Your Copilot Has No Memory — Why RAG Fails and LLM Wiki Fixes It

1
00:00:00,000 --> 00:00:02,360
Last week, your copilot answered a question perfectly.

2
00:00:02,360 --> 00:00:04,200
Someone asked about a project timeline,

3
00:00:04,200 --> 00:00:06,440
and it pulled the right details connected the right dots

4
00:00:06,440 --> 00:00:08,320
gave you an answer that felt almost human.

5
00:00:08,320 --> 00:00:10,160
Today, someone asks the same question,

6
00:00:10,160 --> 00:00:11,880
slightly different wording.

7
00:00:11,880 --> 00:00:14,960
And copilot fumbles the exact connection it made yesterday.

8
00:00:14,960 --> 00:00:17,600
Here's the assumption everyone walks around with.

9
00:00:17,600 --> 00:00:19,600
Copilot remembers your organization.

10
00:00:19,600 --> 00:00:20,120
It doesn't.

11
00:00:20,120 --> 00:00:20,920
It searches.

12
00:00:20,920 --> 00:00:22,920
Every single time, from zero, that's the tension

13
00:00:22,920 --> 00:00:23,920
nobody talks about.

14
00:00:23,920 --> 00:00:25,760
Retrieval and memory are not the same thing.

15
00:00:25,760 --> 00:00:28,800
One finds information, the other builds on what it already knows.

16
00:00:28,800 --> 00:00:31,160
Microsoft built copilot on the first one

17
00:00:31,160 --> 00:00:33,040
and let the marketing imply the second.

18
00:00:33,040 --> 00:00:35,240
There's a pattern circulating right now.

19
00:00:35,240 --> 00:00:37,160
Born out of AI research circles,

20
00:00:37,160 --> 00:00:39,200
that treats knowledge as something you compile,

21
00:00:39,200 --> 00:00:41,160
not something you search for over and over.

22
00:00:41,160 --> 00:00:42,560
We'll name it in a few minutes.

23
00:00:42,560 --> 00:00:44,360
By the end of this episode, you'll understand

24
00:00:44,360 --> 00:00:46,560
why your copilot deployment feels sharp in demos

25
00:00:46,560 --> 00:00:48,600
and shallow in production and what structural change

26
00:00:48,600 --> 00:00:49,880
actually fixes it.

27
00:00:49,880 --> 00:00:51,240
Quick thing before we get into it.

28
00:00:51,240 --> 00:00:53,240
If you're into this channel because you want to understand

29
00:00:53,240 --> 00:00:55,480
the systems behind Microsoft 365,

30
00:00:55,480 --> 00:00:59,480
not just the button clicking, hit subscribe on M365FM podcast.

31
00:00:59,480 --> 00:01:02,040
This episode is exactly the kind of thing we do here.

32
00:01:02,040 --> 00:01:03,720
And I want to set the expectation right now.

33
00:01:03,720 --> 00:01:05,560
This isn't a copilot feature tour.

34
00:01:05,560 --> 00:01:07,360
We're not walking through settings menus.

35
00:01:07,360 --> 00:01:08,840
This is a structural diagnosis,

36
00:01:08,840 --> 00:01:11,320
the kind you do on a system that keeps almost working

37
00:01:11,320 --> 00:01:12,800
but never quite getting there.

38
00:01:12,800 --> 00:01:14,040
So if you came for prompt tips,

39
00:01:14,040 --> 00:01:15,480
this one's going to disappoint you.

40
00:01:15,480 --> 00:01:18,400
We're going under the interface into the architecture.

41
00:01:18,400 --> 00:01:20,000
That's where the real answer lives.

42
00:01:20,000 --> 00:01:22,840
The day one copilot promise.

43
00:01:22,840 --> 00:01:25,000
Think back to how copilot got pitched to you

44
00:01:25,000 --> 00:01:26,920
or to your leadership or in whatever deck

45
00:01:26,920 --> 00:01:28,800
convinced someone to write the check.

46
00:01:28,800 --> 00:01:30,360
Ask questions in plain English.

47
00:01:30,360 --> 00:01:33,040
Get answers grounded in your organization's own data.

48
00:01:33,040 --> 00:01:34,480
No more digging through SharePoint.

49
00:01:34,480 --> 00:01:37,120
No more pinging three people on teams to find one document.

50
00:01:37,120 --> 00:01:38,720
Just ask and copilot knows.

51
00:01:38,720 --> 00:01:40,760
Microsoft backed that pitch with numbers.

52
00:01:40,760 --> 00:01:42,960
11 minutes saved per day per user.

53
00:01:42,960 --> 00:01:46,080
Do the math across a year and that's close to a full work week

54
00:01:46,080 --> 00:01:47,400
handed back to every employee.

55
00:01:47,400 --> 00:01:49,400
That's a real number and for a lot of organizations

56
00:01:49,400 --> 00:01:51,720
it's the number that justified the license cost.

57
00:01:51,720 --> 00:01:53,960
But here's what nobody set out loud during that pitch.

58
00:01:53,960 --> 00:01:56,600
Baked into asked questions and get grounded answers

59
00:01:56,600 --> 00:01:58,440
is an implicit promise.

60
00:01:58,440 --> 00:02:00,560
This thing gets better as you use it.

61
00:02:00,560 --> 00:02:03,280
You feed it more meetings, more documents, more decisions

62
00:02:03,280 --> 00:02:06,120
and it becomes sharper, more attuned to how your organization

63
00:02:06,120 --> 00:02:06,840
actually works.

64
00:02:06,840 --> 00:02:09,200
That's the story everyone heard, even if nobody wrote it

65
00:02:09,200 --> 00:02:10,000
on a slide.

66
00:02:10,000 --> 00:02:11,800
What actually ships is something different.

67
00:02:11,800 --> 00:02:14,280
It's a retrieval system wrapped in a chat interface.

68
00:02:14,280 --> 00:02:16,280
Dressed up just enough to feel like it's learning.

69
00:02:16,280 --> 00:02:17,080
It isn't.

70
00:02:17,080 --> 00:02:19,840
Every time someone types a question, copilot goes and looks,

71
00:02:19,840 --> 00:02:20,760
it doesn't recall.

72
00:02:20,760 --> 00:02:22,720
It searches, finds what looks relevant

73
00:02:22,720 --> 00:02:24,360
and generates a response from that.

74
00:02:24,360 --> 00:02:26,240
Ask again tomorrow and it searches again

75
00:02:26,240 --> 00:02:28,000
as if yesterday never happened.

76
00:02:28,000 --> 00:02:30,560
Now you might be thinking this sounds like a UX complaint,

77
00:02:30,560 --> 00:02:31,720
a minor annoyance.

78
00:02:31,720 --> 00:02:32,560
It's not.

79
00:02:32,560 --> 00:02:34,320
This is a governance and adoption problem

80
00:02:34,320 --> 00:02:36,360
and those are expensive in a completely different way

81
00:02:36,360 --> 00:02:37,800
than a clunky interface.

82
00:02:37,800 --> 00:02:39,800
When an assistant gives inconsistent answers

83
00:02:39,800 --> 00:02:41,480
to the same underlying question,

84
00:02:41,480 --> 00:02:43,440
people stop trusting it for anything that matters.

85
00:02:43,440 --> 00:02:44,760
They keep it around for quick summaries

86
00:02:44,760 --> 00:02:46,160
and calendar questions, sure.

87
00:02:46,160 --> 00:02:48,520
But the moment a decision actually rides on the answer,

88
00:02:48,520 --> 00:02:49,720
they go back to asking a person.

89
00:02:49,720 --> 00:02:50,680
That's not a bug report.

90
00:02:50,680 --> 00:02:52,520
That's an organization quietly deciding

91
00:02:52,520 --> 00:02:54,800
the tool isn't reliable enough to lean on.

92
00:02:54,800 --> 00:02:56,200
And once that decision gets made,

93
00:02:56,200 --> 00:02:58,000
even informally, adoption stores.

94
00:02:58,000 --> 00:02:59,640
Not because people didn't try copilot

95
00:02:59,640 --> 00:03:01,680
because they tried it, got burned once or twice

96
00:03:01,680 --> 00:03:03,040
by an inconsistent answer

97
00:03:03,040 --> 00:03:04,560
and adjusted their behavior around it.

98
00:03:04,560 --> 00:03:05,840
Nobody files a ticket for that.

99
00:03:05,840 --> 00:03:08,240
It just shows up months later as flat usage numbers

100
00:03:08,240 --> 00:03:11,000
and a leadership team asking why the tool they paid for

101
00:03:11,000 --> 00:03:13,200
isn't delivering the return they were promised.

102
00:03:13,200 --> 00:03:16,120
So the gap between the pitch and the product isn't cosmetic.

103
00:03:16,120 --> 00:03:18,840
It's the difference between a tool that compounds in value

104
00:03:18,840 --> 00:03:20,760
and one that just sits there, static,

105
00:03:20,760 --> 00:03:22,400
no matter how long you've had it deployed,

106
00:03:22,400 --> 00:03:24,160
to actually see where that gap comes from.

107
00:03:24,160 --> 00:03:26,280
You have to look at what's happening mechanically

108
00:03:26,280 --> 00:03:28,840
every single time someone types a question into copilot

109
00:03:28,840 --> 00:03:31,160
because the behavior we just described isn't random.

110
00:03:31,160 --> 00:03:32,880
It's the direct, predictable result

111
00:03:32,880 --> 00:03:35,720
of how the system is built underneath the chat window.

112
00:03:35,720 --> 00:03:37,000
What drag actually does?

113
00:03:37,000 --> 00:03:38,200
So let's name it directly.

114
00:03:38,200 --> 00:03:40,320
Our Ag stands for retrieval augmented generation.

115
00:03:40,320 --> 00:03:42,640
Strip the jargon and here's what that actually means.

116
00:03:42,640 --> 00:03:44,720
The system retrieves some information,

117
00:03:44,720 --> 00:03:47,320
then uses that information to help generate an answer.

118
00:03:47,320 --> 00:03:48,920
That's the whole concept.

119
00:03:48,920 --> 00:03:51,480
Retrieval, then generation, glued together.

120
00:03:51,480 --> 00:03:53,400
Here's how it works mechanically, step by step.

121
00:03:53,400 --> 00:03:54,800
First, your documents get chunked.

122
00:03:54,800 --> 00:03:57,720
Every SharePoint file, every Teams transcript,

123
00:03:57,720 --> 00:04:00,400
every policy PDF gets sliced into smaller pieces,

124
00:04:00,400 --> 00:04:02,240
usually a few hundred words at a time.

125
00:04:02,240 --> 00:04:04,000
Second, each of those chunks gets converted

126
00:04:04,000 --> 00:04:05,440
into something called an embedding,

127
00:04:05,440 --> 00:04:07,240
which is just a mathematical fingerprint

128
00:04:07,240 --> 00:04:08,600
of what that chunk means.

129
00:04:08,600 --> 00:04:10,400
Third, when someone types a question,

130
00:04:10,400 --> 00:04:12,640
that question also gets turned into a fingerprint

131
00:04:12,640 --> 00:04:14,600
and the system searches for chunks

132
00:04:14,600 --> 00:04:16,160
whose fingerprints look similar.

133
00:04:16,160 --> 00:04:18,440
Fourth, whichever chunk score highest gets stuffed

134
00:04:18,440 --> 00:04:20,880
into a prompt alongside the original question.

135
00:04:20,880 --> 00:04:22,920
Fifth, the language model reads all of that

136
00:04:22,920 --> 00:04:23,920
and generates an answer.

137
00:04:23,920 --> 00:04:25,280
Now pay attention to that word chunk

138
00:04:25,280 --> 00:04:27,960
because it's doing more work than people give it credit for.

139
00:04:27,960 --> 00:04:29,720
Documents get sliced into fragments

140
00:04:29,720 --> 00:04:31,560
with no awareness of the whole.

141
00:04:31,560 --> 00:04:34,400
A 40-page project plan doesn't get read as a project plan.

142
00:04:34,400 --> 00:04:37,200
It gets cut into 15 or 20 disconnected pieces.

143
00:04:37,200 --> 00:04:39,360
Each one judged purely on whether it resembles

144
00:04:39,360 --> 00:04:40,800
the question being asked.

145
00:04:40,800 --> 00:04:43,000
The system never holds the full document in mind.

146
00:04:43,000 --> 00:04:45,120
It holds fragments and it bets that the right fragments

147
00:04:45,120 --> 00:04:47,680
to stick together will look enough like a coherent answer.

148
00:04:47,680 --> 00:04:49,680
This is a search engine with a language model

149
00:04:49,680 --> 00:04:50,760
bolted on the end.

150
00:04:50,760 --> 00:04:53,440
It's not a thinking system and it was never built to be one.

151
00:04:53,440 --> 00:04:55,240
The intelligence you're seeing when co-pilot

152
00:04:55,240 --> 00:04:58,080
gives you a good answer isn't co-pilot understanding

153
00:04:58,080 --> 00:04:59,120
your organization.

154
00:04:59,120 --> 00:05:01,560
It's a language model doing a genuinely impressive job

155
00:05:01,560 --> 00:05:04,000
of stitching together whatever fragments got handed to it.

156
00:05:04,000 --> 00:05:05,120
That's a real skill.

157
00:05:05,120 --> 00:05:08,080
It's just not memory and it's not comprehension of the whole.

158
00:05:08,080 --> 00:05:09,640
If you want a simple way to picture this,

159
00:05:09,640 --> 00:05:12,280
think about typing a question into SharePoint Search.

160
00:05:12,280 --> 00:05:14,880
You type in a phrase, SharePoint goes and finds documents

161
00:05:14,880 --> 00:05:16,800
that contain similar words or concepts

162
00:05:16,800 --> 00:05:18,360
and it hands you a list of results.

163
00:05:18,360 --> 00:05:19,840
Ragn does almost exactly that.

164
00:05:19,840 --> 00:05:21,560
Except instead of handing you the list,

165
00:05:21,560 --> 00:05:24,360
it reads the top few results and writes you a paragraph summarizing

166
00:05:24,360 --> 00:05:24,760
them.

167
00:05:24,760 --> 00:05:27,880
Same search, same fragments, just a nicer wrapper on the output.

168
00:05:27,880 --> 00:05:29,960
And look, for a single look up, this works fine.

169
00:05:29,960 --> 00:05:31,000
Genuinely fine.

170
00:05:31,000 --> 00:05:33,640
If someone asks, what's our current PTO policy?

171
00:05:33,640 --> 00:05:35,800
Ragn finds the chunk that talks about PTO,

172
00:05:35,800 --> 00:05:37,520
generates a clean summary, and everyone

173
00:05:37,520 --> 00:05:38,480
moves on with their day.

174
00:05:38,480 --> 00:05:42,760
One question, one search, one answer, no complaints.

175
00:05:42,760 --> 00:05:44,440
The failure doesn't show up there.

176
00:05:44,440 --> 00:05:46,240
It shows up the moment you need continuity.

177
00:05:46,240 --> 00:05:47,960
The moment a question depends on something

178
00:05:47,960 --> 00:05:50,000
the system already figured out five minutes ago

179
00:05:50,000 --> 00:05:51,840
or yesterday or last month.

180
00:05:51,840 --> 00:05:54,880
Because Ragn has no concept of already figured out.

181
00:05:54,880 --> 00:05:56,400
It only has search again.

182
00:05:56,400 --> 00:05:59,160
And that's exactly where the cracks start to show.

183
00:05:59,160 --> 00:06:00,520
The stateless trap.

184
00:06:00,520 --> 00:06:02,320
Let's put a real word on what we just described

185
00:06:02,320 --> 00:06:03,400
because it matters.

186
00:06:03,400 --> 00:06:05,440
That behavior is called statelessness.

187
00:06:05,440 --> 00:06:07,760
Every query starts from zero, no matter how many times

188
00:06:07,760 --> 00:06:09,080
you've asked something related before,

189
00:06:09,080 --> 00:06:11,080
no matter how obviously connected today's question

190
00:06:11,080 --> 00:06:12,400
is to yesterday's.

191
00:06:12,400 --> 00:06:14,400
The system has no concept of before.

192
00:06:14,400 --> 00:06:16,360
There's only this query right now treated

193
00:06:16,360 --> 00:06:18,040
like the first one it's ever seen.

194
00:06:18,040 --> 00:06:20,720
Walk through what that actually looks like inside co-pilot.

195
00:06:20,720 --> 00:06:22,440
Someone on a project team asks,

196
00:06:22,440 --> 00:06:24,960
what's the status of the client onboarding for Meridian?

197
00:06:24,960 --> 00:06:27,160
Co-pilot searches, finds the right team's thread,

198
00:06:27,160 --> 00:06:30,080
the right sharepoint update, maybe a status email from last week

199
00:06:30,080 --> 00:06:31,600
and gives a genuinely good answer.

200
00:06:31,600 --> 00:06:33,480
Clean, accurate, useful.

201
00:06:33,480 --> 00:06:35,960
The person moves on with their day, impressed.

202
00:06:35,960 --> 00:06:38,240
Tomorrow, that same person asks a follow-up, something

203
00:06:38,240 --> 00:06:41,240
like, what's blocking us from finishing that onboarding?

204
00:06:41,240 --> 00:06:43,320
Notice what just happened in that sentence.

205
00:06:43,320 --> 00:06:44,960
That onboarding.

206
00:06:44,960 --> 00:06:46,360
A human would hear this and instantly

207
00:06:46,360 --> 00:06:48,040
connected to yesterday's conversation,

208
00:06:48,040 --> 00:06:50,320
no clarification needed, co-pilot doesn't.

209
00:06:50,320 --> 00:06:52,800
It has no idea that onboarding refers to Meridian

210
00:06:52,800 --> 00:06:55,440
because it never stored the fact that Meridian came up yesterday.

211
00:06:55,440 --> 00:06:58,800
So it searches again from scratch, hoping the word onboarding alone

212
00:06:58,800 --> 00:07:00,640
points it back to the right documents.

213
00:07:00,640 --> 00:07:02,480
Sometimes it gets lucky, sometimes it grabs

214
00:07:02,480 --> 00:07:04,680
the different clients onboarding thread entirely

215
00:07:04,680 --> 00:07:06,800
and now you've got a confidently wrong answer

216
00:07:06,800 --> 00:07:09,360
standing right next to yesterday's confidently right one.

217
00:07:09,360 --> 00:07:10,960
Here's the part that actually costs you.

218
00:07:10,960 --> 00:07:11,920
Nothing compounds.

219
00:07:11,920 --> 00:07:14,480
There's no accumulated understanding of the Meridian project

220
00:07:14,480 --> 00:07:16,600
building up over these two conversations,

221
00:07:16,600 --> 00:07:19,520
no growing sense of who's involved, what decisions got made,

222
00:07:19,520 --> 00:07:21,560
what the blockers have historically been.

223
00:07:21,560 --> 00:07:23,080
Each question is an island.

224
00:07:23,080 --> 00:07:24,960
You could ask co-pilot about the same project

225
00:07:24,960 --> 00:07:27,960
50 times over three months and it would know exactly

226
00:07:27,960 --> 00:07:31,080
as much on question 50 as it did on question one

227
00:07:31,080 --> 00:07:34,160
because it never carried anything forward between them.

228
00:07:34,160 --> 00:07:36,880
Now compare that to how an actual colleague would handle it,

229
00:07:36,880 --> 00:07:39,560
say the person who handled Meridian's onboarding last month,

230
00:07:39,560 --> 00:07:42,440
the answer to what's blocking us comes wrapped in memory.

231
00:07:42,440 --> 00:07:43,760
They remember the client pushed back

232
00:07:43,760 --> 00:07:45,920
on the integration timeline two weeks ago.

233
00:07:45,920 --> 00:07:47,640
They remember a decision got made in a meeting

234
00:07:47,640 --> 00:07:48,920
to reprioritize.

235
00:07:48,920 --> 00:07:50,520
They don't need to redrive any of that

236
00:07:50,520 --> 00:07:52,240
because they were there and they kept it.

237
00:07:52,240 --> 00:07:55,280
Every follow-up question they answer builds on the last one,

238
00:07:55,280 --> 00:07:57,760
gets sharper, gets more useful over time.

239
00:07:57,760 --> 00:07:59,800
That's what a second conversation with a person actually

240
00:07:59,800 --> 00:08:02,280
feels like and it's exactly what a second conversation

241
00:08:02,280 --> 00:08:03,440
with co-pilot doesn't.

242
00:08:03,440 --> 00:08:05,240
So name the real cost here plainly

243
00:08:05,240 --> 00:08:06,960
because it's not what people assume.

244
00:08:06,960 --> 00:08:09,240
The cost isn't wrong answers, not usually.

245
00:08:09,240 --> 00:08:10,720
Rags often technically accurate

246
00:08:10,720 --> 00:08:12,520
on any single question you throw at it.

247
00:08:12,520 --> 00:08:15,120
The cost is shallow answers, ones that never deep in no matter

248
00:08:15,120 --> 00:08:16,800
how long you've been using the tool,

249
00:08:16,800 --> 00:08:20,000
no matter how much history exists between you and the system.

250
00:08:20,000 --> 00:08:21,560
You get competence without growth,

251
00:08:21,560 --> 00:08:24,240
fluency without accumulation.

252
00:08:24,240 --> 00:08:26,880
And here's the thing worth sitting with before we go further.

253
00:08:26,880 --> 00:08:29,120
This isn't a training problem, it's not a prompt problem

254
00:08:29,120 --> 00:08:31,240
and it's not a rollout problem you can fix

255
00:08:31,240 --> 00:08:32,640
with better change management.

256
00:08:32,640 --> 00:08:34,960
It's what happens structurally every single time

257
00:08:34,960 --> 00:08:37,680
you take a search mechanism and ask it to do memory's job.

258
00:08:37,680 --> 00:08:40,480
Search was never built to remember, it was built to find.

259
00:08:40,480 --> 00:08:42,080
And no amount of clever phrasing changes

260
00:08:42,080 --> 00:08:44,960
what the system was designed to do underneath.

261
00:08:44,960 --> 00:08:46,920
Why this isn't a prompting problem?

262
00:08:46,920 --> 00:08:50,120
At this point, someone in IT is going to have a reflex reaction

263
00:08:50,120 --> 00:08:51,600
to everything we just said.

264
00:08:51,600 --> 00:08:53,040
They're going to blame the prompts.

265
00:08:53,040 --> 00:08:55,880
Maybe the rollout was rushed, maybe users just don't know how

266
00:08:55,880 --> 00:08:57,400
to ask questions the right way yet

267
00:08:57,400 --> 00:08:59,640
and a few training sessions will close the gap.

268
00:08:59,640 --> 00:09:00,720
That instinct makes sense.

269
00:09:00,720 --> 00:09:02,280
It's usually where the blame lands first

270
00:09:02,280 --> 00:09:03,640
because it's the easiest thing to fix

271
00:09:03,640 --> 00:09:05,520
without touching the underlying system.

272
00:09:05,520 --> 00:09:08,120
Rewrite the prompt library, run a lunch and learn,

273
00:09:08,120 --> 00:09:09,640
tell people to be more specific

274
00:09:09,640 --> 00:09:12,160
and show a better phrasing helps at the margins.

275
00:09:12,160 --> 00:09:13,640
But it doesn't touch the actual problem

276
00:09:13,640 --> 00:09:16,480
and there's a number that makes this hard to ignore.

277
00:09:16,480 --> 00:09:18,000
Research on structured knowledge systems

278
00:09:18,000 --> 00:09:20,840
shows organizations using layered knowledge with LLMs

279
00:09:20,840 --> 00:09:23,080
cut AI error rates by more than 60%

280
00:09:23,080 --> 00:09:25,280
compared to standard rag setups, 60%.

281
00:09:25,280 --> 00:09:26,600
That's not a rounding error

282
00:09:26,600 --> 00:09:28,200
and it's not the kind of gap you close

283
00:09:28,200 --> 00:09:30,320
by teaching people to phrase their questions better.

284
00:09:30,320 --> 00:09:33,000
Think about what that number actually implies.

285
00:09:33,000 --> 00:09:35,240
It's not measuring how well people ask questions.

286
00:09:35,240 --> 00:09:37,240
It's measuring what the system does before a question

287
00:09:37,240 --> 00:09:38,320
ever gets typed.

288
00:09:38,320 --> 00:09:41,360
60% fewer errors didn't come from smarter users.

289
00:09:41,360 --> 00:09:43,440
It came from a completely different relationship

290
00:09:43,440 --> 00:09:45,800
between the AI and the knowledge underneath it,

291
00:09:45,800 --> 00:09:47,600
one where the knowledge had already been organized,

292
00:09:47,600 --> 00:09:50,800
connected and reconciled before anyone opened a chat window.

293
00:09:50,800 --> 00:09:52,920
So here's the reframe and it's worth sitting with

294
00:09:52,920 --> 00:09:54,920
because it changes where you point your effort.

295
00:09:54,920 --> 00:09:56,920
You can't prompt your way out of an architecture

296
00:09:56,920 --> 00:09:59,280
that has no place to store what it learns.

297
00:09:59,280 --> 00:10:01,760
It doesn't matter how carefully worded the question is.

298
00:10:01,760 --> 00:10:03,800
If the system searches fresh every time,

299
00:10:03,800 --> 00:10:06,000
then even a perfect prompt just retrieves a fresh set

300
00:10:06,000 --> 00:10:07,920
of chunks and generates a fresh answer,

301
00:10:07,920 --> 00:10:10,240
disconnected from anything that came before.

302
00:10:10,240 --> 00:10:12,360
A better prompt gets you a better single search.

303
00:10:12,360 --> 00:10:13,880
It doesn't get you memory because memory

304
00:10:13,880 --> 00:10:15,400
was never something a prompt controls.

305
00:10:15,400 --> 00:10:17,200
That's the part worth naming plainly.

306
00:10:17,200 --> 00:10:18,680
The fix doesn't live in the interface.

307
00:10:18,680 --> 00:10:20,400
It doesn't live in how questions get asked.

308
00:10:20,400 --> 00:10:23,080
It lives below all of that in how knowledge gets prepared

309
00:10:23,080 --> 00:10:25,440
and represented before Copilot ever touches it.

310
00:10:25,440 --> 00:10:27,240
If the knowledge sitting behind the assistant

311
00:10:27,240 --> 00:10:29,960
is a pile of raw documents waiting to be chunked on demand,

312
00:10:29,960 --> 00:10:31,640
no amount of prompt engineering changes

313
00:10:31,640 --> 00:10:32,640
what happens next.

314
00:10:32,640 --> 00:10:35,720
The system is built to search and search is what it will keep doing

315
00:10:35,720 --> 00:10:37,600
no matter how the question gets phrased,

316
00:10:37,600 --> 00:10:38,960
which raises the actual question worth

317
00:10:38,960 --> 00:10:40,360
spending the rest of this episode on.

318
00:10:40,360 --> 00:10:42,720
Not how do we ask Copilot better questions.

319
00:10:42,720 --> 00:10:44,840
But what does the system actually look like

320
00:10:44,840 --> 00:10:46,960
when it's built for memory instead of search

321
00:10:46,960 --> 00:10:48,960
because that system exists and it works

322
00:10:48,960 --> 00:10:50,240
on a completely different principle

323
00:10:50,240 --> 00:10:52,560
than the one copilot ships with today?

324
00:10:52,560 --> 00:10:53,880
Enter the LLM Wiki.

325
00:10:53,880 --> 00:10:55,040
Here's that system.

326
00:10:55,040 --> 00:10:56,400
It didn't come out of Redmond

327
00:10:56,400 --> 00:10:58,120
and it didn't come out of a product road map.

328
00:10:58,120 --> 00:11:00,280
It surfaced from AI research circles

329
00:11:00,280 --> 00:11:02,840
and the person most associated with articulating it clearly

330
00:11:02,840 --> 00:11:05,680
is Andre Kapati, one of the more influential voices

331
00:11:05,680 --> 00:11:07,120
in how people think about working

332
00:11:07,120 --> 00:11:08,600
with these models day to day.

333
00:11:08,600 --> 00:11:10,480
He wasn't building an enterprise product.

334
00:11:10,480 --> 00:11:12,160
He was solving his own problem,

335
00:11:12,160 --> 00:11:14,800
building a personal knowledge base for research he cared about

336
00:11:14,800 --> 00:11:17,520
and the pattern he landed on is the one worth stealing.

337
00:11:17,520 --> 00:11:18,680
The core move is simple to say

338
00:11:18,680 --> 00:11:20,160
and genuinely different in practice.

339
00:11:20,160 --> 00:11:21,560
Instead of searching raw documents

340
00:11:21,560 --> 00:11:23,440
every single time a question comes in

341
00:11:23,440 --> 00:11:25,320
and AI reads the sources once

342
00:11:25,320 --> 00:11:28,360
and builds a structured interlinked knowledge base out of them.

343
00:11:28,360 --> 00:11:29,960
Once, not once per question.

344
00:11:29,960 --> 00:11:32,320
Once, period, then it keeps that structure around

345
00:11:32,320 --> 00:11:33,320
and refers back to it.

346
00:11:33,320 --> 00:11:35,360
Picture three layers because this is where it stops

347
00:11:35,360 --> 00:11:37,160
being an abstract idea and starts

348
00:11:37,160 --> 00:11:38,760
being something you could actually build.

349
00:11:38,760 --> 00:11:40,240
Layer one is raw sources.

350
00:11:40,240 --> 00:11:42,400
These are your original documents whatever they are

351
00:11:42,400 --> 00:11:43,640
and they stay read only.

352
00:11:43,640 --> 00:11:45,400
Nobody touches them, nothing re-rides them.

353
00:11:45,400 --> 00:11:47,360
They're ground truth sitting there untouched

354
00:11:47,360 --> 00:11:49,160
the same way a source document should.

355
00:11:49,160 --> 00:11:50,560
Layer two is the Wiki itself.

356
00:11:50,560 --> 00:11:51,960
This is where the actual work happens.

357
00:11:51,960 --> 00:11:53,680
It's a set of pages, the AI writes.

358
00:11:53,680 --> 00:11:55,480
One per concept, one per entity,

359
00:11:55,480 --> 00:11:57,560
one per recurring topic, all cross linked

360
00:11:57,560 --> 00:11:59,200
to each other the way a real Wiki works.

361
00:11:59,200 --> 00:12:00,440
Not a database of chunks.

362
00:12:00,440 --> 00:12:02,800
Pages written in something like plain language

363
00:12:02,800 --> 00:12:04,280
meant to be read as a whole.

364
00:12:04,280 --> 00:12:05,680
Layer three is the schema.

365
00:12:05,680 --> 00:12:07,080
Think of it as the rule book.

366
00:12:07,080 --> 00:12:09,440
It tells the AI how to structure new pages,

367
00:12:09,440 --> 00:12:11,080
how to decide when something updates

368
00:12:11,080 --> 00:12:13,520
an existing page versus when it deserves a new one,

369
00:12:13,520 --> 00:12:15,480
how to format things consistently

370
00:12:15,480 --> 00:12:17,840
so the Wiki doesn't turn into chaos as it grows.

371
00:12:17,840 --> 00:12:21,680
Now here's the part people get wrong on first hearing this.

372
00:12:21,680 --> 00:12:23,960
They assume the Wiki is just static storage,

373
00:12:23,960 --> 00:12:26,640
a folder where documents get copied and organized once.

374
00:12:26,640 --> 00:12:28,160
It's not, it's maintained.

375
00:12:28,160 --> 00:12:30,680
Every time a new source comes in, the AI reads it

376
00:12:30,680 --> 00:12:32,000
and does real work with it.

377
00:12:32,000 --> 00:12:34,600
It updates existing pages if the new source adds detail

378
00:12:34,600 --> 00:12:36,040
to something already compiled.

379
00:12:36,040 --> 00:12:38,280
It creates new pages if the source introduces

380
00:12:38,280 --> 00:12:41,280
a genuinely new concept that didn't exist in the Wiki yet.

381
00:12:41,280 --> 00:12:43,320
And if the new source says something that contradicts

382
00:12:43,320 --> 00:12:46,240
what's already there, it doesn't just override quietly.

383
00:12:46,240 --> 00:12:48,120
It flags it, that flag is the whole point

384
00:12:48,120 --> 00:12:49,920
and we'll come back to why in a bit.

385
00:12:49,920 --> 00:12:52,440
So say the line plainly, because it's the whole argument

386
00:12:52,440 --> 00:12:53,440
in one sentence.

387
00:12:53,440 --> 00:12:55,480
Knowledge gets compiled once then kept current.

388
00:12:55,480 --> 00:12:57,440
That's the exact opposite of what Ragn does.

389
00:12:57,440 --> 00:13:00,560
Ragn rebuilds understanding from scratch on every query.

390
00:13:00,560 --> 00:13:02,840
A Wiki builds it once and then just maintains it,

391
00:13:02,840 --> 00:13:04,800
the way you'd maintain anything that's actually alive

392
00:13:04,800 --> 00:13:07,680
and growing instead of something you keep rebuilding from nothing.

393
00:13:07,680 --> 00:13:10,280
To see why this distinction actually matters for co-pilot

394
00:13:10,280 --> 00:13:13,120
and not just as a nice idea for personal research projects,

395
00:13:13,120 --> 00:13:15,160
you need to put the two systems side by side

396
00:13:15,160 --> 00:13:17,800
and watch where the actual work happens in each one.

397
00:13:17,800 --> 00:13:20,040
Ragn vs LLM Wiki side by side.

398
00:13:20,040 --> 00:13:22,720
Put the two systems next to each other and ask one question,

399
00:13:22,720 --> 00:13:24,480
where does the thinking actually happen?

400
00:13:24,480 --> 00:13:26,280
Not the answering, the thinking.

401
00:13:26,280 --> 00:13:28,640
The part where raw information turns into something

402
00:13:28,640 --> 00:13:30,240
organized enough to be useful.

403
00:13:30,240 --> 00:13:32,160
That question tells you almost everything

404
00:13:32,160 --> 00:13:34,240
you need to know about why one system compounds

405
00:13:34,240 --> 00:13:35,320
and the other doesn't.

406
00:13:35,320 --> 00:13:37,440
In Ragn, the thinking happens at query time.

407
00:13:37,440 --> 00:13:39,440
Every single question triggers the whole process

408
00:13:39,440 --> 00:13:40,360
from the beginning.

409
00:13:40,360 --> 00:13:42,120
Fresh search across the embeddings,

410
00:13:42,120 --> 00:13:44,880
fresh chunking logic deciding which fragments matter.

411
00:13:44,880 --> 00:13:47,840
Fresh synthesis, where the model reads whatever got pulled

412
00:13:47,840 --> 00:13:49,920
and stitches it into an answer on the spot.

413
00:13:49,920 --> 00:13:52,600
None of that work exists until the question shows up.

414
00:13:52,600 --> 00:13:55,240
The system has zero understanding sitting in reserve.

415
00:13:55,240 --> 00:13:58,080
It manufactures understanding, on demand every time

416
00:13:58,080 --> 00:14:00,880
and then throws it away the moment the answer gets delivered.

417
00:14:00,880 --> 00:14:04,000
In the LLM Wiki, the thinking happens at ingestion time.

418
00:14:04,000 --> 00:14:05,520
That's the entire structural difference

419
00:14:05,520 --> 00:14:07,560
and it's easy to miss because it sounds small.

420
00:14:07,560 --> 00:14:10,840
The synthesis is already done before anyone asks the question.

421
00:14:10,840 --> 00:14:14,520
When a new source comes in, the AI does the hard part right then.

422
00:14:14,520 --> 00:14:16,560
Reading it, deciding what it means,

423
00:14:16,560 --> 00:14:18,880
connecting it to what's already compiled,

424
00:14:18,880 --> 00:14:21,240
updating pages, flagging conflicts.

425
00:14:21,240 --> 00:14:24,320
By the time someone actually types a question into the system,

426
00:14:24,320 --> 00:14:25,600
there's nothing left to figure out.

427
00:14:25,600 --> 00:14:27,600
The system isn't generating understanding.

428
00:14:27,600 --> 00:14:30,000
It's retrieving understanding that already exists,

429
00:14:30,000 --> 00:14:31,920
fully formed, sitting on a page.

430
00:14:31,920 --> 00:14:34,680
That difference shows up in a number worth paying attention to.

431
00:14:34,680 --> 00:14:37,400
For focused knowledge basis, LLM Wiki can cut token usage

432
00:14:37,400 --> 00:14:40,480
by roughly 95% compared to naive rag document loading.

433
00:14:40,480 --> 00:14:42,960
95% that's not a marginal efficiency gain

434
00:14:42,960 --> 00:14:44,760
that's a different order of magnitude.

435
00:14:44,760 --> 00:14:46,960
Here's why that number exists and it's not magic.

436
00:14:46,960 --> 00:14:49,920
Rag spends tokens researching and restitching every single time.

437
00:14:49,920 --> 00:14:52,400
It has to pull multiple chunks, feed all of them

438
00:14:52,400 --> 00:14:54,520
into the prompt alongside the question,

439
00:14:54,520 --> 00:14:57,480
and let the model do the work of reconciling fragments

440
00:14:57,480 --> 00:15:00,040
that were never written to sit next to each other.

441
00:15:00,040 --> 00:15:02,800
All of that costs tokens every query, forever.

442
00:15:02,800 --> 00:15:04,120
A Wiki skips that entirely.

443
00:15:04,120 --> 00:15:05,960
You're not researching and restitching.

444
00:15:05,960 --> 00:15:08,080
You're reading a page that was already organized,

445
00:15:08,080 --> 00:15:10,880
already coherent, already written, to be read as a whole.

446
00:15:10,880 --> 00:15:13,640
Less raw material has to get shoved into the model's context window

447
00:15:13,640 --> 00:15:15,880
because someone or something already

448
00:15:15,880 --> 00:15:17,720
did the organizing work in advance.

449
00:15:17,720 --> 00:15:19,960
Now, let's be honest about where this doesn't apply

450
00:15:19,960 --> 00:15:22,720
because overselling it here would undercut the whole argument.

451
00:15:22,720 --> 00:15:26,080
Rag stays the default for large, relatively static repositories

452
00:15:26,080 --> 00:15:27,240
and that's completely fine.

453
00:15:27,240 --> 00:15:29,440
If you're sitting on hundreds of thousands of documents

454
00:15:29,440 --> 00:15:31,600
that don't change much and get searched broadly

455
00:15:31,600 --> 00:15:33,240
across a huge range of topics,

456
00:15:33,240 --> 00:15:37,040
Rag's search first design is doing exactly what it's built for.

457
00:15:37,040 --> 00:15:38,920
The Wiki pattern isn't trying to replace that.

458
00:15:38,920 --> 00:15:40,560
Where the Wiki pattern actually shines

459
00:15:40,560 --> 00:15:44,280
is below roughly 50 to 100,000 tokens of compiled knowledge.

460
00:15:44,280 --> 00:15:45,760
That's a bounded focused domain.

461
00:15:45,760 --> 00:15:47,840
Not everything your organization has ever produced,

462
00:15:47,840 --> 00:15:50,760
a specific slice of it, the stuff people ask about again and again

463
00:15:50,760 --> 00:15:52,320
compiled once in Keptrap.

464
00:15:52,320 --> 00:15:54,400
None of this stays theoretical for long though

465
00:15:54,400 --> 00:15:57,000
because there's a version of exactly this tension

466
00:15:57,000 --> 00:15:59,720
already sitting inside Microsoft 365 right now,

467
00:15:59,720 --> 00:16:01,680
quietly shaping how co-pilot behaves

468
00:16:01,680 --> 00:16:03,760
every time someone opens a chat window.

469
00:16:03,760 --> 00:16:05,960
Where co-pilot's retrieval actually lives,

470
00:16:05,960 --> 00:16:08,960
let's get concrete about where all of this actually plays out

471
00:16:08,960 --> 00:16:11,800
because up to now we've been talking architecture in the abstract.

472
00:16:11,800 --> 00:16:15,120
Inside Microsoft 365, co-pilot indexes SharePoint OneDrive,

473
00:16:15,120 --> 00:16:17,440
Teams Exchange, and whatever other connected sources

474
00:16:17,440 --> 00:16:18,920
your organization has plugged in,

475
00:16:18,920 --> 00:16:21,320
it respects your permission structure while it does this.

476
00:16:21,320 --> 00:16:23,800
So someone only sees what they're already allowed to see.

477
00:16:23,800 --> 00:16:25,440
That parts real and it matters.

478
00:16:25,440 --> 00:16:27,160
But look at what that list actually is.

479
00:16:27,160 --> 00:16:29,320
It's a description of Rag at Enterprise Scale.

480
00:16:29,320 --> 00:16:31,520
Microsoft calls it Graph Grounding.

481
00:16:31,520 --> 00:16:33,240
And the name makes it sound like something new,

482
00:16:33,240 --> 00:16:35,120
something built specifically for this moment.

483
00:16:35,120 --> 00:16:35,920
It isn't.

484
00:16:35,920 --> 00:16:37,520
Strip the branding off and its retrieval

485
00:16:37,520 --> 00:16:39,480
same as we've been describing this whole episode,

486
00:16:39,480 --> 00:16:41,240
just dressed in Enterprise clothing.

487
00:16:41,240 --> 00:16:44,280
Co-pilot reaches into Graph, pulls whatever content looks relevant

488
00:16:44,280 --> 00:16:47,120
to the question and generates an answer from those fragments.

489
00:16:47,120 --> 00:16:49,920
That's the identical mechanism running against a much bigger,

490
00:16:49,920 --> 00:16:52,400
much messier pile of source material.

491
00:16:52,400 --> 00:16:54,440
Now, the permission model deserves real credit here

492
00:16:54,440 --> 00:16:56,400
because it solves a genuinely hard problem.

493
00:16:56,400 --> 00:16:58,560
Making sure someone in finance doesn't accidentally

494
00:16:58,560 --> 00:17:01,240
see HR's compensation planning docs through a chat window

495
00:17:01,240 --> 00:17:03,640
that's not trivial and Microsoft built real infrastructure

496
00:17:03,640 --> 00:17:04,560
to handle it.

497
00:17:04,560 --> 00:17:06,920
But notice exactly what that infrastructure solves.

498
00:17:06,920 --> 00:17:09,520
It solves the access problem, who's allowed to see what.

499
00:17:09,520 --> 00:17:11,840
It does absolutely nothing about the memory problem

500
00:17:11,840 --> 00:17:13,680
we spent the last several sections on.

501
00:17:13,680 --> 00:17:16,920
Permissions decide whether a document is eligible to be retrieved.

502
00:17:16,920 --> 00:17:19,680
They say nothing about whether the system remembers retrieving it

503
00:17:19,680 --> 00:17:22,240
or builds on it or connects it to the conversation

504
00:17:22,240 --> 00:17:23,240
from yesterday.

505
00:17:23,240 --> 00:17:25,760
Access and memory are two entirely separate questions

506
00:17:25,760 --> 00:17:27,520
and solving one doesn't touch the other.

507
00:17:27,520 --> 00:17:29,240
Here's where it gets interesting though.

508
00:17:29,240 --> 00:17:32,240
Microsoft's own guidance quietly admits something worth pausing

509
00:17:32,240 --> 00:17:32,720
on.

510
00:17:32,720 --> 00:17:35,000
Admins can now register specific SharePoint sites

511
00:17:35,000 --> 00:17:36,800
as knowledge sources for Co-pilot.

512
00:17:36,800 --> 00:17:38,000
Read that carefully.

513
00:17:38,000 --> 00:17:40,240
That feature only makes sense if not all content

514
00:17:40,240 --> 00:17:42,360
is equally trustworthy or equally structured

515
00:17:42,360 --> 00:17:44,920
or equally worth trusting Co-pilot to pull from.

516
00:17:44,920 --> 00:17:47,040
If every SharePoint site were equally reliable,

517
00:17:47,040 --> 00:17:49,800
there'd be no reason to let admins flag some of them as special.

518
00:17:49,800 --> 00:17:52,400
The existence of that toggle is Microsoft tacitly admitting

519
00:17:52,400 --> 00:17:53,440
the obvious.

520
00:17:53,440 --> 00:17:55,720
Some of what's sitting in your tenant is good, organized,

521
00:17:55,720 --> 00:17:59,080
current, and some of it is stale, duplicated, or just noise.

522
00:17:59,080 --> 00:18:01,000
And Co-pilot, left to its own devices,

523
00:18:01,000 --> 00:18:02,360
treats all of it the same.

524
00:18:02,360 --> 00:18:04,960
That single feature is Microsoft nudging your organization

525
00:18:04,960 --> 00:18:08,200
toward curation without ever fully naming the underlying issue

526
00:18:08,200 --> 00:18:09,160
out loud.

527
00:18:09,160 --> 00:18:11,600
It's a quiet acknowledgement wrapped in an admin setting.

528
00:18:11,600 --> 00:18:13,120
Nobody in the documentation says,

529
00:18:13,120 --> 00:18:16,400
our retrieval system can't tell good content from bad content,

530
00:18:16,400 --> 00:18:18,000
so we built your workaround.

531
00:18:18,000 --> 00:18:20,440
But that's functionally what registering a knowledge source

532
00:18:20,440 --> 00:18:21,040
does.

533
00:18:21,040 --> 00:18:23,160
It's damage control for a structural limitation

534
00:18:23,160 --> 00:18:24,600
presented as a feature.

535
00:18:24,600 --> 00:18:28,040
And curation, as a starting point, is genuinely useful.

536
00:18:28,040 --> 00:18:30,040
But it's not the same thing as compilation,

537
00:18:30,040 --> 00:18:31,560
and the difference between those two words

538
00:18:31,560 --> 00:18:33,880
is exactly where this conversation needs to go next.

539
00:18:33,880 --> 00:18:35,680
Curation isn't compilation.

540
00:18:35,680 --> 00:18:37,600
So let's draw the line clearly, because it's

541
00:18:37,600 --> 00:18:39,360
easy to conflate these two ideas.

542
00:18:39,360 --> 00:18:42,080
And conflating them is exactly how organizations end up

543
00:18:42,080 --> 00:18:43,920
thinking they've solved the memory problem

544
00:18:43,920 --> 00:18:46,520
when they've only solved the access problem twice.

545
00:18:46,520 --> 00:18:48,760
Marking a SharePoint site as a trusted knowledge source

546
00:18:48,760 --> 00:18:50,120
tells Co-pilot where to look.

547
00:18:50,120 --> 00:18:50,800
That's all it does.

548
00:18:50,800 --> 00:18:51,480
It's a pointer.

549
00:18:51,480 --> 00:18:53,840
It says, when you're searching, wait this location more heavily

550
00:18:53,840 --> 00:18:54,880
than that one.

551
00:18:54,880 --> 00:18:56,760
What it doesn't do, what it can't do,

552
00:18:56,760 --> 00:18:59,040
is tell Co-pilot what it learned from that site.

553
00:18:59,040 --> 00:19:01,040
There's no learning happening in the act

554
00:19:01,040 --> 00:19:02,560
of registering a knowledge source.

555
00:19:02,560 --> 00:19:04,680
There's just a narrower, more confident search.

556
00:19:04,680 --> 00:19:06,160
And that's the part we're sitting with.

557
00:19:06,160 --> 00:19:08,640
Curated sources still get retrieved fresh, chunked fresh,

558
00:19:08,640 --> 00:19:10,440
synthesized fresh every single time someone

559
00:19:10,440 --> 00:19:11,400
asks a question.

560
00:19:11,400 --> 00:19:13,520
Registring a site as trusted doesn't change

561
00:19:13,520 --> 00:19:15,520
the mechanism underneath it at all.

562
00:19:15,520 --> 00:19:17,560
It changes which pile of documents get searched.

563
00:19:17,560 --> 00:19:19,800
It doesn't change the fact that it's still a search starting

564
00:19:19,800 --> 00:19:21,280
from nothing every time.

565
00:19:21,280 --> 00:19:23,520
You've made the haste-axe more and more reliable.

566
00:19:23,520 --> 00:19:26,400
You haven't given the system a memory of what's in it.

567
00:19:26,400 --> 00:19:27,920
The wiki pattern goes one step further,

568
00:19:27,920 --> 00:19:30,000
and this is the actual structural difference.

569
00:19:30,000 --> 00:19:31,280
It doesn't just point at good sources

570
00:19:31,280 --> 00:19:33,200
and say, look here first.

571
00:19:33,200 --> 00:19:35,960
It processes those sources into standing knowledge.

572
00:19:35,960 --> 00:19:38,160
It reads them once and turns them into something

573
00:19:38,160 --> 00:19:40,800
that already exists before the next question arrives.

574
00:19:40,800 --> 00:19:41,920
That's not a better pointer.

575
00:19:41,920 --> 00:19:43,720
That's a different kind of object entirely.

576
00:19:43,720 --> 00:19:45,080
Here's the planist way to say it.

577
00:19:45,080 --> 00:19:46,840
Curation is picking better ingredients.

578
00:19:46,840 --> 00:19:48,440
You've gone through the pantry, thrown out,

579
00:19:48,440 --> 00:19:50,240
what's expired, kept what's fresh,

580
00:19:50,240 --> 00:19:53,440
and told the kitchen, cook from the shelf, not that one.

581
00:19:53,440 --> 00:19:54,520
That's real and it matters.

582
00:19:54,520 --> 00:19:56,760
But compilation is actually cooking the dish once

583
00:19:56,760 --> 00:19:59,040
and keeping it in the fridge ready to serve.

584
00:19:59,040 --> 00:20:00,560
Nobody's back in the kitchen chopping

585
00:20:00,560 --> 00:20:03,240
and searing from scratch every time someone's hungry.

586
00:20:03,240 --> 00:20:04,360
The work happened once.

587
00:20:04,360 --> 00:20:05,800
What's left is just serving it.

588
00:20:05,800 --> 00:20:08,840
Without that second step, even perfectly curated content

589
00:20:08,840 --> 00:20:11,520
still forces co-pilot to start over on every question.

590
00:20:11,520 --> 00:20:13,840
You can register every sharepoint site in your tenant

591
00:20:13,840 --> 00:20:15,760
as a trusted knowledge source, get the pantry

592
00:20:15,760 --> 00:20:17,000
as clean as it's ever going to be

593
00:20:17,000 --> 00:20:19,400
and co-pilot will still chunk those documents fresh,

594
00:20:19,400 --> 00:20:21,400
still search them fresh, still synthesize

595
00:20:21,400 --> 00:20:23,560
an answer fresh every single time.

596
00:20:23,560 --> 00:20:25,560
Curation narrows what gets searched.

597
00:20:25,560 --> 00:20:27,400
It doesn't touch how the searching works

598
00:20:27,400 --> 00:20:28,800
or what happens to the understanding

599
00:20:28,800 --> 00:20:30,280
once the answer's been delivered.

600
00:20:30,280 --> 00:20:33,560
It evaporates, same as always, waiting for the next question

601
00:20:33,560 --> 00:20:34,880
to rebuild it from nothing.

602
00:20:34,880 --> 00:20:36,480
So curation is a real improvement

603
00:20:36,480 --> 00:20:38,400
and worth doing regardless of anything else

604
00:20:38,400 --> 00:20:40,200
in this episode, but it's not the fix.

605
00:20:40,200 --> 00:20:42,600
It's step one of a process that most organizations

606
00:20:42,600 --> 00:20:45,200
stop halfway through mistaking a cleaner haystack

607
00:20:45,200 --> 00:20:46,520
for an actual memory.

608
00:20:46,520 --> 00:20:48,720
This is exactly where the week is three-layer structure

609
00:20:48,720 --> 00:20:51,840
stops being an abstract idea from AI research circles

610
00:20:51,840 --> 00:20:53,680
and starts being something you could genuinely build

611
00:20:53,680 --> 00:20:55,760
inside your own tenant.

612
00:20:55,760 --> 00:20:57,920
The three layers applied to co-pilot.

613
00:20:57,920 --> 00:21:00,720
So let's put actual Microsoft 365 nouns

614
00:21:00,720 --> 00:21:02,200
onto those three layers because that's

615
00:21:02,200 --> 00:21:04,120
where this stops sounding like a research idea

616
00:21:04,120 --> 00:21:05,800
and starts sounding like something you could sketch

617
00:21:05,800 --> 00:21:08,080
on a whiteboard this week.

618
00:21:08,080 --> 00:21:09,800
Layer one raw sources.

619
00:21:09,800 --> 00:21:11,920
In your tenant, that's your SharePoint libraries,

620
00:21:11,920 --> 00:21:14,080
your Teams meeting transcripts, exchange threads

621
00:21:14,080 --> 00:21:16,920
that carry real decisions buried in reply chains.

622
00:21:16,920 --> 00:21:19,520
Viva engage posts where someone announced a policy change

623
00:21:19,520 --> 00:21:21,520
nobody bothered to document properly.

624
00:21:21,520 --> 00:21:23,160
All of it stays exactly where it is.

625
00:21:23,160 --> 00:21:24,520
Nobody touches it, nobody edits it,

626
00:21:24,520 --> 00:21:27,120
nobody deletes it to clean things up.

627
00:21:27,120 --> 00:21:28,880
It's ground truth, sitting untouched,

628
00:21:28,880 --> 00:21:30,720
the same way it should sit, whether or not

629
00:21:30,720 --> 00:21:32,520
you ever build anything on top of it.

630
00:21:32,520 --> 00:21:34,200
Layer two, the wiki itself.

631
00:21:34,200 --> 00:21:36,360
This is the part worth picturing concretely,

632
00:21:36,360 --> 00:21:37,800
not a folder of documents.

633
00:21:37,800 --> 00:21:40,000
A maintained set of pages, one per project,

634
00:21:40,000 --> 00:21:41,920
one per client, one per recurring decision,

635
00:21:41,920 --> 00:21:45,400
your organization keeps having to re-explain to somebody new.

636
00:21:45,400 --> 00:21:47,080
Cross-linked, so the page about a client

637
00:21:47,080 --> 00:21:48,520
connects to the page about the project,

638
00:21:48,520 --> 00:21:50,760
connects to the page about the compliance interpretation

639
00:21:50,760 --> 00:21:52,560
that shaped how that project got scripted.

640
00:21:52,560 --> 00:21:54,800
This is the layer that didn't exist an hour ago

641
00:21:54,800 --> 00:21:57,560
in your tenant and does exist now once someone builds it.

642
00:21:57,560 --> 00:21:59,160
Layer three, the schema.

643
00:21:59,160 --> 00:22:01,480
Think of this as the rule book, the thing that

644
00:22:01,480 --> 00:22:04,040
decides how any of this actually happens

645
00:22:04,040 --> 00:22:05,680
when a new Teams transcript comes in,

646
00:22:05,680 --> 00:22:07,480
does it update an existing client page

647
00:22:07,480 --> 00:22:08,800
or does it spin up a new one?

648
00:22:08,800 --> 00:22:10,480
How does a page get formatted?

649
00:22:10,480 --> 00:22:12,920
So it's consistent with every other page in the wiki.

650
00:22:12,920 --> 00:22:16,000
And critically, what happens when something in that transcript

651
00:22:16,000 --> 00:22:17,680
contradicts what's already written down?

652
00:22:17,680 --> 00:22:19,160
The schema is what answers those questions

653
00:22:19,160 --> 00:22:21,760
before anyone has to answer them manually every single time.

654
00:22:21,760 --> 00:22:24,040
Now here's the thing, none of this should feel foreign

655
00:22:24,040 --> 00:22:27,160
to anyone who spent time in Microsoft 365 governance.

656
00:22:27,160 --> 00:22:28,880
Map it onto what you already understand.

657
00:22:28,880 --> 00:22:31,800
Retention labels already decide what happens to content

658
00:22:31,800 --> 00:22:34,200
based on rules you defined in advance.

659
00:22:34,200 --> 00:22:35,840
Content types already give structure

660
00:22:35,840 --> 00:22:38,760
to what would otherwise be undifferentiated files.

661
00:22:38,760 --> 00:22:40,960
Metadata schemas already tell you what fields

662
00:22:40,960 --> 00:22:43,240
a document needs before it counts as complete.

663
00:22:43,240 --> 00:22:46,000
This isn't a new discipline landing on your desk out of nowhere.

664
00:22:46,000 --> 00:22:48,360
It's an extension of information architecture work.

665
00:22:48,360 --> 00:22:50,720
Your organization has probably already been doing

666
00:22:50,720 --> 00:22:52,160
just aimed at a different output.

667
00:22:52,160 --> 00:22:54,760
Instead of the schema governing how a document gets filed,

668
00:22:54,760 --> 00:22:57,640
it's governing how a wiki page gets written and maintained.

669
00:22:57,640 --> 00:22:59,440
But here's the part that doesn't happen automatically

670
00:22:59,440 --> 00:23:01,120
and it's worth being blunt about it.

671
00:23:01,120 --> 00:23:04,680
Someone or something has to actually do the compiling.

672
00:23:04,680 --> 00:23:06,240
The three layers we just walked through

673
00:23:06,240 --> 00:23:09,640
don't assemble themselves the moment you understand the concept.

674
00:23:09,640 --> 00:23:12,520
Current copilot deployments don't ship with a background process

675
00:23:12,520 --> 00:23:13,960
that reads your SharePoint libraries

676
00:23:13,960 --> 00:23:15,840
and writes structured pages out of them.

677
00:23:15,840 --> 00:23:16,840
That work is real work.

678
00:23:16,840 --> 00:23:18,960
It has to be built or run or triggered

679
00:23:18,960 --> 00:23:21,120
by something that treats it as its job.

680
00:23:21,120 --> 00:23:24,080
Right now, in a standard Microsoft 365 environment,

681
00:23:24,080 --> 00:23:25,440
nothing is doing that job.

682
00:23:25,440 --> 00:23:28,120
Copilot searches it doesn't compile the wiki layer

683
00:23:28,120 --> 00:23:31,160
if it exists at all exists because somebody decided to build it,

684
00:23:31,160 --> 00:23:33,160
not because it came in the box.

685
00:23:33,160 --> 00:23:35,160
Which puts a fairly obvious question on the table

686
00:23:35,160 --> 00:23:36,640
and it's the one worth answering next.

687
00:23:36,640 --> 00:23:38,560
If this layer doesn't ship by default

688
00:23:38,560 --> 00:23:42,000
and it doesn't build itself, who or what is actually supposed

689
00:23:42,000 --> 00:23:44,000
to build it inside a Microsoft ecosystem?

690
00:23:44,000 --> 00:23:45,760
Because the three layer structure only matters

691
00:23:45,760 --> 00:23:47,320
if there's a real answer to that question,

692
00:23:47,320 --> 00:23:49,920
not just a diagram that looks clean on a slide.

693
00:23:49,920 --> 00:23:51,400
Who builds the wiki layer?

694
00:23:51,400 --> 00:23:53,480
Let's be honest about something before we go any further

695
00:23:53,480 --> 00:23:56,160
because glossing over it would make the rest of this episode

696
00:23:56,160 --> 00:23:57,840
sound like a feature announcement.

697
00:23:57,840 --> 00:24:00,280
This isn't a Microsoft ship capability today.

698
00:24:00,280 --> 00:24:02,320
There's no toggle in the admin center labeled

699
00:24:02,320 --> 00:24:04,760
turn on compiled knowledge.

700
00:24:04,760 --> 00:24:06,440
What we've been describing is a pattern,

701
00:24:06,440 --> 00:24:08,920
something you'd implement on top of your existing knowledge

702
00:24:08,920 --> 00:24:10,320
state, not something that arrives

703
00:24:10,320 --> 00:24:11,960
with your next licensing renewal.

704
00:24:11,960 --> 00:24:13,960
So what does implementing it actually look like?

705
00:24:13,960 --> 00:24:16,160
Picture an AI agent and if you've spent any time

706
00:24:16,160 --> 00:24:19,040
in Copilot Studio, the shape of this will feel familiar.

707
00:24:19,040 --> 00:24:20,880
Not unlike how a Copilot Studio agent

708
00:24:20,880 --> 00:24:23,120
gets built to watch a trigger and take an action

709
00:24:23,120 --> 00:24:24,800
or how a custom graph connected agent

710
00:24:24,800 --> 00:24:28,160
gets set up to reach into specific data sources on a schedule.

711
00:24:28,160 --> 00:24:31,240
Same underlying capability pointed at a different job.

712
00:24:31,240 --> 00:24:33,760
Instead of answering a user's question in the moment,

713
00:24:33,760 --> 00:24:36,200
this agent's entire purpose is reading sources

714
00:24:36,200 --> 00:24:38,200
and writing structured pages out of them.

715
00:24:38,200 --> 00:24:40,720
It goes into a SharePoint library, reads what's there,

716
00:24:40,720 --> 00:24:42,480
decides what belongs on an existing page

717
00:24:42,480 --> 00:24:44,640
and what deserves a new one and writes accordingly.

718
00:24:44,640 --> 00:24:45,680
Nobody's chatting with it.

719
00:24:45,680 --> 00:24:47,040
It's not waiting for a prompt.

720
00:24:47,040 --> 00:24:49,360
It's just working quietly in the background

721
00:24:49,360 --> 00:24:51,080
whether or not anyone's logged in that day.

722
00:24:51,080 --> 00:24:52,520
That distinction is worth sitting with

723
00:24:52,520 --> 00:24:55,040
because it's where agente orchestration actually earns

724
00:24:55,040 --> 00:24:56,880
its name instead of just being a buzzword.

725
00:24:56,880 --> 00:24:59,120
This isn't a chatbot answering questions on demand.

726
00:24:59,120 --> 00:25:01,680
It's a background process running on its own schedule,

727
00:25:01,680 --> 00:25:03,160
compiling and maintaining knowledge,

728
00:25:03,160 --> 00:25:05,880
whether or not a single person asks it anything that day.

729
00:25:05,880 --> 00:25:07,360
The value isn't in the conversation.

730
00:25:07,360 --> 00:25:09,160
It's in the fact that the conversation never has

731
00:25:09,160 --> 00:25:11,240
to start from zero because something already did

732
00:25:11,240 --> 00:25:12,960
the work before the question showed up.

733
00:25:12,960 --> 00:25:14,960
And here's the part that should feel encouraging

734
00:25:14,960 --> 00:25:15,800
rather than daunting.

735
00:25:15,800 --> 00:25:18,240
None of the infrastructure to build this is missing.

736
00:25:18,240 --> 00:25:20,840
Copilot Studio already supports custom connectors,

737
00:25:20,840 --> 00:25:23,000
reaching into whatever data source you pointed at.

738
00:25:23,000 --> 00:25:24,720
It already supports scheduled flows,

739
00:25:24,720 --> 00:25:26,480
running processes on a timer without a human

740
00:25:26,480 --> 00:25:27,680
kicking them off manually.

741
00:25:27,680 --> 00:25:29,960
Those two capabilities connectors and schedules

742
00:25:29,960 --> 00:25:31,360
are most of what you need to build

743
00:25:31,360 --> 00:25:33,160
the compiling half of the system.

744
00:25:33,160 --> 00:25:35,000
The pieces exist in your tenant right now.

745
00:25:35,000 --> 00:25:36,840
They're just not assembled this way by default

746
00:25:36,840 --> 00:25:39,040
because nobody's shipped a template that says,

747
00:25:39,040 --> 00:25:40,680
"Here's how you wire a connector

748
00:25:40,680 --> 00:25:42,680
and a schedule together to build a wiki

749
00:25:42,680 --> 00:25:45,160
instead of just automating a workflow."

750
00:25:45,160 --> 00:25:47,960
Which means the real barrier here isn't technical capability.

751
00:25:47,960 --> 00:25:51,200
It's a shift in how IT thinks about what Copilot even is.

752
00:25:51,200 --> 00:25:54,400
Right now, most organizations think of Copilot as one thing,

753
00:25:54,400 --> 00:25:55,680
a chat interface.

754
00:25:55,680 --> 00:25:56,720
You type, it answers.

755
00:25:56,720 --> 00:25:58,080
That's the whole mental model.

756
00:25:58,080 --> 00:26:00,240
But once you see the wiki pattern clearly,

757
00:26:00,240 --> 00:26:02,840
Copilot stops being one system and starts being two.

758
00:26:02,840 --> 00:26:04,720
There's the compiler, the background agent quietly

759
00:26:04,720 --> 00:26:06,680
reading, writing and maintaining structured pages.

760
00:26:06,680 --> 00:26:09,080
And there's the responder, the familiar chat interface

761
00:26:09,080 --> 00:26:11,320
that answers questions, except now it's answering

762
00:26:11,320 --> 00:26:13,800
from compiled pages instead of raw fragments.

763
00:26:13,800 --> 00:26:16,440
Two systems doing two different jobs working together.

764
00:26:16,440 --> 00:26:19,800
Most IT teams today have only built or bought the second one.

765
00:26:19,800 --> 00:26:22,760
Once you see it that way, as two systems instead of one,

766
00:26:22,760 --> 00:26:24,200
a different question starts to matter

767
00:26:24,200 --> 00:26:25,760
more than efficiency ever did.

768
00:26:25,760 --> 00:26:28,040
Because splitting the compiler from the responder

769
00:26:28,040 --> 00:26:30,440
doesn't just make answers cheaper or faster.

770
00:26:30,440 --> 00:26:31,880
It changes something more fundamental

771
00:26:31,880 --> 00:26:34,040
about whether people trust what Copilot tells them

772
00:26:34,040 --> 00:26:35,280
in the first place.

773
00:26:35,280 --> 00:26:37,760
The trust problem, RAAG, can't solve.

774
00:26:37,760 --> 00:26:39,560
There's data behind this, and it's worth putting

775
00:26:39,560 --> 00:26:41,560
a real number on it before we go further.

776
00:26:41,560 --> 00:26:43,640
Over 60% of enterprise IT leaders

777
00:26:43,640 --> 00:26:46,000
said they intended to expand generative AI

778
00:26:46,000 --> 00:26:47,200
for knowledge discovery.

779
00:26:47,200 --> 00:26:48,320
That's not a small commitment.

780
00:26:48,320 --> 00:26:50,520
That's leadership teams planning to lean harder

781
00:26:50,520 --> 00:26:52,960
into this budgeting forward, telling their boards

782
00:26:52,960 --> 00:26:54,520
it's the next phase of the rollout.

783
00:26:54,520 --> 00:26:56,680
And then in a lot of those same organizations adoption

784
00:26:56,680 --> 00:26:59,160
stalls, not because people forgot to use the tool,

785
00:26:59,160 --> 00:27:00,520
because somewhere along the way,

786
00:27:00,520 --> 00:27:03,000
the answer started feeling inconsistent.

787
00:27:03,000 --> 00:27:05,160
And inconsistency is quietly fatal

788
00:27:05,160 --> 00:27:06,920
to whether people keep coming back.

789
00:27:06,920 --> 00:27:09,480
Here's what's worth understanding about that inconsistency,

790
00:27:09,480 --> 00:27:11,960
because it isn't random noise you can chalk up to bad luck

791
00:27:11,960 --> 00:27:13,480
or a model having an off day.

792
00:27:13,480 --> 00:27:14,600
It's structural.

793
00:27:14,600 --> 00:27:17,240
Ask the same underlying question two different ways,

794
00:27:17,240 --> 00:27:19,400
and RAAG doesn't just phrase the answer differently.

795
00:27:19,400 --> 00:27:21,640
It retrieves an entirely different set of chunks,

796
00:27:21,640 --> 00:27:23,320
different search, different fragments pulled

797
00:27:23,320 --> 00:27:25,080
from the index different material handed

798
00:27:25,080 --> 00:27:26,800
to the model to synthesize from.

799
00:27:26,800 --> 00:27:28,280
Two people asking about the same policy,

800
00:27:28,280 --> 00:27:31,240
one saying, what's our remote work policy and the other saying,

801
00:27:31,240 --> 00:27:33,080
can I work from home three days a week?

802
00:27:33,080 --> 00:27:35,480
Might get two answers that don't even agree with each other.

803
00:27:35,480 --> 00:27:38,280
Because under the hood, they triggered two separate searches

804
00:27:38,280 --> 00:27:40,640
that happened to land on different documents.

805
00:27:40,640 --> 00:27:42,240
Users notice this fast.

806
00:27:42,240 --> 00:27:44,040
And here's the part that should worry you more

807
00:27:44,040 --> 00:27:45,360
than a formal complaint would.

808
00:27:45,360 --> 00:27:46,320
They don't complain loudly.

809
00:27:46,320 --> 00:27:49,080
Nobody files a ticket that says co-pilot gave me inconsistent

810
00:27:49,080 --> 00:27:49,880
answers.

811
00:27:49,880 --> 00:27:52,400
What actually happens is quieter and harder to track.

812
00:27:52,400 --> 00:27:55,600
Someone gets burned once, maybe twice, on a question that mattered.

813
00:27:55,600 --> 00:27:57,960
And they just stop asking co-pilot the harder questions.

814
00:27:57,960 --> 00:27:59,520
They still use it for quick summaries,

815
00:27:59,520 --> 00:28:02,480
for drafting an email, for things with no real stakes attached.

816
00:28:02,480 --> 00:28:04,600
But the moment a real decision rides on the answer,

817
00:28:04,600 --> 00:28:06,160
they go find a person instead.

818
00:28:06,160 --> 00:28:08,720
That shift happens silently one person at a time.

819
00:28:08,720 --> 00:28:10,360
And it never shows up as a complaint.

820
00:28:10,360 --> 00:28:12,480
It shows up months later as usage numbers

821
00:28:12,480 --> 00:28:15,200
that look fine on paper, but never touch the questions

822
00:28:15,200 --> 00:28:16,280
that actually mattered.

823
00:28:16,280 --> 00:28:19,120
Now compare that to what a compiled wiki page does.

824
00:28:19,120 --> 00:28:21,680
Ask it the same underlying question five different ways,

825
00:28:21,680 --> 00:28:24,240
phrased five different ways, worded by five different people.

826
00:28:24,240 --> 00:28:25,800
You get the same answer every time.

827
00:28:25,800 --> 00:28:28,160
Because the synthesis already happened once carefully

828
00:28:28,160 --> 00:28:30,680
before any of those five people ever typed anything.

829
00:28:30,680 --> 00:28:32,440
There's no fresh search rolling the dice

830
00:28:32,440 --> 00:28:33,920
on which fragments get pulled.

831
00:28:33,920 --> 00:28:36,320
There's one page already reconciled, already settled,

832
00:28:36,320 --> 00:28:38,840
and every version of the question just gets pointed at it.

833
00:28:38,840 --> 00:28:40,840
This is the part worth naming directly,

834
00:28:40,840 --> 00:28:44,480
because it's easy to file under nice to have and move past.

835
00:28:44,480 --> 00:28:46,200
Consistency isn't a polish feature.

836
00:28:46,200 --> 00:28:47,280
It's a trust mechanism.

837
00:28:47,280 --> 00:28:49,400
It's the actual thing that decides whether someone

838
00:28:49,400 --> 00:28:52,880
is willing to let a tools answer carry weight in a real decision,

839
00:28:52,880 --> 00:28:55,080
or whether they keep it around for the easy stuff

840
00:28:55,080 --> 00:28:58,520
and quietly root anything important through a person instead.

841
00:28:58,520 --> 00:29:00,400
Once trust breaks on the hard questions,

842
00:29:00,400 --> 00:29:02,240
it doesn't come back just because the interface

843
00:29:02,240 --> 00:29:03,640
gets a new code of paint.

844
00:29:03,640 --> 00:29:05,280
And trust is only half the payoff here.

845
00:29:05,280 --> 00:29:07,960
The other half shows up somewhere you might not expect.

846
00:29:07,960 --> 00:29:10,520
And how much redundant work is happening across your organization

847
00:29:10,520 --> 00:29:11,560
right now.

848
00:29:11,560 --> 00:29:14,040
Work that exists purely because nobody trusted search

849
00:29:14,040 --> 00:29:17,120
to surface the answer that already existed.

850
00:29:17,120 --> 00:29:19,080
Killing redundant knowledge creation?

851
00:29:19,080 --> 00:29:20,880
There's another number worth putting on the table

852
00:29:20,880 --> 00:29:22,600
and it points at something most organizations

853
00:29:22,600 --> 00:29:24,280
never think to measure.

854
00:29:24,280 --> 00:29:26,880
Early adopters of AI-based knowledge discovery report

855
00:29:26,880 --> 00:29:30,360
a 25% to 40% reduction in redundant knowledge creation.

856
00:29:30,360 --> 00:29:32,440
That's a strange phrase, redundant knowledge creation.

857
00:29:32,440 --> 00:29:34,000
So let's unpack what it actually means.

858
00:29:34,000 --> 00:29:36,480
Because once you see it, you start noticing it everywhere.

859
00:29:36,480 --> 00:29:38,960
Here's the mechanism, and it's simpler than it sounds.

860
00:29:38,960 --> 00:29:41,840
When nobody trusts search to surface the right answer,

861
00:29:41,840 --> 00:29:43,000
people don't just give up.

862
00:29:43,000 --> 00:29:43,600
They rebuild.

863
00:29:43,600 --> 00:29:46,000
Someone needs a document explaining how a process works.

864
00:29:46,000 --> 00:29:47,680
They run a search, get nothing solid,

865
00:29:47,680 --> 00:29:49,920
and instead of digging further, they just write it again

866
00:29:49,920 --> 00:29:51,080
from scratch.

867
00:29:51,080 --> 00:29:53,120
Except that document already exists.

868
00:29:53,120 --> 00:29:55,840
It's sitting in a team's thread from four months ago

869
00:29:55,840 --> 00:29:58,280
or buried on page three of an old SharePoint site.

870
00:29:58,280 --> 00:29:59,840
Nobody thinks to check anymore.

871
00:29:59,840 --> 00:30:01,080
The information isn't missing.

872
00:30:01,080 --> 00:30:04,040
It's unfindable, which functionally amounts to the same thing.

873
00:30:04,040 --> 00:30:07,360
And unfindable information gets recreated over and over

874
00:30:07,360 --> 00:30:09,240
by different people who have no idea.

875
00:30:09,240 --> 00:30:11,400
They're duplicating work that already happened.

876
00:30:11,400 --> 00:30:13,640
This is where a compiled weeky changes the math.

877
00:30:13,640 --> 00:30:15,240
Instead of a search turning up nothing

878
00:30:15,240 --> 00:30:18,200
useful or five slightly different fragments that half answer

879
00:30:18,200 --> 00:30:20,520
the question, it surfaces the existing answer

880
00:30:20,520 --> 00:30:24,120
directly with a citation pointing back to where it came from.

881
00:30:24,120 --> 00:30:26,520
Someone doesn't spend an afternoon reconstructing a document

882
00:30:26,520 --> 00:30:28,680
that already lives in the company's collective memory.

883
00:30:28,680 --> 00:30:30,560
They find it in minutes because the weeky

884
00:30:30,560 --> 00:30:33,120
already did the work of deciding that this concept exists

885
00:30:33,120 --> 00:30:34,320
and here's where it lives.

886
00:30:34,320 --> 00:30:36,240
And this connects back to something we touched on earlier,

887
00:30:36,240 --> 00:30:38,600
but it's worth stating in its sharpest form here.

888
00:30:38,600 --> 00:30:40,680
Ragtreeze every piece of content is equally hard

889
00:30:40,680 --> 00:30:41,880
to find every single time.

890
00:30:41,880 --> 00:30:44,040
It doesn't matter if someone asked the exact same question

891
00:30:44,040 --> 00:30:45,960
last week and got a great answer.

892
00:30:45,960 --> 00:30:48,280
The system starts over searching the same haystack

893
00:30:48,280 --> 00:30:50,280
with the same odds of missing what it's looking for,

894
00:30:50,280 --> 00:30:51,840
a wiki flips that completely.

895
00:30:51,840 --> 00:30:54,480
It treats concepts as already found permanently.

896
00:30:54,480 --> 00:30:57,440
Once something's compiled onto a page, it stays found.

897
00:30:57,440 --> 00:31:00,280
Nobody has to get lucky with their search terms ever again.

898
00:31:00,280 --> 00:31:02,920
Frame the business case here in the planist terms possible

899
00:31:02,920 --> 00:31:05,320
because this is where a lot of the argument in this episode

900
00:31:05,320 --> 00:31:08,080
stops being architectural and starts being financial.

901
00:31:08,080 --> 00:31:10,160
Less duplicated effort isn't a technical

902
00:31:10,160 --> 00:31:12,120
when you put on a slide about system design.

903
00:31:12,120 --> 00:31:16,000
It's hours, actual hours given back to people doing actual work

904
00:31:16,000 --> 00:31:17,560
instead of quietly rebuilding something

905
00:31:17,560 --> 00:31:19,600
that already existed somewhere in the tenant,

906
00:31:19,600 --> 00:31:22,800
multiply that across a few hundred employees over a year

907
00:31:22,800 --> 00:31:24,680
and you're not talking about a nice efficiency gain.

908
00:31:24,680 --> 00:31:26,480
You're talking about a meaningful chunk

909
00:31:26,480 --> 00:31:28,400
of your organization's collective time

910
00:31:28,400 --> 00:31:30,800
spent solving a problem that had already been solved.

911
00:31:30,800 --> 00:31:32,480
But none of this holds up if the material

912
00:31:32,480 --> 00:31:34,600
feeding the wiki in the first place is a mess

913
00:31:34,600 --> 00:31:36,040
and that's the next real constraint

914
00:31:36,040 --> 00:31:37,600
worth being honest about.

915
00:31:37,600 --> 00:31:39,440
Garbage in, still garbage out.

916
00:31:39,440 --> 00:31:40,640
Let's be direct about something

917
00:31:40,640 --> 00:31:42,800
the whole wiki pattern depends on because it's the part

918
00:31:42,800 --> 00:31:45,640
that's easiest to skip past when you're excited about the architecture.

919
00:31:45,640 --> 00:31:47,680
The wiki doesn't fix bad source material.

920
00:31:47,680 --> 00:31:49,040
It organizes what you give it.

921
00:31:49,040 --> 00:31:51,480
It doesn't audit whether what you gave it is true, current

922
00:31:51,480 --> 00:31:53,040
or even still relevant.

923
00:31:53,040 --> 00:31:54,840
That job never belonged to the compiler

924
00:31:54,840 --> 00:31:55,720
and it never will.

925
00:31:55,720 --> 00:31:57,440
Here's what that means in practice.

926
00:31:57,440 --> 00:32:00,280
If your share point is full of outdated policy documents,

927
00:32:00,280 --> 00:32:02,200
documents nobody archived properly,

928
00:32:02,200 --> 00:32:04,520
documents that describe a process your organization

929
00:32:04,520 --> 00:32:06,480
stopped using two reogs ago,

930
00:32:06,480 --> 00:32:08,480
the wiki will read every one of them faithfully

931
00:32:08,480 --> 00:32:12,000
and compile them into clean, confident looking pages.

932
00:32:12,000 --> 00:32:13,360
That's the uncomfortable part.

933
00:32:13,360 --> 00:32:15,080
The output won't look messy or uncertain.

934
00:32:15,080 --> 00:32:17,280
It'll look organized, it'll look authoritative

935
00:32:17,280 --> 00:32:19,440
and it'll be wrong presented with the same polish

936
00:32:19,440 --> 00:32:21,680
as everything else in the wiki that happens to be right.

937
00:32:21,680 --> 00:32:24,120
This is exactly where content lifecycle management

938
00:32:24,120 --> 00:32:26,120
stops being optional and it's worth pointing out

939
00:32:26,120 --> 00:32:27,440
that Microsoft has been saying this

940
00:32:27,440 --> 00:32:30,720
since its own January 2024 Copilot readiness guidance

941
00:32:30,720 --> 00:32:32,240
that guidance wasn't subtle about it.

942
00:32:32,240 --> 00:32:34,640
Content needs owners, it needs version control,

943
00:32:34,640 --> 00:32:36,440
it needs a process for retiring things

944
00:32:36,440 --> 00:32:38,840
that are no longer accurate instead of letting them sit

945
00:32:38,840 --> 00:32:40,880
indefinitely next to whatever replaced them.

946
00:32:40,880 --> 00:32:43,400
None of that is new advice invented for this episode.

947
00:32:43,400 --> 00:32:45,880
It's been sitting in Microsoft's own documentation

948
00:32:45,880 --> 00:32:47,440
for a while mostly ignored

949
00:32:47,440 --> 00:32:49,120
because it sounded like housekeeping

950
00:32:49,120 --> 00:32:50,680
rather than something that determines

951
00:32:50,680 --> 00:32:53,920
whether your AI investment actually works.

952
00:32:53,920 --> 00:32:56,200
So say it plainly because this is the part organizations

953
00:32:56,200 --> 00:32:57,160
keep skipping.

954
00:32:57,160 --> 00:32:59,800
Structured knowledge work starts before any AI

955
00:32:59,800 --> 00:33:02,080
touches your content retention labels still matter,

956
00:33:02,080 --> 00:33:04,000
content type still matter, clear ownership.

957
00:33:04,000 --> 00:33:05,280
So someone's actually responsible

958
00:33:05,280 --> 00:33:08,400
for knowing whether a document is still true, still matters.

959
00:33:08,400 --> 00:33:09,840
None of that becomes less important

960
00:33:09,840 --> 00:33:11,360
once you add a wiki layer on top.

961
00:33:11,360 --> 00:33:13,400
If anything, it becomes the entire foundation

962
00:33:13,400 --> 00:33:14,520
the wiki is standing on.

963
00:33:14,520 --> 00:33:16,120
And here's the sharper version of the risk

964
00:33:16,120 --> 00:33:17,280
the one worth sitting with.

965
00:33:17,280 --> 00:33:19,520
The wiki pattern raises the stakes on curation

966
00:33:19,520 --> 00:33:21,840
because it makes bad content look more authoritative,

967
00:33:21,840 --> 00:33:22,680
not less.

968
00:33:22,680 --> 00:33:25,240
A messy pile of raw documents, at least looks messy.

969
00:33:25,240 --> 00:33:26,960
People approach it with some skepticism.

970
00:33:26,960 --> 00:33:28,520
A beautifully compiled wiki page

971
00:33:28,520 --> 00:33:30,120
doesn't invite that same skepticism.

972
00:33:30,120 --> 00:33:32,040
It reads like something's already been verified,

973
00:33:32,040 --> 00:33:34,440
organized and settled even when the underlying source

974
00:33:34,440 --> 00:33:35,760
was garbage the whole time.

975
00:33:35,760 --> 00:33:38,120
So assuming your source material actually is solid,

976
00:33:38,120 --> 00:33:39,120
here's what happens next.

977
00:33:39,120 --> 00:33:40,760
Once new content starts arriving

978
00:33:40,760 --> 00:33:43,640
and the system has to decide what to do with it.

979
00:33:43,640 --> 00:33:45,560
When bad knowledge looks authoritative,

980
00:33:45,560 --> 00:33:47,280
here's what that means in practice.

981
00:33:47,280 --> 00:33:50,240
If your share point is full of outdated policy documents,

982
00:33:50,240 --> 00:33:51,880
documents nobody archived properly,

983
00:33:51,880 --> 00:33:54,160
documents that describe a process your organization

984
00:33:54,160 --> 00:33:56,000
stopped using two reoggs ago,

985
00:33:56,000 --> 00:33:58,240
the wiki will read every one of them faithfully

986
00:33:58,240 --> 00:34:01,160
and compile them into clean, confident looking pages.

987
00:34:01,160 --> 00:34:02,520
That's the uncomfortable part.

988
00:34:02,520 --> 00:34:04,640
The output won't look messy or uncertain.

989
00:34:04,640 --> 00:34:06,720
It'll look organized, it'll look authoritative

990
00:34:06,720 --> 00:34:09,040
and it'll be wrong, presented with the same polish

991
00:34:09,040 --> 00:34:11,440
as everything else in the wiki that happens to be right.

992
00:34:11,440 --> 00:34:13,880
This is exactly where content lifecycle management

993
00:34:13,880 --> 00:34:15,120
stops being optional.

994
00:34:15,120 --> 00:34:16,840
And it's worth pointing out that Microsoft

995
00:34:16,840 --> 00:34:20,080
has been saying this since its own January 2024

996
00:34:20,080 --> 00:34:21,760
co-pilot readiness guidance.

997
00:34:21,760 --> 00:34:23,480
That guidance wasn't subtle about it.

998
00:34:23,480 --> 00:34:25,000
Content needs owners.

999
00:34:25,000 --> 00:34:26,200
It needs version controlled.

1000
00:34:26,200 --> 00:34:28,000
It needs a process for retiring things

1001
00:34:28,000 --> 00:34:30,160
that are no longer accurate instead of letting them sit

1002
00:34:30,160 --> 00:34:32,440
indefinitely next to whatever replaced them.

1003
00:34:32,440 --> 00:34:34,800
None of that is new advice invented for this episode.

1004
00:34:34,800 --> 00:34:36,840
It's been sitting in Microsoft's own documentation

1005
00:34:36,840 --> 00:34:40,440
for a while, mostly ignored because it sounded like housekeeping

1006
00:34:40,440 --> 00:34:42,680
rather than something that determines whether your AI

1007
00:34:42,680 --> 00:34:44,560
investment actually works.

1008
00:34:44,560 --> 00:34:46,480
So say it plainly, because this is the part

1009
00:34:46,480 --> 00:34:47,960
organizations keep skipping.

1010
00:34:47,960 --> 00:34:51,680
Structured knowledge work starts before any AI touches your content.

1011
00:34:51,680 --> 00:34:54,480
Retention labels still matter, content types still matter.

1012
00:34:54,480 --> 00:34:56,480
Clear ownership, so someone's actually responsible

1013
00:34:56,480 --> 00:34:58,760
for knowing whether a document is still true still matters.

1014
00:34:58,760 --> 00:35:00,200
None of that becomes less important

1015
00:35:00,200 --> 00:35:01,960
once you add a wiki layer on top.

1016
00:35:01,960 --> 00:35:03,760
If anything, it becomes the entire foundation

1017
00:35:03,760 --> 00:35:05,160
the wiki is standing on.

1018
00:35:05,160 --> 00:35:06,840
And here's the sharper version of the risk

1019
00:35:06,840 --> 00:35:08,080
the one worth sitting with.

1020
00:35:08,080 --> 00:35:09,960
The wiki pattern raises the stakes on curation

1021
00:35:09,960 --> 00:35:12,320
because it makes bad content look more authoritative,

1022
00:35:12,320 --> 00:35:13,000
not less.

1023
00:35:13,000 --> 00:35:15,320
A messy pile of raw documents at least looks messy.

1024
00:35:15,320 --> 00:35:17,320
People approach it with some skepticism.

1025
00:35:17,320 --> 00:35:18,520
A beautifully compiled wiki page

1026
00:35:18,520 --> 00:35:20,160
doesn't invite that same skepticism.

1027
00:35:20,160 --> 00:35:22,560
It reads like something's already been verified, organized

1028
00:35:22,560 --> 00:35:24,640
and settled, even when the underlying source

1029
00:35:24,640 --> 00:35:25,920
was garbage the whole time.

1030
00:35:25,920 --> 00:35:28,320
So assuming your source material actually is solid,

1031
00:35:28,320 --> 00:35:29,960
here's what happens next.

1032
00:35:29,960 --> 00:35:33,680
Once new content starts arriving and the system has to decide

1033
00:35:33,680 --> 00:35:34,560
what to do with it.

1034
00:35:34,560 --> 00:35:37,280
So picture what actually happens when a new source lands

1035
00:35:37,280 --> 00:35:40,080
because this is where the wiki pattern stops being a diagram

1036
00:35:40,080 --> 00:35:42,240
and starts being a repeatable process.

1037
00:35:42,240 --> 00:35:44,240
A document shows up, maybe a policy update,

1038
00:35:44,240 --> 00:35:46,400
maybe a new team's transcript from a client call.

1039
00:35:46,400 --> 00:35:48,520
The system reads it once, not skims it,

1040
00:35:48,520 --> 00:35:50,960
not chunks it into fragments for later retrieval.

1041
00:35:50,960 --> 00:35:52,560
Reads it, the way you'd read something

1042
00:35:52,560 --> 00:35:54,440
if your job was to understand it well enough

1043
00:35:54,440 --> 00:35:56,040
to explain it to someone else.

1044
00:35:56,040 --> 00:35:58,280
Key concepts get pulled out during that read.

1045
00:35:58,280 --> 00:35:59,680
What's this document actually about?

1046
00:35:59,680 --> 00:36:00,520
What does it claim?

1047
00:36:00,520 --> 00:36:01,600
What does it connect to?

1048
00:36:01,600 --> 00:36:03,560
From there, one of two things happens.

1049
00:36:03,560 --> 00:36:05,360
And this is worth being precise about because it's

1050
00:36:05,360 --> 00:36:06,400
the whole mechanism.

1051
00:36:06,400 --> 00:36:08,520
If the new source adds detail to something already

1052
00:36:08,520 --> 00:36:11,040
compiled, an existing page gets updated.

1053
00:36:11,040 --> 00:36:13,200
Say there's already a page about a client's onboarding

1054
00:36:13,200 --> 00:36:15,800
process, and this new transcript adds a wrinkle nobody

1055
00:36:15,800 --> 00:36:17,040
had documented yet.

1056
00:36:17,040 --> 00:36:18,000
That page gets richer.

1057
00:36:18,000 --> 00:36:19,120
It doesn't get duplicated.

1058
00:36:19,120 --> 00:36:21,840
It doesn't sit next to an old version creating confusion.

1059
00:36:21,840 --> 00:36:23,520
It gets updated in place.

1060
00:36:23,520 --> 00:36:25,680
But if the source introduces something genuinely new,

1061
00:36:25,680 --> 00:36:27,640
something that doesn't map onto any page that already

1062
00:36:27,640 --> 00:36:29,400
exists, a new page gets created.

1063
00:36:29,400 --> 00:36:31,520
And whichever of those two things happens,

1064
00:36:31,520 --> 00:36:33,120
related ideas get linked.

1065
00:36:33,120 --> 00:36:35,080
The new page connects to the client it belongs to,

1066
00:36:35,080 --> 00:36:37,720
the project it touches, the policy that shaped it, nothing

1067
00:36:37,720 --> 00:36:40,600
sits alone.

1068
00:36:40,600 --> 00:36:42,360
Here's the part that matters most, though.

1069
00:36:42,360 --> 00:36:45,080
And it's the part, rag has no equivalent for it all.

1070
00:36:45,080 --> 00:36:47,480
What happens when the new information contradicts something

1071
00:36:47,480 --> 00:36:48,920
already sitting in the wiki?

1072
00:36:48,920 --> 00:36:51,160
Say the old page claims a process works one way,

1073
00:36:51,160 --> 00:36:52,960
and the new source says it changed.

1074
00:36:52,960 --> 00:36:54,960
The system doesn't just override the old claim

1075
00:36:54,960 --> 00:36:56,240
and move on like nothing happened.

1076
00:36:56,240 --> 00:36:57,120
It flags it.

1077
00:36:57,120 --> 00:36:58,760
Both versions get preserved with a note

1078
00:36:58,760 --> 00:37:00,440
that something shifted, and a signal

1079
00:37:00,440 --> 00:37:02,480
that someone should look at this before treating either

1080
00:37:02,480 --> 00:37:03,800
version as settled.

1081
00:37:03,800 --> 00:37:05,640
This flagging behavior isn't a nice to have

1082
00:37:05,640 --> 00:37:07,240
detailed buried in the mechanics.

1083
00:37:07,240 --> 00:37:09,720
It matters enormously for actual enterprise use,

1084
00:37:09,720 --> 00:37:11,840
because businesses change constantly.

1085
00:37:11,840 --> 00:37:14,000
And most of that change happens quietly.

1086
00:37:14,000 --> 00:37:17,040
In a policy update here, a strategy pivot there,

1087
00:37:17,040 --> 00:37:19,840
a decision that got reversed without anyone sending out

1088
00:37:19,840 --> 00:37:21,240
a formal memo about it.

1089
00:37:21,240 --> 00:37:23,400
When a policy update contradicts an old assumption

1090
00:37:23,400 --> 00:37:25,000
that's been floating around for months,

1091
00:37:25,000 --> 00:37:27,080
that contradiction doesn't get buried under a pile

1092
00:37:27,080 --> 00:37:29,280
of documents that all look equally current.

1093
00:37:29,280 --> 00:37:32,760
It gets surfaced directly to whoever's supposed to notice it.

1094
00:37:32,760 --> 00:37:34,200
Now hold that against what rag does

1095
00:37:34,200 --> 00:37:36,400
with the exact same situation, because the contrast

1096
00:37:36,400 --> 00:37:37,680
is where this really lands.

1097
00:37:37,680 --> 00:37:39,400
A stale document sits in your share point

1098
00:37:39,400 --> 00:37:41,480
next to the document that replaced it.

1099
00:37:41,480 --> 00:37:44,080
Someone asks a question, and rag retrieves both.

1100
00:37:44,080 --> 00:37:46,680
Not one, both, as two equally weighted chunks,

1101
00:37:46,680 --> 00:37:49,200
stitched together into an answer with absolutely no signal

1102
00:37:49,200 --> 00:37:50,960
about which one is actually current.

1103
00:37:50,960 --> 00:37:52,880
The model doesn't know one superseded the other.

1104
00:37:52,880 --> 00:37:54,880
It just sees two fragments that seem relevant

1105
00:37:54,880 --> 00:37:57,640
and does its best to synthesize something coherent

1106
00:37:57,640 --> 00:37:59,840
out of material that's quietly contradicting itself.

1107
00:37:59,840 --> 00:38:01,040
Nobody flagged anything.

1108
00:38:01,040 --> 00:38:03,320
Nobody got a signal that something needed a second look.

1109
00:38:03,320 --> 00:38:05,800
The contradiction just sits there, silently corrupting

1110
00:38:05,800 --> 00:38:07,640
whatever answer comes out the other end.

1111
00:38:07,640 --> 00:38:10,560
That difference, flagging versus silent averaging,

1112
00:38:10,560 --> 00:38:13,000
points at something bigger than a technical detail

1113
00:38:13,000 --> 00:38:14,440
about how updates get handled.

1114
00:38:14,440 --> 00:38:15,800
It's a concept worth naming directly,

1115
00:38:15,800 --> 00:38:17,480
because once you see it, you can't unsee it

1116
00:38:17,480 --> 00:38:19,600
in every rag deployment you've ever looked at.

1117
00:38:19,600 --> 00:38:22,040
Contradiction as a feature, not a bug.

1118
00:38:22,040 --> 00:38:25,400
Most systems, when they run into contradictory information,

1119
00:38:25,400 --> 00:38:26,720
treat it as noise.

1120
00:38:26,720 --> 00:38:28,680
Something to smooth over, average out,

1121
00:38:28,680 --> 00:38:30,760
resolve into a single, clean answer,

1122
00:38:30,760 --> 00:38:32,560
so nobody has to look at the mess underneath.

1123
00:38:32,560 --> 00:38:35,240
The wiki pattern does something almost nobody expects

1124
00:38:35,240 --> 00:38:36,440
the first time they see it.

1125
00:38:36,440 --> 00:38:38,480
It treats contradiction as a signal,

1126
00:38:38,480 --> 00:38:40,680
worth preserving, not a problem to hide.

1127
00:38:40,680 --> 00:38:42,600
Think about what a contradiction actually represents

1128
00:38:42,600 --> 00:38:43,760
inside a business.

1129
00:38:43,760 --> 00:38:45,120
It's rarely a mistake.

1130
00:38:45,120 --> 00:38:46,800
Most of the time, it's a policy shift somebody

1131
00:38:46,800 --> 00:38:49,040
made three months ago and never fully announced,

1132
00:38:49,040 --> 00:38:52,240
or a strategy pivot that quietly replaced last year's plan,

1133
00:38:52,240 --> 00:38:54,880
or a decision that got reversed after a client pushed back,

1134
00:38:54,880 --> 00:38:57,160
and only half the team ever heard about the reversal.

1135
00:38:57,160 --> 00:38:59,360
Contradictions in an organizational context

1136
00:38:59,360 --> 00:39:02,480
are usually just changed that hasn't finished propagating yet.

1137
00:39:02,480 --> 00:39:05,160
Here's where rag runs into a wall it can't get passed.

1138
00:39:05,160 --> 00:39:08,320
A stateless system has no way to represent this used to be true,

1139
00:39:08,320 --> 00:39:09,720
now it isn't.

1140
00:39:09,720 --> 00:39:11,200
There's no mechanism for that.

1141
00:39:11,200 --> 00:39:13,880
Rag just retrieves, whichever chunk seems more relevant

1142
00:39:13,880 --> 00:39:15,600
at the moment the question gets asked,

1143
00:39:15,600 --> 00:39:18,400
and depending on wording, embedding similarity,

1144
00:39:18,400 --> 00:39:20,400
whatever's closest in the vector space,

1145
00:39:20,400 --> 00:39:23,200
it might hand you the old policy or the new one.

1146
00:39:23,200 --> 00:39:24,840
Neither answer is technically wrong.

1147
00:39:24,840 --> 00:39:26,000
The document exists.

1148
00:39:26,000 --> 00:39:28,520
It's just that rag has no concept of existed,

1149
00:39:28,520 --> 00:39:30,120
but got superseded.

1150
00:39:30,120 --> 00:39:33,120
Every chunk is equally present, equally retrievable,

1151
00:39:33,120 --> 00:39:36,560
equally weighted, forever, regardless of whether it's still true.

1152
00:39:36,560 --> 00:39:39,200
A compiled wiki does something structurally different.

1153
00:39:39,200 --> 00:39:42,520
It can hold both the old state and the new state on the same page,

1154
00:39:42,520 --> 00:39:45,720
with a note attached explaining when the change happened and why,

1155
00:39:45,720 --> 00:39:49,600
if that reason is known, not too competing documents floating in a search index

1156
00:39:49,600 --> 00:39:51,360
with no relationship to each other.

1157
00:39:51,360 --> 00:39:53,600
One page, with history built into it.

1158
00:39:53,600 --> 00:39:58,160
The old assumption is still there, but it's marked as superseded, dated, explained.

1159
00:39:58,160 --> 00:40:00,960
Nothing gets erased, nothing gets hidden, it just gets sequenced.

1160
00:40:00,960 --> 00:40:03,560
This is the actual difference between an assistant that answers

1161
00:40:03,560 --> 00:40:06,360
and an assistant that understands your organization's history.

1162
00:40:06,360 --> 00:40:09,160
Answering means producing something plausible when asked.

1163
00:40:09,160 --> 00:40:10,960
Understanding history means knowing that

1164
00:40:10,960 --> 00:40:12,880
what's true today used to be something else

1165
00:40:12,880 --> 00:40:15,800
and knowing why the shift happened and being able to tell you that story

1166
00:40:15,800 --> 00:40:18,680
instead of just picking aside and hoping it's the right one.

1167
00:40:18,680 --> 00:40:20,600
Understanding history like that is one thing,

1168
00:40:20,600 --> 00:40:23,080
making it usable across an entire enterprise,

1169
00:40:23,080 --> 00:40:25,680
not just one project page or one client file,

1170
00:40:25,680 --> 00:40:27,920
but thousands of them, spanning years,

1171
00:40:27,920 --> 00:40:30,640
across every team, is an entirely different problem.

1172
00:40:30,640 --> 00:40:32,840
And it's the one worth turning to next.

1173
00:40:32,840 --> 00:40:34,080
The scale question.

1174
00:40:34,080 --> 00:40:35,520
So let's be honest about the limit here,

1175
00:40:35,520 --> 00:40:37,840
because this pattern doesn't scale the way you'd wanted to,

1176
00:40:37,840 --> 00:40:39,520
just by wishing hard enough.

1177
00:40:39,520 --> 00:40:42,360
Personal scale wikis, the kind cup he's been describing,

1178
00:40:42,360 --> 00:40:44,920
work well around 100 or so source documents.

1179
00:40:44,920 --> 00:40:48,040
That's the range where one person or one small compiling agent

1180
00:40:48,040 --> 00:40:50,480
can keep everything coherent, cross-linked and current

1181
00:40:50,480 --> 00:40:53,400
without the whole thing turning into its own management problem.

1182
00:40:53,400 --> 00:40:55,040
Your enterprise content estate is not that.

1183
00:40:55,040 --> 00:40:57,320
It's not 100 documents, it's not 1,000.

1184
00:40:57,320 --> 00:41:00,280
Most organizations are sitting on content spread across years,

1185
00:41:00,280 --> 00:41:03,640
teams and systems that nobody's fully audited, let alone compiled.

1186
00:41:03,640 --> 00:41:04,920
So where does that leave the pattern

1187
00:41:04,920 --> 00:41:06,760
once your past personal scale?

1188
00:41:06,760 --> 00:41:09,640
The honest answer is pointing towards 2026 outlooks

1189
00:41:09,640 --> 00:41:11,440
that keep landing on the same conclusion.

1190
00:41:11,440 --> 00:41:13,920
Roughly 75% of enterprise applications

1191
00:41:13,920 --> 00:41:15,960
are moving toward hybrid architectures,

1192
00:41:15,960 --> 00:41:18,200
ones that combine wiki style compilation

1193
00:41:18,200 --> 00:41:21,160
with rag and agentic search, rather than picking one

1194
00:41:21,160 --> 00:41:22,400
and walking away from the other.

1195
00:41:22,400 --> 00:41:24,520
That's not a compromise bone out of indecision.

1196
00:41:24,520 --> 00:41:26,080
It's a recognition that these two systems

1197
00:41:26,080 --> 00:41:29,040
are solving different problems, and neither one covers the other's job.

1198
00:41:29,040 --> 00:41:30,800
Here's why hybrid actually makes sense,

1199
00:41:30,800 --> 00:41:33,320
once you stop thinking of this as a competition.

1200
00:41:33,320 --> 00:41:36,680
Rag remains the default for large, relatively static repositories,

1201
00:41:36,680 --> 00:41:39,200
the sprawling stuff that doesn't get asked about constantly,

1202
00:41:39,200 --> 00:41:42,120
but still needs to be searchable when someone does need it.

1203
00:41:42,120 --> 00:41:44,080
Wiki's excel somewhere else entirely,

1204
00:41:44,080 --> 00:41:47,160
the stable, high value knowledge that gets asked about again and again,

1205
00:41:47,160 --> 00:41:48,680
the stuff worth the effort of compiling

1206
00:41:48,680 --> 00:41:50,440
because people keep coming back to it.

1207
00:41:50,440 --> 00:41:52,920
Frame it as a division of labor because that's really what it is.

1208
00:41:52,920 --> 00:41:54,880
Rag handles breadth, it covers the long tail,

1209
00:41:54,880 --> 00:41:56,480
everything that exists, but doesn't get touched

1210
00:41:56,480 --> 00:41:58,840
often enough to justify compiling it by hand.

1211
00:41:58,840 --> 00:42:01,080
The wiki handles depths on the narrow set of things

1212
00:42:01,080 --> 00:42:04,840
that actually matter enough to deserve careful, maintained synthesis.

1213
00:42:04,840 --> 00:42:06,600
Neither one is trying to do the other's job.

1214
00:42:06,600 --> 00:42:07,640
That's the whole point.

1215
00:42:07,640 --> 00:42:09,440
Worth a cost reality check here, too,

1216
00:42:09,440 --> 00:42:11,600
because none of this is free either direction.

1217
00:42:11,600 --> 00:42:15,440
MVP Rag systems typically run 15 to $40,000 to stand up,

1218
00:42:15,440 --> 00:42:17,080
covering the vector infrastructure,

1219
00:42:17,080 --> 00:42:19,360
the embedding pipeline, the retrieval tooling,

1220
00:42:19,360 --> 00:42:20,720
the wiki layer isn't free,

1221
00:42:20,720 --> 00:42:22,320
but its cost doesn't show up the same way.

1222
00:42:22,320 --> 00:42:24,680
It shows up as ongoing maintenance, not infrastructure.

1223
00:42:24,680 --> 00:42:25,960
You're not buying a bigger system.

1224
00:42:25,960 --> 00:42:27,760
You're paying for the compiling and upkeep

1225
00:42:27,760 --> 00:42:31,360
that keeps the wiki accurate as your organization keeps changing.

1226
00:42:31,360 --> 00:42:32,640
Once you see it that way,

1227
00:42:32,640 --> 00:42:34,960
breadth handled one way, depth handled another,

1228
00:42:34,960 --> 00:42:37,640
the question stops being which architecture do we pick,

1229
00:42:37,640 --> 00:42:39,640
and starts being something much more practical.

1230
00:42:39,640 --> 00:42:41,600
How does this hybrid framing actually change

1231
00:42:41,600 --> 00:42:45,960
what a rollout looks like inside a real Microsoft 365 environment?

1232
00:42:45,960 --> 00:42:47,760
What hybrid looks like for co-pilot?

1233
00:42:47,760 --> 00:42:49,880
So picture what this actually looks like once it's running,

1234
00:42:49,880 --> 00:42:52,280
not as a diagram, but as something sitting inside your tenant

1235
00:42:52,280 --> 00:42:54,040
doing two different jobs at once.

1236
00:42:54,040 --> 00:42:55,360
Co-pilot with two tiers,

1237
00:42:55,360 --> 00:42:57,680
one tier is the broad, graph grounded retrieval

1238
00:42:57,680 --> 00:42:58,800
you already have today,

1239
00:42:58,800 --> 00:43:00,560
reaching across the long tail of everything

1240
00:43:00,560 --> 00:43:02,840
in SharePoint teams exchange all of it,

1241
00:43:02,840 --> 00:43:04,880
searchable the way it's always been searchable.

1242
00:43:04,880 --> 00:43:06,720
The other tier is a compiled wiki layer,

1243
00:43:06,720 --> 00:43:08,200
sitting quietly underneath,

1244
00:43:08,200 --> 00:43:10,920
built specifically for the topics your organization asks about

1245
00:43:10,920 --> 00:43:11,960
again and again,

1246
00:43:11,960 --> 00:43:14,040
and that second tier only earns its keep

1247
00:43:14,040 --> 00:43:15,440
on specific kinds of content.

1248
00:43:15,440 --> 00:43:17,840
So it's worth naming what actually belongs there,

1249
00:43:17,840 --> 00:43:18,840
onboarding knowledge,

1250
00:43:18,840 --> 00:43:21,000
the stuff every new hire needs and every team ends up

1251
00:43:21,000 --> 00:43:24,360
re-explaning verbally because nobody trusts the search results.

1252
00:43:24,360 --> 00:43:26,520
Product positioning, the language your organization

1253
00:43:26,520 --> 00:43:27,720
has already agreed on,

1254
00:43:27,720 --> 00:43:29,640
that keeps drifting because three different people

1255
00:43:29,640 --> 00:43:31,240
describe it three different ways,

1256
00:43:31,240 --> 00:43:32,600
depending on who's asked.

1257
00:43:32,600 --> 00:43:34,040
Recurring client questions,

1258
00:43:34,040 --> 00:43:36,280
the ones that show up every quarter with new phrasing,

1259
00:43:36,280 --> 00:43:38,080
but the same underlying ask.

1260
00:43:38,080 --> 00:43:39,480
Compliance interpretations,

1261
00:43:39,480 --> 00:43:42,400
where getting it slightly wrong isn't a minor inconvenience.

1262
00:43:42,400 --> 00:43:44,240
Project histories, the decisions and reversals

1263
00:43:44,240 --> 00:43:45,960
that shaped how something got built,

1264
00:43:45,960 --> 00:43:48,280
the kind of context a new team member has no way of finding

1265
00:43:48,280 --> 00:43:50,560
without asking someone who's been there for years.

1266
00:43:50,560 --> 00:43:52,000
Notice what all of those have in common.

1267
00:43:52,000 --> 00:43:53,240
They're exactly the categories

1268
00:43:53,240 --> 00:43:55,440
where consistency and trust matter most,

1269
00:43:55,440 --> 00:43:58,280
the ones we spent real time on earlier in this episode,

1270
00:43:58,280 --> 00:43:59,520
and they're exactly the categories

1271
00:43:59,520 --> 00:44:01,800
where redundant recreation waste the most time,

1272
00:44:01,800 --> 00:44:03,600
people rebuilding something that already exists

1273
00:44:03,600 --> 00:44:04,920
because search let them down once

1274
00:44:04,920 --> 00:44:06,000
and they stop trusting it.

1275
00:44:06,000 --> 00:44:07,040
That's not a coincidence.

1276
00:44:07,040 --> 00:44:09,240
Those two lists overlap almost completely,

1277
00:44:09,240 --> 00:44:12,000
which is precisely why this is where compiling earns its cost

1278
00:44:12,000 --> 00:44:14,600
instead of just adding complexity for its own sake.

1279
00:44:14,600 --> 00:44:16,880
Here's the mechanic and it's worth describing plainly,

1280
00:44:16,880 --> 00:44:19,480
instead of dressing it up as something more dramatic than it is.

1281
00:44:19,480 --> 00:44:21,560
An agent process runs periodically,

1282
00:44:21,560 --> 00:44:23,240
not constantly, not on every query,

1283
00:44:23,240 --> 00:44:24,320
just on a schedule,

1284
00:44:24,320 --> 00:44:26,440
reading through relevant, share point libraries

1285
00:44:26,440 --> 00:44:28,560
and teams content tied to whichever domain

1286
00:44:28,560 --> 00:44:29,880
it's responsible for.

1287
00:44:29,880 --> 00:44:32,240
It compiles what it finds into structured pages,

1288
00:44:32,240 --> 00:44:33,640
updating what already exists,

1289
00:44:33,640 --> 00:44:36,760
creating new pages where something genuinely new shows up.

1290
00:44:36,760 --> 00:44:39,320
And then when someone asks Copilot a question

1291
00:44:39,320 --> 00:44:41,440
that touches one of those compiled topics,

1292
00:44:41,440 --> 00:44:44,280
Copilot's studio surfaces those pages preferentially,

1293
00:44:44,280 --> 00:44:46,720
not exclusively, not instead of graph grounding entirely,

1294
00:44:46,720 --> 00:44:48,320
just weighted higher, trusted more,

1295
00:44:48,320 --> 00:44:49,920
pulled first when the topic matches.

1296
00:44:49,920 --> 00:44:51,440
Worth being clear about what this isn't

1297
00:44:51,440 --> 00:44:53,560
because it's easy to oversell a pattern like this

1298
00:44:53,560 --> 00:44:56,080
once you've spent an episode building the case for it.

1299
00:44:56,080 --> 00:44:57,800
This isn't a rip and replace of Copilot.

1300
00:44:57,800 --> 00:44:59,680
You're not tearing out graph grounding

1301
00:44:59,680 --> 00:45:01,000
and swapping in something else.

1302
00:45:01,000 --> 00:45:02,000
It's an added layer.

1303
00:45:02,000 --> 00:45:04,160
Sitting alongside what already exists,

1304
00:45:04,160 --> 00:45:06,920
changing what grounding actually means specifically

1305
00:45:06,920 --> 00:45:09,640
for the topics your organization keeps asking about.

1306
00:45:09,640 --> 00:45:11,840
Everything else still works the way it always has.

1307
00:45:11,840 --> 00:45:13,640
The long tail is still there, still searchable,

1308
00:45:13,640 --> 00:45:15,280
still handle the old way,

1309
00:45:15,280 --> 00:45:17,960
because most of your content doesn't need this kind of treatment

1310
00:45:17,960 --> 00:45:19,040
and never will.

1311
00:45:19,040 --> 00:45:20,120
But none of this runs itself.

1312
00:45:20,120 --> 00:45:21,560
Somebody has to decide what counts

1313
00:45:21,560 --> 00:45:24,400
as a recurring topic worth compiling in the first place.

1314
00:45:24,400 --> 00:45:26,840
Somebody has to define how those pages get formatted,

1315
00:45:26,840 --> 00:45:28,040
what triggers an update,

1316
00:45:28,040 --> 00:45:29,560
when a contradiction gets flagged

1317
00:45:29,560 --> 00:45:31,360
instead of quietly resolved.

1318
00:45:31,360 --> 00:45:32,480
And that's a governance question.

1319
00:45:32,480 --> 00:45:34,480
One IT leader's can't really skip

1320
00:45:34,480 --> 00:45:37,360
once they've decided this layer is worth building at all.

1321
00:45:37,360 --> 00:45:39,360
The governance layer nobody's talking about.

1322
00:45:39,360 --> 00:45:40,600
Somebody has to own this.

1323
00:45:40,600 --> 00:45:42,920
Not in a vague, it will figure it out sense,

1324
00:45:42,920 --> 00:45:45,840
but a real role with a real name attached to it.

1325
00:45:45,840 --> 00:45:47,480
Someone has to own the schema,

1326
00:45:47,480 --> 00:45:49,400
the actual rules for what gets compiled

1327
00:45:49,400 --> 00:45:51,600
and what doesn't, how a page gets formatted,

1328
00:45:51,600 --> 00:45:54,120
so it matches every other page in the wiki.

1329
00:45:54,120 --> 00:45:56,680
And critically, when a contradiction gets escalated

1330
00:45:56,680 --> 00:45:59,440
to an actual human, instead of just sitting flagged

1331
00:45:59,440 --> 00:46:01,080
in a system nobody's watching.

1332
00:46:01,080 --> 00:46:02,800
Without that owner, the schema drifts,

1333
00:46:02,800 --> 00:46:05,480
the pages get inconsistent and the whole thing slowly turns

1334
00:46:05,480 --> 00:46:08,120
into exactly the mess it was supposed to replace.

1335
00:46:08,120 --> 00:46:10,320
Here's the good news, and it's worth saying plainly

1336
00:46:10,320 --> 00:46:12,680
because this part tends to get overcomplicated.

1337
00:46:12,680 --> 00:46:14,440
This role isn't new, it maps directly

1338
00:46:14,440 --> 00:46:16,040
onto jobs that already exist inside

1339
00:46:16,040 --> 00:46:18,320
most Microsoft 365 environments.

1340
00:46:18,320 --> 00:46:19,800
Information architects already think

1341
00:46:19,800 --> 00:46:21,560
about how content gets structured.

1342
00:46:21,560 --> 00:46:24,000
Records managers already think about lifecycle,

1343
00:46:24,000 --> 00:46:25,760
about what stays, what gets retired,

1344
00:46:25,760 --> 00:46:27,160
what needs a review cycle.

1345
00:46:27,160 --> 00:46:29,320
Content owners already carry responsibility

1346
00:46:29,320 --> 00:46:30,520
for whether something's accurate.

1347
00:46:30,520 --> 00:46:32,280
None of these people need a new job title.

1348
00:46:32,280 --> 00:46:34,160
They need this added to a job they're already doing

1349
00:46:34,160 --> 00:46:36,400
because the skill underneath it is identical.

1350
00:46:36,400 --> 00:46:39,640
Deciding what belongs where and who's accountable when it's wrong.

1351
00:46:39,640 --> 00:46:41,400
There's already a preview of where this is heading

1352
00:46:41,400 --> 00:46:43,120
and it's sitting inside Pervue right now.

1353
00:46:43,120 --> 00:46:44,840
The data loss prevention controls,

1354
00:46:44,840 --> 00:46:46,920
Microsoft built for co-pilot's web searches,

1355
00:46:46,920 --> 00:46:49,880
the ones that stop sensitive data from leaking into a prompt,

1356
00:46:49,880 --> 00:46:52,240
while still letting co-pilot ground its answer

1357
00:46:52,240 --> 00:46:54,640
in internal sources, that's the same instinct

1358
00:46:54,640 --> 00:46:56,760
applied to a narrower problem.

1359
00:46:56,760 --> 00:46:59,640
Guard rails around what an AI-driven system is allowed to see

1360
00:46:59,640 --> 00:47:01,800
and by extension, what it's allowed to compile.

1361
00:47:01,800 --> 00:47:03,400
Once you're running an agent that reads

1362
00:47:03,400 --> 00:47:06,240
across your SharePoint libraries and writes structured pages,

1363
00:47:06,240 --> 00:47:08,000
you need the same kind of thinking

1364
00:47:08,000 --> 00:47:10,480
just aimed at compilation instead of search.

1365
00:47:10,480 --> 00:47:13,280
So frame this the right way because how you frame it changes

1366
00:47:13,280 --> 00:47:14,840
whether it actually gets funded.

1367
00:47:14,840 --> 00:47:18,320
This isn't a brand new discipline landing on IT's desk out of nowhere.

1368
00:47:18,320 --> 00:47:19,840
It's an extension of governance work

1369
00:47:19,840 --> 00:47:22,120
that's probably already underway in some form.

1370
00:47:22,120 --> 00:47:24,760
If your organization already has content lifecycle policies,

1371
00:47:24,760 --> 00:47:26,600
retention rules, ownership models,

1372
00:47:26,600 --> 00:47:28,880
you already have the starting point for a Wiki schema.

1373
00:47:28,880 --> 00:47:30,440
You're not building from zero.

1374
00:47:30,440 --> 00:47:32,320
You're extending something that exists.

1375
00:47:32,320 --> 00:47:34,640
And this is really the dividing line between organizations

1376
00:47:34,640 --> 00:47:36,320
that get real value out of this pattern

1377
00:47:36,320 --> 00:47:38,600
and organizations that build something impressive looking

1378
00:47:38,600 --> 00:47:40,880
and then watch it decay within a year.

1379
00:47:40,880 --> 00:47:42,920
The ones that benefit most treat this

1380
00:47:42,920 --> 00:47:44,960
as a knowledge management initiative

1381
00:47:44,960 --> 00:47:46,880
that happens to have an AI component,

1382
00:47:46,880 --> 00:47:49,680
not an AI initiative that happens to touch knowledge,

1383
00:47:49,680 --> 00:47:52,200
that ordering matters more than it sounds like it should.

1384
00:47:52,200 --> 00:47:55,200
One puts governance first and lets the technology serve it.

1385
00:47:55,200 --> 00:47:56,720
The other bolts governance on afterward

1386
00:47:56,720 --> 00:47:58,360
once something's already gone wrong.

1387
00:47:58,360 --> 00:48:00,720
With the mechanics covered and the governance question

1388
00:48:00,720 --> 00:48:02,240
at least named honestly,

1389
00:48:02,240 --> 00:48:04,040
it's worth stepping back from the how

1390
00:48:04,040 --> 00:48:07,760
and naming what's actually different here underneath all of it.

1391
00:48:07,760 --> 00:48:09,080
What actually changes?

1392
00:48:09,080 --> 00:48:12,520
So here's the shift stated as plainly as it deserves to be stated.

1393
00:48:12,520 --> 00:48:15,280
Copilot stops being a search box with better manners,

1394
00:48:15,280 --> 00:48:18,200
one that phrases things politely and summarizes nicely,

1395
00:48:18,200 --> 00:48:19,920
but still starts from zero every time

1396
00:48:19,920 --> 00:48:22,280
and starts becoming a system that actually has

1397
00:48:22,280 --> 00:48:23,720
somewhere to put what it learns.

1398
00:48:23,720 --> 00:48:25,680
That's the whole difference, not a smarter model,

1399
00:48:25,680 --> 00:48:27,880
not a better prompt, a place for knowledge to live

1400
00:48:27,880 --> 00:48:29,760
once it's been figured out instead of a mechanism

1401
00:48:29,760 --> 00:48:31,760
that refigures it out on command forever.

1402
00:48:31,760 --> 00:48:33,560
And that difference doesn't show up on day one.

1403
00:48:33,560 --> 00:48:34,960
It shows up over time,

1404
00:48:34,960 --> 00:48:38,560
which is exactly why so many organizations miss it during a pilot.

1405
00:48:38,560 --> 00:48:40,960
A stateless system is the same tool on day 500

1406
00:48:40,960 --> 00:48:42,400
as it was on day one.

1407
00:48:42,400 --> 00:48:45,480
Ask at the same category of question a year from now

1408
00:48:45,480 --> 00:48:47,320
and it goes through the identical motions

1409
00:48:47,320 --> 00:48:48,800
it went through in week one,

1410
00:48:48,800 --> 00:48:53,160
search, chunk, retrieve, synthesize, forget.

1411
00:48:53,160 --> 00:48:55,160
Nothing about that process improves with age,

1412
00:48:55,160 --> 00:48:57,600
it doesn't get faster, it doesn't get more accurate,

1413
00:48:57,600 --> 00:48:58,600
it just repeats.

1414
00:48:58,600 --> 00:49:00,960
A compiled system by contrast is measurably different

1415
00:49:00,960 --> 00:49:04,120
a year in because every project page, every client history,

1416
00:49:04,120 --> 00:49:06,200
every resolved contradiction from month three

1417
00:49:06,200 --> 00:49:08,320
is still sitting there in month 14,

1418
00:49:08,320 --> 00:49:10,840
ready to be built on instead of rediscovered.

1419
00:49:10,840 --> 00:49:12,120
That's not a theoretical claim.

1420
00:49:12,120 --> 00:49:14,600
It shows up in the numbers we've already touched on earlier

1421
00:49:14,600 --> 00:49:16,640
in this episode around structured knowledge

1422
00:49:16,640 --> 00:49:19,520
and it's worth tying back to the sharpest version of them.

1423
00:49:19,520 --> 00:49:22,120
Enterprise knowledge graphs and structured content approaches

1424
00:49:22,120 --> 00:49:25,720
have delivered results like 287% ROI over three years

1425
00:49:25,720 --> 00:49:28,040
with multi million dollar annual benefits reported

1426
00:49:28,040 --> 00:49:29,440
in some organizations.

1427
00:49:29,440 --> 00:49:30,960
Those aren't numbers from a system

1428
00:49:30,960 --> 00:49:32,280
that got asked more questions,

1429
00:49:32,280 --> 00:49:34,320
they're numbers from a system that got smarter

1430
00:49:34,320 --> 00:49:35,680
at answering the same questions

1431
00:49:35,680 --> 00:49:37,840
because the underlying knowledge kept compounding

1432
00:49:37,840 --> 00:49:39,080
instead of resetting.

1433
00:49:39,080 --> 00:49:41,440
And that's really the mechanism worth naming directly

1434
00:49:41,440 --> 00:49:43,720
because it's easy to miss if you're only looking

1435
00:49:43,720 --> 00:49:45,160
at usage dashboards.

1436
00:49:45,160 --> 00:49:47,120
Compounding knowledge, not repeated search,

1437
00:49:47,120 --> 00:49:49,160
is what actually saves time at scale.

1438
00:49:49,160 --> 00:49:52,080
Repeated search just means people are using the tool more often.

1439
00:49:52,080 --> 00:49:54,600
It says nothing about whether the tool is getting better.

1440
00:49:54,600 --> 00:49:56,600
Compounding knowledge means the thousandth question

1441
00:49:56,600 --> 00:49:58,720
benefits from everything the system learned

1442
00:49:58,720 --> 00:50:00,680
answering the first 999.

1443
00:50:00,680 --> 00:50:02,160
That's a completely different kind of value

1444
00:50:02,160 --> 00:50:04,120
and it's the kind drag structurally can't produce

1445
00:50:04,120 --> 00:50:05,440
no matter how much you use it.

1446
00:50:05,440 --> 00:50:07,000
So say the quiet part plainly

1447
00:50:07,000 --> 00:50:08,800
because it's the sentence this whole episode

1448
00:50:08,800 --> 00:50:10,160
has been building toward.

1449
00:50:10,160 --> 00:50:12,760
An assistant that gets smarter is a fundamentally different

1450
00:50:12,760 --> 00:50:15,080
asset than one that just gets asked more questions.

1451
00:50:15,080 --> 00:50:17,720
One compounds the other just accumulates traffic.

1452
00:50:17,720 --> 00:50:20,240
Confusing the two is exactly how organizations end up

1453
00:50:20,240 --> 00:50:21,920
with impressive adoption metrics

1454
00:50:21,920 --> 00:50:24,320
and a tool that still can't hold a coherent memory

1455
00:50:24,320 --> 00:50:26,640
of its own history, which naturally raises

1456
00:50:26,640 --> 00:50:28,440
the only question left worth answering.

1457
00:50:28,440 --> 00:50:29,760
Not whether this is worth doing

1458
00:50:29,760 --> 00:50:31,680
but where an organization actually starts,

1459
00:50:31,680 --> 00:50:33,680
small and deliberate instead of trying to compile

1460
00:50:33,680 --> 00:50:34,960
everything at once.

1461
00:50:34,960 --> 00:50:36,120
What if you start today?

1462
00:50:36,120 --> 00:50:38,120
So pick one thing, not the whole knowledge estate,

1463
00:50:38,120 --> 00:50:41,120
not every department, one recurring high value domain.

1464
00:50:41,120 --> 00:50:42,480
Maybe it's a single product line

1465
00:50:42,480 --> 00:50:44,120
that support keeps asking about.

1466
00:50:44,120 --> 00:50:46,000
Maybe it's one recurring client type,

1467
00:50:46,000 --> 00:50:47,640
the kind of account whose questions repeat

1468
00:50:47,640 --> 00:50:49,480
every quarter with new phrasing.

1469
00:50:49,480 --> 00:50:51,240
Maybe it's a single compliance area

1470
00:50:51,240 --> 00:50:53,440
where getting the answer wrong actually costs something.

1471
00:50:53,440 --> 00:50:55,320
The domain doesn't matter as much as the discipline

1472
00:50:55,320 --> 00:50:56,360
of picking just one.

1473
00:50:56,360 --> 00:50:58,800
Here's what a first 90 days actually looks like

1474
00:50:58,800 --> 00:51:01,120
without dressing it up as bigger than it is.

1475
00:51:01,120 --> 00:51:03,080
Start by identifying the raw sources,

1476
00:51:03,080 --> 00:51:04,720
whatever SharePoint libraries,

1477
00:51:04,720 --> 00:51:06,600
Teams, Threads and old policy docs

1478
00:51:06,600 --> 00:51:09,320
already touch this domain, then define a schema

1479
00:51:09,320 --> 00:51:11,080
and keep it lightweight on purpose.

1480
00:51:11,080 --> 00:51:13,400
Not a governance framework spanning 50 pages,

1481
00:51:13,400 --> 00:51:15,480
just enough rules to say what a page looks like

1482
00:51:15,480 --> 00:51:17,200
and how contradictions get flagged,

1483
00:51:17,200 --> 00:51:19,480
run an initial compilation pass against that schema.

1484
00:51:19,480 --> 00:51:21,320
Then test it against real questions,

1485
00:51:21,320 --> 00:51:23,000
phrase the way actual people phrase them,

1486
00:51:23,000 --> 00:51:25,240
not the clean version you'd write in a demo script.

1487
00:51:25,240 --> 00:51:27,720
Picture the end of that pilot next to where you started.

1488
00:51:27,720 --> 00:51:30,600
Before, co-pilot retrieves five slightly different chunks

1489
00:51:30,600 --> 00:51:32,640
about the same policy and depending on

1490
00:51:32,640 --> 00:51:34,240
how someone phrased the question,

1491
00:51:34,240 --> 00:51:36,240
they get a different one of those five.

1492
00:51:36,240 --> 00:51:39,000
After, it answers from one compiled current page

1493
00:51:39,000 --> 00:51:41,120
every time, no matter how the question gets asked.

1494
00:51:41,120 --> 00:51:43,920
That's the whole test, not whether the answer sounds smart,

1495
00:51:43,920 --> 00:51:45,560
whether it's the same answer twice,

1496
00:51:45,560 --> 00:51:47,200
worth connecting this back to the research

1497
00:51:47,200 --> 00:51:48,600
we've been leaning on all episode

1498
00:51:48,600 --> 00:51:50,720
because it matters that this isn't theoretical.

1499
00:51:50,720 --> 00:51:53,560
The organization's reporting a 25 to 40% reduction

1500
00:51:53,560 --> 00:51:55,000
in redundant knowledge creation,

1501
00:51:55,000 --> 00:51:57,280
the ones seeing meaningfully lower error rates.

1502
00:51:57,280 --> 00:51:59,440
They didn't get there by rebuilding everything at once.

1503
00:51:59,440 --> 00:52:02,040
They got their one domain at a time, exactly like this.

1504
00:52:02,040 --> 00:52:05,160
Small-bounded, tested against real use before expanding,

1505
00:52:05,160 --> 00:52:06,840
and there's a thread here worth planting

1506
00:52:06,840 --> 00:52:08,680
without pulling on a too hard right now.

1507
00:52:08,680 --> 00:52:09,840
The moment you run this pilot,

1508
00:52:09,840 --> 00:52:12,680
you're going to hit real questions about who owns the schema,

1509
00:52:12,680 --> 00:52:14,840
who decides when a contradiction gets escalated,

1510
00:52:14,840 --> 00:52:17,160
who's accountable when a page goes stale.

1511
00:52:17,160 --> 00:52:18,120
That's governance.

1512
00:52:18,120 --> 00:52:21,240
And it's worth its own conversation rather than a rushed afterthought here.

1513
00:52:21,240 --> 00:52:23,760
So that's the technical case and the strategic case,

1514
00:52:23,760 --> 00:52:26,880
sitting next to each other, time to bring them together.

1515
00:52:26,880 --> 00:52:28,080
Key transformation.

1516
00:52:28,080 --> 00:52:31,320
Here's the transformation, stated as plainly as it deserves.

1517
00:52:31,320 --> 00:52:34,320
Copilot's value ceiling isn't set by the model underneath it.

1518
00:52:34,320 --> 00:52:37,440
It's set by whether your organization's knowledge gets compiled

1519
00:52:37,440 --> 00:52:40,120
or just searched over and over forever.

1520
00:52:40,120 --> 00:52:43,320
Rags will keep giving you fast, plausible, forgettable answers.

1521
00:52:43,320 --> 00:52:44,560
So that's what it's built to do,

1522
00:52:44,560 --> 00:52:45,720
and it'll keep doing it well.

1523
00:52:45,720 --> 00:52:48,480
A wiki layer gives you something structurally different,

1524
00:52:48,480 --> 00:52:51,880
an assistant that remembers what your organization has already figured out

1525
00:52:51,880 --> 00:52:54,800
instead of refiguring it out every time someone asks.

1526
00:52:54,800 --> 00:52:57,400
And here's the part worth sitting with on your way out of this episode.

1527
00:52:57,400 --> 00:53:00,920
The tools inside Microsoft 365 to build this exist today.

1528
00:53:00,920 --> 00:53:01,720
Connectors.

1529
00:53:01,720 --> 00:53:02,960
Scheduled flows.

1530
00:53:02,960 --> 00:53:04,320
Copilot studio agents.

1531
00:53:04,320 --> 00:53:05,800
All of it's already sitting in your tenant.

1532
00:53:05,800 --> 00:53:08,040
They're just not assembled this way by default.

1533
00:53:08,040 --> 00:53:09,720
So here's your challenge for this month.

1534
00:53:09,720 --> 00:53:11,760
Pick one recurring knowledge domain,

1535
00:53:11,760 --> 00:53:14,880
and go find out honestly whether Copilot is searching it

1536
00:53:14,880 --> 00:53:17,600
or actually compiling it, not what the vendor deck says,

1537
00:53:17,600 --> 00:53:19,280
what actually happens when someone asks.

1538
00:53:19,280 --> 00:53:21,920
Here's the test, and it's simple enough to run this week.

1539
00:53:21,920 --> 00:53:24,320
Ask the same question two or three different ways.

1540
00:53:24,320 --> 00:53:26,680
If the answer changes depending on how it's phrased,

1541
00:53:26,680 --> 00:53:29,480
you've found a retrieval problem, not a copilot problem.

1542
00:53:29,480 --> 00:53:31,280
That distinction is the whole episode,

1543
00:53:31,280 --> 00:53:33,240
boiled down to one homework assignment.

1544
00:53:33,240 --> 00:53:35,720
If this changed how you think about what's actually running

1545
00:53:35,720 --> 00:53:37,280
under your copilot deployment,

1546
00:53:37,280 --> 00:53:39,640
subscribe to M365FM podcast.

1547
00:53:39,640 --> 00:53:42,000
Next time we're going deeper into the governance side of this,

1548
00:53:42,000 --> 00:53:44,120
who owns the schema, who owns the wiki,

1549
00:53:44,120 --> 00:53:46,240
and what happens when that ownership isn't clear.

1550
00:53:46,240 --> 00:53:49,360
And if you want more people, more IT pros, more decision makers,

1551
00:53:49,360 --> 00:53:52,520
to run into this conversation about fixing Copilot's memory problem,

1552
00:53:52,520 --> 00:53:53,520
leave a review.

1553
00:53:53,520 --> 00:53:56,200
It's genuinely how this finds the people who need it.

1554
00:53:56,200 --> 00:53:59,000
One last thing, connect with me, Mirko Peters on LinkedIn.

1555
00:53:59,000 --> 00:54:01,800
I want to know what recurring knowledge domain you'd compile first

1556
00:54:01,800 --> 00:54:04,480
and what you think the next episode should cover.