.

1614
01:11:11,180 --> 01:11:13,540
That's the difference between an agent that helps.

1615
01:11:13,540 --> 01:11:16,340
And an agent ecosystem that actually changes throughput.

1616
01:11:16,340 --> 01:11:17,740
Third, auditability.

1617
01:11:17,740 --> 01:11:20,220
This is the metric nobody loves until they need it.

1618
01:11:20,220 --> 01:11:22,900
And then it becomes the only metric that matters.

1619
01:11:22,900 --> 01:11:23,900
Auditability is binary.

1620
01:11:23,900 --> 01:11:27,380
Either you can produce an evidence chain per action or you can't.

1621
01:11:27,380 --> 01:11:29,820
Before, we think the bot did X.

1622
01:11:29,820 --> 01:11:34,540
After, here is the identity, the approval, the tool call and the timestamp.

1623
01:11:34,540 --> 01:11:37,020
That is what keeps the programmer live during scrutiny.

1624
01:11:37,020 --> 01:11:39,580
Because audits don't ask whether the answer sounded correct.

1625
01:11:39,580 --> 01:11:43,300
They ask who performed the action under what authority using what data with what approvals

1626
01:11:43,300 --> 01:11:45,580
and whether controls were bypassed.

1627
01:11:45,580 --> 01:11:49,380
Auditability is where an agent ID stops being an identity feature and becomes a survival

1628
01:11:49,380 --> 01:11:50,380
feature.

1629
01:11:50,380 --> 01:11:51,380
It gives you stable attribution.

1630
01:11:51,380 --> 01:11:53,620
Per view gives you exposure context.

1631
01:11:53,620 --> 01:11:55,340
Defender gives you behavioral signals.

1632
01:11:55,340 --> 01:11:58,540
MCP gives you the tool path that can be inspected in version.

1633
01:11:58,540 --> 01:12:02,260
So the audit story becomes a replay, not a narrative.

1634
01:12:02,260 --> 01:12:04,100
And executives understand this immediately.

1635
01:12:04,100 --> 01:12:06,500
If you can't prove it, you can't scale it.

1636
01:12:06,500 --> 01:12:08,180
Fourth, cost per completed task.

1637
01:12:08,180 --> 01:12:10,180
This is where most agent programs die.

1638
01:12:10,180 --> 01:12:14,060
This team report costs per chat and finance asks the only question that matters.

1639
01:12:14,060 --> 01:12:15,740
What did the chat actually complete?

1640
01:12:15,740 --> 01:12:21,060
A completed task means an outcome that didn't require a second agent, a human retriage,

1641
01:12:21,060 --> 01:12:24,020
a manual rework or an incident cleanup later.

1642
01:12:24,020 --> 01:12:26,780
If you don't define completion, you can't manage cost.

1643
01:12:26,780 --> 01:12:31,740
The organizations that standardized tool contracts and reuse them through MCP often see 25

1644
01:12:31,740 --> 01:12:36,580
to 50% lower cost per resolved request on mature workflows, not because tokens got

1645
01:12:36,580 --> 01:12:40,940
cheaper because the system stopped paying for duplicated integrations, duplicated retries

1646
01:12:40,940 --> 01:12:42,620
and duplicated human escalation.

1647
01:12:42,620 --> 01:12:45,060
And here's the thing that makes all four metrics credible.

1648
01:12:45,060 --> 01:12:46,380
They aren't independent.

1649
01:12:46,380 --> 01:12:50,340
MTTR drops when request to decision time drops because fewer things stall.

1650
01:12:50,340 --> 01:12:54,500
Auditability improves when tool reuse improves because evidence becomes structured.

1651
01:12:54,500 --> 01:12:58,740
Cost per completed task drops when routing becomes deterministic because the system stops

1652
01:12:58,740 --> 01:13:00,340
doing the same work twice.

1653
01:13:00,340 --> 01:13:02,340
So the goal isn't to chase four dashboards.

1654
01:13:02,340 --> 01:13:06,140
The goal is to enforce the control plane so the metrics fall out as a consequence.

1655
01:13:06,140 --> 01:13:09,900
This is also why executives should demand that every agent product has a measurement contract

1656
01:13:09,900 --> 01:13:12,460
before it ships, not will measure later.

1657
01:13:12,460 --> 01:13:13,820
What is the completed task?

1658
01:13:13,820 --> 01:13:15,940
What is the expected MTTR change?

1659
01:13:15,940 --> 01:13:17,260
What decision loop does it compress?

1660
01:13:17,260 --> 01:13:19,660
What audit evidence does it produce by default?

1661
01:13:19,660 --> 01:13:21,980
What is the cost per completion baseline today?

1662
01:13:21,980 --> 01:13:25,140
If a team can't answer those questions, they're shipping an experiment.

1663
01:13:25,140 --> 01:13:28,940
Experiments are fine, but experiments don't get enterprise rollout budgets.

1664
01:13:28,940 --> 01:13:33,340
Next governance gates that don't kill innovation because metrics without enforcement gates

1665
01:13:33,340 --> 01:13:35,900
just produce numbers that lie.

1666
01:13:35,900 --> 01:13:42,060
Governance gates sound like the place innovation goes to die and that's not a branding problem.

1667
01:13:42,060 --> 01:13:44,020
That's history.

1668
01:13:44,020 --> 01:13:48,060
Most enterprises implemented governance as a slow approval queue run by people who don't

1669
01:13:48,060 --> 01:13:52,380
own outcomes, don't understand the work and don't carry the blast radius when something

1670
01:13:52,380 --> 01:13:53,540
ships badly.

1671
01:13:53,540 --> 01:13:56,860
So makers root around it, shadow systems appear entropy wins.

1672
01:13:56,860 --> 01:13:58,300
So the goal isn't to add gates.

1673
01:13:58,300 --> 01:14:02,860
The goal is to add zones with different enforcement and to make the path from experiment to enterprise

1674
01:14:02,860 --> 01:14:03,860
predictable.

1675
01:14:03,860 --> 01:14:06,620
The goal is to add zones, personal departmental enterprise.

1676
01:14:06,620 --> 01:14:10,980
Personal is where people explore, the agent can exist, it can help, it can even be useful,

1677
01:14:10,980 --> 01:14:14,860
but it can't publish broadly, it can't run autonomous triggers and it can't be wired

1678
01:14:14,860 --> 01:14:16,380
into high impact tools.

1679
01:14:16,380 --> 01:14:18,780
That's not punishment, that's containment.

1680
01:14:18,780 --> 01:14:21,140
Personal experimentation stays personal.

1681
01:14:21,140 --> 01:14:23,900
Departmental is where teams solve real work for a defined audience.

1682
01:14:23,900 --> 01:14:25,780
This is where two reuse starts to matter.

1683
01:14:25,780 --> 01:14:30,540
This is where you enforce that the agent has an owner, a scope and a clear purpose.

1684
01:14:30,540 --> 01:14:33,620
Personal agents can connect to tools, but only through reviewed contracts.

1685
01:14:33,620 --> 01:14:37,300
They can read data, but only within defined boundaries and they can ship faster because

1686
01:14:37,300 --> 01:14:39,500
the controls are predefined.

1687
01:14:39,500 --> 01:14:42,260
Enterprise is where the agent becomes part of the operating model.

1688
01:14:42,260 --> 01:14:46,820
That means identity is non-negotiable, tool contracts are standardized, data boundaries

1689
01:14:46,820 --> 01:14:53,140
are enforced and defender monitoring is enabled, not optional, default.

1690
01:14:53,140 --> 01:14:57,300
That zoning model is the first gate, where the agent lives determines what it's allowed

1691
01:14:57,300 --> 01:14:58,300
to do.

1692
01:14:58,300 --> 01:15:01,940
Second gate is the publish gate, this is where most organizations overdo it.

1693
01:15:01,940 --> 01:15:04,780
The publish gate should be brutally short.

1694
01:15:04,780 --> 01:15:07,660
Four yes no checks that map to the control plane.

1695
01:15:07,660 --> 01:15:09,980
One, identity created.

1696
01:15:09,980 --> 01:15:14,900
Entra agent ID exists, ownership is defined, sponsor is defined, life cycle is attached.

1697
01:15:14,900 --> 01:15:18,500
If it can act, it has an identity, if it doesn't, it stays in the sandbox.

1698
01:15:18,500 --> 01:15:20,340
Two, tool contract reviewed.

1699
01:15:20,340 --> 01:15:23,900
MCP tools are used where possible and the actions are explicit.

1700
01:15:23,900 --> 01:15:27,580
No mystery connectors, no bespoke endpoints that nobody can trace.

1701
01:15:27,580 --> 01:15:30,900
And irreversible actions don't exist without an approval pattern.

1702
01:15:30,900 --> 01:15:36,300
Three, data boundary validated, purview signals exist, authoritative sources are defined

1703
01:15:36,300 --> 01:15:39,940
and non-authoritative content is either excluded or marked as such.

1704
01:15:39,940 --> 01:15:43,820
The agent can't become a search engine for your tenant by accident.

1705
01:15:43,820 --> 01:15:45,940
Four, monitoring enabled.

1706
01:15:45,940 --> 01:15:50,340
Defender for AI is watching behavior, baselines exist, containment actions are defined.

1707
01:15:50,340 --> 01:15:54,260
If something drifts you can disable one agent identity without pausing the program, that's

1708
01:15:54,260 --> 01:15:55,260
the publish gate.

1709
01:15:55,260 --> 01:15:57,980
And ask for diagrams, it asks for enforceability.

1710
01:15:57,980 --> 01:15:59,860
Now here's the uncomfortable truth.

1711
01:15:59,860 --> 01:16:03,460
Leased privilege has to become the default, not the aspiration.

1712
01:16:03,460 --> 01:16:08,020
At small scale, people treat least privileged like a best practice they'll implement later

1713
01:16:08,020 --> 01:16:09,940
after the pilot proves value.

1714
01:16:09,940 --> 01:16:14,900
At agent scale, later is where incidents come from, so the gate is denied by design.

1715
01:16:14,900 --> 01:16:19,740
Exceptions become explicit requests, time bound, where possible, and treated as entropy

1716
01:16:19,740 --> 01:16:20,740
generators.

1717
01:16:20,740 --> 01:16:24,260
Not because the organization is mean, because exceptions are the mechanism that converts

1718
01:16:24,260 --> 01:16:27,060
deterministic controls into probabilistic ones.

1719
01:16:27,060 --> 01:16:30,980
And once controls are probabilistic, you don't get predictable outcomes, you get conditional

1720
01:16:30,980 --> 01:16:31,980
chaos.

1721
01:16:31,980 --> 01:16:35,220
Now the next gate, and this is where innovation usually complains, is human in the loop,

1722
01:16:35,220 --> 01:16:37,020
you don't need approvals for everything.

1723
01:16:37,020 --> 01:16:40,660
You need approvals for irreversible actions and sensitive domains.

1724
01:16:40,660 --> 01:16:44,940
Anything that changes access shares data, deletes records, updates, authoritative systems,

1725
01:16:44,940 --> 01:16:48,300
or triggers downstream financial impact needs a human checkpoint.

1726
01:16:48,300 --> 01:16:51,660
And the trick is not making people jump through portals, it's rooting approvals into the

1727
01:16:51,660 --> 01:16:53,060
flow of work.

1728
01:16:53,060 --> 01:16:56,700
And then the data is being structured payloads, so the decision can be fast and accountable

1729
01:16:56,700 --> 01:16:58,100
that pattern scales.

1730
01:16:58,100 --> 01:17:01,900
It also makes audits boring, which is the highest compliment an enterprise can receive.

1731
01:17:01,900 --> 01:17:03,980
Now how does this avoid killing innovation?

1732
01:17:03,980 --> 01:17:07,500
Because gates don't slow teams when the gates are pre-baked into the platform.

1733
01:17:07,500 --> 01:17:13,020
The platform team publishes the tool catalog, security publishes the identity patterns, compliance,

1734
01:17:13,020 --> 01:17:17,260
publishes the authoritative sources, defender publishes the baseline and containment playbooks,

1735
01:17:17,260 --> 01:17:21,580
then makers build inside those constraints without negotiating them every time.

1736
01:17:21,580 --> 01:17:24,260
This is the inversion most enterprises miss.

1737
01:17:24,260 --> 01:17:27,780
Innovation dies when every project has to invent its own governance.

1738
01:17:27,780 --> 01:17:31,940
Innovation survives when governance is a reusable service, and yes you still allow experimentation.

1739
01:17:31,940 --> 01:17:35,020
You just don't let experimentation become production by accident.

1740
01:17:35,020 --> 01:17:37,020
So the governance model isn't, no.

1741
01:17:37,020 --> 01:17:41,700
It's, you can do almost anything in the personal zone, some things in the departmental zone,

1742
01:17:41,700 --> 01:17:44,140
and only enforceable things in the enterprise zone.

1743
01:17:44,140 --> 01:17:46,860
That's how you scale intelligence without scaling chaos.

1744
01:17:46,860 --> 01:17:51,620
Next, the realistic 12 month adoption sequence because this doesn't land by buying licenses.

1745
01:17:51,620 --> 01:17:56,460
It lands by moving in a deliberate order that keeps the program alive under scrutiny.

1746
01:17:56,460 --> 01:17:59,580
The 12 month adoption sequence for the agentic decade.

1747
01:17:59,580 --> 01:18:03,460
Enterprises keep asking for the agent roadmap, like it's a feature checklist.

1748
01:18:03,460 --> 01:18:05,700
It isn't, it's an order of operations problem.

1749
01:18:05,700 --> 01:18:09,340
If you scale capability before you scale enforceability you get sprawl.

1750
01:18:09,340 --> 01:18:13,700
If you scale enforceability before you scale capability you get adoption that survives.

1751
01:18:13,700 --> 01:18:18,180
So here's the 12 month sequence that actually works because it matches how systems decay.

1752
01:18:18,180 --> 01:18:19,980
Month 0 to 3.

1753
01:18:19,980 --> 01:18:21,140
Inventory and identity.

1754
01:18:21,140 --> 01:18:22,380
Not innovation theater.

1755
01:18:22,380 --> 01:18:25,180
The first three months are about admitting what already exists.

1756
01:18:25,180 --> 01:18:27,700
You already have agents, co-pilot prompts.

1757
01:18:27,700 --> 01:18:29,260
Power automate flows.

1758
01:18:29,260 --> 01:18:30,740
Scripted integrations.

1759
01:18:30,740 --> 01:18:33,060
Shared service accounts doing automation.

1760
01:18:33,060 --> 01:18:36,060
The agentic decade doesn't start when you buy agent 365.

1761
01:18:36,060 --> 01:18:39,420
It starts when you measure the non-human actors already touching your tenant.

1762
01:18:39,420 --> 01:18:42,340
So you inventory agents, tools, and automations by domain.

1763
01:18:42,340 --> 01:18:46,700
You create the catalog structure and then you mandate the one enforcement rule that prevents

1764
01:18:46,700 --> 01:18:49,500
identity drift from becoming your first scandal.

1765
01:18:49,500 --> 01:18:52,300
If it can act it must have an entra agent ID.

1766
01:18:52,300 --> 01:18:53,580
Not later now.

1767
01:18:53,580 --> 01:18:57,420
And you attach ownership and sponsorship as part of creation not as a cleaner project.

1768
01:18:57,420 --> 01:19:00,060
If you can't name an owner and sponsor it stays personal.

1769
01:19:00,060 --> 01:19:02,540
It doesn't scale.

1770
01:19:02,540 --> 01:19:04,180
Month 3 to 6.

1771
01:19:04,180 --> 01:19:05,180
Standardize tools.

1772
01:19:05,180 --> 01:19:07,700
Then remove the bespoke connectors you already regret.

1773
01:19:07,700 --> 01:19:11,220
This is where MCP stops being a developer topic and becomes a governance primitive.

1774
01:19:11,220 --> 01:19:13,300
You don't let every team integrate.

1775
01:19:13,300 --> 01:19:15,420
You publish a tool contract catalog.

1776
01:19:15,420 --> 01:19:19,380
What actions exist, what the payload looks like, what the output looks like, and what

1777
01:19:19,380 --> 01:19:21,340
provenance must be included.

1778
01:19:21,340 --> 01:19:24,220
Then you do the anglomerous part, you kill bespoke connectors.

1779
01:19:24,220 --> 01:19:27,460
Not because they're evil, because they multiply your failure modes.

1780
01:19:27,460 --> 01:19:32,060
Every bespoke toolpath is a separate incident response workflow you'll have to run at 2am.

1781
01:19:32,060 --> 01:19:33,380
That is not innovation.

1782
01:19:33,380 --> 01:19:35,100
That's architectural erosion.

1783
01:19:35,100 --> 01:19:39,540
At the same time you set data boundaries in purview that match reality.

1784
01:19:39,540 --> 01:19:45,420
By authoritative sources, identify sensitive repositories and make deny by design the default

1785
01:19:45,420 --> 01:19:47,900
posture for agent identities.

1786
01:19:47,900 --> 01:19:51,980
If a team wants broader data access they're requesting a wider blast radius.

1787
01:19:51,980 --> 01:19:53,540
Make that explicit.

1788
01:19:53,540 --> 01:19:58,580
Month 6 to 9 baselines in containment because drift starts as soon as adoption becomes normal.

1789
01:19:58,580 --> 01:20:02,380
By month 6 the ecosystem starts behaving like an ecosystem.

1790
01:20:02,380 --> 01:20:06,220
People reuse agents, chain agents, and ask agents to do more than they were designed

1791
01:20:06,220 --> 01:20:07,220
for.

1792
01:20:07,220 --> 01:20:08,220
The system doesn't stay stable.

1793
01:20:08,220 --> 01:20:10,580
It accumulates pressure.

1794
01:20:10,580 --> 01:20:14,820
So this phase is defender for AI becoming operational, not just enabled.

1795
01:20:14,820 --> 01:20:20,020
You build baselines for normal per high impact agent, tool call rates, data access patterns,

1796
01:20:20,020 --> 01:20:23,980
unusual retrieval, unusual approvals, anomalous identity behavior.

1797
01:20:23,980 --> 01:20:28,460
Then you write playbooks that contain one agent without punishing the entire program.

1798
01:20:28,460 --> 01:20:33,900
Disable the agent identity, quarantine the tool, restrict the data boundary, force approvals,

1799
01:20:33,900 --> 01:20:35,060
and resume.

1800
01:20:35,060 --> 01:20:39,060
The goal is not perfect prevention, the goal is fast containment with evidence.

1801
01:20:39,060 --> 01:20:43,740
Month 9 to 12 scale the approvals pattern and start deprecation like you mean it.

1802
01:20:43,740 --> 01:20:47,900
This is where the ecosystem either matures or collapses under decision debt.

1803
01:20:47,900 --> 01:20:50,580
The difference is whether you can retire agents cleanly.

1804
01:20:50,580 --> 01:20:55,060
So you standardize the teams plus adaptive cards approval surface for irreversible actions

1805
01:20:55,060 --> 01:20:56,180
and sensitive domains.

1806
01:20:56,180 --> 01:20:59,100
You make it a reusable pattern, not a one off workflow.

1807
01:20:59,100 --> 01:21:03,540
Humans approve outcomes, agents execute bounded steps, and then you do the part nobody wants

1808
01:21:03,540 --> 01:21:04,540
to do.

1809
01:21:04,540 --> 01:21:08,420
You can replicate redundant agents because sprawl is not just too many.

1810
01:21:08,420 --> 01:21:09,860
It's overlapping authority.

1811
01:21:09,860 --> 01:21:13,700
If two agents answer the same policy question differently, you don't have optionality,

1812
01:21:13,700 --> 01:21:17,020
you have inconsistency and inconsistency becomes rework risk and cost.

1813
01:21:17,020 --> 01:21:21,460
So you collapse duplicates, route users to the authoritative agent product, and enforce

1814
01:21:21,460 --> 01:21:22,780
life cycle.

1815
01:21:22,780 --> 01:21:24,540
Build, publish, operate, retire.

1816
01:21:24,540 --> 01:21:28,940
No orphaned identities, no abandoned prompts, no, it still kind of works.

1817
01:21:28,940 --> 01:21:34,300
That's the 12 month sequence, identity first, tool contract second, data boundaries third,

1818
01:21:34,300 --> 01:21:38,180
visual detection fourth, approvals and deprecation as the scaling mechanism.

1819
01:21:38,180 --> 01:21:43,020
And once that order is in place, the organization stops asking how many agents show conclusion,

1820
01:21:43,020 --> 01:21:45,020
the advantage is enforceable intelligence.

1821
01:21:45,020 --> 01:21:47,300
The agentic advantage isn't more intelligence.

1822
01:21:47,300 --> 01:21:51,060
It's intelligence that stays enforceable when it scales under audit, cost, pressure, and

1823
01:21:51,060 --> 01:21:52,340
incident response.

1824
01:21:52,340 --> 01:21:56,460
If you want the practical next step, watch the next episode on building an agent registry

1825
01:21:56,460 --> 01:21:59,700
that prevents sprawl before it becomes your operating model.