The End of Outsourced Judgment: Why Your AI Strategy is Scaling Confusion

1
00:00:00,000 --> 00:00:03,400
Most organizations think their AI strategy is about adoption.

2
00:00:03,400 --> 00:00:06,080
Licenses, prompts may be a champion program.

3
00:00:06,080 --> 00:00:06,880
They are wrong.

4
00:00:06,880 --> 00:00:08,960
The failure mode is simpler and uglier.

5
00:00:08,960 --> 00:00:11,560
They're outsourcing judgment to a probabilistic system

6
00:00:11,560 --> 00:00:13,440
and calling it productivity.

7
00:00:13,440 --> 00:00:15,000
Copilot isn't a faster spreadsheet.

8
00:00:15,000 --> 00:00:17,120
It isn't deterministic software you control

9
00:00:17,120 --> 00:00:18,520
with inputs and outputs.

10
00:00:18,520 --> 00:00:20,920
It's a cognition engine that produces plausible text,

11
00:00:20,920 --> 00:00:23,760
plausible plans, plausible answers at scale.

12
00:00:23,760 --> 00:00:26,320
This episode is about the mindset shift,

13
00:00:26,320 --> 00:00:29,360
from tool adoption to cognitive collaboration.

14
00:00:29,360 --> 00:00:31,080
And the open loop is this.

15
00:00:31,080 --> 00:00:35,120
Outsource judgment scales confusion faster than capability.

16
00:00:35,120 --> 00:00:37,720
Tool metaphors fail because they assume determinism.

17
00:00:37,720 --> 00:00:39,480
Tool metaphors are comforting because they

18
00:00:39,480 --> 00:00:42,040
preserve the old contract between you and the system.

19
00:00:42,040 --> 00:00:42,760
You do a thing.

20
00:00:42,760 --> 00:00:43,640
The tool does the thing.

21
00:00:43,640 --> 00:00:45,440
If it fails, you can point to the broken part.

22
00:00:45,440 --> 00:00:47,920
That contract holds for most of the last 40 years

23
00:00:47,920 --> 00:00:50,800
of enterprise software because those systems are legible.

24
00:00:50,800 --> 00:00:53,000
Repeatable behavior, explainable failures,

25
00:00:53,000 --> 00:00:55,880
and accountability that stays attached to a human.

26
00:00:55,880 --> 00:00:59,320
Spreadsheets execute arithmetic, search retrieves references

27
00:00:59,320 --> 00:01:01,240
automation enforces bounded rules.

28
00:01:01,240 --> 00:01:04,040
Even when they fail, they fail in enumerated ways

29
00:01:04,040 --> 00:01:05,920
you can trace, audit, and govern.

30
00:01:05,920 --> 00:01:09,040
Now enter copilot and watch every one of those metaphors collapse.

31
00:01:09,040 --> 00:01:10,680
Copilot is not deterministic.

32
00:01:10,680 --> 00:01:13,960
It is not a formula engine, a retrieval system, or a rules engine.

33
00:01:13,960 --> 00:01:15,360
It is probabilistic synthesis.

34
00:01:15,360 --> 00:01:17,800
It takes context, partial signals, and patterns

35
00:01:17,800 --> 00:01:19,760
and produces output that looks like an answer.

36
00:01:19,760 --> 00:01:21,120
The output is often useful.

37
00:01:21,120 --> 00:01:21,960
Sometimes it's wrong.

38
00:01:21,960 --> 00:01:24,200
Often it's unprovable without doing actual work.

39
00:01:24,200 --> 00:01:26,360
And it is presented in the shape of certainty.

40
00:01:26,360 --> 00:01:28,080
Fluent, structured, confident.

41
00:01:28,080 --> 00:01:30,240
That answer, shape, text is the trap.

42
00:01:30,240 --> 00:01:32,480
Humans are trained to treat coherent language

43
00:01:32,480 --> 00:01:34,200
as evidence of understanding.

44
00:01:34,200 --> 00:01:35,680
But coherence is not correctness.

45
00:01:35,680 --> 00:01:36,680
Fluency is not truth.

46
00:01:36,680 --> 00:01:38,160
Confidence is not ownership.

47
00:01:38,160 --> 00:01:41,720
Copilot can generate a plan that sounds like your organization,

48
00:01:41,720 --> 00:01:44,960
even when it has no idea what your organization will accept,

49
00:01:44,960 --> 00:01:46,840
tolerate, or legally survive.

50
00:01:46,840 --> 00:01:50,240
So when leaders apply tool-era controls to a cognition system,

51
00:01:50,240 --> 00:01:51,360
they don't get control.

52
00:01:51,360 --> 00:01:52,880
They get conditional chaos.

53
00:01:52,880 --> 00:01:56,040
They create prompt libraries to compensate for unclear strategy.

54
00:01:56,040 --> 00:01:58,360
They add guardrails that assume output is a thing

55
00:01:58,360 --> 00:02:00,000
they can constrain like a macro.

56
00:02:00,000 --> 00:02:03,320
They measure adoption, like it's a new email client rollout.

57
00:02:03,320 --> 00:02:05,800
They treat hallucination as the primary risk,

58
00:02:05,800 --> 00:02:06,920
and they miss the real risk.

59
00:02:06,920 --> 00:02:09,240
The organization starts accepting plausible output

60
00:02:09,240 --> 00:02:10,480
as substitute adjudgment.

61
00:02:10,480 --> 00:02:13,760
Because once the output is in the memo, it becomes the plan.

62
00:02:13,760 --> 00:02:16,800
Once it's in the policy response, it becomes the answer.

63
00:02:16,800 --> 00:02:20,200
Once it's in the incident summary, it becomes what happened.

64
00:02:20,200 --> 00:02:22,160
And now the organization has scaled an artifact

65
00:02:22,160 --> 00:02:24,240
without scaling the responsibility behind it.

66
00:02:24,240 --> 00:02:26,240
This is the foundational misunderstanding.

67
00:02:26,240 --> 00:02:28,520
With deterministic tools, the human decides,

68
00:02:28,520 --> 00:02:29,960
and the tool executes.

69
00:02:29,960 --> 00:02:31,880
With cognitive systems, the tool proposes,

70
00:02:31,880 --> 00:02:33,000
and the human must decide.

71
00:02:33,000 --> 00:02:35,320
If you don't invert that relationship explicitly,

72
00:02:35,320 --> 00:02:38,200
Copilot becomes the silent decision-maker by default.

73
00:02:38,200 --> 00:02:41,080
And default decisions are how enterprises accumulate security

74
00:02:41,080 --> 00:02:43,120
debt, compliance debt, and reputational debt.

75
00:02:43,120 --> 00:02:46,000
So before talking about prompts, governance, or enablement,

76
00:02:46,000 --> 00:02:49,080
the first job is to define what the system actually is.

77
00:02:49,080 --> 00:02:50,920
Because if you keep calling cognition a tool,

78
00:02:50,920 --> 00:02:52,440
you'll keep managing it like one,

79
00:02:52,440 --> 00:02:54,440
and you will keep outsourcing judgment

80
00:02:54,440 --> 00:02:58,280
until the organization forgets who is responsible for outcomes.

81
00:02:58,280 --> 00:03:01,080
Define cognitive collaboration without romance.

82
00:03:01,080 --> 00:03:03,600
Cognitive collaboration is the least exciting definition

83
00:03:03,600 --> 00:03:05,400
you can give it, which is why it's useful.

84
00:03:05,400 --> 00:03:08,120
That means the AI generates candidate thinking

85
00:03:08,120 --> 00:03:09,880
and the human supplies, meaning constraints,

86
00:03:09,880 --> 00:03:11,080
and final responsibility.

87
00:03:11,080 --> 00:03:12,760
The machine expands the option space,

88
00:03:12,760 --> 00:03:15,240
the human collapses it into a decision.

89
00:03:15,240 --> 00:03:17,560
That distinction matters because most organizations

90
00:03:17,560 --> 00:03:20,120
are trying to get the machine to collapse the option space

91
00:03:20,120 --> 00:03:20,640
for them.

92
00:03:20,640 --> 00:03:22,160
They want the answer.

93
00:03:22,160 --> 00:03:24,480
They want a single clean output they can forward,

94
00:03:24,480 --> 00:03:25,800
approve, and move on from.

95
00:03:25,800 --> 00:03:29,080
That is the old tool contract trying to survive in a new system.

96
00:03:29,080 --> 00:03:31,160
In architectural terms, cognitive collaboration

97
00:03:31,160 --> 00:03:34,040
treats Copilot as a distributed idea generator.

98
00:03:34,040 --> 00:03:36,160
It proposes drafts, interpretations, summaries,

99
00:03:36,160 --> 00:03:38,720
and plans based on whatever context you allow it to see.

100
00:03:38,720 --> 00:03:40,240
Sometimes that context is strong.

101
00:03:40,240 --> 00:03:41,200
Often it's weak.

102
00:03:41,200 --> 00:03:43,400
Either way, the output is not truth.

103
00:03:43,400 --> 00:03:44,960
It is a structured possibility.

104
00:03:44,960 --> 00:03:47,080
The human role is not to admire the output.

105
00:03:47,080 --> 00:03:48,120
It is to arbitrate.

106
00:03:48,120 --> 00:03:49,760
Arbitration sounds like a small word,

107
00:03:49,760 --> 00:03:51,440
but it's a high friction job.

108
00:03:51,440 --> 00:03:53,560
Choosing what matters, rejecting what doesn't,

109
00:03:53,560 --> 00:03:55,920
naming trade-offs and accepting consequences.

110
00:03:55,920 --> 00:03:57,080
That is judgment.

111
00:03:57,080 --> 00:03:59,080
And if you don't make that explicit,

112
00:03:59,080 --> 00:04:02,280
you end up with a system that produces infinite plausible

113
00:04:02,280 --> 00:04:04,440
options and nobody willing to own the selection.

114
00:04:04,440 --> 00:04:07,080
So cognitive collaboration has four human responsibilities

115
00:04:07,080 --> 00:04:09,600
that do not disappear just because the pros looks finished.

116
00:04:09,600 --> 00:04:10,640
First intent.

117
00:04:10,640 --> 00:04:12,080
The AI cannot invent your intent.

118
00:04:12,080 --> 00:04:12,920
It can mimic one.

119
00:04:12,920 --> 00:04:14,000
It can infer one.

120
00:04:14,000 --> 00:04:16,400
It can generate something that sounds like intent.

121
00:04:16,400 --> 00:04:18,680
But unless you explicitly name what you are trying

122
00:04:18,680 --> 00:04:20,480
to accomplish and what you refuse to do,

123
00:04:20,480 --> 00:04:22,880
the output will drift toward generic correctness

124
00:04:22,880 --> 00:04:24,840
instead of your actual priorities.

125
00:04:24,840 --> 00:04:26,760
Second, framing.

126
00:04:26,760 --> 00:04:29,160
Framing is where you declare the problem boundary.

127
00:04:29,160 --> 00:04:31,080
What question are we answering for which audience

128
00:04:31,080 --> 00:04:33,800
under which constraints with what definition of success?

129
00:04:33,800 --> 00:04:36,560
Mostly to skip this because they think it's overhead,

130
00:04:36,560 --> 00:04:39,560
then they wonder why the AI produces nice content

131
00:04:39,560 --> 00:04:42,240
that doesn't survive contact with reality.

132
00:04:42,240 --> 00:04:43,280
Third, veto power.

133
00:04:43,280 --> 00:04:45,240
This is where most programs fail quietly.

134
00:04:45,240 --> 00:04:47,800
They train people to get outputs, not to reject them.

135
00:04:47,800 --> 00:04:50,160
A veto is not an emotional reaction.

136
00:04:50,160 --> 00:04:51,560
It is a disciplined refusal.

137
00:04:51,560 --> 00:04:53,000
This claim is unsupported.

138
00:04:53,000 --> 00:04:54,240
This risk is unowned.

139
00:04:54,240 --> 00:04:56,120
This recommendation violates policy.

140
00:04:56,120 --> 00:04:57,840
This tone creates legal exposure.

141
00:04:57,840 --> 00:04:59,280
This conclusion is premature.

142
00:04:59,280 --> 00:05:02,280
If you cannot veto AI output, you are not collaborating.

143
00:05:02,280 --> 00:05:03,440
You are complying.

144
00:05:03,440 --> 00:05:05,280
Fourth, escalation.

145
00:05:05,280 --> 00:05:07,640
When the output touches a high impact decision,

146
00:05:07,640 --> 00:05:09,800
the collaboration must force a human checkpoint,

147
00:05:09,800 --> 00:05:10,960
not a polite suggestion.

148
00:05:10,960 --> 00:05:12,280
A structural checkpoint.

149
00:05:12,280 --> 00:05:14,840
If a system can generate a plausible interpretation

150
00:05:14,840 --> 00:05:17,400
of a security incident or a plausible policy answer

151
00:05:17,400 --> 00:05:19,080
or a plausible change plan,

152
00:05:19,080 --> 00:05:21,160
you need an escalation path that says,

153
00:05:21,160 --> 00:05:23,640
this is the moment where a human must decide

154
00:05:23,640 --> 00:05:25,760
and a workflow must record who did.

155
00:05:25,760 --> 00:05:27,200
Now why does this feel uncomfortable?

156
00:05:27,200 --> 00:05:29,360
Because cognitive collaboration eliminates

157
00:05:29,360 --> 00:05:32,760
the illusion of a single right answer that you can outsource.

158
00:05:32,760 --> 00:05:34,440
Deterministic tools let you pretend

159
00:05:34,440 --> 00:05:36,880
the right configuration produces the right result.

160
00:05:36,880 --> 00:05:39,040
Cognitive systems don't give you that comfort.

161
00:05:39,040 --> 00:05:42,280
They give you a spread of possibilities and a confidence tone.

162
00:05:42,280 --> 00:05:44,800
And they force you to admit the real truth of leadership.

163
00:05:44,800 --> 00:05:46,880
Decisions are made under uncertainty

164
00:05:46,880 --> 00:05:50,280
and responsibility cannot be delegated to a text box.

165
00:05:50,280 --> 00:05:53,120
So let's be explicit about what cognitive collaboration is not.

166
00:05:53,120 --> 00:05:56,080
It is not an assistant because assistance can be held accountable

167
00:05:56,080 --> 00:05:58,360
for execution quality under your direction.

168
00:05:58,360 --> 00:05:59,920
Co-pilot cannot be held accountable.

169
00:05:59,920 --> 00:06:01,280
It has no duty of care.

170
00:06:01,280 --> 00:06:04,400
It is not an oracle because it has no privileged access to truth.

171
00:06:04,400 --> 00:06:07,200
It has access to tokens, patterns, and your data.

172
00:06:07,200 --> 00:06:09,000
Sometimes that correlates with reality.

173
00:06:09,000 --> 00:06:10,200
Sometimes it doesn't.

174
00:06:10,200 --> 00:06:13,040
It is not an intern because an intern learns your context

175
00:06:13,040 --> 00:06:15,960
through consequences, feedback, and social correction.

176
00:06:15,960 --> 00:06:17,760
Co-pilot does not learn your culture

177
00:06:17,760 --> 00:06:20,240
unless you deliberately designed context, constraints,

178
00:06:20,240 --> 00:06:21,320
and enforcement around it.

179
00:06:21,320 --> 00:06:22,560
And it is not an autopilot.

180
00:06:22,560 --> 00:06:24,720
Autopilot's operate inside engineered envelopes

181
00:06:24,720 --> 00:06:26,400
with explicit failover behavior.

182
00:06:26,400 --> 00:06:28,160
Co-pilot operates inside language.

183
00:06:28,160 --> 00:06:29,640
Language is not an envelope.

184
00:06:29,640 --> 00:06:31,120
Language is a persuasion layer.

185
00:06:31,120 --> 00:06:32,880
So the clean definition is this.

186
00:06:32,880 --> 00:06:35,720
Cognitive collaboration is a human control decision loop

187
00:06:35,720 --> 00:06:38,240
where the AI proposes and the human disposes.

188
00:06:38,240 --> 00:06:40,280
If you want to see whether your organization actually

189
00:06:40,280 --> 00:06:44,400
believes that, ask one question in any co-pilot rollout meeting.

190
00:06:44,400 --> 00:06:46,480
When co-pilot produces a plausible output

191
00:06:46,480 --> 00:06:49,400
that leads to a bad outcome, who owns that outcome?

192
00:06:49,400 --> 00:06:52,440
If the answer is the user, but the design doesn't force ownership,

193
00:06:52,440 --> 00:06:55,760
you have just described outsourced judgment with nicer branding.

194
00:06:55,760 --> 00:06:57,800
And that's where the cost starts.

195
00:06:57,800 --> 00:07:02,160
The cost curve, AI scales ambiguity faster than capability.

196
00:07:02,160 --> 00:07:03,920
Here's what most leaders miss.

197
00:07:03,920 --> 00:07:05,760
AI doesn't scale competence first.

198
00:07:05,760 --> 00:07:08,760
It scales whatever is already true in your environment.

199
00:07:08,760 --> 00:07:10,920
If your data is messy, it scales mess.

200
00:07:10,920 --> 00:07:13,560
If your decision rights are unclear, it scales ambiguity.

201
00:07:13,560 --> 00:07:15,440
If your culture avoids accountability,

202
00:07:15,440 --> 00:07:17,560
it scales plausible deniability.

203
00:07:17,560 --> 00:07:19,120
And because the output looks clean,

204
00:07:19,120 --> 00:07:21,600
the organization confuses polish with progress.

205
00:07:21,600 --> 00:07:22,720
That is the cost curve.

206
00:07:22,720 --> 00:07:25,080
AI scales ambiguity faster than capability

207
00:07:25,080 --> 00:07:27,280
because capability takes discipline and time

208
00:07:27,280 --> 00:07:28,920
and ambiguity takes nothing.

209
00:07:28,920 --> 00:07:30,640
You can deploy ambiguity in a week.

210
00:07:30,640 --> 00:07:32,400
The scaling effect is brutally simple.

211
00:07:32,400 --> 00:07:35,080
One unclear assumption, once embedded in a prompt,

212
00:07:35,080 --> 00:07:36,760
a template, a saved co-pilot page,

213
00:07:36,760 --> 00:07:38,240
or a recommended way of working

214
00:07:38,240 --> 00:07:41,160
gets replicated across hundreds or thousands of decisions.

215
00:07:41,160 --> 00:07:42,880
Not because people are malicious.

216
00:07:42,880 --> 00:07:43,960
Because people are busy.

217
00:07:43,960 --> 00:07:46,240
They see something that looks finished and they move.

218
00:07:46,240 --> 00:07:49,480
This is how plausible output becomes institutionalized.

219
00:07:49,480 --> 00:07:51,280
It starts with internal communications.

220
00:07:51,280 --> 00:07:54,280
A leader asks co-pilot to draft a message about a reogg,

221
00:07:54,280 --> 00:07:57,520
a security incident, a policy change, a new benefit.

222
00:07:57,520 --> 00:07:58,600
The output reads well.

223
00:07:58,600 --> 00:07:59,680
It sounds empathetic.

224
00:07:59,680 --> 00:08:00,760
It sounds decisive.

225
00:08:00,760 --> 00:08:03,840
It also quietly invents certainty where there isn't any.

226
00:08:03,840 --> 00:08:04,760
It creates commitments.

227
00:08:04,760 --> 00:08:05,680
Nobody reviewed.

228
00:08:05,680 --> 00:08:06,680
It implies intent.

229
00:08:06,680 --> 00:08:07,720
Nobody approved.

230
00:08:07,720 --> 00:08:09,160
And because the language is coherent,

231
00:08:09,160 --> 00:08:10,640
it becomes the official story.

232
00:08:10,640 --> 00:08:12,560
Then it moves into policy and process.

233
00:08:12,560 --> 00:08:15,760
A manager uses co-pilot to answer an HR question

234
00:08:15,760 --> 00:08:17,520
or to explain a compliance requirement

235
00:08:17,520 --> 00:08:19,120
or to summarize an internal standard.

236
00:08:19,120 --> 00:08:20,160
The answer is plausible.

237
00:08:20,160 --> 00:08:20,920
It's formatted.

238
00:08:20,920 --> 00:08:22,560
It's easier than opening the actual policy.

239
00:08:22,560 --> 00:08:24,040
So the response gets forwarded.

240
00:08:24,040 --> 00:08:24,960
Then it gets copied.

241
00:08:24,960 --> 00:08:26,000
Then it becomes precedent.

242
00:08:26,000 --> 00:08:27,480
Now the organization has drifted

243
00:08:27,480 --> 00:08:29,040
without noticing it drifted.

244
00:08:29,040 --> 00:08:31,440
And then eventually it hits operational decisions.

245
00:08:31,440 --> 00:08:34,520
Incident triage summaries, change impact analyses,

246
00:08:34,520 --> 00:08:37,440
vendor risk writeups, forecast narratives, anything

247
00:08:37,440 --> 00:08:40,280
where a nice paragraph can substitute for actual thinking.

248
00:08:40,280 --> 00:08:42,720
At that point, the AI is not assisting.

249
00:08:42,720 --> 00:08:45,240
It is authoring the artifact that other humans will treat

250
00:08:45,240 --> 00:08:46,880
as evidence that judgment occurred.

251
00:08:46,880 --> 00:08:48,240
This is the rework tax.

252
00:08:48,240 --> 00:08:50,120
And it is the bill nobody budgets for.

253
00:08:50,120 --> 00:08:51,880
The rework tax shows up in three places.

254
00:08:51,880 --> 00:08:53,240
First, verification.

255
00:08:53,240 --> 00:08:54,600
You now have to spend time proving

256
00:08:54,600 --> 00:08:56,360
that the confident claims are grounded.

257
00:08:56,360 --> 00:08:57,880
That time is not optional.

258
00:08:57,880 --> 00:09:01,240
It is the cost of using a probabilistic system responsibly.

259
00:09:01,240 --> 00:09:03,080
But most organizations don't allocate that time.

260
00:09:03,080 --> 00:09:06,320
So verification becomes ad hoc, personal, and uneven.

261
00:09:06,320 --> 00:09:07,440
Second, clean up.

262
00:09:07,440 --> 00:09:09,920
When the output is wrong or misaligned or legally risky,

263
00:09:09,920 --> 00:09:11,520
someone has to unwind it.

264
00:09:11,520 --> 00:09:13,640
That unwind rarely happens in the same meeting

265
00:09:13,640 --> 00:09:14,880
where the output was generated.

266
00:09:14,880 --> 00:09:17,760
It happens later under pressure with partial context,

267
00:09:17,760 --> 00:09:19,720
usually by someone who didn't create the artifact

268
00:09:19,720 --> 00:09:23,000
and doesn't have the authority to correct the underlying intent.

269
00:09:23,000 --> 00:09:25,760
Third, incident response and reputational repair.

270
00:09:25,760 --> 00:09:28,880
When AI generated content causes a downstream problem,

271
00:09:28,880 --> 00:09:30,440
you don't just fix the content.

272
00:09:30,440 --> 00:09:31,200
You fix trust.

273
00:09:31,200 --> 00:09:32,640
You fix stakeholder confidence.

274
00:09:32,640 --> 00:09:33,960
You fix audit narratives.

275
00:09:33,960 --> 00:09:36,360
You fix the next set of questions that start with,

276
00:09:36,360 --> 00:09:37,920
how did this make it out the door?

277
00:09:37,920 --> 00:09:40,280
The hidden bill is cognitive load redistribution.

278
00:09:40,280 --> 00:09:42,760
AI shifts effort from creation to evaluation.

279
00:09:42,760 --> 00:09:44,880
That sounds like a win until you realize evaluation

280
00:09:44,880 --> 00:09:46,560
is harder to scale than creation.

281
00:09:46,560 --> 00:09:47,760
Creation can be delegated.

282
00:09:47,760 --> 00:09:49,640
Evaluation requires judgment context

283
00:09:49,640 --> 00:09:51,240
and the willingness to say no.

284
00:09:51,240 --> 00:09:52,800
It also requires time.

285
00:09:52,800 --> 00:09:55,160
And time is the one resource leadership will not admit

286
00:09:55,160 --> 00:09:56,240
is required.

287
00:09:56,240 --> 00:09:58,920
So the paradox appears, teams generate more outputs

288
00:09:58,920 --> 00:10:00,640
than ever, but decisions get worse.

289
00:10:00,640 --> 00:10:02,840
People feel busier, but outcomes feel less owned.

290
00:10:02,840 --> 00:10:04,800
The organization increases throughput,

291
00:10:04,800 --> 00:10:07,360
but confidence in what is true decreases.

292
00:10:07,360 --> 00:10:10,240
And this is where hallucination becomes a convenient distraction.

293
00:10:10,240 --> 00:10:11,640
Hallucination is visible.

294
00:10:11,640 --> 00:10:12,480
It's a screenshot.

295
00:10:12,480 --> 00:10:13,480
It's a gotcha.

296
00:10:13,480 --> 00:10:15,520
It's something you can blame on the model.

297
00:10:15,520 --> 00:10:18,360
But the real risk is unowned outcomes.

298
00:10:18,360 --> 00:10:20,600
Unowned outcomes look like this.

299
00:10:20,600 --> 00:10:23,040
An AI generated summary becomes the record.

300
00:10:23,040 --> 00:10:26,720
An AI generated recommendation becomes the plan.

301
00:10:26,720 --> 00:10:30,080
An AI generated policy interpretation becomes the rule.

302
00:10:30,080 --> 00:10:32,760
And when it breaks, nobody can identify the decision moment

303
00:10:32,760 --> 00:10:35,400
where a human accepted the trade off and took responsibility.

304
00:10:35,400 --> 00:10:39,040
That is outsourced judgment, not because the AI is malicious,

305
00:10:39,040 --> 00:10:41,240
but because the organization designed a system

306
00:10:41,240 --> 00:10:44,240
where acceptance is effortless and ownership is optional.

307
00:10:44,240 --> 00:10:46,240
So the cost curve isn't a technology curve.

308
00:10:46,240 --> 00:10:47,400
It's an accountability curve.

309
00:10:47,400 --> 00:10:49,680
If you don't embed judgment into the workflow,

310
00:10:49,680 --> 00:10:51,600
AI will scale the absence of it.

311
00:10:51,600 --> 00:10:53,720
And the organization will keep paying quietly

312
00:10:53,720 --> 00:10:55,440
through rework, drift, and incidents

313
00:10:55,440 --> 00:10:58,000
that feel unpredictable only because nobody wrote down

314
00:10:58,000 --> 00:10:59,200
who decided what mattered.

315
00:10:59,200 --> 00:11:01,160
Now we need to talk about why this happens even

316
00:11:01,160 --> 00:11:03,040
in mature organizations.

317
00:11:03,040 --> 00:11:06,440
The automation to augmentation story is a false letter.

318
00:11:06,440 --> 00:11:09,080
Automation, augmentation, collaboration, the false letter.

319
00:11:09,080 --> 00:11:11,680
Most organizations tell themselves a comforting story.

320
00:11:11,680 --> 00:11:14,400
Automation, then augmentation, then collaboration.

321
00:11:14,400 --> 00:11:16,640
And neat maturity letter, a linear climb.

322
00:11:16,640 --> 00:11:18,680
Get a few wins, build confidence,

323
00:11:18,680 --> 00:11:21,440
and eventually you collaborate with AI.

324
00:11:21,440 --> 00:11:23,120
That story is wrong in a specific way.

325
00:11:23,120 --> 00:11:25,840
It assumes the end state is just a more advanced version

326
00:11:25,840 --> 00:11:30,160
of the start state, the same work done faster, with better tooling.

327
00:11:30,160 --> 00:11:32,560
But collaboration isn't more augmentation.

328
00:11:32,560 --> 00:11:34,080
It's a different operating model.

329
00:11:34,080 --> 00:11:35,560
And if you treat it like a ladder,

330
00:11:35,560 --> 00:11:38,320
you will stall exactly where the responsibility shifts.

331
00:11:38,320 --> 00:11:41,280
Automation is the cleanest and most honest form of value.

332
00:11:41,280 --> 00:11:43,440
You define a rule, a boundary, and an outcome.

333
00:11:43,440 --> 00:11:44,960
A workflow roots approvals.

334
00:11:44,960 --> 00:11:46,400
A system enforces a control.

335
00:11:46,400 --> 00:11:48,360
A runbook remediates a known condition.

336
00:11:48,360 --> 00:11:51,400
The RIC stays crisp because the system isn't deciding what matters.

337
00:11:51,400 --> 00:11:53,560
It's executing what you already decided matters.

338
00:11:53,560 --> 00:11:54,880
That is why automation scales.

339
00:11:54,880 --> 00:11:57,040
It turns intent into repeatable consequence.

340
00:11:57,040 --> 00:11:59,720
Augmentation is where the organization starts to get addicted.

341
00:11:59,720 --> 00:12:01,280
Augmentation feels like free speed.

342
00:12:01,280 --> 00:12:03,720
It drafts the email, summarizes the meeting,

343
00:12:03,720 --> 00:12:06,160
generates the slide outline, cleans up the spreadsheet,

344
00:12:06,160 --> 00:12:07,640
writes the status update.

345
00:12:07,640 --> 00:12:11,120
The human still owns the work, but the machine removes friction.

346
00:12:11,120 --> 00:12:13,920
And because the outcomes are low stakes and reversible,

347
00:12:13,920 --> 00:12:16,560
nobody has to confront the accountability problem yet.

348
00:12:16,560 --> 00:12:18,640
If the draft is mediocre, you edit it.

349
00:12:18,640 --> 00:12:20,480
If the summary misses a point, you correct it.

350
00:12:20,480 --> 00:12:23,000
The organization can pretend this is just productivity.

351
00:12:23,000 --> 00:12:25,680
Collaboration is where that pretend contract dies.

352
00:12:25,680 --> 00:12:28,600
Because collaboration is not helping you do the task.

353
00:12:28,600 --> 00:12:31,720
It is co-producing the thinking artifacts that other people will treat

354
00:12:31,720 --> 00:12:33,920
as decisions, commitments and records.

355
00:12:33,920 --> 00:12:37,000
It changes how problems get framed, how options get generated,

356
00:12:37,000 --> 00:12:39,200
and how conclusions get justified.

357
00:12:39,200 --> 00:12:42,400
That means the human role shifts from creator to arbiter.

358
00:12:42,400 --> 00:12:44,720
And arbitration is where leadership discomfort shows up,

359
00:12:44,720 --> 00:12:47,040
because it can't be delegated to a license.

360
00:12:47,040 --> 00:12:48,400
Here's why leaders stall.

361
00:12:48,400 --> 00:12:50,960
They want speed without responsibility shifts.

362
00:12:50,960 --> 00:12:52,920
They want the upside of a cognition engine,

363
00:12:52,920 --> 00:12:56,680
without accepting that cognition generates ambiguity by default.

364
00:12:56,680 --> 00:13:01,080
They want outputs they can forward without having to own the trade-offs embedded in the words.

365
00:13:01,080 --> 00:13:03,720
They want the AI to behave like deterministic software,

366
00:13:03,720 --> 00:13:08,200
because deterministic software lets them enforce compliance with checkboxes and dashboards.

367
00:13:08,200 --> 00:13:10,760
But a probabilistic system doesn't obey checkboxes.

368
00:13:10,760 --> 00:13:12,560
It obeys incentives context and neglect.

369
00:13:12,560 --> 00:13:15,400
So what happens in real deployments is not a smooth climb.

370
00:13:15,400 --> 00:13:16,480
It's a sideways slide.

371
00:13:16,480 --> 00:13:20,560
Organizations skip over collaboration by pretending they're still in augmentation.

372
00:13:20,560 --> 00:13:23,040
They keep the language of assist and productivity,

373
00:13:23,040 --> 00:13:26,640
but they start using AI in places where the output functions as judgment.

374
00:13:26,640 --> 00:13:29,920
Policies, incident narratives, change impact statements,

375
00:13:29,920 --> 00:13:32,760
executive summaries that become direction by implication.

376
00:13:32,760 --> 00:13:34,960
And because nobody redefined accountability,

377
00:13:34,960 --> 00:13:39,120
the organization quietly converts drafting into deciding.

378
00:13:39,120 --> 00:13:41,520
This is why co-pilot exposure is so uncomfortable.

379
00:13:41,520 --> 00:13:43,280
It doesn't just reveal data chaos.

380
00:13:43,280 --> 00:13:44,640
It reveals thinking chaos.

381
00:13:44,640 --> 00:13:46,320
Co-pilot can't fix fuzzy intent.

382
00:13:46,320 --> 00:13:48,760
It can only generate fluent substitutes for it.

383
00:13:48,760 --> 00:13:51,600
If you give it a vague goal like improve customer experience,

384
00:13:51,600 --> 00:13:54,600
it will produce a plausible plan that sounds like every other plan.

385
00:13:54,600 --> 00:13:56,560
If you give it conflicting constraints,

386
00:13:56,560 --> 00:13:59,120
it will produce a compromise that nobody chose.

387
00:13:59,120 --> 00:14:01,760
If you give it a political situation you refuse to name,

388
00:14:01,760 --> 00:14:05,000
it will generate language that avoids conflict while creating risk.

389
00:14:05,000 --> 00:14:06,320
The AI isn't failing there.

390
00:14:06,320 --> 00:14:09,560
You are watching the organization's unspoken assumptions leak into output.

391
00:14:09,560 --> 00:14:12,440
And once that output is shared, it becomes a social object.

392
00:14:12,440 --> 00:14:14,200
People react to it as if it is real.

393
00:14:14,200 --> 00:14:15,120
They align against it.

394
00:14:15,120 --> 00:14:16,320
They execute against it.

395
00:14:16,320 --> 00:14:17,200
They quote it.

396
00:14:17,200 --> 00:14:18,720
They treat it as their decision.

397
00:14:18,720 --> 00:14:19,960
And then when it causes harm,

398
00:14:19,960 --> 00:14:21,960
everyone looks for the person who approved it.

399
00:14:21,960 --> 00:14:23,760
That person often doesn't exist.

400
00:14:23,760 --> 00:14:25,920
This is the moment where the false ladder matters.

401
00:14:25,920 --> 00:14:27,760
Automation has explicit decision points.

402
00:14:27,760 --> 00:14:30,120
Augmentation hides them because stakes are low.

403
00:14:30,120 --> 00:14:32,520
Collaboration requires you to reintroduce them on purpose

404
00:14:32,520 --> 00:14:35,840
because stakes are high and outputs look finished even when the thinking isn't.

405
00:14:35,840 --> 00:14:38,000
So the practical diagnostic is simple.

406
00:14:38,000 --> 00:14:40,760
Where does your organization force a human to declare?

407
00:14:40,760 --> 00:14:43,040
I accept this conclusion and its consequences.

408
00:14:43,040 --> 00:14:45,960
If the answer is nowhere, you are not doing collaboration.

409
00:14:45,960 --> 00:14:48,760
You are doing outsourced judgment with better formatting.

410
00:14:48,760 --> 00:14:51,680
And if leaders keep insisting it's just a productivity tool,

411
00:14:51,680 --> 00:14:52,800
they are not being neutral.

412
00:14:52,800 --> 00:14:54,320
They are choosing a design.

413
00:14:54,320 --> 00:14:56,480
A system where decisions happen by drift,

414
00:14:56,480 --> 00:14:57,960
that is not a maturity ladder,

415
00:14:57,960 --> 00:15:00,240
that is architectural erosion in motion.

416
00:15:00,240 --> 00:15:01,720
Mental model to unlearn.

417
00:15:01,720 --> 00:15:03,200
AI gives answers.

418
00:15:03,200 --> 00:15:05,560
The next mistake is the one leaders defend the hardest

419
00:15:05,560 --> 00:15:07,080
because it feels efficient.

420
00:15:07,080 --> 00:15:08,120
AI gives answers.

421
00:15:08,120 --> 00:15:08,800
They are wrong.

422
00:15:08,800 --> 00:15:10,960
AI gives options that look like answers.

423
00:15:10,960 --> 00:15:14,640
And that distinction matters because answer shaped language triggers closure.

424
00:15:14,640 --> 00:15:17,520
You stop checking, stop challenging, stop asking what changed.

425
00:15:17,520 --> 00:15:19,720
So run the self-test that leadership avoids.

426
00:15:19,720 --> 00:15:21,800
If this AI output turns out to be wrong,

427
00:15:21,800 --> 00:15:23,200
who gets called into the room,

428
00:15:23,200 --> 00:15:25,360
an answer is something you can stake an outcome on.

429
00:15:25,360 --> 00:15:27,080
An answer has an implied warranty.

430
00:15:27,080 --> 00:15:28,840
AI outputs don't come with warranties.

431
00:15:28,840 --> 00:15:30,200
They come with plausible structure

432
00:15:30,200 --> 00:15:32,200
and plausible is not a quality standard.

433
00:15:32,200 --> 00:15:33,480
It's a warning label.

434
00:15:33,480 --> 00:15:35,000
The uncomfortable truth is this.

435
00:15:35,000 --> 00:15:37,480
Copilot is a synthesis engine, not a truth engine.

436
00:15:37,480 --> 00:15:40,440
It compresses context into a narrative that reads well.

437
00:15:40,440 --> 00:15:42,560
The moment you treat that narrative as an answer,

438
00:15:42,560 --> 00:15:44,120
you invert responsibility.

439
00:15:44,120 --> 00:15:46,120
You turn the tool into the decision maker

440
00:15:46,120 --> 00:15:47,680
and the human into the forwarder

441
00:15:47,680 --> 00:15:49,160
and forwarding is not leadership.

442
00:15:49,160 --> 00:15:51,360
Copilot can be wrong without hallucinating.

443
00:15:51,360 --> 00:15:53,240
It just needs missing context,

444
00:15:53,240 --> 00:15:56,120
which is the default state of every organization.

445
00:15:56,120 --> 00:15:58,800
So the right mental model is not AI gives answers.

446
00:15:58,800 --> 00:16:01,040
It's AI generates hypotheses.

447
00:16:01,040 --> 00:16:03,760
If an AI output can move money, change access,

448
00:16:03,760 --> 00:16:06,720
create a commitment, set policy, or become a record,

449
00:16:06,720 --> 00:16:08,600
it is not an answer.

450
00:16:08,600 --> 00:16:10,640
It is a proposal that must be owned.

451
00:16:10,640 --> 00:16:14,560
Owned means someone has to say explicitly, "I accept this."

452
00:16:14,560 --> 00:16:16,240
This is where answer shaped text

453
00:16:16,240 --> 00:16:18,120
does damage at the leadership level.

454
00:16:18,120 --> 00:16:19,480
Executives live in abstraction.

455
00:16:19,480 --> 00:16:21,880
They already operate through summaries, status reports,

456
00:16:21,880 --> 00:16:22,880
and slide decks.

457
00:16:22,880 --> 00:16:26,360
So an AI generated executive summary feels like a perfect fit.

458
00:16:26,360 --> 00:16:28,920
Faster, cleaner, more comprehensive.

459
00:16:28,920 --> 00:16:30,360
Except it isn't more comprehensive.

460
00:16:30,360 --> 00:16:31,600
It is more confident.

461
00:16:31,600 --> 00:16:33,720
And confidence triggers premature closure.

462
00:16:33,720 --> 00:16:35,840
Premature closure is a well-known failure mode

463
00:16:35,840 --> 00:16:37,040
in high stakes work.

464
00:16:37,040 --> 00:16:39,080
You accept the first coherent explanation

465
00:16:39,080 --> 00:16:41,400
and stop exploring alternatives.

466
00:16:41,400 --> 00:16:44,160
In an AI context, premature closure happens

467
00:16:44,160 --> 00:16:46,680
when the output sounds complete enough to ship.

468
00:16:46,680 --> 00:16:49,280
The organization confuses completion with correctness.

469
00:16:49,280 --> 00:16:50,480
So the posture has to change.

470
00:16:50,480 --> 00:16:53,080
Healthy posture is to treat AI outputs as drafts

471
00:16:53,080 --> 00:16:53,960
and hypotheses.

472
00:16:53,960 --> 00:16:56,440
Unhealthy posture is to treat them as decisions.

473
00:16:56,440 --> 00:16:57,840
The difference is not philosophical.

474
00:16:57,840 --> 00:16:59,840
The difference is whether your process forces

475
00:16:59,840 --> 00:17:00,840
a judgment moment.

476
00:17:00,840 --> 00:17:03,040
A judgment moment is the point where a human must do

477
00:17:03,040 --> 00:17:03,920
three things.

478
00:17:03,920 --> 00:17:06,760
Name the intent, name the trade off, and name the owner.

479
00:17:06,760 --> 00:17:09,480
If the output can't survive those three sentences,

480
00:17:09,480 --> 00:17:12,280
it has no business becoming institutional truth.

481
00:17:12,280 --> 00:17:14,560
And here's the final part leaders don't like.

482
00:17:14,560 --> 00:17:17,320
If AI gives options, then someone has to decide

483
00:17:17,320 --> 00:17:18,240
what matters.

484
00:17:18,240 --> 00:17:19,400
Not what is possible.

485
00:17:19,400 --> 00:17:20,240
What matters?

486
00:17:20,240 --> 00:17:21,520
That's the gap AI exposes.

487
00:17:21,520 --> 00:17:23,600
It will happily give you 10 plausible paths.

488
00:17:23,600 --> 00:17:24,840
It will even rank them.

489
00:17:24,840 --> 00:17:27,440
But it cannot carry the moral weight of choosing one path

490
00:17:27,440 --> 00:17:29,880
over another in your context with your constraints,

491
00:17:29,880 --> 00:17:32,040
with your politics and with your risk appetite.

492
00:17:32,040 --> 00:17:34,800
So when a leader asks, what does the AI say?

493
00:17:34,800 --> 00:17:36,400
They are not asking a neutral question.

494
00:17:36,400 --> 00:17:37,800
They are trying to transfer ownership

495
00:17:37,800 --> 00:17:39,240
to an unownable system.

496
00:17:39,240 --> 00:17:41,520
Everything clicked for most experienced architects

497
00:17:41,520 --> 00:17:42,920
when they realized this.

498
00:17:42,920 --> 00:17:44,880
The cost of AI isn't hallucination.

499
00:17:44,880 --> 00:17:47,520
The cost is decision avoidance with better grammar.

500
00:17:47,520 --> 00:17:50,360
And once you see that, the next failure mode becomes obvious.

501
00:17:50,360 --> 00:17:52,840
Leaders respond by trying to brute force better answers

502
00:17:52,840 --> 00:17:53,880
with more prompts.

503
00:17:53,880 --> 00:17:55,200
Mental model to unlearn.

504
00:17:55,200 --> 00:17:57,120
More prompts, it's better results.

505
00:17:57,120 --> 00:17:58,880
So leaders discover the first trap.

506
00:17:58,880 --> 00:18:00,200
AI doesn't give answers.

507
00:18:00,200 --> 00:18:01,920
It gives option shape text.

508
00:18:01,920 --> 00:18:04,080
And instead of changing the operating model,

509
00:18:04,080 --> 00:18:06,720
they reach for the only lever they can see.

510
00:18:06,720 --> 00:18:09,400
More prompts, more templates, more prompt engineering,

511
00:18:09,400 --> 00:18:11,320
more libraries, that's not strategy.

512
00:18:11,320 --> 00:18:12,920
That's avoidance with extra steps.

513
00:18:12,920 --> 00:18:15,720
Prompt volume is usually a proxy for unclear intent

514
00:18:15,720 --> 00:18:17,120
or missing authority.

515
00:18:17,120 --> 00:18:18,720
People keep rewriting prompts because they

516
00:18:18,720 --> 00:18:20,520
can't state what they actually need

517
00:18:20,520 --> 00:18:23,600
or they refuse to commit to the trade-offs they're asking for.

518
00:18:23,600 --> 00:18:26,240
So they keep nudging the system, hoping it will hand them

519
00:18:26,240 --> 00:18:28,360
an output that feels safe to forward.

520
00:18:28,360 --> 00:18:29,360
It won't.

521
00:18:29,360 --> 00:18:32,200
So ask the self-test that collapses the whole fantasy.

522
00:18:32,200 --> 00:18:33,960
If you deleted all your prompts tomorrow,

523
00:18:33,960 --> 00:18:35,680
would the decision still exist?

524
00:18:35,680 --> 00:18:37,680
Prompting can't replace problem definition.

525
00:18:37,680 --> 00:18:39,000
It can't replace constraints.

526
00:18:39,000 --> 00:18:40,440
It can't replace decision rights.

527
00:18:40,440 --> 00:18:43,920
A prompt is not a substitute for knowing what good looks like.

528
00:18:43,920 --> 00:18:46,560
When intent is coherent, you don't need 30 prompts.

529
00:18:46,560 --> 00:18:48,040
You need one.

530
00:18:48,040 --> 00:18:50,240
When intent is fuzzy, you can prompt forever

531
00:18:50,240 --> 00:18:51,880
and still never converge because you

532
00:18:51,880 --> 00:18:54,840
build a culture of try again instead of decide.

533
00:18:54,840 --> 00:18:56,880
Executives love prompt libraries because they

534
00:18:56,880 --> 00:18:58,240
feel like governance.

535
00:18:58,240 --> 00:19:00,000
But a prompt library is not a control plane.

536
00:19:00,000 --> 00:19:01,120
It doesn't enforce anything.

537
00:19:01,120 --> 00:19:04,160
It just makes it easier to generate more unknown artifacts

538
00:19:04,160 --> 00:19:06,840
faster, that is not maturity, that is industrialized

539
00:19:06,840 --> 00:19:07,920
ambiguity.

540
00:19:07,920 --> 00:19:09,640
And yes, prompts matter, of course they do.

541
00:19:09,640 --> 00:19:11,360
If you ask for garbage, you get garbage.

542
00:19:11,360 --> 00:19:13,520
If you give no context, you get generic output.

543
00:19:13,520 --> 00:19:14,880
That's basic system behavior.

544
00:19:14,880 --> 00:19:17,080
But what matters more is whether the organization

545
00:19:17,080 --> 00:19:19,840
has discipline around three things that prompting can't

546
00:19:19,840 --> 00:19:20,320
fix.

547
00:19:20,320 --> 00:19:21,640
First, explicit constraints.

548
00:19:21,640 --> 00:19:23,040
Not be compliant.

549
00:19:23,040 --> 00:19:25,760
Actual constraints, which policy governs, which data is

550
00:19:25,760 --> 00:19:28,040
authoritative, which audience is allowed, which risk

551
00:19:28,040 --> 00:19:30,640
threshold applies, what you refuse to claim, what you refuse

552
00:19:30,640 --> 00:19:31,520
to commit to.

553
00:19:31,520 --> 00:19:32,880
Second, decision ownership.

554
00:19:32,880 --> 00:19:34,720
Who can accept the output as actionable?

555
00:19:34,720 --> 00:19:35,480
Who can approve?

556
00:19:35,480 --> 00:19:36,360
Who can override?

557
00:19:36,360 --> 00:19:37,760
Who gets blamed when it's wrong?

558
00:19:37,760 --> 00:19:39,760
If that sounds harsh, good, harsh is reality.

559
00:19:39,760 --> 00:19:41,360
Accountability is not a vibe.

560
00:19:41,360 --> 00:19:42,960
Third, enforced convergence.

561
00:19:42,960 --> 00:19:45,320
Where does the workflow force a human to stop generating

562
00:19:45,320 --> 00:19:46,360
options and pick one?

563
00:19:46,360 --> 00:19:48,920
If you don't have that, your organization will live inside

564
00:19:48,920 --> 00:19:49,680
drafts forever.

565
00:19:49,680 --> 00:19:52,760
It will generate more content and fewer decisions.

566
00:19:52,760 --> 00:19:55,160
This is why prompt obsession maps so cleanly

567
00:19:55,160 --> 00:19:56,400
to leadership weakness.

568
00:19:56,400 --> 00:19:58,320
Strategy fuzziness becomes prompt fuzziness.

569
00:19:58,320 --> 00:20:00,280
Unclear priorities become prompt bloat.

570
00:20:00,280 --> 00:20:02,400
Political avoidance becomes prompt gymnastics.

571
00:20:02,400 --> 00:20:04,880
The output looks professional, but the thinking stays

572
00:20:04,880 --> 00:20:05,640
uncommitted.

573
00:20:05,640 --> 00:20:08,920
So the mindset to unlearn is not prompts are useful.

574
00:20:08,920 --> 00:20:10,800
It's the belief that prompting is the work.

575
00:20:10,800 --> 00:20:12,280
It isn't.

576
00:20:12,280 --> 00:20:14,440
The work is judgment, framing the problem,

577
00:20:14,440 --> 00:20:17,160
declaring constraints and owning consequences.

578
00:20:17,160 --> 00:20:19,120
Prompts are just how you ask the cognition engine

579
00:20:19,120 --> 00:20:21,560
to propose possibilities inside the box you should have built

580
00:20:21,560 --> 00:20:22,360
first.

581
00:20:22,360 --> 00:20:24,200
And if you don't build the box, co-pilot

582
00:20:24,200 --> 00:20:27,360
will happily build one for you, out of whatever it can infer,

583
00:20:27,360 --> 00:20:29,000
which means you didn't get a better result.

584
00:20:29,000 --> 00:20:31,000
You got a better looking substitute for a decision

585
00:20:31,000 --> 00:20:32,240
you avoided.

586
00:20:32,240 --> 00:20:35,080
Mental model to unlearn will train users later.

587
00:20:35,080 --> 00:20:36,800
After prompt obsession, the next lie

588
00:20:36,800 --> 00:20:38,440
shows up as a scheduling decision.

589
00:20:38,440 --> 00:20:41,040
We'll train users later.

590
00:20:41,040 --> 00:20:42,800
Later is where responsibility goes to die,

591
00:20:42,800 --> 00:20:45,040
because the first two weeks are when habits form.

592
00:20:45,040 --> 00:20:48,240
That's when people learn what works, what gets praised,

593
00:20:48,240 --> 00:20:50,360
what saves time, and what they can get away with.

594
00:20:50,360 --> 00:20:52,920
So ask the only question that matters in week two.

595
00:20:52,920 --> 00:20:55,320
What behavior did you reward in the first two weeks?

596
00:20:55,320 --> 00:20:57,760
AI doesn't just create outputs in those first two weeks.

597
00:20:57,760 --> 00:20:59,680
It creates behavioral defaults.

598
00:20:59,680 --> 00:21:02,880
If the default is copy-paste co-pilot output into email,

599
00:21:02,880 --> 00:21:04,040
that becomes culture.

600
00:21:04,040 --> 00:21:06,720
If the default is forward the summary as the record,

601
00:21:06,720 --> 00:21:08,160
that becomes precedent.

602
00:21:08,160 --> 00:21:10,880
If the default is ask the AI to interpret policy

603
00:21:10,880 --> 00:21:13,840
so I don't have to read it, that becomes the unofficial operating

604
00:21:13,840 --> 00:21:14,360
model.

605
00:21:14,360 --> 00:21:17,080
You don't undo that with a training deck three months later.

606
00:21:17,080 --> 00:21:18,520
This is the uncomfortable truth.

607
00:21:18,520 --> 00:21:21,640
You don't train AI adoption, you condition judgment behavior,

608
00:21:21,640 --> 00:21:23,800
and conditioning happens at the speed of reward.

609
00:21:23,800 --> 00:21:26,000
In most organizations, the reward is simple.

610
00:21:26,000 --> 00:21:26,640
Speed.

611
00:21:26,640 --> 00:21:29,280
You respond faster, you ship faster, you look productive.

612
00:21:29,280 --> 00:21:31,840
Nobody asks how you got there, because they like the output.

613
00:21:31,840 --> 00:21:34,200
So you keep doing it, then someone else copies your pattern.

614
00:21:34,200 --> 00:21:37,080
Then a manager asks why you aren't using co-pilot the same way.

615
00:21:37,080 --> 00:21:39,520
And now you've scaled the behavior before you ever defined

616
00:21:39,520 --> 00:21:40,640
what good looks like.

617
00:21:40,640 --> 00:21:43,280
This is why AI literacy is a misleading label.

618
00:21:43,280 --> 00:21:46,200
AI literacy is knowing that the system is probabilistic,

619
00:21:46,200 --> 00:21:48,120
that it can be wrong, that sources matter,

620
00:21:48,120 --> 00:21:49,480
that sensitive data matters.

621
00:21:49,480 --> 00:21:52,480
Fine, necessary, not sufficient.

622
00:21:52,480 --> 00:21:54,280
Judgment literacy is something else.

623
00:21:54,280 --> 00:21:56,800
The ability to treat AI output as a proposal

624
00:21:56,800 --> 00:21:58,600
to interrogate it, to frame the decision,

625
00:21:58,600 --> 00:21:59,720
and to own the consequences.

626
00:21:59,720 --> 00:22:02,480
Judgment is the bottleneck, not prompts, not features,

627
00:22:02,480 --> 00:22:03,840
not licensing.

628
00:22:03,840 --> 00:22:06,360
And here's the part leadership avoids saying out loud.

629
00:22:06,360 --> 00:22:09,400
They delegate adoption to admins and then blame users.

630
00:22:09,400 --> 00:22:13,120
They assign licenses, run a few enablement sessions,

631
00:22:13,120 --> 00:22:16,560
publish a prompt gallery, and call it change management.

632
00:22:16,560 --> 00:22:19,320
Then they act surprised when users do what humans always do

633
00:22:19,320 --> 00:22:20,920
with low friction systems.

634
00:22:20,920 --> 00:22:22,920
They optimize for speed and social safety.

635
00:22:22,920 --> 00:22:25,200
They avoid being wrong, they avoid being slow,

636
00:22:25,200 --> 00:22:27,960
they avoid taking responsibility for uncertainty.

637
00:22:27,960 --> 00:22:31,320
So they pace the AI output as is, because it sounds competent

638
00:22:31,320 --> 00:22:32,520
and it distributes blame.

639
00:22:32,520 --> 00:22:35,840
If the message is wrong, the user can quietly imply it was the tool.

640
00:22:35,840 --> 00:22:38,120
If the policy interpretation is wrong, they can say,

641
00:22:38,120 --> 00:22:39,600
that's what Copilot said.

642
00:22:39,600 --> 00:22:42,120
If the incident summary is wrong, they can claim

643
00:22:42,120 --> 00:22:44,040
they were just relaying what the system produced.

644
00:22:44,040 --> 00:22:45,120
That's not malice.

645
00:22:45,120 --> 00:22:47,120
That's rational behavior inside a culture

646
00:22:47,120 --> 00:22:50,040
that punishes uncertainty and rewards throughput.

647
00:22:50,040 --> 00:22:53,160
So training later is really, we'll let the system write

648
00:22:53,160 --> 00:22:56,080
our culture for two weeks and then hope a webinar fixes it.

649
00:22:56,080 --> 00:22:56,960
It won't.

650
00:22:56,960 --> 00:22:59,600
Training has to be front loaded because early usage

651
00:22:59,600 --> 00:23:01,960
is where people learn what the organization tolerates.

652
00:23:01,960 --> 00:23:04,120
If nobody teaches verification discipline early,

653
00:23:04,120 --> 00:23:05,400
people won't invent it later.

654
00:23:05,400 --> 00:23:08,400
If nobody teaches that sharing AI output requires framing,

655
00:23:08,400 --> 00:23:10,480
people won't spontaneously start doing it.

656
00:23:10,480 --> 00:23:13,920
If nobody teaches that high stakes outputs require escalation,

657
00:23:13,920 --> 00:23:16,360
people will treat everything like a draft until the day

658
00:23:16,360 --> 00:23:19,880
it becomes evidence in an audit or an incident review.

659
00:23:19,880 --> 00:23:23,600
The other reason later fails is that AI is not a static feature set.

660
00:23:23,600 --> 00:23:25,600
It changes continuously, the model changes,

661
00:23:25,600 --> 00:23:29,160
the UI changes, new grounding mechanisms appear, new agents show up.

662
00:23:29,160 --> 00:23:31,120
So if you think training is a phase,

663
00:23:31,120 --> 00:23:33,160
you are treating an evolving cognition layer

664
00:23:33,160 --> 00:23:34,840
like an email client migration.

665
00:23:34,840 --> 00:23:36,120
That mindset is obsolete.

666
00:23:36,120 --> 00:23:37,840
Training isn't enablement content,

667
00:23:37,840 --> 00:23:40,360
it's behavioral guardrails with repetition.

668
00:23:40,360 --> 00:23:43,480
The minimum training that matters is not here are tips and tricks.

669
00:23:43,480 --> 00:23:45,880
It's three rules that remove deniability.

670
00:23:45,880 --> 00:23:48,320
One, if you share AI output,

671
00:23:48,320 --> 00:23:51,720
you add one sentence that states your intent and confidence level.

672
00:23:51,720 --> 00:23:55,560
Draft for discussion, recommendation, pending validation,

673
00:23:55,560 --> 00:23:57,720
confirmed against policy X,

674
00:23:57,720 --> 00:24:00,440
anything that reattaches ownership to a human.

675
00:24:00,440 --> 00:24:03,400
Two, if the output affects access, money, policy,

676
00:24:03,400 --> 00:24:07,480
or risk posture, you escalate, not by etiquette, by design.

677
00:24:07,480 --> 00:24:11,360
Three, if you can't explain why the output is correct,

678
00:24:11,360 --> 00:24:12,560
you're not allowed to act on it.

679
00:24:12,560 --> 00:24:14,240
These rules feel strict because they are.

680
00:24:14,240 --> 00:24:17,400
Strictness is how you stop judgment from evaporating.

681
00:24:17,400 --> 00:24:18,760
So the unlearn is simple.

682
00:24:18,760 --> 00:24:21,680
You don't get to postpone training in a probabilistic system.

683
00:24:21,680 --> 00:24:23,600
The system trains your users immediately.

684
00:24:23,600 --> 00:24:25,200
If you don't define the discipline,

685
00:24:25,200 --> 00:24:26,760
the platform defines the habit,

686
00:24:26,760 --> 00:24:28,920
and the habit will always drift towards speed

687
00:24:28,920 --> 00:24:30,000
and away from ownership.

688
00:24:30,000 --> 00:24:31,560
And once that drift exists,

689
00:24:31,560 --> 00:24:34,120
governance becomes the next lie people tell themselves

690
00:24:34,120 --> 00:24:36,080
is mental model to unlearn.

691
00:24:36,080 --> 00:24:38,080
Governance slows innovation.

692
00:24:38,080 --> 00:24:40,120
Governance slows innovation is what people say

693
00:24:40,120 --> 00:24:41,600
when they want the benefits of scale

694
00:24:41,600 --> 00:24:43,040
without the cost of control.

695
00:24:43,040 --> 00:24:45,160
It is not a philosophy, it's a confession.

696
00:24:45,160 --> 00:24:47,160
So ask the self-test before you repeat it

697
00:24:47,160 --> 00:24:48,800
in your next steering committee.

698
00:24:48,800 --> 00:24:50,440
Where does the system physically prevent

699
00:24:50,440 --> 00:24:51,840
a bad decision from shipping?

700
00:24:51,840 --> 00:24:54,360
In the tool era, you could get away with weak governance

701
00:24:54,360 --> 00:24:56,400
because most failures stayed local.

702
00:24:56,400 --> 00:24:57,960
A bad spreadsheet breaks a forecast,

703
00:24:57,960 --> 00:24:59,840
a bad workflow breaks a process.

704
00:24:59,840 --> 00:25:01,600
The blast radius is annoying but bounded.

705
00:25:01,600 --> 00:25:03,320
Cognitive systems don't fail locally.

706
00:25:03,320 --> 00:25:04,640
They fail socially.

707
00:25:04,640 --> 00:25:07,040
They produce language that moves through the organization

708
00:25:07,040 --> 00:25:09,080
faster than any ticketing system,

709
00:25:09,080 --> 00:25:10,640
faster than any policy update,

710
00:25:10,640 --> 00:25:12,920
faster than any corrective communication.

711
00:25:12,920 --> 00:25:16,040
And once people repeat it, it becomes what we believe.

712
00:25:16,040 --> 00:25:18,800
So governance in an AI environment is not bureaucracy.

713
00:25:18,800 --> 00:25:20,040
It is entropy management,

714
00:25:20,040 --> 00:25:21,720
and if you refuse to manage entropy,

715
00:25:21,720 --> 00:25:24,320
you don't get innovation, you get drift.

716
00:25:24,320 --> 00:25:25,840
Here's the uncomfortable truth.

717
00:25:25,840 --> 00:25:27,440
Experimentation without guardrails

718
00:25:27,440 --> 00:25:28,960
is just uncontrolled variance.

719
00:25:28,960 --> 00:25:31,360
It creates a hundred ways of doing the same thing.

720
00:25:31,360 --> 00:25:33,880
None of them documented, none of them accountable.

721
00:25:33,880 --> 00:25:36,120
And all of them capable of becoming precedent.

722
00:25:36,120 --> 00:25:37,280
That is not a learning culture.

723
00:25:37,280 --> 00:25:39,240
That is an organization producing noise

724
00:25:39,240 --> 00:25:40,640
and calling it progress.

725
00:25:40,640 --> 00:25:43,400
This is where executives reach for the wrong kind of governance.

726
00:25:43,400 --> 00:25:45,040
They write an acceptable use policy.

727
00:25:45,040 --> 00:25:46,640
They do a quick training module.

728
00:25:46,640 --> 00:25:49,000
They publish a list of do's and don'ts.

729
00:25:49,000 --> 00:25:52,120
They tell people not to paste confidential data into prompts

730
00:25:52,120 --> 00:25:54,160
and then they declare governance done.

731
00:25:54,160 --> 00:25:55,240
That's not governance.

732
00:25:55,240 --> 00:25:57,000
That's paperwork.

733
00:25:57,000 --> 00:26:00,000
Real governance in this context is enforced intent.

734
00:26:00,000 --> 00:26:03,160
Explicit decision rights, data boundaries, auditability

735
00:26:03,160 --> 00:26:06,360
and escalation paths that activate when the output matters.

736
00:26:06,360 --> 00:26:07,480
The key shift is this.

737
00:26:07,480 --> 00:26:09,360
Governance isn't about controlling the model.

738
00:26:09,360 --> 00:26:11,400
It's about controlling what the organization is allowed

739
00:26:11,400 --> 00:26:12,600
to treat as true.

740
00:26:12,600 --> 00:26:14,880
Because AI will generate plausible statements

741
00:26:14,880 --> 00:26:16,400
about anything you let it touch.

742
00:26:16,400 --> 00:26:18,760
Your policies, your customers, your risk posture,

743
00:26:18,760 --> 00:26:20,480
your incidents, your strategy.

744
00:26:20,480 --> 00:26:21,760
And if you don't build a mechanism

745
00:26:21,760 --> 00:26:24,840
that forces humans to validate own and record decisions,

746
00:26:24,840 --> 00:26:27,720
those plausible statements turn into institutional behavior.

747
00:26:27,720 --> 00:26:30,240
Over time, policies drift away, missing policies

748
00:26:30,240 --> 00:26:32,920
create obvious gaps, drifting policies create ambiguity.

749
00:26:32,920 --> 00:26:34,600
Ambiguity is where accountability dies

750
00:26:34,600 --> 00:26:36,400
and AI accelerates that death

751
00:26:36,400 --> 00:26:38,520
because it produces reasonable language

752
00:26:38,520 --> 00:26:42,240
that fills the gap between what you meant and what you enforced.

753
00:26:42,240 --> 00:26:44,280
So let's talk about the thing people really mean

754
00:26:44,280 --> 00:26:45,720
when they complain about governance.

755
00:26:45,720 --> 00:26:46,720
They mean friction.

756
00:26:46,720 --> 00:26:49,680
They mean someone might have to slow down, ask permission,

757
00:26:49,680 --> 00:26:52,040
document a rational or admit uncertainty.

758
00:26:52,040 --> 00:26:53,280
Yes, that's the point.

759
00:26:54,800 --> 00:26:57,960
Innovation at enterprise scale is not fast output.

760
00:26:57,960 --> 00:26:59,160
It is safe change.

761
00:26:59,160 --> 00:27:01,160
It is the ability to try something new

762
00:27:01,160 --> 00:27:03,800
without destroying trust and trust requires constraints

763
00:27:03,800 --> 00:27:06,160
that hold under pressure, not vibes that collapse

764
00:27:06,160 --> 00:27:07,520
when the first incident happens.

765
00:27:07,520 --> 00:27:10,200
This is why shadow AI is not a user problem.

766
00:27:10,200 --> 00:27:12,480
It is the predictable outcome of weak enforcement.

767
00:27:12,480 --> 00:27:15,920
If official tooling feels slow and unofficial tooling feels easy,

768
00:27:15,920 --> 00:27:18,760
people will root around controls, not because they are evil,

769
00:27:18,760 --> 00:27:20,560
because the organization rewarded speed

770
00:27:20,560 --> 00:27:23,240
and didn't enforce boundaries, the system behaved as designed.

771
00:27:23,240 --> 00:27:24,880
You just didn't like the consequences.

772
00:27:24,880 --> 00:27:26,520
So the governance reframe is simple.

773
00:27:26,520 --> 00:27:28,480
Governance is how you scale trust.

774
00:27:28,480 --> 00:27:29,960
Trust isn't a marketing claim.

775
00:27:29,960 --> 00:27:31,400
Trust is an operational property.

776
00:27:31,400 --> 00:27:33,320
It emerges when three things are true.

777
00:27:33,320 --> 00:27:34,960
First, boundaries are clear.

778
00:27:34,960 --> 00:27:37,320
What data is in scope, what is out of scope,

779
00:27:37,320 --> 00:27:39,000
and what must never be inferred.

780
00:27:39,000 --> 00:27:40,240
Confidential is not a boundary.

781
00:27:40,240 --> 00:27:41,080
It's a label.

782
00:27:41,080 --> 00:27:42,800
Boundaries are enforced by access controls,

783
00:27:42,800 --> 00:27:44,920
data classification and workflow rules

784
00:27:44,920 --> 00:27:46,960
that prevent accidental leakage.

785
00:27:46,960 --> 00:27:48,640
Second, decision rights are explicit.

786
00:27:48,640 --> 00:27:51,000
Who can interpret policy, who can approve exceptions,

787
00:27:51,000 --> 00:27:52,360
who can change risk posture,

788
00:27:52,360 --> 00:27:54,240
who can accept blast radius.

789
00:27:54,240 --> 00:27:56,520
If an AI output touches one of those areas,

790
00:27:56,520 --> 00:27:59,400
the workflow must force a named human to accept it.

791
00:27:59,400 --> 00:28:01,360
Not a team, not the business,

792
00:28:01,360 --> 00:28:03,960
a human with a calendar and a manager.

793
00:28:03,960 --> 00:28:05,400
Third, evidence exists.

794
00:28:05,400 --> 00:28:07,720
Logs, decision records and audit trails

795
00:28:07,720 --> 00:28:09,680
that survive the next incident review.

796
00:28:09,680 --> 00:28:11,960
If you can't reconstruct why a decision was made,

797
00:28:11,960 --> 00:28:12,960
you didn't govern it.

798
00:28:12,960 --> 00:28:14,000
You hoped it would work.

799
00:28:14,000 --> 00:28:16,000
The weird part is that this kind of governance

800
00:28:16,000 --> 00:28:17,240
is not anti-innovation.

801
00:28:17,240 --> 00:28:19,360
It is what makes innovation repeatable.

802
00:28:19,360 --> 00:28:21,720
It turns experimentation into learning.

803
00:28:21,720 --> 00:28:24,080
Because learning requires traceability.

804
00:28:24,080 --> 00:28:24,920
What did we try?

805
00:28:24,920 --> 00:28:25,800
Why did we try it?

806
00:28:25,800 --> 00:28:27,680
What happened and who owned the trade-off?

807
00:28:27,680 --> 00:28:29,200
Without that, you don't have innovation.

808
00:28:29,200 --> 00:28:31,560
You have uncontrolled behavior and selective memory.

809
00:28:31,560 --> 00:28:34,520
So when someone says governance slows innovation,

810
00:28:34,520 --> 00:28:36,440
the correct response is no.

811
00:28:36,440 --> 00:28:38,520
Governance slows fantasy.

812
00:28:38,520 --> 00:28:40,720
Thinking without enforcement is fantasy.

813
00:28:40,720 --> 00:28:43,600
Enforcement without thinking is bureaucracy.

814
00:28:43,600 --> 00:28:45,280
Your job is to build the middle,

815
00:28:45,280 --> 00:28:48,000
where AI can propose, humans can judge,

816
00:28:48,000 --> 00:28:49,960
and the system can enforce consequences

817
00:28:49,960 --> 00:28:51,600
with no deniability.

818
00:28:51,600 --> 00:28:54,920
The triad, cognition, action, judgment.

819
00:28:54,920 --> 00:28:56,400
So here's the model that stops this

820
00:28:56,400 --> 00:28:57,960
from turning into a moral lecture

821
00:28:57,960 --> 00:29:00,440
about using AI responsibly.

822
00:29:00,440 --> 00:29:01,360
It's a systems model.

823
00:29:01,360 --> 00:29:03,200
Three systems, three jobs, no romance,

824
00:29:03,200 --> 00:29:05,720
system of cognition, system of action, system of judgment.

825
00:29:05,720 --> 00:29:07,720
If you can't name which one you're operating in,

826
00:29:07,720 --> 00:29:08,760
you're already drifting.

827
00:29:08,760 --> 00:29:10,800
And drift is how organizations end up

828
00:29:10,800 --> 00:29:13,040
in post-incident meetings pretending nobody

829
00:29:13,040 --> 00:29:14,000
could have seen it coming.

830
00:29:14,000 --> 00:29:15,520
Start with the system of cognition.

831
00:29:15,520 --> 00:29:17,400
This is where M365 Copilot lives.

832
00:29:17,400 --> 00:29:20,040
Chat, summaries, drafts, synthesis

833
00:29:20,040 --> 00:29:22,160
across email, meetings, documents,

834
00:29:22,160 --> 00:29:24,560
and whatever content the graph can legally expose.

835
00:29:24,560 --> 00:29:26,000
Cognition is not execution.

836
00:29:26,000 --> 00:29:28,040
Cognition is possibility generation,

837
00:29:28,040 --> 00:29:30,200
interpretations, options, narratives,

838
00:29:30,200 --> 00:29:31,440
and proposed next steps.

839
00:29:31,440 --> 00:29:33,840
It is fast, it is wide, it is persuasive.

840
00:29:33,840 --> 00:29:35,080
And it has one built in flow.

841
00:29:35,080 --> 00:29:36,560
It will produce coherent language

842
00:29:36,560 --> 00:29:38,680
even when the underlying reality is incoherent.

843
00:29:38,680 --> 00:29:40,760
That's not a bug, that's what it's designed to do.

844
00:29:40,760 --> 00:29:42,280
So the output of cognition

845
00:29:42,280 --> 00:29:45,200
must always be treated as provisional, candidate thinking.

846
00:29:45,200 --> 00:29:47,440
A draft artifact that still needs judgment.

847
00:29:47,440 --> 00:29:50,000
If you treat cognition output as the decision,

848
00:29:50,000 --> 00:29:51,640
you have already collapsed the triad

849
00:29:51,640 --> 00:29:54,080
into a single fragile point of failure.

850
00:29:54,080 --> 00:29:55,200
The model said so.

851
00:29:55,200 --> 00:29:56,600
Now the system of action.

852
00:29:56,600 --> 00:29:58,320
This is where consequences get enforced.

853
00:29:58,320 --> 00:30:00,440
Workflows, approvals, access changes,

854
00:30:00,440 --> 00:30:03,720
containment actions, policy exceptions, audit trails.

855
00:30:03,720 --> 00:30:05,920
In most enterprises, this is some combination

856
00:30:05,920 --> 00:30:07,680
of ticketing, workflow engines,

857
00:30:07,680 --> 00:30:10,400
and the systems that actually touch production reality.

858
00:30:10,400 --> 00:30:11,880
Service now is an obvious example,

859
00:30:11,880 --> 00:30:13,000
but the brand doesn't matter.

860
00:30:13,000 --> 00:30:13,960
The function matters.

861
00:30:13,960 --> 00:30:15,720
Action is the part of the organization

862
00:30:15,720 --> 00:30:18,080
that turns intent into irreversible effects.

863
00:30:18,080 --> 00:30:19,800
Action systems are allowed to be boring.

864
00:30:19,800 --> 00:30:20,480
They should be.

865
00:30:20,480 --> 00:30:22,800
They exist to create friction in the right places,

866
00:30:22,800 --> 00:30:25,040
to force acknowledgement, to root decisions,

867
00:30:25,040 --> 00:30:27,760
to authorised owners, to record evidence,

868
00:30:27,760 --> 00:30:29,280
to stop looks fine thinking

869
00:30:29,280 --> 00:30:31,400
from becoming production downtime.

870
00:30:31,400 --> 00:30:32,880
And then there is the system of judgment.

871
00:30:32,880 --> 00:30:34,880
This is the human layer, not as a slogan.

872
00:30:34,880 --> 00:30:37,040
As a constraint, the enterprise cannot escape.

873
00:30:37,040 --> 00:30:38,640
Judgment is where intent is framed,

874
00:30:38,640 --> 00:30:41,240
trade-offs are chosen, and responsibility is assigned

875
00:30:41,240 --> 00:30:43,360
to a person who can be held accountable.

876
00:30:43,360 --> 00:30:44,560
Judgment is the only system

877
00:30:44,560 --> 00:30:46,800
that can legitimately answer questions like,

878
00:30:46,800 --> 00:30:47,880
"What matters right now?

879
00:30:47,880 --> 00:30:49,040
"What is acceptable risk?

880
00:30:49,040 --> 00:30:49,880
"What is fair?

881
00:30:49,880 --> 00:30:50,800
"What is defensible?

882
00:30:50,800 --> 00:30:52,520
"What does the organization refuse to do

883
00:30:52,520 --> 00:30:54,080
"even if it is efficient?

884
00:30:54,080 --> 00:30:55,560
"AI cannot answer those questions.

885
00:30:55,560 --> 00:30:56,520
"They can mimic answers.

886
00:30:56,520 --> 00:30:58,600
"It can produce language that resembles them,

887
00:30:58,600 --> 00:31:00,080
"but it cannot own the consequences.

888
00:31:00,080 --> 00:31:01,560
"And if it can't own consequences,

889
00:31:01,560 --> 00:31:03,080
"it is not judgment.

890
00:31:03,080 --> 00:31:04,520
"That distinction matters."

891
00:31:04,520 --> 00:31:08,040
Because most failed AI strategies accidentally invert the triad.

892
00:31:08,040 --> 00:31:10,800
They treat the system of cognition as if it is judgment,

893
00:31:10,800 --> 00:31:12,280
and they treat the system of action

894
00:31:12,280 --> 00:31:13,800
as if it is optional bureaucracy.

895
00:31:13,800 --> 00:31:15,680
So they get lots of thinking artifacts

896
00:31:15,680 --> 00:31:17,440
and very little enforced reality.

897
00:31:17,440 --> 00:31:19,880
The organization becomes a factory of plausible language

898
00:31:19,880 --> 00:31:21,280
with no operational gravity.

899
00:31:21,280 --> 00:31:23,120
This is the diagnostic line you already have,

900
00:31:23,120 --> 00:31:24,320
and it is not negotiable.

901
00:31:24,320 --> 00:31:27,320
Thinking without enforcement is fantasy.

902
00:31:27,320 --> 00:31:29,120
Enforcement without thinking is bureaucracy.

903
00:31:29,120 --> 00:31:30,720
Now translate that into behavior.

904
00:31:30,720 --> 00:31:32,800
If copilot produces a summary of an incident,

905
00:31:32,800 --> 00:31:33,760
that's cognition.

906
00:31:33,760 --> 00:31:35,360
If a human classifies severity

907
00:31:35,360 --> 00:31:37,840
and chooses a response posture, that's judgment.

908
00:31:37,840 --> 00:31:39,840
If a workflow enforces containment steps,

909
00:31:39,840 --> 00:31:41,880
approvals, notifications, and evidence capture,

910
00:31:41,880 --> 00:31:42,560
that's action.

911
00:31:42,560 --> 00:31:45,040
If any one of those three is missing, you don't have a system.

912
00:31:45,040 --> 00:31:45,880
You have a vibe.

913
00:31:45,880 --> 00:31:47,480
Cognition without judgment is noise.

914
00:31:47,480 --> 00:31:49,160
It produces options nobody chooses.

915
00:31:49,160 --> 00:31:50,720
Judgment without action is theater.

916
00:31:50,720 --> 00:31:52,720
It produces decisions, nobody implements.

917
00:31:52,720 --> 00:31:54,360
Action without judgment is dangerous.

918
00:31:54,360 --> 00:31:56,360
It produces consequences, nobody intended.

919
00:31:56,360 --> 00:31:58,680
And here's the part that makes this model operational,

920
00:31:58,680 --> 00:32:00,000
instead of philosophical.

921
00:32:00,000 --> 00:32:02,280
The hand-offs are where organizations fail.

922
00:32:02,280 --> 00:32:04,920
The output of cognition must enter a judgment moment.

923
00:32:04,920 --> 00:32:07,160
Not someone should review this.

924
00:32:07,160 --> 00:32:09,760
A design checkpoint where a named human must accept

925
00:32:09,760 --> 00:32:11,640
or reject the interpretation,

926
00:32:11,640 --> 00:32:13,800
and where the acceptance triggers action pathways

927
00:32:13,800 --> 00:32:15,280
with recorded ownership.

928
00:32:15,280 --> 00:32:17,200
That is how you stop outsourced judgment,

929
00:32:17,200 --> 00:32:18,920
not by telling people to be careful,

930
00:32:18,920 --> 00:32:20,280
not by publishing guidance,

931
00:32:20,280 --> 00:32:22,920
by engineering a reality where the organization cannot move

932
00:32:22,920 --> 00:32:24,160
from possible to done,

933
00:32:24,160 --> 00:32:26,360
without a human decision owner being visible.

934
00:32:26,360 --> 00:32:27,760
This also explains why governance

935
00:32:27,760 --> 00:32:30,200
that lives only in M365 feels weak.

936
00:32:30,200 --> 00:32:33,200
M365 can guide cognition, it can restrict data exposure,

937
00:32:33,200 --> 00:32:34,240
it can log usage,

938
00:32:34,240 --> 00:32:37,320
but it cannot, by itself, enforce enterprise consequences.

939
00:32:37,320 --> 00:32:39,360
That enforcement lives in action systems.

940
00:32:39,360 --> 00:32:40,600
That's where you put the brakes.

941
00:32:40,600 --> 00:32:41,840
That's where you put the receipts.

942
00:32:41,840 --> 00:32:44,000
So when you hear someone describe an AI strategy

943
00:32:44,000 --> 00:32:46,920
as we rolled out co-pilot and trained users,

944
00:32:46,920 --> 00:32:48,200
you should hear what they didn't say.

945
00:32:48,200 --> 00:32:49,680
They didn't say where judgment lives,

946
00:32:49,680 --> 00:32:51,600
they didn't say where action gets enforced,

947
00:32:51,600 --> 00:32:54,000
they didn't say how ownership is recorded,

948
00:32:54,000 --> 00:32:55,960
which means they built a cognition layer

949
00:32:55,960 --> 00:32:57,920
on top of an accountability vacuum.

950
00:32:57,920 --> 00:32:59,440
And now, to make this real,

951
00:32:59,440 --> 00:33:01,280
we're going to run the triad through scenarios

952
00:33:01,280 --> 00:33:03,480
everyone recognizes, starting with the one

953
00:33:03,480 --> 00:33:06,880
that fails the fastest, security incident triage.

954
00:33:06,880 --> 00:33:09,080
Scenario one, security incident triage.

955
00:33:09,080 --> 00:33:11,200
Security incident triage is the cleanest place

956
00:33:11,200 --> 00:33:13,320
to watch this fail because the organization

957
00:33:13,320 --> 00:33:14,960
already believes it has process.

958
00:33:14,960 --> 00:33:16,440
It already believes it has governance,

959
00:33:16,440 --> 00:33:18,720
it already believes it has accountability.

960
00:33:18,720 --> 00:33:20,920
Then co-pilot shows up and turns analysis

961
00:33:20,920 --> 00:33:24,080
into a decorative layer that nobody operationalizes.

962
00:33:24,080 --> 00:33:26,200
Start with the system of cognition.

963
00:33:26,200 --> 00:33:28,480
Co-pilot can ingest the mess humans can't.

964
00:33:28,480 --> 00:33:31,560
Alert summaries, email threads, teams chat fragments,

965
00:33:31,560 --> 00:33:33,920
meeting notes, defender notifications,

966
00:33:33,920 --> 00:33:35,920
a pile of half finished work items

967
00:33:35,920 --> 00:33:39,600
and the one key sentence someone dropped in a channel at 2.13 a.m.

968
00:33:39,600 --> 00:33:42,320
It can synthesize those signals into a narrative.

969
00:33:42,320 --> 00:33:43,840
What happened, what might be happening,

970
00:33:43,840 --> 00:33:45,360
what changed, who touched what

971
00:33:45,360 --> 00:33:47,480
and what it thinks the likely root cause is.

972
00:33:47,480 --> 00:33:49,040
That synthesis is valuable.

973
00:33:49,040 --> 00:33:50,200
It compresses time.

974
00:33:50,200 --> 00:33:53,320
It reduces the cognitive load of finding the story.

975
00:33:53,320 --> 00:33:56,360
It proposes interpretations a tired human might miss,

976
00:33:56,360 --> 00:33:57,640
but it is still cognition.

977
00:33:57,640 --> 00:33:59,080
It is still candidate thinking.

978
00:33:59,080 --> 00:34:01,680
So a well designed triage flow treats co-pilot output

979
00:34:01,680 --> 00:34:03,840
the same way it treats an analyst's first pass.

980
00:34:03,840 --> 00:34:05,920
Provisional, biased by incomplete data

981
00:34:05,920 --> 00:34:09,040
and requiring explicit judgment before consequence.

982
00:34:09,040 --> 00:34:11,120
The system of judgment is where a human does the work

983
00:34:11,120 --> 00:34:12,680
nobody wants to admit his work.

984
00:34:12,680 --> 00:34:14,040
Severity is not a number.

985
00:34:14,040 --> 00:34:15,600
It is an organizational decision

986
00:34:15,600 --> 00:34:18,720
about blast radius, tolerable risk, regulatory exposure

987
00:34:18,720 --> 00:34:19,840
and response posture.

988
00:34:19,840 --> 00:34:21,160
Intent is not a pattern match.

989
00:34:21,160 --> 00:34:23,280
Intent is a claim you will defend later

990
00:34:23,280 --> 00:34:25,360
in a post-incident review in an audit

991
00:34:25,360 --> 00:34:27,320
and possibly in a legal conversation.

992
00:34:27,320 --> 00:34:31,000
So the judgment step is where someone has to say out loud,

993
00:34:31,000 --> 00:34:32,720
this is a probable fishing compromise

994
00:34:32,720 --> 00:34:34,400
or this is credential stuffing

995
00:34:34,400 --> 00:34:36,360
or this is an internal misconfiguration

996
00:34:36,360 --> 00:34:37,800
with external symptoms.

997
00:34:37,800 --> 00:34:39,640
And then commit to a response posture.

998
00:34:39,640 --> 00:34:41,560
Contain now an accept disruption

999
00:34:41,560 --> 00:34:43,960
or observe longer and accept uncertainty.

1000
00:34:43,960 --> 00:34:45,560
That isn't something co-pilot can own.

1001
00:34:45,560 --> 00:34:46,640
That is a human decision

1002
00:34:46,640 --> 00:34:48,880
because the organization is the one paying the trade off.

1003
00:34:48,880 --> 00:34:51,320
Now the system of action has to make the decision real.

1004
00:34:51,320 --> 00:34:53,040
This is where most organizations fail.

1005
00:34:53,040 --> 00:34:56,000
They let cognition produce a beautiful triage summary.

1006
00:34:56,000 --> 00:34:57,840
They let a human nod at it

1007
00:34:57,840 --> 00:35:00,760
and then they drop back into chat driven execution.

1008
00:35:00,760 --> 00:35:02,640
Can someone reset the account?

1009
00:35:02,640 --> 00:35:04,400
Did we block the IP?

1010
00:35:04,400 --> 00:35:06,160
Who owns comms?

1011
00:35:06,160 --> 00:35:08,040
Are we notifying legal?

1012
00:35:08,040 --> 00:35:09,880
All those questions are action questions

1013
00:35:09,880 --> 00:35:11,800
and if they aren't enforced through workflow

1014
00:35:11,800 --> 00:35:13,400
they become optional under pressure.

1015
00:35:13,400 --> 00:35:15,920
So the action system needs to compile consequence.

1016
00:35:15,920 --> 00:35:17,920
Severity classification triggers a required

1017
00:35:17,920 --> 00:35:20,160
containment playbook, required approvals,

1018
00:35:20,160 --> 00:35:23,040
required communications, required evidence collection

1019
00:35:23,040 --> 00:35:25,440
and required post-incident review.

1020
00:35:25,440 --> 00:35:27,160
Not because service now is magic

1021
00:35:27,160 --> 00:35:28,720
because enforcement is the only way

1022
00:35:28,720 --> 00:35:30,240
to prevent the organization

1023
00:35:30,240 --> 00:35:33,000
from improvising itself into inconsistency.

1024
00:35:33,000 --> 00:35:35,200
Here's what the failure mode looks like in the real world.

1025
00:35:35,200 --> 00:35:36,800
Co-pilot flags risk.

1026
00:35:36,800 --> 00:35:39,280
It surfaces three plausible interpretations.

1027
00:35:39,280 --> 00:35:41,520
Everyone agrees it looks concerning.

1028
00:35:41,520 --> 00:35:43,200
And then nothing is enforced.

1029
00:35:43,200 --> 00:35:44,960
No one is named as the decision owner.

1030
00:35:44,960 --> 00:35:46,920
No one records the rationale for why

1031
00:35:46,920 --> 00:35:49,440
it was treated as low severity versus high.

1032
00:35:49,440 --> 00:35:51,600
No one triggers mandatory containment steps.

1033
00:35:51,600 --> 00:35:54,240
So the organization generates a lot of analysis artifacts

1034
00:35:54,240 --> 00:35:57,280
and a lot of chat messages, but no operational gravity.

1035
00:35:57,280 --> 00:35:58,240
Alert fatigue scales

1036
00:35:58,240 --> 00:36:00,240
because if every alert gets a better narrative

1037
00:36:00,240 --> 00:36:02,240
but still doesn't produce enforced decisions

1038
00:36:02,240 --> 00:36:04,840
you've improved the story and preserved the paralysis.

1039
00:36:04,840 --> 00:36:06,440
People stop trusting the summaries

1040
00:36:06,440 --> 00:36:07,360
not because they're wrong

1041
00:36:07,360 --> 00:36:10,040
but because the organization never does anything decisive

1042
00:36:10,040 --> 00:36:11,080
with them.

1043
00:36:11,080 --> 00:36:12,400
This is the subtle corrosion.

1044
00:36:12,400 --> 00:36:14,160
The SOC starts to look busy

1045
00:36:14,160 --> 00:36:16,400
while the risk posture stays unchanged.

1046
00:36:16,400 --> 00:36:18,320
Leadership sees dashboards and summaries

1047
00:36:18,320 --> 00:36:20,200
and assumes governance is functioning.

1048
00:36:20,200 --> 00:36:22,880
In reality, the accountability pathway is missing.

1049
00:36:22,880 --> 00:36:24,400
The decision moment was never designed.

1050
00:36:24,400 --> 00:36:26,840
So when the same pattern repeats nobody can say

1051
00:36:26,840 --> 00:36:29,640
last time we decided X because of Y and here's the record.

1052
00:36:29,640 --> 00:36:33,080
And this is where the line lands every single time.

1053
00:36:33,080 --> 00:36:35,080
Thinking without enforcement is fantasy.

1054
00:36:35,080 --> 00:36:37,520
If the triage summary doesn't force a choice

1055
00:36:37,520 --> 00:36:39,400
and the choice doesn't trigger consequence

1056
00:36:39,400 --> 00:36:41,000
then Copa-la didn't make you safer.

1057
00:36:41,000 --> 00:36:43,800
It made you more articulate while you remained indecisive.

1058
00:36:43,800 --> 00:36:44,920
The hard truth is this.

1059
00:36:44,920 --> 00:36:46,680
Analysis without action pathways

1060
00:36:46,680 --> 00:36:48,360
becomes decorative intelligence.

1061
00:36:48,360 --> 00:36:49,800
It produces confidence theater.

1062
00:36:49,800 --> 00:36:51,640
It produces the feeling that the organization

1063
00:36:51,640 --> 00:36:53,640
is handling risk because it can describe risk

1064
00:36:53,640 --> 00:36:54,680
in nice paragraphs.

1065
00:36:54,680 --> 00:36:56,040
But security is not narrative.

1066
00:36:56,040 --> 00:36:57,320
Security is consequence.

1067
00:36:57,320 --> 00:36:59,240
So the proper handoff is brutally simple.

1068
00:36:59,240 --> 00:37:01,760
Copa-la proposes interpretations.

1069
00:37:01,760 --> 00:37:04,200
A human selects intent and severity

1070
00:37:04,200 --> 00:37:06,440
and becomes the named owner of that decision.

1071
00:37:06,440 --> 00:37:08,440
The action system enforces the response path

1072
00:37:08,440 --> 00:37:10,000
and records evidence that records

1073
00:37:10,000 --> 00:37:12,120
survives the next incident review.

1074
00:37:12,120 --> 00:37:13,840
It also survives leadership denial

1075
00:37:13,840 --> 00:37:16,600
because it removes the ability to pretend nobody decided.

1076
00:37:16,600 --> 00:37:18,600
The AI didn't decide what mattered.

1077
00:37:18,600 --> 00:37:20,040
It decided what was possible.

1078
00:37:20,040 --> 00:37:21,800
The organization decided nothing.

1079
00:37:21,800 --> 00:37:23,400
And so nothing meaningful happened.

1080
00:37:23,400 --> 00:37:25,960
Scenario two, HR policy interpretation.

1081
00:37:25,960 --> 00:37:28,840
Security triage fails fast because pressure forces the cracks

1082
00:37:28,840 --> 00:37:30,800
to show HR policy fails differently.

1083
00:37:30,800 --> 00:37:33,800
It fails quietly, politely, and at scale.

1084
00:37:33,800 --> 00:37:35,560
Because HR policy is where organization

1085
00:37:35,560 --> 00:37:38,360
store ambiguity on purpose, exceptions, discretion,

1086
00:37:38,360 --> 00:37:41,360
manager judgment, union context, local law differences,

1087
00:37:41,360 --> 00:37:44,240
and the uncomfortable reality that fairness isn't a formula.

1088
00:37:44,240 --> 00:37:47,200
So when you drop a cognitive system into that environment,

1089
00:37:47,200 --> 00:37:48,840
you don't just get helpful drafts.

1090
00:37:48,840 --> 00:37:50,240
You get doctrine by accident.

1091
00:37:50,240 --> 00:37:51,960
Start again with the system of cognition.

1092
00:37:51,960 --> 00:37:55,120
Copa-la can read policy documents prior HR cases,

1093
00:37:55,120 --> 00:37:57,760
email threads, teams, chats, and whatever precedent

1094
00:37:57,760 --> 00:37:59,640
exists in tickets and knowledge articles.

1095
00:37:59,640 --> 00:38:01,240
It can summarize eligibility rules.

1096
00:38:01,240 --> 00:38:02,960
It can propose response language.

1097
00:38:02,960 --> 00:38:04,760
It can even generate a decision tree

1098
00:38:04,760 --> 00:38:06,440
that looks reasonable and calm.

1099
00:38:06,440 --> 00:38:09,240
This is the part leaders love because it feels like consistency.

1100
00:38:09,240 --> 00:38:10,280
It feels like scale.

1101
00:38:10,280 --> 00:38:13,840
It feels like standard answers, replacing messy human variation.

1102
00:38:13,840 --> 00:38:15,520
But the output is still cognition.

1103
00:38:15,520 --> 00:38:17,840
It is still a plausible interpretation of text.

1104
00:38:17,840 --> 00:38:20,800
The system of judgment is the manager, HR business partner,

1105
00:38:20,800 --> 00:38:22,600
or people leader who has to decide

1106
00:38:22,600 --> 00:38:25,040
what the organization is actually willing to stand behind.

1107
00:38:25,040 --> 00:38:27,800
That includes things policy text rarely captures,

1108
00:38:27,800 --> 00:38:31,240
context, intent, equity, cultural precedent,

1109
00:38:31,240 --> 00:38:32,600
and second order effects.

1110
00:38:32,600 --> 00:38:34,800
A classic example is exception handling,

1111
00:38:34,800 --> 00:38:36,520
an employee asks for a deviation.

1112
00:38:36,520 --> 00:38:38,960
Extended leave, remote work outside the policy,

1113
00:38:38,960 --> 00:38:41,280
and accommodation, a one off schedule change,

1114
00:38:41,280 --> 00:38:43,680
a benefit interpretation that doesn't fit the template.

1115
00:38:43,680 --> 00:38:46,640
The AI will generate a neat answer because neatness is its job.

1116
00:38:46,640 --> 00:38:47,720
It will cite sections.

1117
00:38:47,720 --> 00:38:49,320
It will propose language that sounds

1118
00:38:49,320 --> 00:38:51,400
considerate while staying compliant.

1119
00:38:51,400 --> 00:38:52,960
And this is where the failure begins

1120
00:38:52,960 --> 00:38:54,720
because exceptions are not technical.

1121
00:38:54,720 --> 00:38:57,560
Exceptions are moral and organizational decisions.

1122
00:38:57,560 --> 00:38:59,800
They set precedent, they change expectations,

1123
00:38:59,800 --> 00:39:02,920
they alter trust, they become stories employees tell each other,

1124
00:39:02,920 --> 00:39:05,400
which is how culture gets enforced in the real world.

1125
00:39:05,400 --> 00:39:08,840
So the judgment moment is not what does the policy say.

1126
00:39:08,840 --> 00:39:11,880
The judgment moment is, do we apply the policy strictly here?

1127
00:39:11,880 --> 00:39:13,000
Do we allow an exception?

1128
00:39:13,000 --> 00:39:14,800
And if we do, what boundary are we setting?

1129
00:39:14,800 --> 00:39:16,360
So this doesn't become the new rule.

1130
00:39:16,360 --> 00:39:18,200
If you don't force that decision to be explicit,

1131
00:39:18,200 --> 00:39:19,600
co-pilot will produce an answer

1132
00:39:19,600 --> 00:39:21,360
that becomes policy by repetition.

1133
00:39:21,360 --> 00:39:22,480
Now the system of action.

1134
00:39:22,480 --> 00:39:25,160
In HR, action is not created ticket.

1135
00:39:25,160 --> 00:39:28,240
Action is, record the decision, apply downstream effects,

1136
00:39:28,240 --> 00:39:30,320
and make the exception visible to governance.

1137
00:39:30,320 --> 00:39:32,240
That means the workflow has to capture

1138
00:39:32,240 --> 00:39:34,040
who decided what rationale they used,

1139
00:39:34,040 --> 00:39:36,640
what policy they referenced, what exception they granted,

1140
00:39:36,640 --> 00:39:38,400
and what review is required.

1141
00:39:38,400 --> 00:39:41,000
It also needs to trigger whatever consequences follow.

1142
00:39:41,000 --> 00:39:43,640
Payroll changes, access adjustments, manager approvals,

1143
00:39:43,640 --> 00:39:46,440
legal review, or a case review board, if required.

1144
00:39:46,440 --> 00:39:48,920
This is where thinking without enforcement is fantasy

1145
00:39:48,920 --> 00:39:49,720
gets sharper.

1146
00:39:49,720 --> 00:39:51,640
If co-pilot drafts a perfect response,

1147
00:39:51,640 --> 00:39:53,560
but the organization doesn't force the manager

1148
00:39:53,560 --> 00:39:55,160
to log the decision and own it,

1149
00:39:55,160 --> 00:39:57,520
then the only thing that scaled was deniability.

1150
00:39:57,520 --> 00:40:00,800
The manager can say, I followed what the system suggested.

1151
00:40:00,800 --> 00:40:02,800
HR can say, we didn't approve that.

1152
00:40:02,800 --> 00:40:05,360
Legal can say, that's not our interpretation.

1153
00:40:05,360 --> 00:40:07,200
And the employee hears one message.

1154
00:40:07,200 --> 00:40:09,720
The organization has no coherent owner for fairness.

1155
00:40:09,720 --> 00:40:10,960
That is not a people problem.

1156
00:40:10,960 --> 00:40:12,280
It is a system design problem.

1157
00:40:12,280 --> 00:40:14,320
Here's the predictable failure mode.

1158
00:40:14,320 --> 00:40:15,760
A manager pings co-pilot.

1159
00:40:15,760 --> 00:40:18,640
Employee is requesting X based on policy Y.

1160
00:40:18,640 --> 00:40:19,880
What should I say?

1161
00:40:19,880 --> 00:40:21,600
Co-pilot produces a clean answer.

1162
00:40:21,600 --> 00:40:22,800
The manager forwards it.

1163
00:40:22,800 --> 00:40:24,920
The employee accepts it as an official position.

1164
00:40:24,920 --> 00:40:27,560
A week later, another employee requests the same thing.

1165
00:40:27,560 --> 00:40:30,440
Someone else asks co-pilot, the wording is slightly different,

1166
00:40:30,440 --> 00:40:31,880
but the intent is similar.

1167
00:40:31,880 --> 00:40:33,520
Now you have inconsistency.

1168
00:40:33,520 --> 00:40:36,040
Or worse, you have a de facto rule that was never approved.

1169
00:40:36,040 --> 00:40:37,080
Then it escalates.

1170
00:40:37,080 --> 00:40:38,560
An employee disputes treatment.

1171
00:40:38,560 --> 00:40:39,680
HR investigates.

1172
00:40:39,680 --> 00:40:42,040
They ask, who approved this exception?

1173
00:40:42,040 --> 00:40:42,640
Nobody knows.

1174
00:40:42,640 --> 00:40:43,320
There's no record.

1175
00:40:43,320 --> 00:40:44,000
There was a chat.

1176
00:40:44,000 --> 00:40:44,680
There was an email.

1177
00:40:44,680 --> 00:40:46,440
There was a paragraph that sounded official.

1178
00:40:46,440 --> 00:40:47,640
But there was no decision log.

1179
00:40:47,640 --> 00:40:49,640
No owner, no rationale, no review.

1180
00:40:49,920 --> 00:40:53,360
So the system said so, which is the most corrosive phrase in any organization

1181
00:40:53,360 --> 00:40:57,880
because it converts leadership into bureaucracy and bureaucracy into moral abdication.

1182
00:40:57,880 --> 00:41:01,320
The uncomfortable point is that AI didn't create the ambiguity.

1183
00:41:01,320 --> 00:41:02,920
HR policy already contains it.

1184
00:41:02,920 --> 00:41:08,400
AI just makes it fast to manufacture official sounding interpretations of that ambiguity.

1185
00:41:08,400 --> 00:41:12,240
So the operational fix isn't tell managers not to use co-pilot.

1186
00:41:12,240 --> 00:41:13,200
That's fantasy too.

1187
00:41:13,200 --> 00:41:15,240
They will use it because speed wins.

1188
00:41:15,240 --> 00:41:16,640
The fix is to design the hand of.

1189
00:41:16,640 --> 00:41:20,040
Co-pilot can propose language and surface president patterns.

1190
00:41:20,040 --> 00:41:21,880
The manager must select intent.

1191
00:41:21,880 --> 00:41:25,600
Strict policy application approved exception or escalation required.

1192
00:41:25,600 --> 00:41:27,480
Then the workflow enforces consequence.

1193
00:41:27,480 --> 00:41:31,560
It records the decision, triggers approvals and creates an orderable trail.

1194
00:41:31,560 --> 00:41:35,440
And the line that matters here is the same one as security, just quieter.

1195
00:41:35,440 --> 00:41:37,360
The AI didn't decide what was fair.

1196
00:41:37,360 --> 00:41:38,560
It decided what was possible.

1197
00:41:38,560 --> 00:41:40,440
The organization decided nothing.

1198
00:41:40,440 --> 00:41:43,000
And then everyone acted like a decision happened anyway.

1199
00:41:43,000 --> 00:41:45,480
Scenario 3, IT change management.

1200
00:41:45,480 --> 00:41:49,640
IT change management is where outsourced judgment stops being a philosophical concern

1201
00:41:49,640 --> 00:41:51,960
and becomes an outage with a timestamp.

1202
00:41:51,960 --> 00:41:56,520
And it's where AI makes the failure cleaner, faster and harder to argue with because the

1203
00:41:56,520 --> 00:41:59,920
paperwork looks better right up until production breaks.

1204
00:41:59,920 --> 00:42:01,840
Start with the system of cognition.

1205
00:42:01,840 --> 00:42:04,600
Co-pilot can draft what humans hate writing.

1206
00:42:04,600 --> 00:42:08,920
Impact analysis, dependency lists, comms plans, rollback steps, stakeholder summaries and

1207
00:42:08,920 --> 00:42:11,240
the ritual language of risk and mitigation.

1208
00:42:11,240 --> 00:42:15,360
It can scan related tickets, past incidents, meeting notes and the change description itself

1209
00:42:15,360 --> 00:42:19,160
and produce something that sounds like a mature engineering organization.

1210
00:42:19,160 --> 00:42:20,240
This is the seductive part.

1211
00:42:20,240 --> 00:42:21,440
The artifact looks complete.

1212
00:42:21,440 --> 00:42:22,600
The tone is confident.

1213
00:42:22,600 --> 00:42:24,400
The structure matches what cab expects.

1214
00:42:24,400 --> 00:42:26,160
It reads like someone did the work.

1215
00:42:26,160 --> 00:42:27,720
But cognition is still not judgment.

1216
00:42:27,720 --> 00:42:30,400
It is still a proposal generator wearing a suit.

1217
00:42:30,400 --> 00:42:35,240
The system of judgment in change management is the part most organizations pretend is automated

1218
00:42:35,240 --> 00:42:36,240
already.

1219
00:42:36,240 --> 00:42:37,520
Acceptable blast radius.

1220
00:42:37,520 --> 00:42:39,400
Because blast radius is not a technical measurement.

1221
00:42:39,400 --> 00:42:43,560
It is an organizational decision about what the business is willing to lose today.

1222
00:42:43,560 --> 00:42:48,080
Reliability, performance, customer trust, operational focus so a change can happen.

1223
00:42:48,080 --> 00:42:49,760
The AI can list potential impacts.

1224
00:42:49,760 --> 00:42:51,280
It can propose mitigation steps.

1225
00:42:51,280 --> 00:42:53,040
It can recommend a maintenance window.

1226
00:42:53,040 --> 00:42:56,800
What it cannot do is decide which failure mode is tolerable, which stakeholder gets to

1227
00:42:56,800 --> 00:43:00,200
be angry and which risk is worth accepting to land the change.

1228
00:43:00,200 --> 00:43:04,640
That's the architects job or the change owners job or whoever will be sitting in the post-incident

1229
00:43:04,640 --> 00:43:07,760
review explaining why this was reasonable.

1230
00:43:07,760 --> 00:43:12,200
So the judgment moment is where someone must say explicitly, we are changing X.

1231
00:43:12,200 --> 00:43:16,880
We believe the blast radius is Y, the rollback posture is Z and we accept the residual

1232
00:43:16,880 --> 00:43:17,880
risk.

1233
00:43:17,880 --> 00:43:19,520
Without that sentence you don't have a change.

1234
00:43:19,520 --> 00:43:22,480
You have a document now the system of action is where this becomes real.

1235
00:43:22,480 --> 00:43:26,120
A change workflow should enforce the consequences of that judgment.

1236
00:43:26,120 --> 00:43:31,680
Approvals, blackout windows, evidence attachments, pre-change checks, implementation steps and

1237
00:43:31,680 --> 00:43:36,320
mandatory post implementation review if certain conditions are met.

1238
00:43:36,320 --> 00:43:39,320
Action systems exist because humans under pressure improvise.

1239
00:43:39,320 --> 00:43:40,320
They skip steps.

1240
00:43:40,320 --> 00:43:41,720
They rely on memory.

1241
00:43:41,720 --> 00:43:43,440
They just do it real quick.

1242
00:43:43,440 --> 00:43:46,600
Workflow exists to stop improvisation from becoming policy.

1243
00:43:46,600 --> 00:43:48,560
Here's the failure mode AI introduces.

1244
00:43:48,560 --> 00:43:50,600
Copilot produces a polished impact analysis.

1245
00:43:50,600 --> 00:43:52,120
It proposes a rollback plan.

1246
00:43:52,120 --> 00:43:53,600
It suggests comms language.

1247
00:43:53,600 --> 00:43:57,880
The change record looks good enough that reviewers assume the hard thinking already happened.

1248
00:43:57,880 --> 00:43:59,440
The cab meeting moves faster.

1249
00:43:59,440 --> 00:44:04,000
Approvals happen by momentum and the organization confuses completeness with correctness.

1250
00:44:04,000 --> 00:44:05,480
Then the change lands.

1251
00:44:05,480 --> 00:44:07,120
Something unexpected happens.

1252
00:44:07,120 --> 00:44:11,280
Not necessarily a hallucination problem, a context problem, a hidden dependency, a permission

1253
00:44:11,280 --> 00:44:15,920
edge case, a replication lag, an integration that was out of scope in the mind of the person

1254
00:44:15,920 --> 00:44:19,240
who requested the change but very much in scope in production reality.

1255
00:44:19,240 --> 00:44:22,920
Now you have an outage and you start asking the only question that matters.

1256
00:44:22,920 --> 00:44:24,880
Who decided this was acceptable?

1257
00:44:24,880 --> 00:44:28,720
In too many organizations the answer is a shrug disguised as process.

1258
00:44:28,720 --> 00:44:32,960
The change record has text, it has risk language, it has mitigation bullets, it even has a rollback

1259
00:44:32,960 --> 00:44:37,400
section, but nobody can point to the judgment moment where a human accepted the blast radius

1260
00:44:37,400 --> 00:44:38,600
and owned the trade-off.

1261
00:44:38,600 --> 00:44:41,920
Because the artifact got treated as the decision and that's outsourced judgment in its

1262
00:44:41,920 --> 00:44:46,040
purest form, the document looks like governance, so governance is assumed to exist.

1263
00:44:46,040 --> 00:44:49,400
This is why looks fine is one of the most expensive phrases in IT.

1264
00:44:49,400 --> 00:44:52,280
AI increases the number of things that look fine.

1265
00:44:52,280 --> 00:44:53,280
That's the problem.

1266
00:44:53,280 --> 00:44:56,800
The correct handoff is not copilot writes the change record and we move on.

1267
00:44:56,800 --> 00:45:00,880
The correct handoff is copilot proposes the thinking artifact, a human selects intent

1268
00:45:00,880 --> 00:45:05,800
and declares risk posture and the workflow enforces that posture with irreversible constraints.

1269
00:45:05,800 --> 00:45:10,920
For example, if the human selects high blast radius, the action system should require senior

1270
00:45:10,920 --> 00:45:15,760
approval, enforce a tighter window, require a tested rollback and force a post implementation

1271
00:45:15,760 --> 00:45:17,360
review.

1272
00:45:17,360 --> 00:45:22,120
If the human selects low blast radius, the workflow can be lighter, but it should still

1273
00:45:22,120 --> 00:45:24,160
record who made that classification.

1274
00:45:24,160 --> 00:45:28,240
That is how you turn change management from paper compliance into operational gravity.

1275
00:45:28,240 --> 00:45:32,120
And if you want to diagnostic that cuts through the theatre, ask this, can you reconstruct

1276
00:45:32,120 --> 00:45:36,040
the decision owner and their rationale for every outage causing change in the last year?

1277
00:45:36,040 --> 00:45:37,880
If you can't, you don't have change management.

1278
00:45:37,880 --> 00:45:41,840
You have a calendar, so the uncomfortable point stands even here, the AI didn't cause

1279
00:45:41,840 --> 00:45:43,000
the outage.

1280
00:45:43,000 --> 00:45:44,240
Unowned judgment did.

1281
00:45:44,240 --> 00:45:47,880
The AI generated what was possible, the organization never forced a human to decide what

1282
00:45:47,880 --> 00:45:52,680
was acceptable and the system enforced nothing until production enforced it for you.

1283
00:45:52,680 --> 00:45:55,080
The skills AI makes more valuable.

1284
00:45:55,080 --> 00:45:56,080
Judgment

1285
00:45:56,080 --> 00:45:58,880
After the scenarios, the pattern should feel obvious.

1286
00:45:58,880 --> 00:46:00,680
AI didn't break the organization.

1287
00:46:00,680 --> 00:46:04,880
It revealed what was missing and what's missing almost everywhere is judgment.

1288
00:46:04,880 --> 00:46:08,040
Not intelligence, not information, not output capacity.

1289
00:46:08,040 --> 00:46:12,480
Judgment, the ability to choose under uncertainty, to name trade-offs and to own consequences

1290
00:46:12,480 --> 00:46:13,480
in public.

1291
00:46:13,480 --> 00:46:15,680
AI increases the volume of plausible options.

1292
00:46:15,680 --> 00:46:19,280
That means it increases the volume of uncertainty you have to arbitrate.

1293
00:46:19,280 --> 00:46:23,040
If your organization was already weak at arbitration, AI doesn't help.

1294
00:46:23,040 --> 00:46:24,480
It just accelerates the collapse.

1295
00:46:24,480 --> 00:46:25,680
This is the uncomfortable truth.

1296
00:46:25,680 --> 00:46:27,680
AI doesn't replace decision making.

1297
00:46:27,680 --> 00:46:29,800
It industrializes decision pressure.

1298
00:46:29,800 --> 00:46:31,480
Before the bottleneck was production.

1299
00:46:31,480 --> 00:46:35,360
Writing the draft, building the deck, pulling the data, creating the status report, AI makes

1300
00:46:35,360 --> 00:46:36,360
those cheap.

1301
00:46:36,360 --> 00:46:40,160
Which means the scarce resource becomes the part nobody can automate.

1302
00:46:40,160 --> 00:46:44,640
Deciding what matters, what is acceptable, and what you are willing to be accountable for.

1303
00:46:44,640 --> 00:46:49,400
And when judgment becomes the scarce resource, organizations do what they always do with scarcity.

1304
00:46:49,400 --> 00:46:50,400
They try to outsource it.

1305
00:46:50,400 --> 00:46:55,120
They treat the AI's output as if it is a decision because the output looks like a decision.

1306
00:46:55,120 --> 00:46:59,040
It has structure, it has confidence, it has bullet points, it has a neat conclusion.

1307
00:46:59,040 --> 00:47:01,480
In a culture trained to worship throughput, that's enough.

1308
00:47:01,480 --> 00:47:03,400
The artifact gets mistaken for ownership.

1309
00:47:03,400 --> 00:47:04,720
But ownership is not a format.

1310
00:47:04,720 --> 00:47:05,840
Ownership is a person.

1311
00:47:05,840 --> 00:47:07,440
Judgment is not being smart.

1312
00:47:07,440 --> 00:47:10,120
It's arbitration under constraints you didn't choose.

1313
00:47:10,120 --> 00:47:11,120
Time.

1314
00:47:11,120 --> 00:47:12,120
Politics.

1315
00:47:12,120 --> 00:47:13,120
Risk.

1316
00:47:13,120 --> 00:47:14,120
Unclear data.

1317
00:47:14,120 --> 00:47:15,120
Conflicting incentives.

1318
00:47:15,120 --> 00:47:18,320
The human job is to take that mess and still pick a direction with a rationale that

1319
00:47:18,320 --> 00:47:20,560
survives contact with reality.

1320
00:47:20,560 --> 00:47:24,240
AI cannot do that job because AI cannot pay the cost of being wrong.

1321
00:47:24,240 --> 00:47:25,240
So here's the shift.

1322
00:47:25,240 --> 00:47:29,000
In an AI rich environment, judgment becomes the primary differentiator between teams that

1323
00:47:29,000 --> 00:47:31,800
scale capability and teams that scale confusion.

1324
00:47:31,800 --> 00:47:34,760
Strong judgment looks boring in the way auditors love.

1325
00:47:34,760 --> 00:47:35,760
It names assumptions.

1326
00:47:35,760 --> 00:47:37,600
It identifies what's unknown.

1327
00:47:37,600 --> 00:47:39,520
It states what would change the decision.

1328
00:47:39,520 --> 00:47:40,720
It documents the trade off.

1329
00:47:40,720 --> 00:47:43,240
It assigns a decision owner who is not a committee.

1330
00:47:43,240 --> 00:47:45,320
It doesn't hide behind wheel monitor.

1331
00:47:45,320 --> 00:47:48,520
It defines what monitoring means and what triggers action.

1332
00:47:48,520 --> 00:47:50,480
Weak judgment in contrast looks productive.

1333
00:47:50,480 --> 00:47:51,480
It ships more drafts.

1334
00:47:51,480 --> 00:47:52,640
It generates more options.

1335
00:47:52,640 --> 00:47:54,440
It produces more analysis artifacts.

1336
00:47:54,440 --> 00:47:55,440
It uses more tools.

1337
00:47:55,440 --> 00:47:56,520
It schedules more meetings.

1338
00:47:56,520 --> 00:47:58,040
It creates more alignment.

1339
00:47:58,040 --> 00:48:01,880
It creates activity that feels like work while avoiding the moment where someone says,

1340
00:48:01,880 --> 00:48:04,960
"I choose this and I accept the consequences."

1341
00:48:04,960 --> 00:48:07,200
That is why AI makes weak leadership worse.

1342
00:48:07,200 --> 00:48:10,600
It gives weak leadership more ways to avoid choosing while still looking busy.

1343
00:48:10,600 --> 00:48:14,840
Now take the executive lens because executives are the ones most tempted to outsource judgment.

1344
00:48:14,840 --> 00:48:15,840
They live in abstraction.

1345
00:48:15,840 --> 00:48:16,840
They receive summaries.

1346
00:48:16,840 --> 00:48:19,440
They make decisions based on compressed narratives.

1347
00:48:19,440 --> 00:48:22,760
So an AI generated summary feels like a perfect fit.

1348
00:48:22,760 --> 00:48:25,920
Better compression, more coverage, fewer gaps.

1349
00:48:25,920 --> 00:48:27,360
Except it doesn't reduce gaps.

1350
00:48:27,360 --> 00:48:28,360
It hides them.

1351
00:48:28,360 --> 00:48:31,440
The KPI for leadership in an AI environment can't be throughput.

1352
00:48:31,440 --> 00:48:32,440
Thruput will always go up.

1353
00:48:32,440 --> 00:48:36,640
The KPI has to be decision quality and decision quality is visible in outcomes.

1354
00:48:36,640 --> 00:48:39,320
And in the trail of reasoning you can defend later.

1355
00:48:39,320 --> 00:48:42,840
The hard part is that judgment discipline requires admitting uncertainty.

1356
00:48:42,840 --> 00:48:46,440
Publicly, in writing, with your name attached, most organizations punish that.

1357
00:48:46,440 --> 00:48:47,720
They reward certainty theatre.

1358
00:48:47,720 --> 00:48:52,200
They reward confidence even when it's fake because fake confidence moves meetings forward.

1359
00:48:52,200 --> 00:48:56,360
AI produces fake confidence on demand so the temptation becomes structural.

1360
00:48:56,360 --> 00:48:59,440
This is why outsource judgment is the real failure mode.

1361
00:48:59,440 --> 00:49:01,000
Hallucinations are easy to blame.

1362
00:49:01,000 --> 00:49:03,800
Decision avoidance is harder because it implicates everyone.

1363
00:49:03,800 --> 00:49:07,720
So if you want a practical definition of judgment in this model, here it is.

1364
00:49:07,720 --> 00:49:10,440
Judgment is selecting a trade-off and making it explicit.

1365
00:49:10,440 --> 00:49:12,920
Everything else is commentary.

1366
00:49:12,920 --> 00:49:17,120
When a security lead classifies an incident as low severity, they are choosing a trade-off.

1367
00:49:17,120 --> 00:49:19,400
Test disruption now, more risk later.

1368
00:49:19,400 --> 00:49:22,840
When a manager grants an HR exception, they are choosing a trade-off.

1369
00:49:22,840 --> 00:49:25,360
Compassion now, precedent later.

1370
00:49:25,360 --> 00:49:28,240
When an architect approves a change, they are choosing a trade-off.

1371
00:49:28,240 --> 00:49:30,600
Progress now, blast radius later.

1372
00:49:30,600 --> 00:49:33,880
When finance holds forecast guidance, they are choosing a trade-off.

1373
00:49:33,880 --> 00:49:35,880
Stability now, credibility later.

1374
00:49:35,880 --> 00:49:37,440
AI can list those trade-offs.

1375
00:49:37,440 --> 00:49:40,680
It cannot choose one that aligns with your values and constraints.

1376
00:49:40,680 --> 00:49:41,920
That choice is leadership.

1377
00:49:41,920 --> 00:49:45,600
And the organization needs to start treating that as the work, not as an overhead.

1378
00:49:45,600 --> 00:49:46,800
So the skill stack shifts.

1379
00:49:46,800 --> 00:49:50,600
The person who can arbitrate under uncertainty becomes more valuable than the person who

1380
00:49:50,600 --> 00:49:52,080
can draft under pressure.

1381
00:49:52,080 --> 00:49:56,160
The team that can frame intent and enforce ownership will outperform the team that can

1382
00:49:56,160 --> 00:49:58,080
generate 100 plausible paths.

1383
00:49:58,080 --> 00:50:01,800
The enterprise that builds judgment moments into workflows will scale trust.

1384
00:50:01,800 --> 00:50:03,880
The enterprise that doesn't will scale deniability.

1385
00:50:03,880 --> 00:50:05,200
And that's the final irony.

1386
00:50:05,200 --> 00:50:06,760
AI will solve this productivity.

1387
00:50:06,760 --> 00:50:10,800
What it actually does is expose whether your organization is capable of responsible choice.

1388
00:50:10,800 --> 00:50:12,880
If you are, AI amplifies clarity.

1389
00:50:12,880 --> 00:50:15,040
If you aren't, it amplifies entropy.

1390
00:50:15,040 --> 00:50:18,000
The skills AI makes more valuable problem framing.

1391
00:50:18,000 --> 00:50:22,040
If judgment is the scarce resource, problem framing is the control surface that determines

1392
00:50:22,040 --> 00:50:24,360
whether judgment even has a chance.

1393
00:50:24,360 --> 00:50:26,840
Most organizations treat framing like preamble.

1394
00:50:26,840 --> 00:50:30,760
A few sentences at the top of a dock, then they rush to the real work.

1395
00:50:30,760 --> 00:50:32,640
AI punishes that habit immediately.

1396
00:50:32,640 --> 00:50:34,840
Because if you don't frame the problem, the model will.

1397
00:50:34,840 --> 00:50:38,680
And it will frame it using whatever scraps of context it can infer.

1398
00:50:38,680 --> 00:50:43,400
Your wording, your org's historical artifacts, and the ambient assumptions embedded in your

1399
00:50:43,400 --> 00:50:44,400
data.

1400
00:50:44,400 --> 00:50:47,000
And the collaboration that's abdication.

1401
00:50:47,000 --> 00:50:51,560
Problem framing is the act of declaring intent, constraints, and success conditions before

1402
00:50:51,560 --> 00:50:53,160
you generate options.

1403
00:50:53,160 --> 00:50:57,440
It is the difference between, help me write something and help me decide something.

1404
00:50:57,440 --> 00:51:00,080
AI is excellent at generating possibilities.

1405
00:51:00,080 --> 00:51:04,680
It is useless at deciding relevance unless you tell it what relevance means here.

1406
00:51:04,680 --> 00:51:07,240
This is why AI often produces polished nonsense.

1407
00:51:07,240 --> 00:51:11,320
Not because the model is broken, but because the question was, a bad frame produces outputs

1408
00:51:11,320 --> 00:51:14,200
that are internally coherent and externally wrong.

1409
00:51:14,200 --> 00:51:17,920
They sound reasonable, but they solve a different problem than the one you actually have.

1410
00:51:17,920 --> 00:51:19,600
They optimize for the wrong stakeholder.

1411
00:51:19,600 --> 00:51:20,960
They assume the wrong constraints.

1412
00:51:20,960 --> 00:51:22,800
They move fast in the wrong direction.

1413
00:51:22,800 --> 00:51:25,960
And the organization accepts them because they read well and nobody wants to admit the

1414
00:51:25,960 --> 00:51:27,280
initial ask was vague.

1415
00:51:27,280 --> 00:51:28,680
That's the core failure.

1416
00:51:28,680 --> 00:51:30,960
Vague intent scales into confident artifacts.

1417
00:51:30,960 --> 00:51:34,400
Good framing fixes that by doing three things up front.

1418
00:51:34,400 --> 00:51:35,960
First it declares the decision type.

1419
00:51:35,960 --> 00:51:40,200
Is this a recommendation, a policy interpretation, a communication draft, an escalation, or an

1420
00:51:40,200 --> 00:51:41,200
irreversible action?

1421
00:51:41,200 --> 00:51:44,280
If you can't name the decision type, you can't govern the output.

1422
00:51:44,280 --> 00:51:46,440
You can't decide what validation is required.

1423
00:51:46,440 --> 00:51:48,040
You can't decide who needs to own it.

1424
00:51:48,040 --> 00:51:49,520
You're just generating text.

1425
00:51:49,520 --> 00:51:51,320
Second it declares constraints.

1426
00:51:51,320 --> 00:51:52,960
Not aspirational constraints.

1427
00:51:52,960 --> 00:51:54,240
Real constraints.

1428
00:51:54,240 --> 00:51:55,400
What must be true?

1429
00:51:55,400 --> 00:51:56,400
What must not happen?

1430
00:51:56,400 --> 00:51:57,880
What sources are authoritative?

1431
00:51:57,880 --> 00:51:59,160
What time window matters?

1432
00:51:59,160 --> 00:52:00,640
What risk threshold applies?

1433
00:52:00,640 --> 00:52:02,120
What you are not allowed to claim?

1434
00:52:02,120 --> 00:52:03,320
Constraints are the rails.

1435
00:52:03,320 --> 00:52:05,280
Without rails you get plausible drift.

1436
00:52:05,280 --> 00:52:07,040
Third it declares success conditions.

1437
00:52:07,040 --> 00:52:09,640
What will a good output enable a human to do?

1438
00:52:09,640 --> 00:52:14,080
Decide severity, notify stakeholders approve a change, reject an exception.

1439
00:52:14,080 --> 00:52:16,720
If the output doesn't change an action, it's entertainment.

1440
00:52:16,720 --> 00:52:18,400
And here's the uncomfortable part.

1441
00:52:18,400 --> 00:52:20,200
Framing is work that can't be outsourced.

1442
00:52:20,200 --> 00:52:22,720
It's upstream judgment in its purest form.

1443
00:52:22,720 --> 00:52:26,400
It forces leaders to say what they want, what they believe, and what they're willing to

1444
00:52:26,400 --> 00:52:27,400
trade.

1445
00:52:27,400 --> 00:52:28,400
That's why most leaders skip it.

1446
00:52:28,400 --> 00:52:33,000
They prefer to ask for a strategy or a plan and let the AI fill the void.

1447
00:52:33,000 --> 00:52:37,280
Then they act surprised when the plan is generic or optimistic or politically incoherent.

1448
00:52:37,280 --> 00:52:41,560
The AI didn't fail, the leader refused to specify what reality they are operating in.

1449
00:52:41,560 --> 00:52:45,200
So the discipline has to be explicit and it has to be small enough that people will actually

1450
00:52:45,200 --> 00:52:46,200
do it.

1451
00:52:46,200 --> 00:52:48,480
Here's the simplest habit that changes everything.

1452
00:52:48,480 --> 00:52:51,640
Before you share any AI output write a one sentence frame.

1453
00:52:51,640 --> 00:52:53,800
One sentence, not a template, not a workshop.

1454
00:52:53,800 --> 00:52:57,640
A sentence that names intent, constraints, and the next action.

1455
00:52:57,640 --> 00:52:58,640
Examples sound like this.

1456
00:52:58,640 --> 00:52:59,640
Draft for discussion.

1457
00:52:59,640 --> 00:53:04,000
Propose three response options to this incident summary, assuming we prioritize containment

1458
00:53:04,000 --> 00:53:07,160
over uptime and we must preserve ordered evidence.

1459
00:53:07,160 --> 00:53:08,160
Enterpretation.

1460
00:53:08,160 --> 00:53:13,160
Summarize how policy X applies to this request and flag where manager discretion or escalation

1461
00:53:13,160 --> 00:53:16,360
is required without implying an approval.

1462
00:53:16,360 --> 00:53:17,360
Change assessment.

1463
00:53:17,360 --> 00:53:21,400
List dependencies and rollback steps for this deployment, assuming a two hour window and

1464
00:53:21,400 --> 00:53:23,440
zero tolerance for data loss.

1465
00:53:23,440 --> 00:53:24,440
Notice what that does.

1466
00:53:24,440 --> 00:53:26,400
It forces the human to state the trade off.

1467
00:53:26,400 --> 00:53:28,120
It forces constraints into the open.

1468
00:53:28,120 --> 00:53:31,640
It makes it harder to forward the output as if it is a finished decision.

1469
00:53:31,640 --> 00:53:35,280
It reattaches cognition to intent and it also makes the AI better.

1470
00:53:35,280 --> 00:53:38,680
Not because you found a better prompt, because you finally told the system what game you're

1471
00:53:38,680 --> 00:53:39,680
playing.

1472
00:53:39,680 --> 00:53:42,680
This is the bridge between the prompt obsession chapter and reality.

1473
00:53:42,680 --> 00:53:44,600
Prompts are downstream, framing is upstream.

1474
00:53:44,600 --> 00:53:47,560
If the upstream is incoherent, downstream effort is just trash.

1475
00:53:47,560 --> 00:53:52,720
In other words, good framing makes fewer prompts necessary, because it collapses ambiguity

1476
00:53:52,720 --> 00:53:54,240
before it becomes output.

1477
00:53:54,240 --> 00:53:58,320
And framing is also how you stop the social failure mode where answer-shaped text becomes

1478
00:53:58,320 --> 00:53:59,560
truth by repetition.

1479
00:53:59,560 --> 00:54:03,720
When the first line of an email says draft for discussion, the organization has a chance

1480
00:54:03,720 --> 00:54:05,160
to treat it as a draft.

1481
00:54:05,160 --> 00:54:07,720
And it says nothing, the organization treats it as policy.

1482
00:54:07,720 --> 00:54:11,720
So if you want the high leverage mental shift, stop asking AI for answers and stop asking

1483
00:54:11,720 --> 00:54:13,360
your people for prompts.

1484
00:54:13,360 --> 00:54:14,680
Ask for frames.

1485
00:54:14,680 --> 00:54:17,680
Because the frame determines what the organization is actually doing.

1486
00:54:17,680 --> 00:54:20,080
Thinking, deciding, or just producing language.

1487
00:54:20,080 --> 00:54:23,280
And if you can't tell which one it is, the organization can't either.

1488
00:54:23,280 --> 00:54:26,880
The skills AI makes more valuable, context and ethics.

1489
00:54:26,880 --> 00:54:29,080
Problem framing gets you into the right room.

1490
00:54:29,080 --> 00:54:31,880
Context tells you what not to say once you're in it.

1491
00:54:31,880 --> 00:54:36,000
And AI is terrible at that because context is mostly made of things humans don't write

1492
00:54:36,000 --> 00:54:37,000
down.

1493
00:54:37,000 --> 00:54:42,120
History, power, timing, reputational risk and the quiet rules about what the organization

1494
00:54:42,120 --> 00:54:43,600
is willing to admit.

1495
00:54:43,600 --> 00:54:46,840
Copilot can summarize what happened, it can infer what might be happening.

1496
00:54:46,840 --> 00:54:50,600
It can propose what could be said, it cannot reliably know what this organization is

1497
00:54:50,600 --> 00:54:54,480
allowed to say to whom, right now, with these stakeholders watching.

1498
00:54:54,480 --> 00:54:55,480
That's context.

1499
00:54:55,480 --> 00:54:59,440
It's local, it's situational, it's political, it's often ethical, and it's where most

1500
00:54:59,440 --> 00:55:01,200
real world damage happens.

1501
00:55:01,200 --> 00:55:03,840
This is why just draft the email is not a safe ask.

1502
00:55:03,840 --> 00:55:08,600
The email isn't just words, it's a commitment, it creates expectations, it becomes evidence,

1503
00:55:08,600 --> 00:55:11,240
it shapes how people interpret intent.

1504
00:55:11,240 --> 00:55:15,280
And AI will happily draft something that is technically well written and strategically

1505
00:55:15,280 --> 00:55:18,800
catastrophic because it doesn't feel the cost of being wrong.

1506
00:55:18,800 --> 00:55:22,800
Context awareness is the human skill of mapping an output to the actual environment.

1507
00:55:22,800 --> 00:55:23,960
Who is impacted?

1508
00:55:23,960 --> 00:55:24,960
What is at stake?

1509
00:55:24,960 --> 00:55:28,720
What is irreversible and what will be misunderstood on first read?

1510
00:55:28,720 --> 00:55:33,600
Most organizations don't teach this explicitly because they assume seniority equals context.

1511
00:55:33,600 --> 00:55:35,800
It doesn't, seniority often equals abstraction.

1512
00:55:35,800 --> 00:55:40,360
AI makes that worse by making it easy to operate at the level of summary without ever touching

1513
00:55:40,360 --> 00:55:41,360
reality.

1514
00:55:41,360 --> 00:55:45,680
So you need a discipline before you act on AI output, you ask a context question that the

1515
00:55:45,680 --> 00:55:47,120
model can't answer for you.

1516
00:55:47,120 --> 00:55:49,520
What is the consequence of being wrong here?

1517
00:55:49,520 --> 00:55:51,960
Who gets harmed if we sound confident and we're wrong?

1518
00:55:51,960 --> 00:55:54,640
What downstream systems will treat this as authoritative?

1519
00:55:54,640 --> 00:55:58,240
What parts of this statement are unverifiable even if they sound reasonable?

1520
00:55:58,240 --> 00:55:59,440
It's not paranoia.

1521
00:55:59,440 --> 00:56:00,760
That's governance of meaning.

1522
00:56:00,760 --> 00:56:06,200
Now add ethics because people like to pretend ethics is optional until it becomes a headline.

1523
00:56:06,200 --> 00:56:08,440
Ethical reasoning is not a compliance module.

1524
00:56:08,440 --> 00:56:12,400
It's the ability to notice when a decision is being disguised as a suggestion, when an

1525
00:56:12,400 --> 00:56:16,640
exception is being normalized or when a probabilistic summary is being treated as truth because

1526
00:56:16,640 --> 00:56:18,040
it feels convenient.

1527
00:56:18,040 --> 00:56:20,200
The dangerous part of AI isn't that it lies.

1528
00:56:20,200 --> 00:56:23,720
The dangerous part is that it speaks fluently in the language of legitimacy.

1529
00:56:23,720 --> 00:56:26,240
If it drafts an HR response, it sounds like HR.

1530
00:56:26,240 --> 00:56:28,600
If it drafts a security update, it sounds like security.

1531
00:56:28,600 --> 00:56:31,440
If it drafts an executive statement, it sounds like leadership.

1532
00:56:31,440 --> 00:56:34,200
Tone becomes camouflage.

1533
00:56:34,200 --> 00:56:36,240
That's how bias and unfairness scale.

1534
00:56:36,240 --> 00:56:39,840
Not through obvious malice but through plausible defaults that nobody challenges because they

1535
00:56:39,840 --> 00:56:41,040
read well.

1536
00:56:41,040 --> 00:56:43,280
And bias is rarely just demographic.

1537
00:56:43,280 --> 00:56:45,960
Organizational bias shows up as invisible priorities.

1538
00:56:45,960 --> 00:56:50,600
Protect the schedule, protect the narrative, protect the executive, minimize escalation, avoid

1539
00:56:50,600 --> 00:56:52,120
admitting uncertainty.

1540
00:56:52,120 --> 00:56:56,120
AI absorbs those priorities from the artifacts you already produced.

1541
00:56:56,120 --> 00:56:59,080
And it amplifies them because that's what Pat and synthesis does.

1542
00:56:59,080 --> 00:57:01,360
So the ethical posture has to be explicit.

1543
00:57:01,360 --> 00:57:02,360
Humans own outcomes.

1544
00:57:02,360 --> 00:57:03,360
Tools do not.

1545
00:57:03,360 --> 00:57:05,360
Not in principle, in practice.

1546
00:57:05,360 --> 00:57:09,440
That means you can't let the system set so become the justification for a decision that

1547
00:57:09,440 --> 00:57:12,880
affects someone's access, pay, job, health or reputation.

1548
00:57:12,880 --> 00:57:16,440
If your organization accepts that excuse, you've built a moral escape hatch.

1549
00:57:16,440 --> 00:57:19,120
And once people have an escape hatch, responsibility diffuses.

1550
00:57:19,120 --> 00:57:20,520
Here's the operational test.

1551
00:57:20,520 --> 00:57:23,160
Can a human defend the decision without mentioning the AI?

1552
00:57:23,160 --> 00:57:25,280
If the answer is no, you didn't make a decision.

1553
00:57:25,280 --> 00:57:29,480
You delegated one and you delegated it to a system that cannot be accountable.

1554
00:57:29,480 --> 00:57:33,200
This also ties back to the triad because ethics lives in the handoff between cognition and

1555
00:57:33,200 --> 00:57:34,200
action.

1556
00:57:34,200 --> 00:57:35,200
Cognition proposes.

1557
00:57:35,200 --> 00:57:36,200
Fine.

1558
00:57:36,200 --> 00:57:39,560
But the moment you operationalize an output, send the email, deny the request, contain

1559
00:57:39,560 --> 00:57:44,040
the account, publish the policy interpretation you moved into action and action creates consequences

1560
00:57:44,040 --> 00:57:45,720
whether you meant to or not.

1561
00:57:45,720 --> 00:57:50,360
So the judgment moment must include an ethics check that is almost offensively simple.

1562
00:57:50,360 --> 00:57:51,360
Who is affected?

1563
00:57:51,360 --> 00:57:52,360
What is the harm?

1564
00:57:52,360 --> 00:57:54,560
If this is wrong, what is the appeal path if we got it wrong?

1565
00:57:54,560 --> 00:57:58,680
If you can't answer those questions, you are not ready to operationalize the output.

1566
00:57:58,680 --> 00:58:00,520
You're still thinking or pretending you are.

1567
00:58:00,520 --> 00:58:04,760
This is why the AI mindset shift is not about being nicer to the machine or learning more

1568
00:58:04,760 --> 00:58:05,760
prompt tricks.

1569
00:58:05,760 --> 00:58:09,880
It's about reasserting human agency where the platform makes agency easy to forget.

1570
00:58:09,880 --> 00:58:11,720
AI will generate text.

1571
00:58:11,720 --> 00:58:12,720
That's the cheap part now.

1572
00:58:12,720 --> 00:58:14,520
The expensive part is context.

1573
00:58:14,520 --> 00:58:17,160
Knowing what the text means in the world you actually live in.

1574
00:58:17,160 --> 00:58:19,480
And the most expensive part is ethics.

1575
00:58:19,480 --> 00:58:22,000
Owning who gets hurt when you pretend the machine made the call.

1576
00:58:22,000 --> 00:58:26,560
If you can hold context in ethics together, you can use AI without dissolving responsibility.

1577
00:58:26,560 --> 00:58:30,960
If you can't, the organization will keep scaling fluency and calling it competence.

1578
00:58:30,960 --> 00:58:35,760
And that's how judgment erodes quietly until the day it becomes undeniable.

1579
00:58:35,760 --> 00:58:38,800
The atrophy pattern, why junior roles break first?

1580
00:58:38,800 --> 00:58:42,680
If you want to see the future of AI in your organization, don't watch the executives.

1581
00:58:42,680 --> 00:58:43,880
Watch the junior roles.

1582
00:58:43,880 --> 00:58:45,280
They fail first.

1583
00:58:45,280 --> 00:58:49,280
Not because they're less capable, but because they sit closest to the work that AI replaces

1584
00:58:49,280 --> 00:58:55,680
cleanly, drafting, summarizing, formatting, translating chaos into something that looks coherent.

1585
00:58:55,680 --> 00:58:57,600
And that work was never just output.

1586
00:58:57,600 --> 00:58:58,920
It was apprenticeship.

1587
00:58:58,920 --> 00:59:02,200
Junior roles learn judgment by doing the boring parts under supervision.

1588
00:59:02,200 --> 00:59:06,320
They learn what a good answer looks like because they had to produce 10 bad ones and get

1589
00:59:06,320 --> 00:59:07,320
corrected.

1590
00:59:07,320 --> 00:59:10,360
They learn what matters because someone read line they're draft and told them why.

1591
00:59:10,360 --> 00:59:14,120
They learn what's defensible because their manager asked, "How do you know that?"

1592
00:59:14,120 --> 00:59:15,600
And they had to point to evidence.

1593
00:59:15,600 --> 00:59:19,440
AI short circuits that entire loop because now the junior can generate a competent looking

1594
00:59:19,440 --> 00:59:23,560
artifact on the first try, not a correct artifact, a competent looking one.

1595
00:59:23,560 --> 00:59:27,720
And the organization being addicted to speed will accept looks fine as proof of progress,

1596
00:59:27,720 --> 00:59:30,400
especially when the writing is clean and the bullets are crisp.

1597
00:59:30,400 --> 00:59:33,160
This is how capability collapses while throughput rises.

1598
00:59:33,160 --> 00:59:37,800
The junior never builds the muscle for framing the problem, selecting sources, checking assumptions

1599
00:59:37,800 --> 00:59:39,120
or defending trade-offs.

1600
00:59:39,120 --> 00:59:42,840
They build a different muscle, prompt iteration and output polishing.

1601
00:59:42,840 --> 00:59:43,840
That is not judgment.

1602
00:59:43,840 --> 00:59:45,360
That is interface literacy.

1603
00:59:45,360 --> 00:59:49,080
And then leadership acts surprised when the next generation can't run the system without

1604
00:59:49,080 --> 00:59:50,080
the tool.

1605
00:59:50,080 --> 00:59:51,960
Here's the structural reason this hits junior's harder.

1606
00:59:51,960 --> 00:59:57,000
Their work is template shaped, status updates, notes, first drafts, customer emails, requirements,

1607
00:59:57,000 --> 00:59:58,680
summaries, basic analysis.

1608
00:59:58,680 --> 01:00:00,760
AI is an infinite template engine.

1609
01:00:00,760 --> 01:00:05,280
So the junior's role gets absorbed into the machine's lowest cost capability and the organization

1610
01:00:05,280 --> 01:00:06,800
calls it efficiency.

1611
01:00:06,800 --> 01:00:09,960
But apprenticeship isn't efficient, it's expensive by design.

1612
01:00:09,960 --> 01:00:15,240
It forces slow thinking, it forces error, it forces correction, it forces the junior

1613
01:00:15,240 --> 01:00:17,760
to internalize standards instead of borrowing them.

1614
01:00:17,760 --> 01:00:20,720
When you remove that friction you don't create a faster junior.

1615
01:00:20,720 --> 01:00:24,480
You create a person who can move text around without understanding what it means and you'll

1616
01:00:24,480 --> 01:00:28,320
still promote them because the surface level signals look good.

1617
01:00:28,320 --> 01:00:31,480
Now add the second failure mode, overconfidence.

1618
01:00:31,480 --> 01:00:35,200
Fluent text triggers authority bias, people read something well written and assume it was

1619
01:00:35,200 --> 01:00:36,200
well thought.

1620
01:00:36,200 --> 01:00:41,560
AI exploits that bias perfectly because it produces tone, structure and certainty on demand.

1621
01:00:41,560 --> 01:00:44,320
So the junior becomes dangerous in a way they weren't before.

1622
01:00:44,320 --> 01:00:47,560
They can ship plausible decisions faster than they can recognize risk.

1623
01:00:47,560 --> 01:00:48,720
That's not a moral indictment.

1624
01:00:48,720 --> 01:00:51,840
It's a predictable outcome of how humans evaluate language.

1625
01:00:51,840 --> 01:00:54,800
And this is where the atrophy pattern becomes visible.

1626
01:00:54,800 --> 01:00:56,120
Strong performers get stronger.

1627
01:00:56,120 --> 01:00:57,360
Weak performers get hidden.

1628
01:00:57,360 --> 01:00:59,520
A strong junior uses AI as scaffolding.

1629
01:00:59,520 --> 01:01:00,720
They ask better questions.

1630
01:01:00,720 --> 01:01:01,720
They validate outputs.

1631
01:01:01,720 --> 01:01:04,000
They treat the model as a sparring partner.

1632
01:01:04,000 --> 01:01:07,320
They improve faster because they have a base layer of judgment and they use the tool

1633
01:01:07,320 --> 01:01:08,320
to extend it.

1634
01:01:08,320 --> 01:01:11,080
A weak junior uses AI as an authority proxy.

1635
01:01:11,080 --> 01:01:12,880
They accept the first plausible answer.

1636
01:01:12,880 --> 01:01:15,160
They stop learning the underlying domain.

1637
01:01:15,160 --> 01:01:16,840
They confuse speed with competence.

1638
01:01:16,840 --> 01:01:20,600
And because the artifacts look professional, the organization can't tell the difference until

1639
01:01:20,600 --> 01:01:23,720
the person is in a role where mistakes have blast radius.

1640
01:01:23,720 --> 01:01:25,800
That widening gap is the real workforce risk.

1641
01:01:25,800 --> 01:01:28,040
AI doesn't democratize expertise.

1642
01:01:28,040 --> 01:01:30,320
It amplifies existing judgment disparities.

1643
01:01:30,320 --> 01:01:32,920
You already see this pattern in every other system.

1644
01:01:32,920 --> 01:01:34,240
Tools don't fix bad thinking.

1645
01:01:34,240 --> 01:01:35,240
They scale it.

1646
01:01:35,240 --> 01:01:38,920
AI is just the first tool that produces outputs that look like thinking.

1647
01:01:38,920 --> 01:01:40,960
So what happens organizationally?

1648
01:01:40,960 --> 01:01:45,680
Many people stop delegating work to juniors and start delegating judgment to the AI.

1649
01:01:45,680 --> 01:01:48,960
Juniors stop being trained and start being throughput amplifiers.

1650
01:01:48,960 --> 01:01:52,760
Managers stop coaching and start reviewing polish drafts that contain hidden errors.

1651
01:01:52,760 --> 01:01:54,400
The feedback loop degrades.

1652
01:01:54,400 --> 01:01:55,760
Correction becomes sporadic.

1653
01:01:55,760 --> 01:01:56,760
Standards drift.

1654
01:01:56,760 --> 01:01:59,320
And the organization quietly loses its bench strengths.

1655
01:01:59,320 --> 01:02:03,560
Then six months later leadership complains about AI skills gaps and buys more training.

1656
01:02:03,560 --> 01:02:05,320
But the gap isn't AI usage.

1657
01:02:05,320 --> 01:02:06,320
It's judgment development.

1658
01:02:06,320 --> 01:02:09,800
And this is why the will train later lie is especially toxic for junior roles.

1659
01:02:09,800 --> 01:02:13,680
The first two weeks set defaults and juniors are the ones most likely to adopt defaults

1660
01:02:13,680 --> 01:02:14,680
as doctrine.

1661
01:02:14,680 --> 01:02:16,800
If the default is copy paste, they will copy paste.

1662
01:02:16,800 --> 01:02:20,240
If the default is the AI knows, they will believe it.

1663
01:02:20,240 --> 01:02:23,680
If the default is speed gets rewarded, they will optimize for speed.

1664
01:02:23,680 --> 01:02:24,760
That's not rebellion.

1665
01:02:24,760 --> 01:02:26,280
That's adaptation.

1666
01:02:26,280 --> 01:02:30,560
So if you care about talent, you have to treat AI as a threat to apprenticeship, not just

1667
01:02:30,560 --> 01:02:31,640
a productivity win.

1668
01:02:31,640 --> 01:02:33,640
You need design judgment moments.

1669
01:02:33,640 --> 01:02:36,280
Not only for risk and compliance, but for learning.

1670
01:02:36,280 --> 01:02:41,280
This is where juniors must explain rationale, site sources and state confidence before their

1671
01:02:41,280 --> 01:02:42,880
work becomes action.

1672
01:02:42,880 --> 01:02:47,040
Otherwise you will build an organization that ships faster, sounds smarter and understands

1673
01:02:47,040 --> 01:02:48,040
less.

1674
01:02:48,040 --> 01:02:49,040
And that is not transformation.

1675
01:02:49,040 --> 01:02:52,080
That is atrophy with better formatting.

1676
01:02:52,080 --> 01:02:53,080
Organizational readiness.

1677
01:02:53,080 --> 01:02:55,160
AI, psychological safety.

1678
01:02:55,160 --> 01:02:59,080
So the junior roles break first, but the organization fails for a different reason.

1679
01:02:59,080 --> 01:03:03,280
It fails because people stop speaking clearly when they don't feel safe to be wrong.

1680
01:03:03,280 --> 01:03:05,560
AI, psychological safety isn't a soft concept.

1681
01:03:05,560 --> 01:03:09,520
It's an operational requirement because cognitive collaboration requires a behavior most enterprises

1682
01:03:09,520 --> 01:03:12,160
actively punish, saying, "I don't know, I disagree."

1683
01:03:12,160 --> 01:03:14,080
Or, "This output feels wrong."

1684
01:03:14,080 --> 01:03:17,360
And if people can't say that out loud, the system will drift until it hits a wall.

1685
01:03:17,360 --> 01:03:21,560
There are two kinds of fear that show up immediately when you introduce co-pilot at scale.

1686
01:03:21,560 --> 01:03:24,600
The first is the obvious one, fear of being replaced.

1687
01:03:24,600 --> 01:03:29,680
That fear makes people hide work, hoard context, and avoid experimenting in ways that would

1688
01:03:29,680 --> 01:03:31,760
make them look less necessary.

1689
01:03:31,760 --> 01:03:35,720
They reduce transparency because transparency feels like volunteering for redundancy.

1690
01:03:35,720 --> 01:03:39,200
The second fear is quieter and more corrosive.

1691
01:03:39,200 --> 01:03:43,080
Fear of being wrong on behalf of the AI.

1692
01:03:43,080 --> 01:03:47,200
Because when you send an AI drafted email or share an AI-generated summary, you aren't

1693
01:03:47,200 --> 01:03:48,520
just wrong privately.

1694
01:03:48,520 --> 01:03:50,200
You are wrong in a way that looks official.

1695
01:03:50,200 --> 01:03:51,800
You are wrong in a way that spreads.

1696
01:03:51,800 --> 01:03:55,600
You are wrong in a way that can be screenshot, forwarded, and quoted later.

1697
01:03:55,600 --> 01:03:56,600
So people adapt.

1698
01:03:56,600 --> 01:03:59,080
They stop experimenting in high visibility channels.

1699
01:03:59,080 --> 01:04:01,000
They use AI privately and quietly.

1700
01:04:01,000 --> 01:04:04,760
They paste the output with minimal edits because editing takes time.

1701
01:04:04,760 --> 01:04:07,440
And if they are going to be blamed anyway, they might as well move fast.

1702
01:04:07,440 --> 01:04:10,880
Or they avoid AI entirely and resent the people who use it.

1703
01:04:10,880 --> 01:04:15,480
Because the cultural signal becomes, speed is rewarded, mistakes are punished, and ambiguity

1704
01:04:15,480 --> 01:04:16,640
is your problem.

1705
01:04:16,640 --> 01:04:18,320
That is not readiness.

1706
01:04:18,320 --> 01:04:20,200
That is a pressure cooker.

1707
01:04:20,200 --> 01:04:22,640
Now add the enterprise's favorite habit.

1708
01:04:22,640 --> 01:04:23,840
Perfection culture.

1709
01:04:23,840 --> 01:04:27,720
Most organizations have spent decades training people to pretend certainty.

1710
01:04:27,720 --> 01:04:32,240
They reward the person who sounds confident, not the person who asks the better question.

1711
01:04:32,240 --> 01:04:36,360
They reward the deck that looks complete, not the decision that's defensible.

1712
01:04:36,360 --> 01:04:38,280
They treat ambiguity as incompetence.

1713
01:04:38,280 --> 01:04:40,560
AI turns that culture into a liability.

1714
01:04:40,560 --> 01:04:43,840
Because AI produces certainty theatre at industrial scale.

1715
01:04:43,840 --> 01:04:48,120
If the organization already struggles to tolerate uncertainty, it will accept confident output

1716
01:04:48,120 --> 01:04:49,120
as relief.

1717
01:04:49,120 --> 01:04:50,400
Not because it's right.

1718
01:04:50,400 --> 01:04:53,400
Because it ends the uncomfortable pause where someone has to think.

1719
01:04:53,400 --> 01:04:56,880
So psychological safety in an AI environment is not about protecting feelings.

1720
01:04:56,880 --> 01:04:58,560
It's about protecting dissent.

1721
01:04:58,560 --> 01:05:02,680
It's about creating permission structurally for people to say three things without career

1722
01:05:02,680 --> 01:05:03,680
damage.

1723
01:05:03,680 --> 01:05:04,960
This is a draft.

1724
01:05:04,960 --> 01:05:06,800
This is a hypothesis.

1725
01:05:06,800 --> 01:05:08,760
This needs escalation.

1726
01:05:08,760 --> 01:05:13,000
If those sentences don't exist in your culture, co-pilot will become a narrative engine

1727
01:05:13,000 --> 01:05:15,920
that nobody challenges until the consequences arrive.

1728
01:05:15,920 --> 01:05:17,520
And the consequences always arrive.

1729
01:05:17,520 --> 01:05:19,280
Here's what most organizations get wrong.

1730
01:05:19,280 --> 01:05:21,400
They treat safety as a communications problem.

1731
01:05:21,400 --> 01:05:22,840
They run an awareness campaign.

1732
01:05:22,840 --> 01:05:24,920
They say it's okay to experiment.

1733
01:05:24,920 --> 01:05:26,680
They tell people there are no stupid questions.

1734
01:05:26,680 --> 01:05:29,600
They publish a Viva-Engaged community with weekly tips.

1735
01:05:29,600 --> 01:05:30,600
They make posters.

1736
01:05:30,600 --> 01:05:31,600
They do prompt-a-thons.

1737
01:05:31,600 --> 01:05:33,280
They do all the visible things.

1738
01:05:33,280 --> 01:05:36,320
Then the first person gets burned for a mistake made with AI.

1739
01:05:36,320 --> 01:05:38,240
Maybe they send a draft with the wrong number.

1740
01:05:38,240 --> 01:05:40,280
Maybe they summarize the meeting incorrectly.

1741
01:05:40,280 --> 01:05:42,600
Maybe they use the wrong tone in a sensitive message.

1742
01:05:42,600 --> 01:05:44,680
Maybe they interpreted policy too confidently.

1743
01:05:44,680 --> 01:05:45,880
It doesn't matter.

1744
01:05:45,880 --> 01:05:47,200
What matters is the outcome.

1745
01:05:47,200 --> 01:05:50,680
The organization punishes the mistake as if it was traditional negligence instead of

1746
01:05:50,680 --> 01:05:54,480
recognizing it as a predictable failure mode of probabilistic cognition.

1747
01:05:54,480 --> 01:05:56,440
And at that moment experimentation stops.

1748
01:05:56,440 --> 01:05:58,120
Not officially, quietly.

1749
01:05:58,120 --> 01:05:59,920
People don't need a memo to learn what is safe.

1750
01:05:59,920 --> 01:06:03,160
They watch what happens to the first person who gets embarrassed in public.

1751
01:06:03,160 --> 01:06:05,480
The culture sets the rule in a single incident.

1752
01:06:05,480 --> 01:06:10,080
So readiness requires something more disciplined, a shared language for good AI usage that teams

1753
01:06:10,080 --> 01:06:12,080
can use to self-correct without shame.

1754
01:06:12,080 --> 01:06:13,840
Not use co-pilot more.

1755
01:06:13,840 --> 01:06:15,320
Use these sentences.

1756
01:06:15,320 --> 01:06:17,040
My confidence is low.

1757
01:06:17,040 --> 01:06:18,480
Here are the sources I used.

1758
01:06:18,480 --> 01:06:20,000
Here's what I didn't verify.

1759
01:06:20,000 --> 01:06:21,520
Here's the decision owner.

1760
01:06:21,520 --> 01:06:25,040
Those statements are how you keep cognition from becoming doctrine.

1761
01:06:25,040 --> 01:06:29,120
They also reduce the load on managers because managers can't review everything.

1762
01:06:29,120 --> 01:06:32,720
They need signals that tell them when to trust, when to dig, and when to escalate.

1763
01:06:32,720 --> 01:06:36,200
Now connect this to governance because safety without guardrails becomes chaos.

1764
01:06:36,200 --> 01:06:39,280
You can't tell people to experiment and then give them no boundaries.

1765
01:06:39,280 --> 01:06:41,800
That creates fear too because nobody knows what is allowed.

1766
01:06:41,800 --> 01:06:44,160
People either freeze or they root around controls.

1767
01:06:44,160 --> 01:06:46,720
So psychological safety has a minimum viable boundary.

1768
01:06:46,720 --> 01:06:47,880
Make judgment visible.

1769
01:06:47,880 --> 01:06:49,160
That's the whole requirement.

1770
01:06:49,160 --> 01:06:53,040
If someone uses AI to produce an output that will influence a decision, they must attach

1771
01:06:53,040 --> 01:06:54,400
a judgment statement.

1772
01:06:54,400 --> 01:06:56,760
But it is what it's for and who owns it.

1773
01:06:56,760 --> 01:06:58,240
If they can't, they don't ship it.

1774
01:06:58,240 --> 01:06:59,920
This is not a training initiative.

1775
01:06:59,920 --> 01:07:01,560
It is behavioral design.

1776
01:07:01,560 --> 01:07:05,760
And it's how you create a culture where people can collaborate with AI without surrendering

1777
01:07:05,760 --> 01:07:06,920
responsibility.

1778
01:07:06,920 --> 01:07:09,260
Because the goal isn't to make everyone comfortable.

1779
01:07:09,260 --> 01:07:13,080
The goal is to make avoidance impossible and descent safe enough to surface before the

1780
01:07:13,080 --> 01:07:15,760
incident review does it for you.

1781
01:07:15,760 --> 01:07:19,080
Minimal irreversible prescriptions that remove deniability.

1782
01:07:19,080 --> 01:07:21,600
So here's where most episodes like this go off the rails.

1783
01:07:21,600 --> 01:07:25,080
They hear judgment and they immediately build a maturity model.

1784
01:07:25,080 --> 01:07:29,400
A road map, a center of excellence, a 12 week adoption program with badges and a sharepoint

1785
01:07:29,400 --> 01:07:32,040
hub and a congratulatory town hall.

1786
01:07:32,040 --> 01:07:35,040
That is how organizations turn a real problem into a safe ceremony.

1787
01:07:35,040 --> 01:07:36,480
This needs the opposite.

1788
01:07:36,480 --> 01:07:38,800
Minimal prescriptions that remove deniability.

1789
01:07:38,800 --> 01:07:41,320
Things you can't agree with and then ignore.

1790
01:07:41,320 --> 01:07:44,120
Things that make avoidance visible immediately in normal work.

1791
01:07:44,120 --> 01:07:45,120
Three of them.

1792
01:07:45,120 --> 01:07:46,120
That's it.

1793
01:07:46,120 --> 01:07:47,120
First, decision logs.

1794
01:07:47,120 --> 01:07:48,440
Not a repository, not a weekly project.

1795
01:07:48,440 --> 01:07:51,640
A single page artifact that exists because memory is not governance.

1796
01:07:51,640 --> 01:07:55,280
Every meaningful decision influenced by AI gets a decision log.

1797
01:07:55,280 --> 01:07:56,280
One page.

1798
01:07:56,280 --> 01:07:57,280
Short.

1799
01:07:57,280 --> 01:07:58,280
Owned by a named human.

1800
01:07:58,280 --> 01:08:00,080
Stored where the action system can reference it.

1801
01:08:00,080 --> 01:08:03,000
It contains five fields and none of them are optional.

1802
01:08:03,000 --> 01:08:04,000
What is the decision?

1803
01:08:04,000 --> 01:08:05,000
Who owns it?

1804
01:08:05,000 --> 01:08:06,240
What options were considered?

1805
01:08:06,240 --> 01:08:07,240
What tradeoff was chosen?

1806
01:08:07,240 --> 01:08:08,600
What would cause a reversal?

1807
01:08:08,600 --> 01:08:10,800
That last one matters more than people admit.

1808
01:08:10,800 --> 01:08:11,960
Because it forces honesty.

1809
01:08:11,960 --> 01:08:15,560
It makes you state what evidence would change your mind, which is how you prevent post-hoc

1810
01:08:15,560 --> 01:08:16,720
rationalization.

1811
01:08:16,720 --> 01:08:20,840
The log explicitly says whether AI was used for cognition, summarization, drafting, option

1812
01:08:20,840 --> 01:08:22,800
generation, not to shame anyone.

1813
01:08:22,800 --> 01:08:27,000
To make the epistemology explicit, so later when someone asks why did we do this, the answer

1814
01:08:27,000 --> 01:08:29,440
isn't because the email sounded reasonable.

1815
01:08:29,440 --> 01:08:32,040
Second, judgment moments embedded in workflow.

1816
01:08:32,040 --> 01:08:36,200
This is the part that most organizations avoid because it feels like slowing down.

1817
01:08:36,200 --> 01:08:38,400
It does slow down on purpose.

1818
01:08:38,400 --> 01:08:42,240
A judgment moment is a force checkpoint where a human must classify intent before the

1819
01:08:42,240 --> 01:08:48,960
system allows action to proceed, not reviewed, not looks good, classification, severity,

1820
01:08:48,960 --> 01:08:53,400
risk posture, exception approved or denied, escalation required or not, customer impact

1821
01:08:53,400 --> 01:08:57,440
statement accepted or rejected, change blast radius accepted or rejected.

1822
01:08:57,440 --> 01:08:59,360
If you can't classify, you can't proceed.

1823
01:08:59,360 --> 01:09:02,400
This is what turns thinking into accountable movement.

1824
01:09:02,400 --> 01:09:06,440
It's also what stops Copilot from becoming a narrative engine that quietly drives decisions

1825
01:09:06,440 --> 01:09:07,440
through momentum.

1826
01:09:07,440 --> 01:09:09,000
And no, this is not bureaucracy.

1827
01:09:09,000 --> 01:09:10,560
This is entropy management.

1828
01:09:10,560 --> 01:09:14,320
Because the organization will always drift toward the easiest path.

1829
01:09:14,320 --> 01:09:19,240
Fast outputs, unclear ownership and plausible deniability when it fails.

1830
01:09:19,240 --> 01:09:21,680
Judgment moments are how you make that drift expensive.

1831
01:09:21,680 --> 01:09:25,640
Third, name decision owners, individuals, not teams.

1832
01:09:25,640 --> 01:09:30,320
The phrase the business decided is a compliance evasion tactic, so is IT approved it, so is

1833
01:09:30,320 --> 01:09:36,480
HR said, teams execute, committees advise, tools propose only an individual can be held

1834
01:09:36,480 --> 01:09:37,480
accountable.

1835
01:09:37,480 --> 01:09:41,840
So every judgment moment needs a person attached, their role can be delegated, their decision

1836
01:09:41,840 --> 01:09:43,120
can be informed by others.

1837
01:09:43,120 --> 01:09:46,920
But their name is the anchor that stops responsibility from diffusing into the org chart.

1838
01:09:46,920 --> 01:09:49,520
Now if you do those three things, here's what changes.

1839
01:09:49,520 --> 01:09:53,760
AI stops being a magic answer machine and becomes what it actually is, a cognitive surface

1840
01:09:53,760 --> 01:09:58,200
that generates options and your organization becomes what it has to be, a place where choices

1841
01:09:58,200 --> 01:10:00,480
are owned, recorded and enforced.

1842
01:10:00,480 --> 01:10:04,880
This is also where the M365 to service now handoff stops being a vague integration story

1843
01:10:04,880 --> 01:10:06,760
and becomes a governance mechanism.

1844
01:10:06,760 --> 01:10:09,800
Co-pilot produces synthesis and proposed interpretations.

1845
01:10:09,800 --> 01:10:14,240
Fine, but the only thing that can move an organization is a decision that triggers enforced

1846
01:10:14,240 --> 01:10:15,240
consequences.

1847
01:10:15,240 --> 01:10:19,120
That means co-pilot output must land in a workflow that contains a judgment moment,

1848
01:10:19,120 --> 01:10:21,160
a named owner and a decision log reference.

1849
01:10:21,160 --> 01:10:23,800
For example, incident triage.

1850
01:10:23,800 --> 01:10:26,640
Co-pilot proposes three plausible interpretations.

1851
01:10:26,640 --> 01:10:31,200
Credential theft, misconfigured conditional access or benign automation gone noisy.

1852
01:10:31,200 --> 01:10:32,720
It provides evidence links.

1853
01:10:32,720 --> 01:10:33,720
Great.

1854
01:10:33,720 --> 01:10:35,680
Then the human selects intent.

1855
01:10:35,680 --> 01:10:37,960
Credential, credential theft and select severity.

1856
01:10:37,960 --> 01:10:38,960
High.

1857
01:10:38,960 --> 01:10:42,720
That selection writes the decision owner into the record and forces a rationale field.

1858
01:10:42,720 --> 01:10:46,200
Now the action system enforces containment steps and approvals.

1859
01:10:46,200 --> 01:10:47,520
It creates the audit trail.

1860
01:10:47,520 --> 01:10:49,760
It schedules the post-incident review.

1861
01:10:49,760 --> 01:10:53,400
It blocks the organization from improvising itself into inconsistency.

1862
01:10:53,400 --> 01:10:54,640
That's the handoff.

1863
01:10:54,640 --> 01:10:56,280
Cognition proposes possibilities.

1864
01:10:56,280 --> 01:10:57,760
Judgment selects intent.

1865
01:10:57,760 --> 01:10:59,600
Action enforces consequence.

1866
01:10:59,600 --> 01:11:01,200
And notice what disappears.

1867
01:11:01,200 --> 01:11:03,160
The ability to pretend.

1868
01:11:03,160 --> 01:11:06,320
You can't say the system said so because the system didn't decide.

1869
01:11:06,320 --> 01:11:07,640
A human did.

1870
01:11:07,640 --> 01:11:10,600
You can't say nobody approved this because the owner is recorded.

1871
01:11:10,600 --> 01:11:14,440
You can't say we didn't know because the options and evidence are attached.

1872
01:11:14,440 --> 01:11:16,760
This is why the prescriptions must be irreversible.

1873
01:11:16,760 --> 01:11:18,240
They don't teach people to be better.

1874
01:11:18,240 --> 01:11:22,520
They force the enterprise to behave like it believes its own risk post-gematters.

1875
01:11:22,520 --> 01:11:25,840
You want the simplest test for whether your AI program is real?

1876
01:11:25,840 --> 01:11:28,880
Pick any AI influence decision from last week and ask.

1877
01:11:28,880 --> 01:11:30,120
Where is the judgment moment?

1878
01:11:30,120 --> 01:11:31,120
Who owned it?

1879
01:11:31,120 --> 01:11:32,120
And where is the log?

1880
01:11:32,120 --> 01:11:35,520
You can't answer in under 60 seconds you didn't scale capability.

1881
01:11:35,520 --> 01:11:37,320
You scaled plausible deniability?

1882
01:11:37,320 --> 01:11:38,320
The future of work.

1883
01:11:38,320 --> 01:11:40,240
From execution to evaluation.

1884
01:11:40,240 --> 01:11:44,800
So once you accept the triad, cognition, judgment, action, you also accept the future of work

1885
01:11:44,800 --> 01:11:46,680
you are trying to avoid.

1886
01:11:46,680 --> 01:11:48,000
Execution becomes cheap.

1887
01:11:48,000 --> 01:11:49,760
Evaluation becomes expensive.

1888
01:11:49,760 --> 01:11:50,920
That isn't a slogan.

1889
01:11:50,920 --> 01:11:52,920
It's the new cost model of knowledge work.

1890
01:11:52,920 --> 01:11:55,880
Most organizations build their identity around execution.

1891
01:11:55,880 --> 01:11:59,160
Shipping documents, shipping decks, shipping tickets, shipping updates.

1892
01:11:59,160 --> 01:12:03,480
They rewarded the person who could produce the artifact fastest because artifacts were expensive.

1893
01:12:03,480 --> 01:12:05,480
AI collapses that scarcity.

1894
01:12:05,480 --> 01:12:07,320
The artifact is no longer the proof of work.

1895
01:12:07,320 --> 01:12:09,440
It's the receipt that work might have happened.

1896
01:12:09,440 --> 01:12:12,640
And that distinction matters because enterprises don't run on receipts.

1897
01:12:12,640 --> 01:12:13,720
They run on consequences.

1898
01:12:13,720 --> 01:12:15,160
So the future looks like this.

1899
01:12:15,160 --> 01:12:17,880
Meeting stop being places where people create content in real time.

1900
01:12:17,880 --> 01:12:18,880
AI will do that.

1901
01:12:18,880 --> 01:12:23,160
Notes, summaries, action items, option lists, draft narratives, done.

1902
01:12:23,160 --> 01:12:26,880
The meeting becomes the place where someone has to choose what matters, what changes,

1903
01:12:26,880 --> 01:12:28,280
and what gets enforced.

1904
01:12:28,280 --> 01:12:30,800
Which means the real meeting isn't the calendar invite.

1905
01:12:30,800 --> 01:12:34,920
The real meeting is the judgment moment embedded in the workflow afterward.

1906
01:12:34,920 --> 01:12:38,520
That's where the organization either becomes accountable or becomes theatrical.

1907
01:12:38,520 --> 01:12:39,520
Documents change too.

1908
01:12:39,520 --> 01:12:41,440
A document is no longer a deliverable.

1909
01:12:41,440 --> 01:12:43,480
It's a dialogue between cognition and judgment.

1910
01:12:43,480 --> 01:12:45,600
AI produces drafts.

1911
01:12:45,600 --> 01:12:48,080
Humans state intent, constraints, and ownership.

1912
01:12:48,080 --> 01:12:51,720
The value shifts from did we write it to, can we defend it?

1913
01:12:51,720 --> 01:12:53,080
So the writing isn't the work.

1914
01:12:53,080 --> 01:12:54,080
The reasoning is.

1915
01:12:54,080 --> 01:12:56,560
And the most uncomfortable shift is for leadership.

1916
01:12:56,560 --> 01:13:01,520
Leadership has spent years outsourcing reasoning to process committees and status reporting.

1917
01:13:01,520 --> 01:13:03,160
AI makes that outsourcing easier.

1918
01:13:03,160 --> 01:13:04,520
That's why this episode exists.

1919
01:13:04,520 --> 01:13:07,560
In the AI era, leaders don't get judged on throughput.

1920
01:13:07,560 --> 01:13:09,160
Thruput is free.

1921
01:13:09,160 --> 01:13:11,240
They get judged on decision quality.

1922
01:13:11,240 --> 01:13:13,000
Decision quality shows up in three ways.

1923
01:13:13,000 --> 01:13:16,320
Fewer reversals, fewer surprises, and fewer unowned outcomes.

1924
01:13:16,320 --> 01:13:20,800
It shows up when incident response has a record, when policy exceptions have a rationale,

1925
01:13:20,800 --> 01:13:25,320
when changes have an owner, when forecast trigger actual constraints instead of more slides.

1926
01:13:25,320 --> 01:13:27,320
And if that sounds like bureaucracy, good.

1927
01:13:27,320 --> 01:13:29,200
That's the part you were already living with.

1928
01:13:29,200 --> 01:13:33,520
The difference is that now bureaucracy is optional and judgment is mandatory.

1929
01:13:33,520 --> 01:13:37,040
Because if you don't embed judgment into the system, you don't get fast innovation

1930
01:13:37,040 --> 01:13:38,040
here.

1931
01:13:38,040 --> 01:13:39,040
You get fast entropy.

1932
01:13:39,040 --> 01:13:42,160
You get people shipping answer-shaped text into the organization without any enforcement

1933
01:13:42,160 --> 01:13:46,200
pathway attached and then acting surprised when it becomes doctrine by repetition.

1934
01:13:46,200 --> 01:13:49,480
This is also where agents become dangerous in a very specific way.

1935
01:13:49,480 --> 01:13:50,480
Agents will act.

1936
01:13:50,480 --> 01:13:51,480
They will call APIs.

1937
01:13:51,480 --> 01:13:52,480
They will execute tasks.

1938
01:13:52,480 --> 01:13:53,480
They will move data.

1939
01:13:53,480 --> 01:13:54,480
They will send messages.

1940
01:13:54,480 --> 01:13:55,480
They will close tickets.

1941
01:13:55,480 --> 01:14:00,240
They will do all the things your organization currently does slowly and inconsistently.

1942
01:14:00,240 --> 01:14:03,120
But an agent cannot own the decision that justified the action.

1943
01:14:03,120 --> 01:14:07,400
So if you scale agents without scaling judgment moments, you don't create autonomy.

1944
01:14:07,400 --> 01:14:09,360
You create automation without accountability.

1945
01:14:09,360 --> 01:14:12,200
You create conditional chaos with a nicer user interface.

1946
01:14:12,200 --> 01:14:16,840
Over time, the organization drifts into a probabilistic operating model.

1947
01:14:16,840 --> 01:14:17,840
Outputs happen.

1948
01:14:17,840 --> 01:14:19,640
Actions happen and nobody can explain why.

1949
01:14:19,640 --> 01:14:20,640
That is not modern work.

1950
01:14:20,640 --> 01:14:22,040
That is unmanaged delegation.

1951
01:14:22,040 --> 01:14:26,400
So the only sustainable pattern is this, cognition proposes, judgment selects action

1952
01:14:26,400 --> 01:14:27,400
and forces.

1953
01:14:27,400 --> 01:14:29,080
And the trick is not building more intelligence.

1954
01:14:29,080 --> 01:14:31,440
The trick is building operational gravity.

1955
01:14:31,440 --> 01:14:35,400
Constraints that force a human to declare intent before the system is allowed to move.

1956
01:14:35,400 --> 01:14:40,640
That is what separates an enterprise that scales trust from an enterprise that scales confusion.

1957
01:14:40,640 --> 01:14:43,680
Because AI doesn't eliminate work, it relocates it.

1958
01:14:43,680 --> 01:14:48,160
It moves the effort from "make the thing" to "decide the thing".

1959
01:14:48,160 --> 01:14:51,680
From "write the answer to own the consequences".

1960
01:14:51,680 --> 01:14:54,560
And produce the artifact to defend the rationale.

1961
01:14:54,560 --> 01:14:58,180
And if your organization refuses that shift, it will keep buying intelligence and keep

1962
01:14:58,180 --> 01:15:02,080
suffering the same failures just faster and with better formatting.

1963
01:15:02,080 --> 01:15:03,080
Conclusion

1964
01:15:03,080 --> 01:15:04,080
The behavior shift

1965
01:15:04,080 --> 01:15:06,240
Reintroduce judgment into the system.

1966
01:15:06,240 --> 01:15:08,000
AI scales whatever you are.

1967
01:15:08,000 --> 01:15:10,480
Clarity or confusion because it can't own your decisions.

1968
01:15:10,480 --> 01:15:13,640
If you do one thing after this, stop asking what does the AI say?

1969
01:15:13,640 --> 01:15:15,640
To start asking who owns this decision?

1970
01:15:15,640 --> 01:15:16,640
Subscribe?

1971
01:15:16,640 --> 01:15:22,080
Go to the next episode on building judgment moments into the M365 service now operating model.