Your SharePoint Data is a Liability: Fixing the Metadata Gap

1
00:00:00,000 --> 00:00:02,560
You've been organizing SharePoint like it's 2015,

2
00:00:02,560 --> 00:00:05,480
and that filing logic is about to become a liability,

3
00:00:05,480 --> 00:00:07,400
but in reality, it's not a storage problem.

4
00:00:07,400 --> 00:00:09,720
It's a structural failure that kills search,

5
00:00:09,720 --> 00:00:13,120
breaks compliance, and undermines every co-pilot rollout.

6
00:00:13,120 --> 00:00:15,640
Industry audits consistently find the same pattern.

7
00:00:15,640 --> 00:00:17,360
Organizations only discover this gap

8
00:00:17,360 --> 00:00:19,920
after a failed audit or a co-pilot hallucination.

9
00:00:19,920 --> 00:00:22,880
Custom AI document models close that gap automatically,

10
00:00:22,880 --> 00:00:24,960
and in 2026, the tool that does it

11
00:00:24,960 --> 00:00:27,840
just got renamed for the fourth time in six years.

12
00:00:27,840 --> 00:00:29,720
Your SharePoint data is a liability.

13
00:00:29,720 --> 00:00:31,760
Your IT team probably thinks your SharePoint estate

14
00:00:31,760 --> 00:00:35,440
has a storage problem, more files, more sites, more noise.

15
00:00:35,440 --> 00:00:37,480
But that diagnosis misses the actual disease.

16
00:00:37,480 --> 00:00:39,200
The real problem is that most of those files

17
00:00:39,200 --> 00:00:41,000
carry no business metadata at all.

18
00:00:41,000 --> 00:00:43,280
They're unstructured, untagged, and invisible

19
00:00:43,280 --> 00:00:45,480
to every system that needs to act on them.

20
00:00:45,480 --> 00:00:47,880
Organizations create over two million SharePoint sites

21
00:00:47,880 --> 00:00:49,040
every single day.

22
00:00:49,040 --> 00:00:51,120
They upload more than two billion files daily,

23
00:00:51,120 --> 00:00:53,160
and most of those uploads land with nothing more

24
00:00:53,160 --> 00:00:54,960
than a file name and a timestamp.

25
00:00:54,960 --> 00:00:57,120
No contract type, no expiration date,

26
00:00:57,120 --> 00:00:59,800
no sensitivity label, no retention category,

27
00:00:59,800 --> 00:01:01,600
just a document sitting in a folder,

28
00:01:01,600 --> 00:01:04,360
hoping a human remembers what it is and where it belongs.

29
00:01:04,360 --> 00:01:05,280
That isn't messy.

30
00:01:05,280 --> 00:01:07,360
That's a liability hiding in plain sight.

31
00:01:07,360 --> 00:01:09,080
When documents lack consistent metadata,

32
00:01:09,080 --> 00:01:11,800
employees spend hours hunting for the latest approved version

33
00:01:11,800 --> 00:01:14,680
or recreating information they cannot find.

34
00:01:14,680 --> 00:01:16,080
Analysts have measured this friction

35
00:01:16,080 --> 00:01:17,880
as a direct productivity loss.

36
00:01:17,880 --> 00:01:19,440
In distributed organizations,

37
00:01:19,440 --> 00:01:21,400
multiple teams maintain parallel versions

38
00:01:21,400 --> 00:01:23,280
of the same materials in siloed sites

39
00:01:23,280 --> 00:01:25,000
or personal one-drive accounts.

40
00:01:25,000 --> 00:01:26,720
The same contract lives in three places,

41
00:01:26,720 --> 00:01:27,960
each with different edits,

42
00:01:27,960 --> 00:01:29,960
and nobody knows which one is authoritative.

43
00:01:29,960 --> 00:01:31,640
Somebody rewrites the work from scratch

44
00:01:31,640 --> 00:01:33,120
because they couldn't find the original.

45
00:01:33,120 --> 00:01:35,480
That's not laziness, that's a metadata failure.

46
00:01:35,480 --> 00:01:38,120
The folder-based mental model is the root cause.

47
00:01:38,120 --> 00:01:40,520
Folders assume that people know what they're looking for

48
00:01:40,520 --> 00:01:41,640
and where it lives.

49
00:01:41,640 --> 00:01:43,160
They assume a single hierarchy

50
00:01:43,160 --> 00:01:44,680
can represent every possible way

51
00:01:44,680 --> 00:01:46,360
a document might be discovered.

52
00:01:46,360 --> 00:01:48,520
But modern work doesn't start with navigation.

53
00:01:48,520 --> 00:01:49,760
It starts with context.

54
00:01:49,760 --> 00:01:51,720
A project manager needs every contract

55
00:01:51,720 --> 00:01:53,440
that expires in the next 90 days,

56
00:01:53,440 --> 00:01:55,000
regardless of which site stores it.

57
00:01:55,000 --> 00:01:57,200
A legal team needs every document containing

58
00:01:57,200 --> 00:01:59,920
a liability clause above a specific dollar threshold

59
00:01:59,920 --> 00:02:01,720
without opening each file manually.

60
00:02:01,720 --> 00:02:04,280
Folders can't answer those questions metadata can.

61
00:02:04,280 --> 00:02:07,320
But the problem runs deeper than lost time and duplicated work.

62
00:02:07,320 --> 00:02:10,560
Microsoft 365 co-pilot and other generative AI tools

63
00:02:10,560 --> 00:02:12,440
rely on underlying content signals

64
00:02:12,440 --> 00:02:14,720
to return relevant trustworthy responses

65
00:02:14,720 --> 00:02:16,440
when documents like accurate metadata

66
00:02:16,440 --> 00:02:18,280
or live in haphazard locations,

67
00:02:18,280 --> 00:02:20,960
semantic search and AI-driven retrieval struggle

68
00:02:20,960 --> 00:02:23,160
to surface the right context.

69
00:02:23,160 --> 00:02:25,400
Every missing or incorrect metadata field

70
00:02:25,400 --> 00:02:27,560
degrades the signal that AI systems use

71
00:02:27,560 --> 00:02:29,000
to reason about your content.

72
00:02:29,000 --> 00:02:31,520
Your co-pilot investment becomes inconsistent

73
00:02:31,520 --> 00:02:32,800
and difficult to trust,

74
00:02:32,800 --> 00:02:34,280
not because the model is broken

75
00:02:34,280 --> 00:02:35,800
but because the ground beneath it is.

76
00:02:35,800 --> 00:02:38,280
The metadata gap is also a genuine compliance

77
00:02:38,280 --> 00:02:39,360
and risk concern.

78
00:02:39,360 --> 00:02:41,800
Regulators and courts expect defensible retention,

79
00:02:41,800 --> 00:02:44,280
disposition and e-discovery practices.

80
00:02:44,280 --> 00:02:46,440
Retention policies and labels in Microsoft Perview

81
00:02:46,440 --> 00:02:48,600
are designed to help organizations retain content

82
00:02:48,600 --> 00:02:51,360
for required periods or deleted when it is no longer needed.

83
00:02:51,360 --> 00:02:53,560
But these controls rely on being able to identify

84
00:02:53,560 --> 00:02:56,160
which content falls into which category.

85
00:02:56,160 --> 00:02:57,640
When documents lack classification

86
00:02:57,640 --> 00:02:59,640
that differentiates records from working copies

87
00:02:59,640 --> 00:03:01,880
or regulated data from non-critical content,

88
00:03:01,880 --> 00:03:04,520
retention settings may be applied broadly or not at all.

89
00:03:04,520 --> 00:03:06,280
That creates both over retention risk

90
00:03:06,280 --> 00:03:08,160
and premature deletion risk.

91
00:03:08,160 --> 00:03:10,080
In e-discovery and litigation holds,

92
00:03:10,080 --> 00:03:12,400
metadata becomes even more critical.

93
00:03:12,400 --> 00:03:15,200
Metadata provides creation and modification timestamps,

94
00:03:15,200 --> 00:03:17,600
authorship, file paths and message headers,

95
00:03:17,600 --> 00:03:19,840
all of which are needed to reconstruct events

96
00:03:19,840 --> 00:03:22,240
and demonstrate chain of custody.

97
00:03:22,240 --> 00:03:24,360
Best practices for defensible e-discovery

98
00:03:24,360 --> 00:03:27,720
emphasize preserving metadata using forensic grade tools,

99
00:03:27,720 --> 00:03:29,360
documenting the chain of custody

100
00:03:29,360 --> 00:03:32,440
and auditing metadata before and after processing.

101
00:03:32,440 --> 00:03:34,880
If SharePoint content spreads across uncontrolled sites,

102
00:03:34,880 --> 00:03:37,920
personal one drive locations and shadow repositories.

103
00:03:37,920 --> 00:03:40,600
And if the metadata is inconsistent or incomplete,

104
00:03:40,600 --> 00:03:43,360
legal teams may struggle to identify, preserve

105
00:03:43,360 --> 00:03:46,520
and produce relevant materials in a defensible manner.

106
00:03:46,520 --> 00:03:49,240
Security and privacy risks are similarly amplified.

107
00:03:49,240 --> 00:03:52,280
SharePoint and one drive provide strong security capabilities,

108
00:03:52,280 --> 00:03:54,560
including encryption in transit and at rest,

109
00:03:54,560 --> 00:03:56,640
two-factor authentication, conditional access

110
00:03:56,640 --> 00:03:58,400
and data loss prevention policies.

111
00:03:58,400 --> 00:04:00,360
However, DLP rules, sensitivity labels

112
00:04:00,360 --> 00:04:02,720
and conditional access rely on understanding

113
00:04:02,720 --> 00:04:04,920
which documents contain sensitive information.

114
00:04:04,920 --> 00:04:06,480
Without reliable classification,

115
00:04:06,480 --> 00:04:09,160
organizations may either over-restrict access

116
00:04:09,160 --> 00:04:12,640
and hamper collaboration or under-protect sensitive documents

117
00:04:12,640 --> 00:04:14,560
and lead to accidental exposure.

118
00:04:14,560 --> 00:04:16,400
In an environment with weak metadata,

119
00:04:16,400 --> 00:04:18,760
even powerful tooling cannot reliably enforce

120
00:04:18,760 --> 00:04:20,120
the principle of least privilege.

121
00:04:20,120 --> 00:04:21,880
Hybrid work has fundamentally altered

122
00:04:21,880 --> 00:04:24,200
how teams create and share information

123
00:04:24,200 --> 00:04:26,480
and traditional folder hierarchies no longer align

124
00:04:26,480 --> 00:04:29,120
with how people work in teams, SharePoint

125
00:04:29,120 --> 00:04:31,240
and other collaborative tools.

126
00:04:31,240 --> 00:04:33,720
Modern collaboration centers on projects, channels

127
00:04:33,720 --> 00:04:36,320
or topics that cut across organizational structures,

128
00:04:36,320 --> 00:04:38,640
leading to documents stored in multiple contexts

129
00:04:38,640 --> 00:04:40,040
simultaneously.

130
00:04:40,040 --> 00:04:42,120
Users expect to find information through search,

131
00:04:42,120 --> 00:04:44,160
recommendations or AI assistance,

132
00:04:44,160 --> 00:04:46,960
rather than working through deeply nested folder trees.

133
00:04:46,960 --> 00:04:48,720
This shift exposes the limitations

134
00:04:48,720 --> 00:04:50,800
of purely location-based organization

135
00:04:50,800 --> 00:04:54,160
and shows why content must carry rich descriptive metadata

136
00:04:54,160 --> 00:04:56,640
that allows discovery and reuse regardless

137
00:04:56,640 --> 00:04:58,880
of where it physically resides.

138
00:04:58,880 --> 00:05:02,120
At the same time, AI is becoming embedded in everyday work

139
00:05:02,120 --> 00:05:05,600
via co-pilot experiences in Microsoft 365,

140
00:05:05,600 --> 00:05:07,560
which use an intelligence layer sometimes

141
00:05:07,560 --> 00:05:10,000
described as work IQ to connect individual

142
00:05:10,000 --> 00:05:11,640
and organizational knowledge.

143
00:05:11,640 --> 00:05:14,560
This intelligence layer leverages existing metadata,

144
00:05:14,560 --> 00:05:17,800
signals and user interactions to provide relevant suggestions,

145
00:05:17,800 --> 00:05:19,080
summaries and answers.

146
00:05:19,080 --> 00:05:21,240
If that underlying content is poorly governed

147
00:05:21,240 --> 00:05:23,920
with inconsistent labels or incomplete metadata,

148
00:05:23,920 --> 00:05:27,280
AI systems may surface obsolete or inappropriate documents,

149
00:05:27,280 --> 00:05:29,680
undermining user trust and potentially exposing

150
00:05:29,680 --> 00:05:30,880
sensitive information.

151
00:05:30,880 --> 00:05:32,720
The metadata gap becomes a barrier

152
00:05:32,720 --> 00:05:36,160
to realizing the promise of AI accelerated hybrid work,

153
00:05:36,160 --> 00:05:39,240
reinforcing the argument that unmanaged SharePoint data

154
00:05:39,240 --> 00:05:41,280
represents not only a latent liability,

155
00:05:41,280 --> 00:05:43,040
but also an opportunity cost.

156
00:05:43,040 --> 00:05:45,200
And one level deeper, the problem isn't even the absence

157
00:05:45,200 --> 00:05:45,960
of metadata.

158
00:05:45,960 --> 00:05:47,800
It's the absence of a metadata strategy.

159
00:05:47,800 --> 00:05:49,720
Most SharePoint environments have columns.

160
00:05:49,720 --> 00:05:51,080
They have content types.

161
00:05:51,080 --> 00:05:53,720
They have term sets that somebody built three years ago

162
00:05:53,720 --> 00:05:54,760
and abandoned.

163
00:05:54,760 --> 00:05:57,160
What they don't have is a living governed business-aligned

164
00:05:57,160 --> 00:06:00,040
taxonomy that connects document content to automated action.

165
00:06:00,040 --> 00:06:03,400
Without that strategy, any AI tool you deploy will be guessing.

166
00:06:03,400 --> 00:06:05,240
And guessing at scale is expensive.

167
00:06:05,240 --> 00:06:07,960
Consider what happens when a new employee joins your legal team.

168
00:06:07,960 --> 00:06:10,760
They need to find every active contract with a specific vendor.

169
00:06:10,760 --> 00:06:12,760
They open SharePoint, go to the legal site

170
00:06:12,760 --> 00:06:15,400
and find a folder called Contracts2024.

171
00:06:15,400 --> 00:06:17,760
Inside that folder are subfolders for each quarter.

172
00:06:17,760 --> 00:06:21,440
Inside Q2 are files named ContractFinalV2, DockerX,

173
00:06:21,440 --> 00:06:25,280
ContractFinalV2 are Mike, DockerX, and ContractFinalScient,

174
00:06:25,280 --> 00:06:26,400
at PDF.

175
00:06:26,400 --> 00:06:29,080
None of these files have metadata columns populated.

176
00:06:29,080 --> 00:06:32,640
The new employee opens each one, reads through pages of legal text,

177
00:06:32,640 --> 00:06:34,560
and manually compiles a list.

178
00:06:34,560 --> 00:06:36,200
This process takes four hours.

179
00:06:36,200 --> 00:06:37,440
It should take four minutes.

180
00:06:37,440 --> 00:06:40,080
The same scenario plays out in finance when month and close

181
00:06:40,080 --> 00:06:43,000
requires reconciling invoices against purchase orders.

182
00:06:43,000 --> 00:06:45,640
The accounts payable team searches through email attachments,

183
00:06:45,640 --> 00:06:47,560
teams chats, and personal one-drive folders

184
00:06:47,560 --> 00:06:50,920
because the SharePoint library has no vendor name, no PO number,

185
00:06:50,920 --> 00:06:53,080
and no approval status in the metadata.

186
00:06:53,080 --> 00:06:55,680
They open each PDF, read the header, and copy numbers

187
00:06:55,680 --> 00:06:56,560
into a spreadsheet.

188
00:06:56,560 --> 00:06:58,920
The process is not just slow, it is fragile.

189
00:06:58,920 --> 00:07:01,280
One wrong copy paste creates a reconciliation error

190
00:07:01,280 --> 00:07:03,000
that takes days to untangle.

191
00:07:03,000 --> 00:07:04,920
IT teams often respond to these complaints

192
00:07:04,920 --> 00:07:06,960
by adding more storage or faster search,

193
00:07:06,960 --> 00:07:08,920
but faster search of bad metadata still

194
00:07:08,920 --> 00:07:10,840
returns bad results.

195
00:07:10,840 --> 00:07:13,720
A search for Contract in a 10 terabyte SharePoint estate

196
00:07:13,720 --> 00:07:16,640
returns 10,000 files, most of them irrelevant.

197
00:07:16,640 --> 00:07:19,680
The user adds more keywords, narrows the date range,

198
00:07:19,680 --> 00:07:22,280
and still cannot find the specific document they need.

199
00:07:22,280 --> 00:07:24,480
Search relevance depends on metadata density.

200
00:07:24,480 --> 00:07:27,960
Without metadata, search is just a faster way to browse garbage,

201
00:07:27,960 --> 00:07:29,440
the cost compounds over time.

202
00:07:29,440 --> 00:07:31,400
A contract that auto-renews because nobody

203
00:07:31,400 --> 00:07:33,640
tracked the expiration date costs real money,

204
00:07:33,640 --> 00:07:36,680
a sensitive document that leaks because it was never classified,

205
00:07:36,680 --> 00:07:39,080
costs reputation and regulatory fines.

206
00:07:39,080 --> 00:07:41,360
A co-pilot responds that sites in obsolete policy

207
00:07:41,360 --> 00:07:44,560
because the current version lacked metadata costs user trust.

208
00:07:44,560 --> 00:07:45,880
These are not edge cases.

209
00:07:45,880 --> 00:07:48,200
They are the daily operating reality of organizations

210
00:07:48,200 --> 00:07:49,920
that treat SharePoint as a filing cabinet

211
00:07:49,920 --> 00:07:51,120
instead of a knowledge system.

212
00:07:51,120 --> 00:07:54,160
The opportunity cost is harder to measure, but just as real.

213
00:07:54,160 --> 00:07:55,920
Organizations that govern their metadata

214
00:07:55,920 --> 00:07:58,520
can query their document estate like a database.

215
00:07:58,520 --> 00:08:01,200
They can ask which vendors represent the highest liability

216
00:08:01,200 --> 00:08:03,680
exposure, they can ask which contracts expire

217
00:08:03,680 --> 00:08:06,280
in the next quarter, and require renewal decisions.

218
00:08:06,280 --> 00:08:08,960
They can ask which policies have not been reviewed in two years

219
00:08:08,960 --> 00:08:10,120
and need updating.

220
00:08:10,120 --> 00:08:12,080
These questions transform document management

221
00:08:12,080 --> 00:08:14,760
from a cost center into a strategic capability.

222
00:08:14,760 --> 00:08:17,320
The metadata gap is what prevents most organizations

223
00:08:17,320 --> 00:08:19,000
from making that transformation.

224
00:08:19,000 --> 00:08:21,440
It keeps SharePoint in the filing cabinet era

225
00:08:21,440 --> 00:08:24,480
while competitors move into the intelligence era.

226
00:08:24,480 --> 00:08:25,840
The metadata foundation.

227
00:08:25,840 --> 00:08:27,520
Before you can fix the structure,

228
00:08:27,520 --> 00:08:30,000
you need to understand what the structure actually is.

229
00:08:30,000 --> 00:08:32,800
Most people define metadata as data about data.

230
00:08:32,800 --> 00:08:35,760
That definition is technically accurate and practically useless.

231
00:08:35,760 --> 00:08:37,680
For governance and automation purposes,

232
00:08:37,680 --> 00:08:40,520
metadata is better understood as the structured descriptors

233
00:08:40,520 --> 00:08:42,600
that let systems and people understand,

234
00:08:42,600 --> 00:08:44,960
organize, and act on information.

235
00:08:44,960 --> 00:08:47,880
For a document, this includes simple attributes like title,

236
00:08:47,880 --> 00:08:49,200
author, and creation date,

237
00:08:49,200 --> 00:08:51,880
as well as business-specific properties like contract type,

238
00:08:51,880 --> 00:08:55,680
client name, region, sensitivity level, or retention category.

239
00:08:55,680 --> 00:08:58,120
Metadata provides the context that allows documents

240
00:08:58,120 --> 00:09:00,440
to be grouped, searched, filtered, and processed

241
00:09:00,440 --> 00:09:02,640
in a consistent way, especially when stored at scale

242
00:09:02,640 --> 00:09:04,480
in repositories like SharePoint.

243
00:09:04,480 --> 00:09:07,920
Without a coherent metadata strategy, content becomes opaque.

244
00:09:07,920 --> 00:09:10,120
Even basic questions become impossible to answer

245
00:09:10,120 --> 00:09:11,440
without manual review.

246
00:09:11,440 --> 00:09:13,800
Which contracts expire in the next 90 days?

247
00:09:13,800 --> 00:09:16,920
Which proposals contain governing law clauses for California?

248
00:09:16,920 --> 00:09:19,880
Which HR files are subject to a seven-year retention rule?

249
00:09:19,880 --> 00:09:22,680
Folders can't answer those questions only metadata can.

250
00:09:22,680 --> 00:09:24,560
There are many ways to classify metadata,

251
00:09:24,560 --> 00:09:27,120
but one useful distinction is between technical metadata

252
00:09:27,120 --> 00:09:28,800
and business metadata.

253
00:09:28,800 --> 00:09:31,320
Technical metadata describes system-level properties

254
00:09:31,320 --> 00:09:32,720
like file size or encoding,

255
00:09:32,720 --> 00:09:35,200
and it is often generated automatically by systems.

256
00:09:35,200 --> 00:09:36,920
Business metadata captures meaning

257
00:09:36,920 --> 00:09:38,880
from the perspective of business processes

258
00:09:38,880 --> 00:09:40,400
and regulatory obligations,

259
00:09:40,400 --> 00:09:43,240
and it typically requires explicit design and governance,

260
00:09:43,240 --> 00:09:45,400
including agreed vocabularies and rules for how

261
00:09:45,400 --> 00:09:47,320
and when to apply labels.

262
00:09:47,320 --> 00:09:49,120
Standard schemers, such as Dublin Core,

263
00:09:49,120 --> 00:09:51,760
offer generic elements that conservers a starting point,

264
00:09:51,760 --> 00:09:54,320
but organizations usually need custom schemers

265
00:09:54,320 --> 00:09:57,040
that reflect their specific use cases, industries,

266
00:09:57,040 --> 00:09:59,120
and regulatory requirements.

267
00:09:59,120 --> 00:10:02,440
This is where SharePoints managed metadata infrastructure

268
00:10:02,440 --> 00:10:05,160
and intelligent document processing create real value.

269
00:10:05,160 --> 00:10:08,600
SharePoint has long provided a rich set of metadata mechanisms,

270
00:10:08,600 --> 00:10:10,800
including site columns, content types,

271
00:10:10,800 --> 00:10:12,520
and the managed metadata service,

272
00:10:12,520 --> 00:10:14,240
which allows organizations to define

273
00:10:14,240 --> 00:10:17,080
and centrally manage vocabularies of terms.

274
00:10:17,080 --> 00:10:19,440
Managed metadata enables the creation of terms sets

275
00:10:19,440 --> 00:10:21,320
that can be shared across site collections,

276
00:10:21,320 --> 00:10:24,200
providing consistent choices for fields like department,

277
00:10:24,200 --> 00:10:26,640
product line, or document category.

278
00:10:26,640 --> 00:10:29,320
This shared taxonomy not only improves user experience

279
00:10:29,320 --> 00:10:32,000
when tagging content but also supports more reliable search

280
00:10:32,000 --> 00:10:34,800
refiners, navigation, and policy targeting.

281
00:10:34,800 --> 00:10:37,200
Content types extend this model by bundling together

282
00:10:37,200 --> 00:10:39,560
a set of columns, templates, and workflows

283
00:10:39,560 --> 00:10:42,360
for a specific type of content such as a customer contract

284
00:10:42,360 --> 00:10:44,040
or standard operating procedure,

285
00:10:44,040 --> 00:10:47,160
which can then be reused across libraries and sites.

286
00:10:47,160 --> 00:10:49,120
Taxonomy governance is critical to ensuring

287
00:10:49,120 --> 00:10:53,200
that these structures remain usable and relevant over time.

288
00:10:53,200 --> 00:10:54,800
Frameworks for taxonomy governance

289
00:10:54,800 --> 00:10:56,960
emphasize the need for cross-functional involvement

290
00:10:56,960 --> 00:11:00,040
from business units, IT and information management,

291
00:11:00,040 --> 00:11:02,600
clear ownership of terms sets, and defined processes

292
00:11:02,600 --> 00:11:05,600
for adding, modifying, or deprecating terms.

293
00:11:05,600 --> 00:11:07,760
If terms sets grow organically without oversight,

294
00:11:07,760 --> 00:11:09,840
they can become cluttered and inconsistent,

295
00:11:09,840 --> 00:11:11,840
eroding the benefits of structured metadata

296
00:11:11,840 --> 00:11:13,640
and confusing end-users.

297
00:11:13,640 --> 00:11:17,040
Conversely, a well-governed taxonomy aligned to business domains

298
00:11:17,040 --> 00:11:19,640
can form the backbone of automated classification

299
00:11:19,640 --> 00:11:21,560
and retention policies, particularly

300
00:11:21,560 --> 00:11:23,480
when paired with AI-based systems

301
00:11:23,480 --> 00:11:27,040
that can infer metadata from document content.

302
00:11:27,040 --> 00:11:27,960
But here's the problem.

303
00:11:27,960 --> 00:11:30,320
Most organizations skip the taxonomy step

304
00:11:30,320 --> 00:11:31,840
and jump straight to the AI.

305
00:11:31,840 --> 00:11:33,800
They think a machine learning model will somehow

306
00:11:33,800 --> 00:11:35,560
invent their business logic from scratch.

307
00:11:35,560 --> 00:11:36,320
It won't.

308
00:11:36,320 --> 00:11:38,000
AI can recognize patterns and documents,

309
00:11:38,000 --> 00:11:40,280
but it needs a target vocabulary to populate.

310
00:11:40,280 --> 00:11:42,040
It can extract an expiration date,

311
00:11:42,040 --> 00:11:44,240
but it needs to know which column to drop it into.

312
00:11:44,240 --> 00:11:46,120
It can classify a document as sensitive,

313
00:11:46,120 --> 00:11:48,760
but it needs to know which sensitivity label your organization

314
00:11:48,760 --> 00:11:49,440
uses.

315
00:11:49,440 --> 00:11:50,960
The model isn't the hard part.

316
00:11:50,960 --> 00:11:52,800
The taxonomy is, and if your taxonomy is broken,

317
00:11:52,800 --> 00:11:54,000
your model will be too.

318
00:11:54,000 --> 00:11:56,000
Managed metadata service is the engine that

319
00:11:56,000 --> 00:11:58,440
makes taxonomy scalable across your tenant.

320
00:11:58,440 --> 00:12:01,120
It stores term sets in a central service application,

321
00:12:01,120 --> 00:12:03,160
making them available to any site collection that

322
00:12:03,160 --> 00:12:03,960
subscribes to them.

323
00:12:03,960 --> 00:12:06,840
When your legal team defines a term set for contract types,

324
00:12:06,840 --> 00:12:08,880
every SharePoint site can use those same terms

325
00:12:08,880 --> 00:12:10,120
without recreating them.

326
00:12:10,120 --> 00:12:12,120
When your HR team updates a department list,

327
00:12:12,120 --> 00:12:13,600
the change propagates automatically

328
00:12:13,600 --> 00:12:15,360
to every library that references it.

329
00:12:15,360 --> 00:12:17,440
This centralization is what prevents

330
00:12:17,440 --> 00:12:19,960
the organic sprawl of inconsistent labels

331
00:12:19,960 --> 00:12:22,240
that destroys metadata quality over time.

332
00:12:22,240 --> 00:12:24,360
Content types add another layer of structure.

333
00:12:24,360 --> 00:12:26,360
A content type bundles columns, templates,

334
00:12:26,360 --> 00:12:28,440
and workflows into a reusable package.

335
00:12:28,440 --> 00:12:30,400
When you define a customer contract content type

336
00:12:30,400 --> 00:12:32,760
with columns for counterparty effective date and governing

337
00:12:32,760 --> 00:12:34,680
law, you can attach that content type

338
00:12:34,680 --> 00:12:36,360
to any library in your tenant.

339
00:12:36,360 --> 00:12:38,360
Users see the same column structure,

340
00:12:38,360 --> 00:12:41,200
the same document template, and the same retention workflow

341
00:12:41,200 --> 00:12:43,160
regardless of which site they are in.

342
00:12:43,160 --> 00:12:45,240
This consistency is what makes enterprise governance

343
00:12:45,240 --> 00:12:46,080
possible.

344
00:12:46,080 --> 00:12:47,960
Without content types, every site owner

345
00:12:47,960 --> 00:12:50,480
invents their own column names, their own data formats,

346
00:12:50,480 --> 00:12:51,920
and their own business rules.

347
00:12:51,920 --> 00:12:53,640
The result is a metadata tower of Bable

348
00:12:53,640 --> 00:12:56,440
where vendor in one site means counterparty in another,

349
00:12:56,440 --> 00:12:58,720
and no system can correlate them.

350
00:12:58,720 --> 00:13:00,680
Site columns are the individual fields

351
00:13:00,680 --> 00:13:03,280
that make up a content type or library schema.

352
00:13:03,280 --> 00:13:06,200
Each site column has a data type, a default value,

353
00:13:06,200 --> 00:13:07,880
and a set of allowed values.

354
00:13:07,880 --> 00:13:09,760
When you define a site column for contract value

355
00:13:09,760 --> 00:13:11,560
as currency with a minimum of 0,

356
00:13:11,560 --> 00:13:14,160
you prevent users from entering text or negative numbers.

357
00:13:14,160 --> 00:13:16,680
When you define a site column for contract type,

358
00:13:16,680 --> 00:13:20,040
as a managed metadata field tied to a closed term set,

359
00:13:20,040 --> 00:13:22,560
you prevent users from inventing new contract types

360
00:13:22,560 --> 00:13:23,480
on the fly.

361
00:13:23,480 --> 00:13:25,400
These constraints are not limitations.

362
00:13:25,400 --> 00:13:27,640
They are the guardrails that keep metadata

363
00:13:27,640 --> 00:13:29,120
clean enough for automation.

364
00:13:29,120 --> 00:13:31,160
The relationship between these elements matters.

365
00:13:31,160 --> 00:13:32,920
Terms sets define the vocabulary.

366
00:13:32,920 --> 00:13:35,160
Site columns define the field structure.

367
00:13:35,160 --> 00:13:37,880
Content types bundle them into reusable packages.

368
00:13:37,880 --> 00:13:39,880
Libraries apply them to real documents

369
00:13:39,880 --> 00:13:42,160
and syntax models populate them automatically.

370
00:13:42,160 --> 00:13:45,000
If any layer is missing or broken, the whole stack fails.

371
00:13:45,000 --> 00:13:46,880
A beautiful model with no target columns

372
00:13:46,880 --> 00:13:48,760
has nowhere to write its extractions.

373
00:13:48,760 --> 00:13:50,240
A perfect taxonomy with no model

374
00:13:50,240 --> 00:13:52,040
has no way to populate itself.

375
00:13:52,040 --> 00:13:53,800
A governance policy with no classification

376
00:13:53,800 --> 00:13:54,840
has nothing to enforce.

377
00:13:54,840 --> 00:13:57,320
These four layers, taxonomy, columns, content types,

378
00:13:57,320 --> 00:13:59,520
and models must be designed together.

379
00:13:59,520 --> 00:14:02,000
Unfortunately, most organizations design them separately

380
00:14:02,000 --> 00:14:03,480
if they design them at all.

381
00:14:03,480 --> 00:14:05,840
The SharePoint admin creates generic libraries

382
00:14:05,840 --> 00:14:08,000
with default document content types.

383
00:14:08,000 --> 00:14:10,680
The business user uploads files without tagging them.

384
00:14:10,680 --> 00:14:12,760
The compliance officer writes retention policies

385
00:14:12,760 --> 00:14:15,320
that apply to everything because there is no classification

386
00:14:15,320 --> 00:14:16,040
to target.

387
00:14:16,040 --> 00:14:17,920
The AI team deploys co-pilot and wonders

388
00:14:17,920 --> 00:14:19,800
why it returns irrelevant results.

389
00:14:19,800 --> 00:14:22,040
Each group is doing their job in isolation

390
00:14:22,040 --> 00:14:25,280
and the isolation is what creates the gap.

391
00:14:25,280 --> 00:14:28,040
Governance, compliance, and co-pilot readiness.

392
00:14:28,040 --> 00:14:30,800
The metadata gap isn't only a productivity issue.

393
00:14:30,800 --> 00:14:33,040
It is a genuine compliance and risk concern

394
00:14:33,040 --> 00:14:35,400
that sits at the intersection of governance trends

395
00:14:35,400 --> 00:14:36,640
and AI readiness.

396
00:14:36,640 --> 00:14:39,240
Regulators and courts increasingly expect organizations

397
00:14:39,240 --> 00:14:41,760
to demonstrate defensible retention, disposition,

398
00:14:41,760 --> 00:14:44,000
and e-discovery practices, particularly

399
00:14:44,000 --> 00:14:45,840
in regulated industries or jurisdictions

400
00:14:45,840 --> 00:14:51,320
governed by frameworks like GDPR, CCPA, and ISO 27701.

401
00:14:51,320 --> 00:14:53,680
Retention policies and labels in Microsoft PerView

402
00:14:53,680 --> 00:14:56,160
are designed to help organizations retain content

403
00:14:56,160 --> 00:14:59,680
for required periods or delete it when it is no longer needed,

404
00:14:59,680 --> 00:15:02,720
but these controls rely on being able to identify which

405
00:15:02,720 --> 00:15:04,920
content falls into which category.

406
00:15:04,920 --> 00:15:06,440
When documents lack classification

407
00:15:06,440 --> 00:15:08,680
that differentiates records from working copies

408
00:15:08,680 --> 00:15:11,160
or regulated data from non-critical content,

409
00:15:11,160 --> 00:15:14,240
retention settings may be applied broadly or not at all.

410
00:15:14,240 --> 00:15:17,000
That increases both over-attention and premature deletion risk.

411
00:15:17,000 --> 00:15:19,640
In the context of e-discovery and litigation holds,

412
00:15:19,640 --> 00:15:21,840
metadata becomes even more critical.

413
00:15:21,840 --> 00:15:24,760
Metadata provides details such as creation and modification

414
00:15:24,760 --> 00:15:27,960
timestamps, authorship, file paths, and message headers,

415
00:15:27,960 --> 00:15:29,680
which are needed to reconstruct events

416
00:15:29,680 --> 00:15:31,720
and demonstrate chain of custody.

417
00:15:31,720 --> 00:15:33,600
Best practices for defensible e-discovery

418
00:15:33,600 --> 00:15:37,120
emphasize preserving metadata using forensic grade tools,

419
00:15:37,120 --> 00:15:39,000
documenting the chain of custody,

420
00:15:39,000 --> 00:15:41,880
and auditing metadata before and after processing

421
00:15:41,880 --> 00:15:44,160
to avoid challenges to evidence integrity.

422
00:15:44,160 --> 00:15:46,920
If SharePoint content is spread across uncontrolled sites,

423
00:15:46,920 --> 00:15:50,000
personal OneDrive locations and shadow repositories,

424
00:15:50,000 --> 00:15:52,480
and if the metadata is inconsistent or incomplete,

425
00:15:52,480 --> 00:15:54,840
legal teams may struggle to identify, preserve,

426
00:15:54,840 --> 00:15:58,000
and produce relevant materials in a defensible manner.

427
00:15:58,000 --> 00:16:01,280
So the lack of structured metadata increases litigation risk

428
00:16:01,280 --> 00:16:03,360
and the cost of discovery exercises.

429
00:16:03,360 --> 00:16:06,080
It also amplifies security and privacy risks.

430
00:16:06,080 --> 00:16:09,240
SharePoint and OneDrive provide strong security capabilities,

431
00:16:09,240 --> 00:16:12,640
but DLP rules, sensitivity labels, and conditional access

432
00:16:12,640 --> 00:16:16,640
rely on understanding which documents contain sensitive information.

433
00:16:16,640 --> 00:16:18,040
Without reliable classification,

434
00:16:18,040 --> 00:16:19,960
organizations may either over-restrict access

435
00:16:19,960 --> 00:16:23,000
and hamper collaboration or under-protect sensitive documents

436
00:16:23,000 --> 00:16:24,680
and lead to accidental exposure.

437
00:16:24,680 --> 00:16:27,600
Governance practices in the Microsoft 365 environment

438
00:16:27,600 --> 00:16:30,520
have been evolving to address both the scale of content

439
00:16:30,520 --> 00:16:33,520
and the emergence of AI-powered experiences.

440
00:16:33,520 --> 00:16:35,680
Emerging trends include more granular controls

441
00:16:35,680 --> 00:16:37,920
for regulating which content can be surfaced

442
00:16:37,920 --> 00:16:40,240
by AI assistance and tenant-wide search,

443
00:16:40,240 --> 00:16:42,040
reflecting both privacy concerns,

444
00:16:42,040 --> 00:16:44,480
and the need to respect regulatory boundaries.

445
00:16:44,480 --> 00:16:47,520
One notable feature is restricted content discovery,

446
00:16:47,520 --> 00:16:49,600
which allows administrators to prevent content

447
00:16:49,600 --> 00:16:51,640
from specified SharePoint sites from appearing

448
00:16:51,640 --> 00:16:54,600
in organization-wide search and Microsoft 365

449
00:16:54,600 --> 00:16:55,920
co-pilot business chat,

450
00:16:55,920 --> 00:16:58,560
while still permitting users to access documents they own

451
00:16:58,560 --> 00:17:00,400
or have recently interacted with.

452
00:17:00,400 --> 00:17:02,720
This capability reflects a broader recognition

453
00:17:02,720 --> 00:17:06,040
that AI experiences introduce new governance dimensions.

454
00:17:06,040 --> 00:17:08,320
Organizations must manage not only who can access

455
00:17:08,320 --> 00:17:09,600
a document directly,

456
00:17:09,600 --> 00:17:12,200
but also how content can be discovered or summarized

457
00:17:12,200 --> 00:17:13,880
indirectly by AI tools.

458
00:17:13,880 --> 00:17:16,560
Governance strategies must cover both traditional controls

459
00:17:16,560 --> 00:17:19,920
around permissions and newer controls around AI visibility,

460
00:17:19,920 --> 00:17:22,280
ensuring that sensitive or highly regulated content

461
00:17:22,280 --> 00:17:25,160
is not surfaced in unintended contexts.

462
00:17:25,160 --> 00:17:27,440
At the same time, Microsoft messaging around hybrid work

463
00:17:27,440 --> 00:17:29,920
and AI emphasizes that organizations

464
00:17:29,920 --> 00:17:32,440
which invest in governing and organizing their information

465
00:17:32,440 --> 00:17:34,280
will gain a competitive advantage

466
00:17:34,280 --> 00:17:36,080
as they can more effectively use AI

467
00:17:36,080 --> 00:17:38,720
to drive productivity and innovation.

468
00:17:38,720 --> 00:17:41,200
The metadata gap thus sits at the intersection

469
00:17:41,200 --> 00:17:44,080
of governance trends and AI readiness.

470
00:17:44,080 --> 00:17:46,160
It is the single biggest reason your co-pilot rollout

471
00:17:46,160 --> 00:17:48,440
will underwhelm, not because the AI is weak,

472
00:17:48,440 --> 00:17:51,120
but because the ground beneath it is incomplete.

473
00:17:51,120 --> 00:17:53,520
Co-pilot relies on metadata, permissions,

474
00:17:53,520 --> 00:17:55,840
and information architecture to return relevant,

475
00:17:55,840 --> 00:17:57,200
trustworthy responses.

476
00:17:57,200 --> 00:17:59,280
When documents lack accurate metadata

477
00:17:59,280 --> 00:18:01,280
or are stored in haphazard locations,

478
00:18:01,280 --> 00:18:03,680
semantic search and AI-driven retrieval struggle

479
00:18:03,680 --> 00:18:05,280
to surface the right context.

480
00:18:05,280 --> 00:18:06,760
The metadata gap becomes a barrier

481
00:18:06,760 --> 00:18:09,840
to realizing the promise of AI accelerated hybrid work

482
00:18:09,840 --> 00:18:13,080
and unmanaged SharePoint data quietly becomes a liability

483
00:18:13,080 --> 00:18:14,840
that most boards do not see coming.

484
00:18:14,840 --> 00:18:17,040
There's a secret weapon Microsoft built for this

485
00:18:17,040 --> 00:18:18,720
and most tenants haven't turned it on yet.

486
00:18:18,720 --> 00:18:20,760
Restricted content discovery is one example

487
00:18:20,760 --> 00:18:24,040
of how governance is evolving to address AI-specific risks.

488
00:18:24,040 --> 00:18:26,040
Administrators can configure this feature

489
00:18:26,040 --> 00:18:27,920
to prevent specified SharePoint sites

490
00:18:27,920 --> 00:18:29,800
from appearing in organization wide search

491
00:18:29,800 --> 00:18:33,000
and co-pilot business chat while still allowing direct access

492
00:18:33,000 --> 00:18:36,200
to users who own or have recently interacted with the content.

493
00:18:36,200 --> 00:18:37,800
This is not about blocking access,

494
00:18:37,800 --> 00:18:39,320
it is about controlling discovery.

495
00:18:39,320 --> 00:18:41,320
A site containing merger negotiations

496
00:18:41,320 --> 00:18:43,240
should not surface in a co-pilot query

497
00:18:43,240 --> 00:18:45,800
from an intern who happens to have red permissions.

498
00:18:45,800 --> 00:18:47,240
The metadata and permissions layer

499
00:18:47,240 --> 00:18:49,480
must govern not just who can open a document,

500
00:18:49,480 --> 00:18:51,200
but who can discover it through AI.

501
00:18:51,200 --> 00:18:55,040
ISO 27701, GDPR and CCPA all impose obligations

502
00:18:55,040 --> 00:18:57,800
on data minimization, purpose limitation, and transparency.

503
00:18:57,800 --> 00:18:59,440
These frameworks require organizations

504
00:18:59,440 --> 00:19:01,120
to know what personal data they hold,

505
00:19:01,120 --> 00:19:03,320
where it resides, and how long they keep it.

506
00:19:03,320 --> 00:19:06,600
A SharePoint environment with no metadata cannot answer those questions.

507
00:19:06,600 --> 00:19:09,160
When a data subject requests a raja under GDPR,

508
00:19:09,160 --> 00:19:11,400
the organization must identify every document

509
00:19:11,400 --> 00:19:13,320
containing that person's information.

510
00:19:13,320 --> 00:19:15,760
Without metadata tagging for personal data types,

511
00:19:15,760 --> 00:19:18,560
this becomes a manual search across millions of files,

512
00:19:18,560 --> 00:19:22,320
the cost of compliance scales directly with the metadata gap.

513
00:19:22,320 --> 00:19:24,400
The same logic applies to records management.

514
00:19:24,400 --> 00:19:26,360
A record is not just any old document,

515
00:19:26,360 --> 00:19:28,880
it is evidence of a business transaction or decision

516
00:19:28,880 --> 00:19:31,200
that must be preserved for a defined period

517
00:19:31,200 --> 00:19:33,480
and disposed of according to a legal authority.

518
00:19:33,480 --> 00:19:35,320
Records management requires classification

519
00:19:35,320 --> 00:19:37,520
that distinguishes records from working copies,

520
00:19:37,520 --> 00:19:40,200
transitory communications, and reference materials.

521
00:19:40,200 --> 00:19:44,040
Without that classification, organizations either over-retain everything,

522
00:19:44,040 --> 00:19:46,600
accumulating storage cost and litigation risk

523
00:19:46,600 --> 00:19:49,320
or under-retain destroying evidence they later need.

524
00:19:49,320 --> 00:19:52,440
Metadata is the mechanism that makes records management precise,

525
00:19:52,440 --> 00:19:54,200
defensible, and auditable.

526
00:19:54,200 --> 00:19:57,800
Even basic information security relies on metadata.

527
00:19:57,800 --> 00:19:59,560
Conditional access policies can enforce

528
00:19:59,560 --> 00:20:02,320
multi-factor authentication for sensitive documents,

529
00:20:02,320 --> 00:20:05,440
but only if the system knows which documents are sensitive.

530
00:20:05,440 --> 00:20:08,800
Encryption policies can apply stronger algorithms to regulated data,

531
00:20:08,800 --> 00:20:11,880
but only if the system can identify regulated data.

532
00:20:11,880 --> 00:20:13,600
The principle of least privilege,

533
00:20:13,600 --> 00:20:16,320
granting users the minimum access necessary for their role,

534
00:20:16,320 --> 00:20:19,760
depends on understanding what each document contains and who should see it.

535
00:20:19,760 --> 00:20:22,360
Metadata makes those distinctions machine readable.

536
00:20:22,360 --> 00:20:24,840
Without it, security policies become blanket rules

537
00:20:24,840 --> 00:20:28,320
that either block legitimate work or leave gaps that attack as exploit.

538
00:20:28,320 --> 00:20:32,400
Copilot readiness is the final and most urgent reason to close the metadata gap.

539
00:20:32,400 --> 00:20:36,600
Microsoft has enhanced copilot's ability to reason over SharePoint metadata

540
00:20:36,600 --> 00:20:38,880
as of Ignite 2025.

541
00:20:38,880 --> 00:20:41,240
Copilot can now distinguish between similar documents

542
00:20:41,240 --> 00:20:42,880
by understanding their classification,

543
00:20:42,880 --> 00:20:46,080
department, project associations, and custom tags.

544
00:20:46,080 --> 00:20:49,000
When grounded on libraries with populated metadata,

545
00:20:49,000 --> 00:20:52,920
organizations see tangible improvements in response, quality, and relevance.

546
00:20:52,920 --> 00:20:57,240
For example, asking copilot to show the latest client proposals for the healthcare vertical

547
00:20:57,240 --> 00:21:00,520
now returns accurate results because it understands both the content

548
00:21:00,520 --> 00:21:04,840
and the metadata tags indicating industry vertical, status, and date.

549
00:21:04,840 --> 00:21:08,360
Without metadata, copilot is just a fast way to find the wrong answer.

550
00:21:08,360 --> 00:21:11,080
What intelligent document processing actually means?

551
00:21:11,080 --> 00:21:13,920
The weapon isn't copilot itself, it's the layer underneath.

552
00:21:13,920 --> 00:21:16,200
Intelligent document processing or IDP

553
00:21:16,200 --> 00:21:20,120
refers to software capabilities that capture, transform, and process data

554
00:21:20,120 --> 00:21:23,600
from documents using AI techniques such as computer vision,

555
00:21:23,600 --> 00:21:26,760
optical character recognition, natural language processing,

556
00:21:26,760 --> 00:21:28,240
and machine learning.

557
00:21:28,240 --> 00:21:31,400
IDP solutions ingest documents in various formats including emails,

558
00:21:31,400 --> 00:21:34,840
PDFs, word files, and scanned images and convert them into structured data

559
00:21:34,840 --> 00:21:39,440
that can be analyzed, categorized, and integrated into downstream systems and workflows.

560
00:21:39,440 --> 00:21:42,240
A key goal is to automate the extraction of relevant information

561
00:21:42,240 --> 00:21:45,560
such as invoice numbers, contract dates, or customer details

562
00:21:45,560 --> 00:21:48,520
and to classify documents according to type or business process

563
00:21:48,520 --> 00:21:52,160
which reduces manual data entry and classification work.

564
00:21:52,160 --> 00:21:55,600
At a technical level, IDP systems combine several components,

565
00:21:55,600 --> 00:22:00,080
optical character recognition transforms images of text into machine readable text,

566
00:22:00,080 --> 00:22:04,560
enabling the processing of scanned documents or photographs of physical forms.

567
00:22:04,560 --> 00:22:08,640
Natural language processing techniques interpret the semantics and structure of the text,

568
00:22:08,640 --> 00:22:12,160
identifying entities, relationships, and topics that help classify documents

569
00:22:12,160 --> 00:22:14,280
or extract key value pairs.

570
00:22:14,280 --> 00:22:16,960
Machine learning models, often trained on labeled examples,

571
00:22:16,960 --> 00:22:19,920
learn to recognize patterns in document layouts and content,

572
00:22:19,920 --> 00:22:24,000
allowing them to handle structured, semi-structured, and unstructured documents.

573
00:22:24,000 --> 00:22:26,880
Over time, models can be tuned to specific domains,

574
00:22:26,880 --> 00:22:30,160
improving accuracy for particular types of documents or industries.

575
00:22:30,160 --> 00:22:32,720
Understanding the nature of the documents being processed

576
00:22:32,720 --> 00:22:35,240
is needed to design effective IDP solutions.

577
00:22:35,240 --> 00:22:39,600
Structured documents such as standardized forms and invoices with consistent layouts

578
00:22:39,600 --> 00:22:41,840
are relatively straightforward for extraction engines

579
00:22:41,840 --> 00:22:45,280
because fields appear in predictable locations and formats.

580
00:22:45,280 --> 00:22:49,760
For these, structured document processing models can map specific regions of a page

581
00:22:49,760 --> 00:22:54,320
to named fields, achieving high accuracy with relatively little training data.

582
00:22:54,320 --> 00:22:57,360
Semi-structured documents combine fixed and variable elements.

583
00:22:57,360 --> 00:23:00,000
Examples include purchase orders where line items vary,

584
00:23:00,000 --> 00:23:03,600
but the header structure is consistent, or invoices from multiple vendors

585
00:23:03,600 --> 00:23:06,640
that follow similar patterns but differ in detail.

586
00:23:06,640 --> 00:23:10,000
Unstructured documents such as contracts, policies, memos, and correspondence

587
00:23:10,000 --> 00:23:13,440
lack fixed layouts and rely heavily on linguistic cues and context,

588
00:23:13,440 --> 00:23:16,880
making them more challenging to pass reliably.

589
00:23:16,880 --> 00:23:21,600
The business case for intelligent document processing is grounded in both cost savings and risk reduction.

590
00:23:21,600 --> 00:23:26,080
By automating repetitive tasks such as data entry from invoices or manual tagging of documents,

591
00:23:26,080 --> 00:23:29,200
IDP reduces human effort and speeds up business processes

592
00:23:29,200 --> 00:23:31,440
from accounts payable to contract onboarding.

593
00:23:31,440 --> 00:23:34,960
AI-driven classification and extraction can also improve accuracy

594
00:23:34,960 --> 00:23:38,320
relative to manual processes, which are prone to errors,

595
00:23:38,320 --> 00:23:40,960
especially when volume and complexity are high.

596
00:23:40,960 --> 00:23:43,920
Over time, the data extracted from documents becomes a valuable source

597
00:23:43,920 --> 00:23:45,840
for business intelligence and analytics,

598
00:23:45,840 --> 00:23:50,320
enabling organizations to derive insights from previously inaccessible unstructured content.

599
00:23:50,320 --> 00:23:55,040
Studies have shown how AI-driven automation can transform employee productivity

600
00:23:55,040 --> 00:23:58,000
and accelerate business results by freeing knowledge workers

601
00:23:58,000 --> 00:23:59,760
to focus on higher value tasks.

602
00:23:59,760 --> 00:24:02,400
However, these benefits are not automatic,

603
00:24:02,400 --> 00:24:04,000
and there are important caveats.

604
00:24:04,000 --> 00:24:07,360
AI systems must be trained and validated on representative data

605
00:24:07,360 --> 00:24:10,720
and their outputs need to be monitored for accuracy and bias,

606
00:24:10,720 --> 00:24:15,200
particularly when used for decisions that carry financial, legal, or ethical implications.

607
00:24:15,920 --> 00:24:18,960
Organizations cannot simply assume AI outputs are correct.

608
00:24:18,960 --> 00:24:22,480
Best practice is to treat AI-generated classifications and extractions

609
00:24:22,480 --> 00:24:25,520
as draft work that should be reviewed by subject matter experts,

610
00:24:25,520 --> 00:24:27,280
especially during initial deployment.

611
00:24:27,280 --> 00:24:30,800
Overreliance on AI without proper governance may introduce new risks,

612
00:24:30,800 --> 00:24:34,480
such as misclassification of sensitive data or incorrect metadata

613
00:24:34,480 --> 00:24:37,040
that propagates through retention and DLP policies.

614
00:24:37,040 --> 00:24:41,440
IDP does not eliminate the need for clear process and information architecture.

615
00:24:41,440 --> 00:24:45,520
If business processes are poorly defined or content is created in a chaotic manner,

616
00:24:45,520 --> 00:24:48,080
even the best models will struggle to impose order.

617
00:24:48,080 --> 00:24:50,160
AI amplifies whatever patterns exist.

618
00:24:50,160 --> 00:24:51,680
If those patterns are inconsistent,

619
00:24:51,680 --> 00:24:53,840
the generated metadata will be inconsistent

620
00:24:53,840 --> 00:24:56,240
and governance outcomes will remain unreliable.

621
00:24:56,240 --> 00:24:58,720
IDP should be positioned not as a silver bullet,

622
00:24:58,720 --> 00:25:01,120
but as a powerful enabler within a broader program

623
00:25:01,120 --> 00:25:02,960
that includes taxonomy design,

624
00:25:02,960 --> 00:25:05,600
governance frameworks, and change management.

625
00:25:05,600 --> 00:25:07,440
Within the Microsoft product suite,

626
00:25:07,440 --> 00:25:10,080
IDP capabilities are available across several products

627
00:25:10,080 --> 00:25:11,840
that serve different layers of the stack.

628
00:25:11,840 --> 00:25:16,240
Microsoft syntax now integrated into SharePoint Premium and Document Processing

629
00:25:16,240 --> 00:25:18,960
operates directly on documents stored in SharePoint,

630
00:25:18,960 --> 00:25:20,960
enabling organizations to understand,

631
00:25:20,960 --> 00:25:25,280
classify and extract information from their content within Microsoft 365.

632
00:25:25,280 --> 00:25:28,560
Syntax uses document processing models that can be configured

633
00:25:28,560 --> 00:25:31,680
to recognize structured forms, semi-structured documents,

634
00:25:31,680 --> 00:25:33,200
or unstructured content,

635
00:25:33,200 --> 00:25:36,880
and it integrates with SharePoint libraries to automatically apply metadata,

636
00:25:36,880 --> 00:25:40,640
content types, retention labels, and other governance artifacts.

637
00:25:40,640 --> 00:25:43,760
It also supports related capabilities such as content assembly,

638
00:25:43,760 --> 00:25:47,040
which can generate standardized documents from templates and data sources.

639
00:25:47,040 --> 00:25:50,080
As your document intelligence, now part of as your content understanding

640
00:25:50,080 --> 00:25:53,680
in Foundry tools provides IDP capabilities as a cloud service

641
00:25:53,680 --> 00:25:57,200
that can process documents at scale and integrate with custom applications.

642
00:25:57,200 --> 00:26:00,640
It applies advanced AI models to extract text,

643
00:26:00,640 --> 00:26:05,040
key value pairs, tables, and document structure from files such as PDFs,

644
00:26:05,040 --> 00:26:09,200
images and forms with strong performance on structured and templated documents.

645
00:26:09,200 --> 00:26:12,640
The service exposes rest APIs for developers to analyze documents

646
00:26:12,640 --> 00:26:16,080
using pre-built analyzers such as a general purpose document analyzer

647
00:26:16,080 --> 00:26:20,000
that extracts text and layout elements like paragraphs and tables.

648
00:26:20,000 --> 00:26:22,880
This makes it suitable for scenarios where document processing needs

649
00:26:22,880 --> 00:26:25,680
to be tightly integrated with line of business applications,

650
00:26:25,680 --> 00:26:27,760
data pipelines, or external storage.

651
00:26:27,760 --> 00:26:30,720
Azure content understanding extends these capabilities

652
00:26:30,720 --> 00:26:32,960
beyond documents to other modalities,

653
00:26:32,960 --> 00:26:35,440
including images, audio, and video,

654
00:26:35,440 --> 00:26:38,400
providing pre-built analyzers for extracting transcripts,

655
00:26:38,400 --> 00:26:40,160
keyframes, and descriptions.

656
00:26:40,160 --> 00:26:43,600
A developer can send a file URL to a content understanding endpoint,

657
00:26:43,600 --> 00:26:48,080
specify an analyzer such as pre-built document analyzer or pre-built video analyzer,

658
00:26:48,080 --> 00:26:50,160
and then pull for analysis results,

659
00:26:50,160 --> 00:26:53,200
integrating the extracted data into downstream workflows.

660
00:26:53,200 --> 00:26:57,120
This positions content understanding as a general multimodal extraction layer

661
00:26:57,120 --> 00:26:58,880
within an AI architecture,

662
00:26:58,880 --> 00:27:03,120
complementing syntaxes focus on SharePoint content and user-facing automation.

663
00:27:03,120 --> 00:27:07,040
Azure Databricks in turn offers an intelligent document processing pattern

664
00:27:07,040 --> 00:27:10,640
built on its lakehouse platform using natively composable AI functions

665
00:27:10,640 --> 00:27:13,440
to implement end-to-end IDP pipelines.

666
00:27:13,440 --> 00:27:16,240
This approach involves ingesting raw documents into the lakehouse,

667
00:27:16,240 --> 00:27:20,480
passing them into structured representations using functions like AI-pass document,

668
00:27:20,480 --> 00:27:24,560
enriching them via functions such as AI extract and AI classify,

669
00:27:24,560 --> 00:27:28,480
and then leveraging the results for analytics, retrieval augmented generation,

670
00:27:28,480 --> 00:27:30,080
or agent workflows.

671
00:27:30,080 --> 00:27:32,640
Because each stage of the pipeline, ingestion, passing,

672
00:27:32,640 --> 00:27:35,120
enrichment, and analysis is unified on the lakehouse,

673
00:27:35,120 --> 00:27:38,800
organizations can avoid complex integration or data movement,

674
00:27:38,800 --> 00:27:41,360
while benefiting from data bricks, scalability,

675
00:27:41,360 --> 00:27:43,760
and advanced analytics tooling.

676
00:27:43,760 --> 00:27:46,000
For most SharePoint-centric organizations,

677
00:27:46,000 --> 00:27:47,760
the choice is not between these platforms,

678
00:27:47,760 --> 00:27:49,520
but about how they complement each other.

679
00:27:49,520 --> 00:27:51,600
Syntax handles the SharePoint native automation.

680
00:27:51,600 --> 00:27:54,960
Azure Document Intelligence handles the API-driven bulk processing,

681
00:27:54,960 --> 00:27:57,920
content understanding handles the multimodal enrichment.

682
00:27:57,920 --> 00:28:01,840
Databricks handles the enterprise scale analytics and machine learning refinement.

683
00:28:01,840 --> 00:28:04,800
The metadata schema is the common thread that ties them together.

684
00:28:04,800 --> 00:28:06,800
When every platform writes to the same textonomy,

685
00:28:06,800 --> 00:28:08,960
the document becomes portable across systems,

686
00:28:08,960 --> 00:28:12,080
and the metadata becomes the single source of truth.

687
00:28:12,080 --> 00:28:15,600
A practical integration pattern starts with Syntax as the intake layer.

688
00:28:15,600 --> 00:28:19,360
Documents arrive in SharePoint and Syntax extracts the core metadata.

689
00:28:19,360 --> 00:28:23,440
For simple automation, Power Automate reacts to the extracted values directly.

690
00:28:23,440 --> 00:28:27,520
For complex analytics, the metadata and document references flow to Azure Data Lake,

691
00:28:27,520 --> 00:28:30,560
via Power BI Data Flows or Azure Synapse link.

692
00:28:30,560 --> 00:28:32,480
For advanced machine learning refinement,

693
00:28:32,480 --> 00:28:34,480
Data Bricks reads the label documents,

694
00:28:34,480 --> 00:28:37,120
compares model predictions against human corrections,

695
00:28:37,120 --> 00:28:40,720
and generates improved training sets that feed back into Syntax.

696
00:28:40,720 --> 00:28:41,840
The loop is continuous.

697
00:28:41,840 --> 00:28:43,360
The document stays in SharePoint.

698
00:28:43,360 --> 00:28:44,960
The metadata serves every system.

699
00:28:44,960 --> 00:28:47,680
The taxonomy ensures consistency across the pipeline.

700
00:28:47,680 --> 00:28:51,120
The choice of processing platform also depends on latency requirements.

701
00:28:51,120 --> 00:28:53,520
Syntax processes documents synchronously

702
00:28:53,520 --> 00:28:55,120
as they arrive in SharePoint,

703
00:28:55,120 --> 00:28:57,280
making it ideal for real-time automation.

704
00:28:57,440 --> 00:29:00,800
Azure document intelligence can process large batches asynchronously,

705
00:29:00,800 --> 00:29:02,960
making it ideal for backlogs and migrations.

706
00:29:02,960 --> 00:29:05,200
Data Bricks handles scheduled analytics jobs

707
00:29:05,200 --> 00:29:06,960
that don't need real-time results.

708
00:29:06,960 --> 00:29:09,120
Architects who understand these latency profiles

709
00:29:09,120 --> 00:29:12,400
can design pipelines that use each platform for what it does best,

710
00:29:12,400 --> 00:29:15,040
rather than forcing one tool to handle every scenario.

711
00:29:15,040 --> 00:29:18,960
Syntax, SharePoint Premium, and Document Processing,

712
00:29:18,960 --> 00:29:20,240
the rebrand story.

713
00:29:20,240 --> 00:29:24,960
Microsoft has been building IDP capabilities directly into SharePoint for years.

714
00:29:24,960 --> 00:29:27,200
The problem is that they keep renaming the product,

715
00:29:27,200 --> 00:29:28,480
and if you're keeping score,

716
00:29:28,480 --> 00:29:31,200
this marks the fourth rebrand in just six years.

717
00:29:31,200 --> 00:29:33,120
Back in 2019, while still in development,

718
00:29:33,120 --> 00:29:34,800
it was called Project Cortex.

719
00:29:34,800 --> 00:29:37,280
A year later, it became SharePoint Syntax,

720
00:29:37,280 --> 00:29:39,920
focused on AI-powered document understanding.

721
00:29:39,920 --> 00:29:42,640
Then in 2022, it shifted to Microsoft Syntax

722
00:29:42,640 --> 00:29:45,040
as the company pushed content AI across workloads,

723
00:29:45,040 --> 00:29:46,240
not just SharePoint.

724
00:29:46,240 --> 00:29:48,320
At ignite in November 2023,

725
00:29:48,320 --> 00:29:50,720
Jeff Taper introduced SharePoint Premium,

726
00:29:50,720 --> 00:29:54,080
a suite designed to help you manage ground and use content for AI.

727
00:29:54,080 --> 00:29:55,200
The idea was simple.

728
00:29:55,200 --> 00:29:56,800
One brand everyone could recognize,

729
00:29:56,800 --> 00:29:59,760
bringing together all the extra features on top of base SharePoint.

730
00:29:59,760 --> 00:30:02,320
SharePoint Premium was organized into three pillars,

731
00:30:02,320 --> 00:30:04,640
Experiences, Processes, and Governance.

732
00:30:04,640 --> 00:30:07,120
For content experiences, it included brand new features

733
00:30:07,120 --> 00:30:10,400
like the Agreements app in Teams, SharePoint e-signature,

734
00:30:10,400 --> 00:30:12,720
and the Documents Hub for Customers and Partners.

735
00:30:12,720 --> 00:30:15,280
The content processing pillar carried forward Syntax Classics

736
00:30:15,280 --> 00:30:16,640
like Autofill columns,

737
00:30:16,640 --> 00:30:18,480
Taxonomy tagging, Content Query,

738
00:30:18,480 --> 00:30:20,880
Translation for both Documents and Videos,

739
00:30:20,880 --> 00:30:22,880
PDF annotations, and more.

740
00:30:22,880 --> 00:30:25,040
This is where you transform content with AI,

741
00:30:25,040 --> 00:30:26,880
not just to improve user experience,

742
00:30:26,880 --> 00:30:29,280
but also to prepare your content for co-pilot.

743
00:30:29,280 --> 00:30:30,640
The better your content quality,

744
00:30:30,640 --> 00:30:32,720
the better your co-pilot rollout will be.

745
00:30:32,720 --> 00:30:35,360
The governance pillar introduced advanced admin capabilities

746
00:30:35,360 --> 00:30:37,040
like SharePoint Advanced Management,

747
00:30:37,040 --> 00:30:40,720
Microsoft 365 Archive, and Microsoft 365 Backup.

748
00:30:40,720 --> 00:30:42,960
But now, almost two years later, here we are again.

749
00:30:42,960 --> 00:30:44,960
The funny part is that the rename from Syntax

750
00:30:44,960 --> 00:30:47,360
to SharePoint Premium wasn't even fully completed.

751
00:30:47,360 --> 00:30:48,640
As of late 2025,

752
00:30:48,640 --> 00:30:50,800
you could still see Syntax mentioned everywhere

753
00:30:50,800 --> 00:30:53,200
in the documentation and inside the admin center.

754
00:30:53,200 --> 00:30:54,960
Microsoft decided to split things up.

755
00:30:54,960 --> 00:30:57,280
SharePoint Advanced Management, Backup, and Archive

756
00:30:57,280 --> 00:30:58,800
are now standalone products.

757
00:30:58,800 --> 00:31:00,080
They're no longer under an umbrella,

758
00:31:00,080 --> 00:31:01,120
they're their own thing.

759
00:31:01,120 --> 00:31:03,120
Everything that used to be Syntax is now bundled

760
00:31:03,120 --> 00:31:05,360
under a new umbrella called Document Processing

761
00:31:05,360 --> 00:31:06,720
for Microsoft 365,

762
00:31:06,720 --> 00:31:09,360
and yes, Microsoft is very particular about branding.

763
00:31:09,360 --> 00:31:11,120
In the documentation, you'll notice it's written

764
00:31:11,120 --> 00:31:13,360
with a lowercase DNP because it's a category,

765
00:31:13,360 --> 00:31:14,560
not a product name.

766
00:31:14,560 --> 00:31:16,400
If you check Microsoft Learn right now,

767
00:31:16,400 --> 00:31:19,600
you'll see that this rename helps clarify their priorities.

768
00:31:19,600 --> 00:31:22,960
At the Microsoft 365 Conference in May 2026,

769
00:31:22,960 --> 00:31:26,000
they named four services that will get the most attention

770
00:31:26,000 --> 00:31:26,800
going forward.

771
00:31:26,800 --> 00:31:29,840
Autofill columns, document translation, OCR, and e-signatures.

772
00:31:29,840 --> 00:31:31,360
These are their priorities.

773
00:31:31,360 --> 00:31:33,440
Meanwhile, older features from Syntax,

774
00:31:33,440 --> 00:31:36,080
like document processing models, content assembly,

775
00:31:36,080 --> 00:31:38,400
taxonomy tagging, and image tagging,

776
00:31:38,400 --> 00:31:40,960
have been pushed into the past generations bucket.

777
00:31:40,960 --> 00:31:42,640
In the documentation, they're tucked away

778
00:31:42,640 --> 00:31:44,960
under other document processing services.

779
00:31:44,960 --> 00:31:46,800
Microsoft clearly doesn't want these to be seen

780
00:31:46,800 --> 00:31:47,840
as main services anymore.

781
00:31:47,840 --> 00:31:50,880
They're still available, but they're not part of the future roadmap.

782
00:31:50,880 --> 00:31:53,200
To be clear, Microsoft has not announced any plans

783
00:31:53,200 --> 00:31:55,600
to deprecate or stop supporting these older features.

784
00:31:55,600 --> 00:31:57,440
So if you're using them in production,

785
00:31:57,440 --> 00:31:59,840
don't worry, there's no emergency to move off them.

786
00:31:59,840 --> 00:32:02,000
But it's obvious which features will get the most love

787
00:32:02,000 --> 00:32:04,000
and investment moving forward.

788
00:32:04,000 --> 00:32:06,640
If you're relying heavily on older machine learning-based features

789
00:32:06,640 --> 00:32:09,040
like unstructured document processing models

790
00:32:09,040 --> 00:32:10,000
or content assembly,

791
00:32:10,000 --> 00:32:12,400
this is a good moment to start planning for the future.

792
00:32:12,400 --> 00:32:15,120
Look into whether newer tools like Autofill columns

793
00:32:15,120 --> 00:32:17,120
or co-pilot can replace those features.

794
00:32:17,120 --> 00:32:19,680
It's not urgent, but moving in the same direction as Microsoft

795
00:32:19,680 --> 00:32:21,280
usually means you'll get newer features,

796
00:32:21,280 --> 00:32:24,240
a smoother experience, and maybe even cost savings.

797
00:32:24,240 --> 00:32:25,840
The operating model is changing,

798
00:32:25,840 --> 00:32:27,840
and architects need to understand what that means

799
00:32:27,840 --> 00:32:29,280
for existing investments.

800
00:32:29,280 --> 00:32:31,280
If you've already invested in Microsoft syntax

801
00:32:31,280 --> 00:32:33,040
and built models, bound them to libraries

802
00:32:33,040 --> 00:32:35,280
and wired power automate flows around them,

803
00:32:35,280 --> 00:32:36,320
you're not behind.

804
00:32:36,320 --> 00:32:38,720
You're actually in a better spot than tenant starting from scratch

805
00:32:38,720 --> 00:32:41,280
because the work you did is still valid and can be converted.

806
00:32:41,280 --> 00:32:43,680
But the operating model for that work is changing,

807
00:32:43,680 --> 00:32:46,720
and you need to develop a plan for what you can do moving forward.

808
00:32:46,720 --> 00:32:48,240
Nothing you built is wasted,

809
00:32:48,240 --> 00:32:49,920
but nothing you built is finished either.

810
00:32:49,920 --> 00:32:52,320
Treat this as a refresh, not a re-platforming.

811
00:32:52,320 --> 00:32:54,320
The first step in that plan is inventory.

812
00:32:54,320 --> 00:32:56,800
Pull a list of every syntax model in your tenant,

813
00:32:56,800 --> 00:32:57,840
what it is bound to,

814
00:32:57,840 --> 00:32:59,760
and what depends on its downstream output,

815
00:32:59,760 --> 00:33:02,480
flows, custom apps, and third-party integrations.

816
00:33:02,480 --> 00:33:04,320
Microsoft's re-use story is real,

817
00:33:04,320 --> 00:33:06,240
but it is a re-use story for the model.

818
00:33:06,240 --> 00:33:08,800
The shift from a human-trained model will not be one-to-one

819
00:33:08,800 --> 00:33:11,280
when moving to a large language-model-based classification

820
00:33:11,280 --> 00:33:12,160
and extraction.

821
00:33:12,160 --> 00:33:14,240
You need to know which models are business critical,

822
00:33:14,240 --> 00:33:15,360
which are experimental,

823
00:33:15,360 --> 00:33:16,640
and which have been abandoned.

824
00:33:16,640 --> 00:33:18,400
The second step is governance review.

825
00:33:18,400 --> 00:33:20,720
The libraries where you turned syntax on were,

826
00:33:20,720 --> 00:33:24,000
by definition, libraries with high-value, structured content.

827
00:33:24,000 --> 00:33:26,160
Those are now the libraries, co-pilot, and AI,

828
00:33:26,160 --> 00:33:28,560
in SharePoint will reason over by default.

829
00:33:28,560 --> 00:33:31,280
Re-ask the questions that were used to build those models.

830
00:33:31,280 --> 00:33:33,680
Decide whether there are better AI in SharePoint features

831
00:33:33,680 --> 00:33:35,600
to execute the same governance actions.

832
00:33:35,600 --> 00:33:39,280
A model that classifies contracts might be replaced by an autofill column,

833
00:33:39,280 --> 00:33:42,320
a model that extracts invoice data might be replaced by a skill.

834
00:33:42,320 --> 00:33:45,280
The function might survive even if the implementation changes.

835
00:33:45,280 --> 00:33:48,000
The third step is mapping by feature.

836
00:33:48,000 --> 00:33:50,560
Use the one-to-one mapping from Microsoft documentation

837
00:33:50,560 --> 00:33:54,000
to figure out what you need to enable and learn with AI in SharePoint.

838
00:33:54,000 --> 00:33:57,600
Syntax Autofill columns map to the new create Autofill columns experience.

839
00:33:57,600 --> 00:33:59,920
Document translation remains a metered content service

840
00:33:59,920 --> 00:34:02,720
with co-pilot providing conversational entry points.

841
00:34:02,720 --> 00:34:06,400
OCR becomes part of the baseline AI-ready content pipeline.

842
00:34:06,400 --> 00:34:10,080
Content assembly shifts to template driven generation in content AI.

843
00:34:10,080 --> 00:34:13,600
Image tagging and taxonomy tagging are internalized into the broader

844
00:34:13,600 --> 00:34:18,640
AI-ready content concept pre-built, structured, freeform and unstructured models

845
00:34:18,640 --> 00:34:22,160
all form the processing substrate for higher-level agent actions.

846
00:34:22,160 --> 00:34:25,360
Understanding this mapping prevents you from rebuilding what you already have.

847
00:34:25,360 --> 00:34:28,000
The fourth step is piloting skills somewhere boring.

848
00:34:28,000 --> 00:34:28,960
Skills are interesting.

849
00:34:28,960 --> 00:34:32,400
Agent-driven orchestration is genuinely new and the agent assets library

850
00:34:32,400 --> 00:34:36,800
deserves a real governance review before you let business critical workflows live there.

851
00:34:36,800 --> 00:34:40,080
Pick a low-stakes site, an internal team that will not be heard

852
00:34:40,080 --> 00:34:43,440
if something gets edited, deleted or executed unexpectedly

853
00:34:43,440 --> 00:34:44,720
and learn there.

854
00:34:44,720 --> 00:34:47,120
The interesting skills are the ones that touch content.

855
00:34:47,120 --> 00:34:49,840
The interesting failures are the ones that touch the wrong content.

856
00:34:49,840 --> 00:34:51,680
The fifth step is keeping the receipts.

857
00:34:51,680 --> 00:34:55,040
Don't tear down your Microsoft Syntax configuration documentation,

858
00:34:55,040 --> 00:34:57,840
your model training notes or your cost baselines yet.

859
00:34:57,840 --> 00:35:00,240
We're in the middle of a transition, the meter is still bill

860
00:35:00,240 --> 00:35:04,720
and Microsoft moved, my cheese is going to be a recurring conversation with your finance partner.

861
00:35:04,720 --> 00:35:07,680
The documentation you have now is an asset you'll want later.

862
00:35:07,680 --> 00:35:11,280
It proves what you built, how much it cost and what it delivered.

863
00:35:11,280 --> 00:35:14,240
That evidence is what secures budget for the next phase.

864
00:35:14,240 --> 00:35:17,440
Model types, structured, freeform and unstructured.

865
00:35:17,440 --> 00:35:20,480
Behind all the branding changes, the core technology hasn't changed.

866
00:35:20,480 --> 00:35:21,520
It's still models.

867
00:35:21,520 --> 00:35:25,360
Microsoft defines several model types tuned to different classes of documents.

868
00:35:25,360 --> 00:35:29,760
Understanding which type fits your content is the first real decision you need to make.

869
00:35:29,760 --> 00:35:33,840
Structured document processing is for forms and layouts with predictable fields.

870
00:35:33,840 --> 00:35:38,640
Invoices, purchase orders, tax forms and standardized applications all fall into this category

871
00:35:38,640 --> 00:35:41,520
because the fields appear in consistent locations and formats,

872
00:35:41,520 --> 00:35:45,600
structured models can map specific regions of a page to named columns,

873
00:35:45,600 --> 00:35:49,120
achieving high accuracy with relatively little training data.

874
00:35:49,120 --> 00:35:52,400
You upload a few sample documents, draw boxes around the fields you want,

875
00:35:52,400 --> 00:35:55,280
and the model learns to find those same fields on new documents.

876
00:35:55,280 --> 00:35:57,280
This is the fastest path to value.

877
00:35:57,280 --> 00:36:00,080
If your documents are mostly structured, start here,

878
00:36:00,080 --> 00:36:02,080
freeform document processing sits in the middle.

879
00:36:02,080 --> 00:36:06,560
These are documents that contain fields in variable positions but with recognizable patterns.

880
00:36:06,560 --> 00:36:10,400
A statement of work might list deliverables, timelines and budgets in different places

881
00:36:10,400 --> 00:36:13,440
depending on the client, but the language patterns are consistent.

882
00:36:13,440 --> 00:36:18,160
A health insurance claim might mix structured tables with freeform provider notes.

883
00:36:18,160 --> 00:36:21,760
Freeform models are backed by AI builder and use a combination of layout,

884
00:36:21,760 --> 00:36:23,360
understanding and language recognition.

885
00:36:23,360 --> 00:36:26,080
They require more training examples than structured models,

886
00:36:26,080 --> 00:36:29,200
but handle variability that rigid templates cannot capture.

887
00:36:29,200 --> 00:36:33,360
Unstructured document processing is the most complex and the most valuable for governance.

888
00:36:33,360 --> 00:36:35,840
Contracts, policies, memos and correspondence,

889
00:36:35,840 --> 00:36:39,360
lack fixed layouts and rely heavily on linguistic cues in context.

890
00:36:39,360 --> 00:36:41,600
An effective date might appear after the phrase,

891
00:36:41,600 --> 00:36:46,000
this agreement is effective as of in one contract and after commencement date in another.

892
00:36:46,000 --> 00:36:48,800
A liability cap might be expressed as a dollar amount,

893
00:36:48,800 --> 00:36:51,760
a percentage of revenue or a multiple of fees.

894
00:36:51,760 --> 00:36:55,040
Unstructured models use natural language processing and machine learning

895
00:36:55,040 --> 00:36:57,680
to infirm meaning from context, not just location.

896
00:36:57,680 --> 00:37:01,920
They need more training examples, more careful field definition and more iterative refinement,

897
00:37:01,920 --> 00:37:05,200
but they are the ones that transform document dumps into governed,

898
00:37:05,200 --> 00:37:06,880
searchable actionable libraries.

899
00:37:06,880 --> 00:37:11,440
For common business scenarios, Microsoft provides pre-built models that do not require training.

900
00:37:11,440 --> 00:37:13,440
These include models for contract processing,

901
00:37:13,440 --> 00:37:16,960
invoices, receipts, sensitive information and simple documents.

902
00:37:16,960 --> 00:37:21,120
These pre-built models aim to accelerate value by addressing everyday document types

903
00:37:21,120 --> 00:37:24,400
without requiring organizations to label their own training sets,

904
00:37:24,400 --> 00:37:27,520
while still allowing customization where domain specific needs arise.

905
00:37:27,520 --> 00:37:30,160
If your documents match one of these types,

906
00:37:30,160 --> 00:37:33,600
the pre-built model can get you live in hours instead of weeks.

907
00:37:33,600 --> 00:37:37,840
From a machine learning standpoint, unstructured document classification and extraction

908
00:37:37,840 --> 00:37:42,800
often use natural language processing models that encode text into numerical representations,

909
00:37:42,800 --> 00:37:46,960
using techniques such as term frequency or modern transformer based embeddings.

910
00:37:46,960 --> 00:37:51,520
Models including naive bays, support vector machines,

911
00:37:51,520 --> 00:37:54,240
logistic regression and deep learning architectures,

912
00:37:54,240 --> 00:37:57,040
like transformers or convolutional neural networks,

913
00:37:57,040 --> 00:38:02,160
can then be trained to map these representations to labels such as document type or sensitivity

914
00:38:02,160 --> 00:38:04,960
or to extract spans corresponding to entities.

915
00:38:04,960 --> 00:38:09,360
As organizations accumulate labeled examples from human review or syntax feedback,

916
00:38:09,360 --> 00:38:11,520
they can iteratively improve these models,

917
00:38:11,520 --> 00:38:15,520
whether through built-in training capabilities or external machine learning platforms.

918
00:38:15,520 --> 00:38:17,040
The choice depends on your documents.

919
00:38:17,040 --> 00:38:19,920
Start with pre-built if your documents match a standard type,

920
00:38:19,920 --> 00:38:22,080
move to structured if your forms are consistent,

921
00:38:22,080 --> 00:38:24,560
use free form for semi-structured variability,

922
00:38:24,560 --> 00:38:28,640
reserve unstructured for the complex narrative documents that carry the most business risk.

923
00:38:28,640 --> 00:38:30,640
And remember, you don't need a data science degree.

924
00:38:30,640 --> 00:38:35,200
The no-code interfaces in syntax and AI builder-led subject matter experts label documents

925
00:38:35,200 --> 00:38:37,920
defined fields and train models without writing code.

926
00:38:37,920 --> 00:38:40,720
The model learns from your expertise, not the other way around.

927
00:38:40,720 --> 00:38:43,200
Let's walk through a concrete decision scenario.

928
00:38:43,200 --> 00:38:47,280
Your finance department receives 500 invoices per month from 40 different vendors.

929
00:38:47,280 --> 00:38:49,200
Some vendors use their own templates.

930
00:38:49,200 --> 00:38:52,080
Other sent PDFs generated from accounting software.

931
00:38:52,080 --> 00:38:54,240
A few still send scanned paper invoices.

932
00:38:54,240 --> 00:38:55,920
The header information is always present,

933
00:38:55,920 --> 00:38:58,320
but the line items vary in number and layout.

934
00:38:58,320 --> 00:39:00,720
The total amount might appear at the bottom right,

935
00:39:00,720 --> 00:39:03,600
in a summary box, or in a payment instructions block.

936
00:39:03,600 --> 00:39:05,280
This is a semi-structured problem.

937
00:39:05,280 --> 00:39:08,000
A structured model would fail on the vendor-specific variations.

938
00:39:08,000 --> 00:39:10,960
An unstructured model would be overkill and slower to train.

939
00:39:10,960 --> 00:39:14,400
A free form model, possibly starting from the pre-built invoice processor,

940
00:39:14,400 --> 00:39:15,360
is the right fit.

941
00:39:15,360 --> 00:39:16,880
Now consider your legal department.

942
00:39:16,880 --> 00:39:18,720
They manage 2,000 active contracts,

943
00:39:18,720 --> 00:39:20,400
including master service agreements,

944
00:39:20,400 --> 00:39:23,200
statements of work, amendments, and non-disclosure agreements.

945
00:39:23,200 --> 00:39:25,280
Each law firm uses its own template.

946
00:39:25,280 --> 00:39:26,880
Some contracts are 50 pages.

947
00:39:26,880 --> 00:39:27,760
Others are 5.

948
00:39:27,760 --> 00:39:30,000
Govning law might appear in a dedicated section,

949
00:39:30,000 --> 00:39:33,440
in a footer, or in a governing law rider attached as a separate document.

950
00:39:33,440 --> 00:39:36,480
There are no fixed layouts, no consistent field positions.

951
00:39:36,480 --> 00:39:39,200
The language varies between plain English and dense legalese.

952
00:39:39,200 --> 00:39:40,720
This is an unstructured problem.

953
00:39:40,720 --> 00:39:44,160
Only an unstructured document processing model can extract effective dates,

954
00:39:44,160 --> 00:39:46,320
termination clauses, liability caps,

955
00:39:46,320 --> 00:39:48,640
and renewal terms across this variability.

956
00:39:48,640 --> 00:39:51,360
The pre-built contract model gives you a head start,

957
00:39:51,360 --> 00:39:55,680
but you will need to customize it heavily for your organization's specific contract language.

958
00:39:55,680 --> 00:39:59,920
The machine learning behind these models is more accessible than most people assume.

959
00:39:59,920 --> 00:40:01,920
For structured and free form documents,

960
00:40:01,920 --> 00:40:04,400
AI Builder uses a form recognizer

961
00:40:04,400 --> 00:40:07,120
that learns layout patterns from your labeled examples.

962
00:40:07,120 --> 00:40:10,080
It creates a geometric map of where fields typically appear

963
00:40:10,080 --> 00:40:12,480
and refines that map as it sees more documents.

964
00:40:12,480 --> 00:40:15,680
For unstructured documents, syntax uses natural language processing models

965
00:40:15,680 --> 00:40:18,240
that encode text into numerical representations.

966
00:40:18,240 --> 00:40:21,680
These representations capture semantic meaning not just keyword matching.

967
00:40:21,680 --> 00:40:24,320
The model learns that effective date and commencement date

968
00:40:24,320 --> 00:40:27,760
and start date are semantically equivalent in contract contexts.

969
00:40:27,760 --> 00:40:30,880
It learns that a dollar amount following liability shall not exceed

970
00:40:30,880 --> 00:40:32,160
is a liability cap,

971
00:40:32,160 --> 00:40:35,920
while a dollar amount following total amount due is an invoice total.

972
00:40:35,920 --> 00:40:39,120
This semantic understanding is what makes unstructured extraction powerful

973
00:40:39,120 --> 00:40:42,320
and it is why the training data must include linguistic diversity

974
00:40:42,320 --> 00:40:43,680
not just layout diversity.

975
00:40:43,680 --> 00:40:46,960
Model accuracy follows a predictable curve.

976
00:40:46,960 --> 00:40:51,120
With structured documents, you might reach 80% accuracy with 20 labeled examples

977
00:40:51,120 --> 00:40:52,800
and 95% with 50.

978
00:40:52,800 --> 00:40:54,880
With free form documents, the curve is flatter.

979
00:40:54,880 --> 00:40:59,040
You might need 100 examples to reach 80% and 300 to reach 95.

980
00:40:59,040 --> 00:41:01,920
With unstructured documents, the curve is steeper and longer.

981
00:41:01,920 --> 00:41:05,040
You might need 200 examples to reach 70% accuracy

982
00:41:05,040 --> 00:41:07,760
and continuous iteration to push past 85.

983
00:41:07,760 --> 00:41:10,240
The organizations that succeed plan for this curve

984
00:41:10,240 --> 00:41:12,880
they allocate time for multiple training cycles.

985
00:41:12,880 --> 00:41:15,680
They build feedback loops where users correct extractions

986
00:41:15,680 --> 00:41:17,520
and those corrections feedback into the model.

987
00:41:17,520 --> 00:41:19,680
They measure accuracy per field, not per document

988
00:41:19,680 --> 00:41:21,680
because some fields are harder than others.

989
00:41:21,680 --> 00:41:24,720
A model that extracts invoice totals at 98% accuracy

990
00:41:24,720 --> 00:41:28,800
but vendor names at 70% accuracy needs more training on vendor name variants

991
00:41:28,800 --> 00:41:30,320
not more training on totals.

992
00:41:30,320 --> 00:41:33,040
AI agents, skills and the new operating model

993
00:41:33,040 --> 00:41:35,120
but choosing the model type is just the warm-up.

994
00:41:35,120 --> 00:41:36,800
The real work starts when you train it

995
00:41:36,800 --> 00:41:40,160
and in 2026 the training environment itself just changed.

996
00:41:40,160 --> 00:41:42,480
The architectural shift that Microsoft is pushing

997
00:41:42,480 --> 00:41:45,200
is a move from feature by feature syntax configuration

998
00:41:45,200 --> 00:41:48,080
to conversational intent-based AI agents.

999
00:41:48,080 --> 00:41:49,840
Every SharePoint site now has an agent.

1000
00:41:49,840 --> 00:41:52,800
The SharePoint admin agent monitors governance automatically,

1001
00:41:52,800 --> 00:41:56,000
identifying inactive sites, flagging overshared content,

1002
00:41:56,000 --> 00:41:59,040
tracking permissions sprawl and highlighting high activity sites

1003
00:41:59,040 --> 00:42:02,000
that may need additional attention as co-pilot adoption grows.

1004
00:42:02,000 --> 00:42:04,640
Rather than requiring manual intervention,

1005
00:42:04,640 --> 00:42:06,880
the admin agent automatically applies policies

1006
00:42:06,880 --> 00:42:08,720
such as archiving inactive sites

1007
00:42:08,720 --> 00:42:11,840
or adjusting access permissions to reduce security risks.

1008
00:42:11,840 --> 00:42:15,840
Site-specific agents create instant knowledge bases for every SharePoint site.

1009
00:42:15,840 --> 00:42:19,360
These agents understand the content, structure and purpose of their respective sites

1010
00:42:19,360 --> 00:42:22,160
allowing employees to quickly tap into relevant information

1011
00:42:22,160 --> 00:42:24,320
without complex searches or navigation.

1012
00:42:24,320 --> 00:42:26,000
Site agents can answer questions like

1013
00:42:26,000 --> 00:42:28,480
"What are the key deliverables for the current quarter?"

1014
00:42:28,480 --> 00:42:32,000
By synthesizing information from documents, lists, and pages within their site.

1015
00:42:32,000 --> 00:42:35,120
Organizations can also build custom agents directly

1016
00:42:35,120 --> 00:42:37,600
within SharePoint using co-pilot studio integration.

1017
00:42:37,600 --> 00:42:39,440
These agents connect to SharePoint lists,

1018
00:42:39,440 --> 00:42:42,080
document libraries and pages as knowledge sources,

1019
00:42:42,080 --> 00:42:44,400
providing real-time access to current data.

1020
00:42:44,400 --> 00:42:47,440
Custom agents handle specific business processes such as employee

1021
00:42:47,440 --> 00:42:49,680
onboarding assistance, policy question answering,

1022
00:42:49,680 --> 00:42:52,240
project status updates, or equipment request management,

1023
00:42:52,240 --> 00:42:55,040
all without requiring extensive development resources.

1024
00:42:55,040 --> 00:42:56,480
Skills are the new automation layer.

1025
00:42:56,480 --> 00:43:00,160
A skill is a reusable AI workflow stored as a markdown definition

1026
00:43:00,160 --> 00:43:02,240
in the site's agent-assets library.

1027
00:43:02,240 --> 00:43:05,840
Skills describe repeatable AI workflows over the site's content,

1028
00:43:05,840 --> 00:43:08,560
such as review, enrich or root operations.

1029
00:43:08,560 --> 00:43:10,240
You write a skill in natural language,

1030
00:43:10,240 --> 00:43:11,840
describing what you want to happen,

1031
00:43:11,840 --> 00:43:13,840
and the agent executes it across the site.

1032
00:43:13,840 --> 00:43:16,160
For example, you might write a skill that says

1033
00:43:16,160 --> 00:43:18,720
"For every invoice uploaded to the finance library,

1034
00:43:18,720 --> 00:43:21,920
extract the vendor name, invoice number, and total amount,

1035
00:43:21,920 --> 00:43:24,560
then send a notification to the accounts payable team

1036
00:43:24,560 --> 00:43:26,320
if the total exceeds $10,000."

1037
00:43:26,320 --> 00:43:29,440
The skill chain's extraction, metadata update,

1038
00:43:29,440 --> 00:43:32,640
list operation and notification into a single automated workflow.

1039
00:43:32,640 --> 00:43:35,120
This sounds like the future and in many ways it is,

1040
00:43:35,120 --> 00:43:37,120
but there are drawbacks nobody is talking about.

1041
00:43:37,120 --> 00:43:39,200
Skills have no version control and no rollback.

1042
00:43:39,200 --> 00:43:40,960
They are stored as markdown in SharePoint,

1043
00:43:40,960 --> 00:43:43,200
edit and save, and it's live immediately.

1044
00:43:43,200 --> 00:43:46,080
There is no git history, no undo and no sandbox environment.

1045
00:43:46,080 --> 00:43:49,280
You write the skill, submit it, and it runs on real data.

1046
00:43:49,280 --> 00:43:51,680
One mistake and you've modified thousands of records,

1047
00:43:51,680 --> 00:43:53,600
and by default everyone can create a skill

1048
00:43:53,600 --> 00:43:55,440
that is a governance problem waiting to happen.

1049
00:43:55,440 --> 00:43:58,000
Autofill columns represent another major shift.

1050
00:43:58,000 --> 00:43:59,840
In the syntax era, you build a model,

1051
00:43:59,840 --> 00:44:02,400
bound it to a library and configured extractors.

1052
00:44:02,400 --> 00:44:06,080
In the AI era, users just ask co-pilot to fill library columns,

1053
00:44:06,080 --> 00:44:08,400
no model galleries, no complex binding,

1054
00:44:08,400 --> 00:44:10,000
just a natural language prompt.

1055
00:44:10,000 --> 00:44:12,000
But Autofill has its own limitations.

1056
00:44:12,000 --> 00:44:13,920
Configurations live in your document library

1057
00:44:13,920 --> 00:44:15,840
and can't be reused in another library.

1058
00:44:15,840 --> 00:44:18,560
Each library requires its own separate setup.

1059
00:44:18,560 --> 00:44:20,560
Testing is limited to one file at a time,

1060
00:44:20,560 --> 00:44:22,560
with thousands of files you're flying blind.

1061
00:44:22,560 --> 00:44:24,240
And when you run batch prompt testing,

1062
00:44:24,240 --> 00:44:26,320
there's no preview of what's actually happening.

1063
00:44:26,320 --> 00:44:28,160
You submit the batch and hope it works.

1064
00:44:28,160 --> 00:44:30,400
No visibility into errors or extraction quality

1065
00:44:30,400 --> 00:44:31,600
until it's too late.

1066
00:44:31,600 --> 00:44:33,600
There is also a critical architectural gap.

1067
00:44:33,600 --> 00:44:36,320
There is no native ability to perform sensitivity

1068
00:44:36,320 --> 00:44:39,520
and retention labeling with AI in SharePoint currently.

1069
00:44:39,520 --> 00:44:41,440
The agents and skills can extract metadata,

1070
00:44:41,440 --> 00:44:43,440
classified documents and update columns,

1071
00:44:43,440 --> 00:44:45,120
but they do not natively apply

1072
00:44:45,120 --> 00:44:48,080
per view sensitivity labels or retention policies.

1073
00:44:48,080 --> 00:44:50,480
That means organizations still need SharePoint's

1074
00:44:50,480 --> 00:44:53,360
native governance controls, power automate orchestration,

1075
00:44:53,360 --> 00:44:55,200
or custom integrations to bridge the gap

1076
00:44:55,200 --> 00:44:58,240
between extracted metadata and enforced compliance.

1077
00:44:58,240 --> 00:45:01,280
The AI layer handles extraction and classification,

1078
00:45:01,280 --> 00:45:03,280
per view handles enforcement.

1079
00:45:03,280 --> 00:45:04,960
And the connection between them is still manual.

1080
00:45:04,960 --> 00:45:07,120
The one-to-one mapping from syntax to AI

1081
00:45:07,120 --> 00:45:08,960
in SharePoint is helpful for architects

1082
00:45:08,960 --> 00:45:10,480
trying to manage the transition.

1083
00:45:10,480 --> 00:45:12,240
Syntax Autofill columns map directly

1084
00:45:12,240 --> 00:45:14,240
to the new create Autofill columns experience

1085
00:45:14,240 --> 00:45:15,760
in AI enabled libraries.

1086
00:45:15,760 --> 00:45:18,240
Document translation remains a metered content service

1087
00:45:18,240 --> 00:45:20,480
with copilot providing conversational entry points

1088
00:45:20,480 --> 00:45:21,840
for ad hoc translation.

1089
00:45:21,840 --> 00:45:25,360
OCR becomes part of the baseline AI ready content pipeline,

1090
00:45:25,360 --> 00:45:27,520
triggered when users initiate interactions

1091
00:45:27,520 --> 00:45:30,000
rather than running automatically on every upload.

1092
00:45:30,000 --> 00:45:32,960
Content assembly shifts to template driven document generation

1093
00:45:32,960 --> 00:45:36,560
in content AI combined with copilot driven content generation.

1094
00:45:36,560 --> 00:45:39,040
Image tagging and taxonomy tagging are internalized

1095
00:45:39,040 --> 00:45:41,280
into the broader AI ready content concept,

1096
00:45:41,280 --> 00:45:43,520
surfaced through improved search and Autofill

1097
00:45:43,520 --> 00:45:45,600
rather than standalone configurations.

1098
00:45:45,600 --> 00:45:48,640
Prebuilt, structured, freeform, and unstructured models

1099
00:45:48,640 --> 00:45:50,240
all form the processing substrate

1100
00:45:50,240 --> 00:45:51,920
for higher level agent actions.

1101
00:45:51,920 --> 00:45:54,800
What this means in practice is that your existing syntax models

1102
00:45:54,800 --> 00:45:55,920
are not wasted.

1103
00:45:55,920 --> 00:45:58,160
They still build through Azure payers you go meters.

1104
00:45:58,160 --> 00:46:00,160
They still run, they still extract,

1105
00:46:00,160 --> 00:46:02,720
but their role is shifting from primary user interface

1106
00:46:02,720 --> 00:46:04,560
to back and substrate for agent actions.

1107
00:46:04,560 --> 00:46:07,120
The user-facing experience is now conversational.

1108
00:46:07,120 --> 00:46:10,160
The configuration experience is now intent-based

1109
00:46:10,160 --> 00:46:12,080
and the governance experience is now distributed

1110
00:46:12,080 --> 00:46:13,840
across every site in your tenant.

1111
00:46:13,840 --> 00:46:16,400
That is a structural change, not a cosmetic one.

1112
00:46:16,400 --> 00:46:18,080
Architects who understand this shift

1113
00:46:18,080 --> 00:46:20,800
will design solutions that survive the transition.

1114
00:46:20,800 --> 00:46:23,120
Those who don't will find their carefully trained models

1115
00:46:23,120 --> 00:46:26,480
often behind a user interface that no longer exists.

1116
00:46:26,480 --> 00:46:28,880
The knowledge agent is another major piece of this puzzle

1117
00:46:28,880 --> 00:46:31,600
that entered public preview in September 2025

1118
00:46:31,600 --> 00:46:34,480
and reached general availability in early 2026.

1119
00:46:34,480 --> 00:46:37,360
It is included with Microsoft 365 co-pilot licenses

1120
00:46:37,360 --> 00:46:39,040
at no extra cost.

1121
00:46:39,040 --> 00:46:40,960
Knowledge agent uses artificial intelligence

1122
00:46:40,960 --> 00:46:44,000
to automatically tag and classify documents with metadata,

1123
00:46:44,000 --> 00:46:46,080
eliminating hours of manual data entry.

1124
00:46:46,080 --> 00:46:48,400
When you upload financial reports or project plans,

1125
00:46:48,400 --> 00:46:51,920
the agent analyzes content and suggests relevant metadata columns

1126
00:46:51,920 --> 00:46:55,360
ensuring consistent organization across libraries and sites.

1127
00:46:55,360 --> 00:46:56,720
Early adopters like Mars Inc.

1128
00:46:56,720 --> 00:46:59,120
called the Autofill metadata capability

1129
00:46:59,120 --> 00:47:02,160
a total game-changer drastically reducing time spent

1130
00:47:02,160 --> 00:47:03,600
on manual categorization.

1131
00:47:03,600 --> 00:47:06,160
The agent also continuously monitors sharepoint environments

1132
00:47:06,160 --> 00:47:07,760
to maintain content quality.

1133
00:47:07,760 --> 00:47:10,800
It identifies broken links, flags outdated pages,

1134
00:47:10,800 --> 00:47:12,400
detects duplicate content,

1135
00:47:12,400 --> 00:47:15,120
and locates missing information that users are searching for,

1136
00:47:15,120 --> 00:47:16,720
but cannot find.

1137
00:47:16,720 --> 00:47:18,640
Rather than waiting for manual audits,

1138
00:47:18,640 --> 00:47:20,800
knowledge agent proactively keeps your sharepoint

1139
00:47:20,800 --> 00:47:22,400
instance clean and current.

1140
00:47:22,400 --> 00:47:25,200
A floating button appears throughout sharepoint surfaces,

1141
00:47:25,200 --> 00:47:27,200
providing role-specific actions based

1142
00:47:27,200 --> 00:47:28,320
on what you're working on.

1143
00:47:28,320 --> 00:47:31,200
Site owners see options to improve their sites and fill content gaps

1144
00:47:31,200 --> 00:47:32,960
while content creators receive suggestions

1145
00:47:32,960 --> 00:47:35,600
for metadata enrichment and automated workflows.

1146
00:47:35,600 --> 00:47:39,200
Automated workflow creation is where the agent gets genuinely powerful.

1147
00:47:39,200 --> 00:47:41,040
You describe what you need in natural language,

1148
00:47:41,040 --> 00:47:44,320
such as identifying new invoices exceeding a specific threshold

1149
00:47:44,320 --> 00:47:47,200
and knowledge agent builds the workflow automatically.

1150
00:47:47,200 --> 00:47:49,440
This removes the technical barrier to automation,

1151
00:47:49,440 --> 00:47:52,000
allowing business users to create sophisticated processes

1152
00:47:52,000 --> 00:47:53,200
without coding knowledge.

1153
00:47:53,200 --> 00:47:55,280
But this power comes with the same governance warnings

1154
00:47:55,280 --> 00:47:56,560
that apply to skills.

1155
00:47:56,560 --> 00:47:59,920
A workflow built by an agent is still a workflow running on real data.

1156
00:47:59,920 --> 00:48:02,480
If the description is ambiguous, the result might be wrong.

1157
00:48:02,480 --> 00:48:04,560
If the threshold is mistated, the notification

1158
00:48:04,560 --> 00:48:05,760
goes to the wrong person.

1159
00:48:05,760 --> 00:48:08,160
Natural language is fast but imprecise.

1160
00:48:08,160 --> 00:48:10,000
Precision requires governance.

1161
00:48:10,000 --> 00:48:11,920
The SharePoint page agent enables users

1162
00:48:11,920 --> 00:48:14,320
with Microsoft 365 co-pilot licenses

1163
00:48:14,320 --> 00:48:17,760
to create and refine SharePoint pages using natural language.

1164
00:48:17,760 --> 00:48:19,840
Rather than manually building page layouts

1165
00:48:19,840 --> 00:48:23,200
and adding web parts, content creators describe what they need,

1166
00:48:23,200 --> 00:48:25,920
and the agent generates professionally designed pages

1167
00:48:25,920 --> 00:48:28,720
that match organizational branding and standards.

1168
00:48:28,720 --> 00:48:30,880
This accelerates content creation dramatically

1169
00:48:30,880 --> 00:48:34,560
but it also raises questions about content quality and consistency.

1170
00:48:34,560 --> 00:48:36,400
An AI-generated page might look correct

1171
00:48:36,400 --> 00:48:38,480
while containing outdated policy language

1172
00:48:38,480 --> 00:48:40,480
or incorrect contact information.

1173
00:48:40,480 --> 00:48:42,800
Human review remains needed, especially for pages

1174
00:48:42,800 --> 00:48:45,440
that serve as authoritative sources for co-pilot grounding.

1175
00:48:45,440 --> 00:48:48,720
Custom no-code agents built through co-pilot studio integration

1176
00:48:48,720 --> 00:48:51,520
represent the most flexible layer of this new architecture.

1177
00:48:51,520 --> 00:48:54,640
Organizations can create agents that connect to SharePoint lists,

1178
00:48:54,640 --> 00:48:57,440
document libraries and pages as knowledge sources,

1179
00:48:57,440 --> 00:48:59,760
providing real-time access to current data.

1180
00:48:59,760 --> 00:49:01,840
These agents handle specific business processes

1181
00:49:01,840 --> 00:49:04,640
such as employee onboarding assistance, policy question answering,

1182
00:49:04,640 --> 00:49:07,360
project status updates, or equipment request management.

1183
00:49:07,360 --> 00:49:09,520
Because they live inside the SharePoint environment,

1184
00:49:09,520 --> 00:49:10,880
they can reference current data

1185
00:49:10,880 --> 00:49:13,120
without the latency of external integrations.

1186
00:49:13,120 --> 00:49:15,600
But they also inherit the permissions and visibility rules

1187
00:49:15,600 --> 00:49:16,960
of the sites they access.

1188
00:49:16,960 --> 00:49:19,520
An agent that answers questions about salary ranges

1189
00:49:19,520 --> 00:49:22,400
must respect the permissions of the HR site at queries.

1190
00:49:22,400 --> 00:49:24,560
An agent that summarizes project status

1191
00:49:24,560 --> 00:49:26,960
must not expose confidential client information

1192
00:49:26,960 --> 00:49:28,480
from a restricted library.

1193
00:49:28,480 --> 00:49:30,160
Agent governance is site governance

1194
00:49:30,160 --> 00:49:32,480
and the two cannot be separated.

1195
00:49:32,480 --> 00:49:34,960
Building your first custom document processing model.

1196
00:49:34,960 --> 00:49:37,280
Building a custom model that survives this transition

1197
00:49:37,280 --> 00:49:38,880
is less about machine learning theory

1198
00:49:38,880 --> 00:49:40,720
and more about document archaeology.

1199
00:49:40,720 --> 00:49:43,680
You need to understand your content before the machine can.

1200
00:49:43,680 --> 00:49:45,760
Step one is auditing your documents.

1201
00:49:45,760 --> 00:49:48,160
Pull a representative sample of every document type

1202
00:49:48,160 --> 00:49:49,120
you want to process.

1203
00:49:49,120 --> 00:49:50,640
If you're starting with contracts,

1204
00:49:50,640 --> 00:49:52,400
gather master service agreements,

1205
00:49:52,400 --> 00:49:53,600
statements of work,

1206
00:49:53,600 --> 00:49:56,000
non-disclosure agreements, and amendment documents.

1207
00:49:56,000 --> 00:49:57,840
Include the good scans, the bad scans,

1208
00:49:57,840 --> 00:49:59,520
the digital word files, and the PDFs

1209
00:49:59,520 --> 00:50:01,440
that somebody printed and scanned back in.

1210
00:50:01,440 --> 00:50:02,880
The model needs to see the mess.

1211
00:50:02,880 --> 00:50:04,400
Not just the clean examples.

1212
00:50:04,400 --> 00:50:06,080
Diversity matters more than volume.

1213
00:50:06,080 --> 00:50:09,440
A hundred varied documents teach more than a thousand identical ones.

1214
00:50:09,440 --> 00:50:11,440
Step two is designing the taxonomy first.

1215
00:50:11,440 --> 00:50:14,240
Before you open the model builder, open a whiteboard.

1216
00:50:14,240 --> 00:50:16,240
List every column your library needs.

1217
00:50:16,240 --> 00:50:18,560
Contract type, counterparty name, effective date,

1218
00:50:18,560 --> 00:50:20,560
expiration date, governing law jurisdiction,

1219
00:50:20,560 --> 00:50:22,560
liability cap, auto renewal flag,

1220
00:50:22,560 --> 00:50:24,320
termination notice period.

1221
00:50:24,320 --> 00:50:26,400
Each of these fields maps to a SharePoint column

1222
00:50:26,400 --> 00:50:28,960
and each column needs a data type and a vocabulary.

1223
00:50:28,960 --> 00:50:30,640
If you're using managed metadata,

1224
00:50:30,640 --> 00:50:32,160
define your term sets now.

1225
00:50:32,160 --> 00:50:34,560
If you're using choice columns, list the options.

1226
00:50:34,560 --> 00:50:35,760
The model isn't the hard part.

1227
00:50:35,760 --> 00:50:36,800
The taxonomy is.

1228
00:50:36,800 --> 00:50:39,440
And if your taxonomy changes after you've trained the model,

1229
00:50:39,440 --> 00:50:40,400
you're retraining.

1230
00:50:40,400 --> 00:50:43,200
Step three is selecting and uploading your training set.

1231
00:50:43,200 --> 00:50:46,880
In the syntax model builder or AI builder interface,

1232
00:50:46,880 --> 00:50:49,680
you upload your sample documents and begin labeling.

1233
00:50:49,680 --> 00:50:51,040
For structured and freeform models,

1234
00:50:51,040 --> 00:50:54,080
you draw bounding boxes around the fields you want to extract.

1235
00:50:54,080 --> 00:50:55,360
For unstructured models,

1236
00:50:55,360 --> 00:50:58,240
you highlight text spans and assign them to named entities.

1237
00:50:58,240 --> 00:51:00,560
The interface is visual and requires no code.

1238
00:51:00,560 --> 00:51:03,120
A contract manager or legal analyst can do this work.

1239
00:51:03,120 --> 00:51:04,880
The model learns from their expertise.

1240
00:51:04,880 --> 00:51:07,520
Step four is defining your entities with precision.

1241
00:51:07,520 --> 00:51:09,840
An expiration date is not just any date on the page.

1242
00:51:09,840 --> 00:51:11,120
It's the date after the phrase,

1243
00:51:11,120 --> 00:51:12,880
"This agreement shall terminate on"

1244
00:51:12,880 --> 00:51:14,560
or the date in the term section.

1245
00:51:14,560 --> 00:51:17,120
A liability cap is not just any dollar amount.

1246
00:51:17,120 --> 00:51:21,200
It's the number after liability shall not exceed or cap on damages.

1247
00:51:21,200 --> 00:51:24,800
The more precisely you define the linguistic context around each field,

1248
00:51:24,800 --> 00:51:27,200
the more accurately the model will extract it.

1249
00:51:27,200 --> 00:51:29,600
Vague field definitions produce vague results.

1250
00:51:29,600 --> 00:51:31,920
Step five is training, testing and validating.

1251
00:51:31,920 --> 00:51:36,240
The model builder splits your label documents into training and test sets automatically.

1252
00:51:36,240 --> 00:51:39,360
After training, you review the confidence scores for each extraction.

1253
00:51:39,360 --> 00:51:44,400
A confidence score above 80% is generally reliable for automated processing.

1254
00:51:44,400 --> 00:51:47,440
A score between 50 and 80% should flag for human review.

1255
00:51:47,440 --> 00:51:51,440
Below 50%, you need more training examples or better field definitions.

1256
00:51:51,440 --> 00:51:53,760
Test on documents that weren't in your training set.

1257
00:51:53,760 --> 00:51:54,960
Test on the edge cases.

1258
00:51:54,960 --> 00:51:57,280
Test on the blurry scans and the odd formatting.

1259
00:51:57,280 --> 00:51:59,600
If the model fails there, it will fail in production.

1260
00:51:59,600 --> 00:52:01,040
The no-code promise is real.

1261
00:52:01,040 --> 00:52:04,080
Subject matter experts can build these models without writing Python

1262
00:52:04,080 --> 00:52:06,160
or understanding neural network architecture.

1263
00:52:06,160 --> 00:52:08,720
But the no-code label is misleading in one important way.

1264
00:52:08,720 --> 00:52:09,920
It doesn't mean no expertise.

1265
00:52:09,920 --> 00:52:11,760
It means the expertise is domain expertise.

1266
00:52:11,760 --> 00:52:12,720
Not data science.

1267
00:52:12,720 --> 00:52:15,680
The person labeling contracts needs to understand contract structure.

1268
00:52:15,680 --> 00:52:19,760
The person defining entity extractors needs to know where exploration dates hide

1269
00:52:19,760 --> 00:52:21,280
in their organization's documents.

1270
00:52:21,280 --> 00:52:23,120
The model learns from human pattern recognition.

1271
00:52:23,120 --> 00:52:25,760
If the human doesn't recognize the pattern, the machine won't either.

1272
00:52:25,760 --> 00:52:30,320
Microsoft provides pre-built models that accelerate this process for common scenarios.

1273
00:52:30,320 --> 00:52:34,480
The pre-built contract processing model can identify contract type, parties,

1274
00:52:34,480 --> 00:52:38,320
dates and other standard elements without requiring extensive training.

1275
00:52:38,320 --> 00:52:41,360
The pre-built invoice and receipt models extract totals,

1276
00:52:41,360 --> 00:52:43,040
dates and vendor information.

1277
00:52:43,040 --> 00:52:44,880
If your documents match these patterns,

1278
00:52:44,880 --> 00:52:47,520
start with the pre-built model and customize from there.

1279
00:52:47,520 --> 00:52:51,200
You'll get to production faster and you'll have a baseline for accuracy

1280
00:52:51,200 --> 00:52:52,960
that you can improve iteratively.

1281
00:52:52,960 --> 00:52:56,560
Pay as you go billing means you only pay for what you process.

1282
00:52:56,560 --> 00:53:00,880
Each page, sheet, slide or file transaction is meted through your attached

1283
00:53:00,880 --> 00:53:02,160
as your subscription.

1284
00:53:02,160 --> 00:53:04,720
This makes it feasible to run proof-of-concept projects

1285
00:53:04,720 --> 00:53:06,960
without committing to large upfront licenses.

1286
00:53:06,960 --> 00:53:09,440
Process a few hundred documents, measure the accuracy,

1287
00:53:09,440 --> 00:53:12,640
calculate the time savings and build the business case from real data.

1288
00:53:12,640 --> 00:53:15,600
The meters continue to exist behind the scenes in the AI era

1289
00:53:15,600 --> 00:53:18,720
and Azure Cost Management lets you monitor consumption as you scale.

1290
00:53:18,720 --> 00:53:21,040
There is a critical limitation you need to plan for.

1291
00:53:21,040 --> 00:53:24,320
Autofill configurations in the AI era live in your document library.

1292
00:53:24,320 --> 00:53:26,160
You can't reuse them in another library.

1293
00:53:26,160 --> 00:53:28,400
Each library requires its own separate setup.

1294
00:53:28,400 --> 00:53:30,320
There is no sharing, no inheritance.

1295
00:53:30,320 --> 00:53:32,800
If you have 50 contract libraries across your tenant,

1296
00:53:32,800 --> 00:53:34,800
you may need 50 separate configurations.

1297
00:53:34,800 --> 00:53:35,920
That is not a scaling model.

1298
00:53:35,920 --> 00:53:37,280
It is a point solution.

1299
00:53:37,280 --> 00:53:40,240
For enterprise deployment, you need to think about whether the new

1300
00:53:40,240 --> 00:53:42,880
autofill approach fits your architecture or whether you should stick

1301
00:53:42,880 --> 00:53:47,280
with classic syntax model binding that can be applied across libraries via content types.

1302
00:53:47,280 --> 00:53:51,680
Power Automate remains the connective tissue between extracted metadata

1303
00:53:51,680 --> 00:53:52,960
and downstream action.

1304
00:53:52,960 --> 00:53:55,440
When a syntax model extracts an expiration date,

1305
00:53:55,440 --> 00:53:57,840
a power automate flow can read that column value,

1306
00:53:57,840 --> 00:54:01,120
calculate the 90-day warning threshold, create a planner task

1307
00:54:01,120 --> 00:54:03,040
and send an email to the contract owner.

1308
00:54:03,040 --> 00:54:06,000
When a model classifies a document as containing personal data,

1309
00:54:06,000 --> 00:54:08,560
a flow can trigger a sensitivity label application,

1310
00:54:08,560 --> 00:54:10,640
log the event to a compliance register,

1311
00:54:10,640 --> 00:54:12,480
and notify the data protection officer.

1312
00:54:12,480 --> 00:54:15,360
These orchestrations are not optional extras.

1313
00:54:15,360 --> 00:54:18,720
They are what turns extracted metadata into business process automation.

1314
00:54:18,720 --> 00:54:23,280
Error handling in production workflows is often overlooked during pilot deployments.

1315
00:54:23,280 --> 00:54:26,080
Your production flows need branches for model failures,

1316
00:54:26,080 --> 00:54:29,840
low-confidence scores, corrupted files, and password protected PDFs.

1317
00:54:29,840 --> 00:54:32,800
A missing expiration date should not create a blank task.

1318
00:54:32,800 --> 00:54:35,360
A missing expiration date should not create a blank task.

1319
00:54:35,360 --> 00:54:38,640
A low-confidence extraction should root to a human reviewer queue.

1320
00:54:38,640 --> 00:54:41,280
A processing failure should log to an admin dashboard,

1321
00:54:41,280 --> 00:54:42,960
not silently drop the document.

1322
00:54:42,960 --> 00:54:45,120
Build these exception paths before you scale,

1323
00:54:45,120 --> 00:54:47,600
because exceptions become common at volume.

1324
00:54:47,600 --> 00:54:51,040
Document quality is another production reality that pilots often ignore.

1325
00:54:51,040 --> 00:54:53,920
A training set of pristine digital PDFs produces a model

1326
00:54:53,920 --> 00:54:56,080
that fails on the fact scanned, coffee stains,

1327
00:54:56,080 --> 00:54:58,560
sideways-oriented documents that arrive in production.

1328
00:54:58,560 --> 00:55:00,720
Your intake process needs quality gates.

1329
00:55:00,720 --> 00:55:03,120
Reject scans below 300 dots per inch.

1330
00:55:03,120 --> 00:55:05,040
Flag documents with missing pages.

1331
00:55:05,040 --> 00:55:08,000
Alert users when a file is encrypted or password protected.

1332
00:55:08,000 --> 00:55:10,480
The model cannot process what it cannot read

1333
00:55:10,480 --> 00:55:13,600
and failed processing costs money without delivering value.

1334
00:55:13,600 --> 00:55:18,480
A small amount of intake validation prevents a large amount of downstream rework.

1335
00:55:18,480 --> 00:55:22,080
Integration with external systems is where the architecture gets interesting.

1336
00:55:22,080 --> 00:55:27,280
Documents arrive through email, third-party portals, migration batches, and API uploads.

1337
00:55:27,280 --> 00:55:30,400
Each of these channels needs to land the file in a SharePoint library,

1338
00:55:30,400 --> 00:55:33,600
trigger the model, and handle the metadata output consistently.

1339
00:55:33,600 --> 00:55:36,960
A common pattern is to create a staging library for incoming documents,

1340
00:55:36,960 --> 00:55:39,680
run the model there, and then move the classified document

1341
00:55:39,680 --> 00:55:43,280
to its final destination library based on the extracted metadata.

1342
00:55:43,280 --> 00:55:45,200
This decouples intake from storage,

1343
00:55:45,200 --> 00:55:47,360
allowing you to apply different processing rules

1344
00:55:47,360 --> 00:55:51,200
to different document sources without complicating the destination libraries.

1345
00:55:51,200 --> 00:55:54,720
Monitoring and telemetry complete the production picture.

1346
00:55:54,720 --> 00:55:57,280
Azure application insights can track model performance,

1347
00:55:57,280 --> 00:55:59,520
processing latency, and error rates.

1348
00:55:59,520 --> 00:56:01,840
SharePoint audit logs record who uploaded what,

1349
00:56:01,840 --> 00:56:04,240
when the model ran, and what metadata was applied.

1350
00:56:04,240 --> 00:56:08,720
Power BI dashboards can visualize throughput, accuracy trends, and cost per document.

1351
00:56:08,720 --> 00:56:11,120
These monitoring layers are not just for troubleshooting,

1352
00:56:11,120 --> 00:56:12,640
they are for continuous improvement.

1353
00:56:12,640 --> 00:56:14,560
When accuracy drops, the dashboard tells you.

1354
00:56:14,560 --> 00:56:16,320
When costs spike the alert fires,

1355
00:56:16,320 --> 00:56:18,480
when a new document type appears in the intake,

1356
00:56:18,480 --> 00:56:21,040
the pattern detection flags it for model retraining.

1357
00:56:21,040 --> 00:56:23,920
Entity extraction, deep dive, the technical core.

1358
00:56:23,920 --> 00:56:26,800
Entity extractors are where beginners become professionals.

1359
00:56:26,800 --> 00:56:29,680
Simple field mapping finds text in a fixed location.

1360
00:56:29,680 --> 00:56:33,200
Entity extraction finds text based on context, pattern, and meaning.

1361
00:56:33,200 --> 00:56:36,080
That difference is everything when you're processing contracts,

1362
00:56:36,080 --> 00:56:37,680
policies, and legal documents.

1363
00:56:37,680 --> 00:56:39,360
Let's start with a concrete example.

1364
00:56:39,360 --> 00:56:42,400
A master service agreement might contain an effective date.

1365
00:56:42,400 --> 00:56:47,760
In one document, it appears after this agreement is effective as of January 15, 2024.

1366
00:56:47,760 --> 00:56:51,760
In another, it appears after commencement date 15th January 2024.

1367
00:56:51,760 --> 00:56:55,920
In a third, it appears in a table labeled Key Terms with a row for start date.

1368
00:56:55,920 --> 00:56:58,880
A simple field map I would miss two of these three instances.

1369
00:56:58,880 --> 00:57:03,360
An entity extractor trained on the linguistic patterns around effective dates would catch all three.

1370
00:57:03,360 --> 00:57:07,120
The extractor works by learning the relationship between a label and a value.

1371
00:57:07,120 --> 00:57:10,960
You highlight multiple examples of the same entity across different documents

1372
00:57:10,960 --> 00:57:12,640
and the model generalizes the pattern.

1373
00:57:12,640 --> 00:57:15,520
It learns that effective dates often follow phrases like

1374
00:57:15,520 --> 00:57:20,080
effective as of, commences on, start date, or term begins.

1375
00:57:20,080 --> 00:57:23,040
It learns that expiration dates follow terminates on,

1376
00:57:23,040 --> 00:57:25,680
end date, expiration, or renewal deadline.

1377
00:57:26,640 --> 00:57:30,480
The more examples you provide, the stronger the pattern recognition becomes.

1378
00:57:30,480 --> 00:57:33,360
Liability clauses are another high value extraction target.

1379
00:57:33,360 --> 00:57:35,840
A liability cap might be expressed as a fixed dollar amount,

1380
00:57:35,840 --> 00:57:38,800
a percentage of fees paid, a multiple of annual charges,

1381
00:57:38,800 --> 00:57:40,240
or an insurance coverage limit.

1382
00:57:40,240 --> 00:57:44,320
The extractor needs to understand that liability shall not exceed $500,000

1383
00:57:44,320 --> 00:57:49,920
and total liability capped at $500K and damages limited to the greater of $500,000,

1384
00:57:49,920 --> 00:57:53,840
or 12 months of fees, or represent the same conceptual entity.

1385
00:57:53,840 --> 00:57:58,960
You train this by labeling each variant and letting the model learn the semantic equivalence.

1386
00:57:58,960 --> 00:58:01,760
Termination clauses require similar semantic understanding.

1387
00:58:01,760 --> 00:58:03,760
Some contracts terminate on a fixed date,

1388
00:58:03,760 --> 00:58:05,520
others terminate after a notice period,

1389
00:58:05,520 --> 00:58:08,400
others terminate upon a material breach, others auto-renew,

1390
00:58:08,400 --> 00:58:10,560
unless one party provides written notice.

1391
00:58:10,560 --> 00:58:13,120
Each of these represents a different termination pattern

1392
00:58:13,120 --> 00:58:16,240
and your taxonomy needs columns that capture the distinction.

1393
00:58:16,240 --> 00:58:18,080
A single end date column is insufficient.

1394
00:58:18,080 --> 00:58:20,880
You need termination type, notice period,

1395
00:58:20,880 --> 00:58:24,160
auto-renewal flag, and renewal notice deadline.

1396
00:58:24,160 --> 00:58:27,680
The richness of your taxonomy determines the richness of your extraction.

1397
00:58:27,680 --> 00:58:30,160
Confidence scores are your quality control mechanism.

1398
00:58:30,160 --> 00:58:31,680
When the model extracts a value,

1399
00:58:31,680 --> 00:58:34,720
it assigns a confidence score between 0 and 100.

1400
00:58:34,720 --> 00:58:36,960
You need to set thresholds that match your risk tolerance.

1401
00:58:36,960 --> 00:58:40,480
For low-risk extractions like document type classification,

1402
00:58:40,480 --> 00:58:42,720
an 80% threshold might be acceptable.

1403
00:58:42,720 --> 00:58:45,600
For high-risk extractions like liability caps or governing law,

1404
00:58:45,600 --> 00:58:50,400
you might want 95% confidence before auto accepting anything below that flags for human review.

1405
00:58:50,400 --> 00:58:52,240
Don't set one threshold for everything.

1406
00:58:52,240 --> 00:58:55,280
Match the threshold to the business consequence of an error.

1407
00:58:55,280 --> 00:58:58,560
Handling variability is where most models fail in production.

1408
00:58:58,560 --> 00:59:01,040
Documents that look clean in training arrive in production

1409
00:59:01,040 --> 00:59:03,360
with handwritten annotations, stamp overlays,

1410
00:59:03,360 --> 00:59:05,360
fax headers or rotated pages.

1411
00:59:05,360 --> 00:59:07,840
The model needs to see these variants during training,

1412
00:59:07,840 --> 00:59:09,840
include at least 10% of your training set

1413
00:59:09,840 --> 00:59:11,600
as edge cases and bad examples.

1414
00:59:11,600 --> 00:59:13,360
If you only train on pristine documents,

1415
00:59:13,360 --> 00:59:15,440
your model will be a pristine failure.

1416
00:59:15,440 --> 00:59:17,600
Scanned PDFs present a specific challenge.

1417
00:59:17,600 --> 00:59:19,920
Optical character recognition transforms the image

1418
00:59:19,920 --> 00:59:22,240
into text before the model processes it.

1419
00:59:22,240 --> 00:59:24,480
But OCR quality varies with scan resolution,

1420
00:59:24,480 --> 00:59:26,480
font clarity and page rotation.

1421
00:59:26,480 --> 00:59:28,400
A document scanned at 200 dots per inch

1422
00:59:28,400 --> 00:59:31,440
might produce garbled text where 300 dots per inch scan

1423
00:59:31,440 --> 00:59:32,800
produces perfect extraction.

1424
00:59:32,800 --> 00:59:36,400
If your document pipeline includes a lot of scan paper,

1425
00:59:36,400 --> 00:59:39,280
test your OCR layer separately from your extraction layer,

1426
00:59:39,280 --> 00:59:41,360
bad OCR produces bad extractions,

1427
00:59:41,360 --> 00:59:43,600
and the extraction model gets blamed for a problem

1428
00:59:43,600 --> 00:59:44,800
that happened upstream.

1429
00:59:44,800 --> 00:59:47,440
Common training mistakes follow predictable patterns.

1430
00:59:47,440 --> 00:59:51,280
Two few edge cases produces a model that only works on ideal documents,

1431
00:59:51,280 --> 00:59:53,920
overfitting to one template produces a model that fails

1432
00:59:53,920 --> 00:59:56,000
when a new vendor sends a different layout.

1433
00:59:56,000 --> 00:59:57,520
Ignoring negative examples,

1434
00:59:57,520 --> 00:59:59,920
documents that don't contain the entity at all

1435
00:59:59,920 --> 01:00:02,880
produces a model that hallucinates extractions when none exists.

1436
01:00:02,880 --> 01:00:05,680
The cure for all of these is more representative training data

1437
01:00:05,680 --> 01:00:07,440
and iterative validation.

1438
01:00:07,440 --> 01:00:08,960
Iteration is not a sign of failure.

1439
01:00:08,960 --> 01:00:10,000
It is the process.

1440
01:00:10,000 --> 01:00:11,760
Deploy your model to a pilot library,

1441
01:00:11,760 --> 01:00:13,520
process a few hundred real documents,

1442
01:00:13,520 --> 01:00:15,760
review the extractions and feed the corrections

1443
01:00:15,760 --> 01:00:17,200
back into the training set.

1444
01:00:17,200 --> 01:00:20,080
Every corrected extraction makes the model smarter.

1445
01:00:20,080 --> 01:00:22,480
Every false positive teaches it what to avoid.

1446
01:00:22,480 --> 01:00:25,520
Plan for at least three training cycles

1447
01:00:25,520 --> 01:00:27,360
before you call a model production ready.

1448
01:00:27,360 --> 01:00:29,360
The organizations that get the best results

1449
01:00:29,360 --> 01:00:32,400
treat model training as a continuous improvement process,

1450
01:00:32,400 --> 01:00:34,160
not a one time setup task.

1451
01:00:34,160 --> 01:00:36,400
Multi-page documents present extraction challenges

1452
01:00:36,400 --> 01:00:38,160
that single-page models do not.

1453
01:00:38,160 --> 01:00:40,640
A contract might have the effective date on page one,

1454
01:00:40,640 --> 01:00:42,240
the liability clause on page eight,

1455
01:00:42,240 --> 01:00:44,160
and the signature block on page 12.

1456
01:00:44,160 --> 01:00:46,000
The model needs to understand that entities

1457
01:00:46,000 --> 01:00:47,760
can appear anywhere in the document,

1458
01:00:47,760 --> 01:00:49,040
not just on the first page.

1459
01:00:49,040 --> 01:00:50,720
When you label training examples,

1460
01:00:50,720 --> 01:00:53,040
make sure to label entities on every page,

1461
01:00:53,040 --> 01:00:55,760
not just the page where they most commonly appear.

1462
01:00:55,760 --> 01:00:58,080
A model trained only on page one extractions

1463
01:00:58,080 --> 01:01:00,080
will miss page eight data in production.

1464
01:01:00,080 --> 01:01:01,600
Tables are another common pitfall.

1465
01:01:01,600 --> 01:01:03,200
Contracts often include schedules,

1466
01:01:03,200 --> 01:01:05,600
pricing tables and responsibility matrices.

1467
01:01:05,600 --> 01:01:08,640
A simple entity extractor might see a table cell

1468
01:01:08,640 --> 01:01:10,800
containing a date and incorrectly label it

1469
01:01:10,800 --> 01:01:11,920
as the effective date

1470
01:01:11,920 --> 01:01:14,880
when it is actually a delivery milestone from a schedule.

1471
01:01:14,880 --> 01:01:17,760
You need to train the model to recognize table boundaries

1472
01:01:17,760 --> 01:01:20,000
and distinguish between header rows, data rows,

1473
01:01:20,000 --> 01:01:21,200
and footer summaries.

1474
01:01:21,200 --> 01:01:23,840
Some extraction problems are better solved by table extraction

1475
01:01:23,840 --> 01:01:25,280
than by entity extraction.

1476
01:01:25,280 --> 01:01:27,840
If your document contains a recurring table structure,

1477
01:01:27,840 --> 01:01:29,520
such as a monthly payment schedule,

1478
01:01:29,520 --> 01:01:31,680
consider whether a structured or freeform model

1479
01:01:31,680 --> 01:01:33,840
would capture that table more reliably

1480
01:01:33,840 --> 01:01:35,920
than an unstructured entity extractor

1481
01:01:35,920 --> 01:01:37,920
hunting for isolated values.

1482
01:01:37,920 --> 01:01:39,600
Signatures and handwritten annotations

1483
01:01:39,600 --> 01:01:41,520
break many extraction pipelines.

1484
01:01:41,520 --> 01:01:44,640
A scan signature might cover a date field making OCR fail,

1485
01:01:44,640 --> 01:01:46,160
a handwritten note in the margin

1486
01:01:46,160 --> 01:01:47,840
might contain a critical amendment

1487
01:01:47,840 --> 01:01:51,360
that the model ignores because it only processes printed text.

1488
01:01:51,360 --> 01:01:54,480
If your document pipeline includes heavily annotated documents,

1489
01:01:54,480 --> 01:01:57,600
you need to decide whether to extract the annotation layer separately

1490
01:01:57,600 --> 01:02:00,160
or to train the model to recognize handwriting.

1491
01:02:00,160 --> 01:02:01,600
Neither approach is trivial

1492
01:02:01,600 --> 01:02:03,680
and both require higher quality scans

1493
01:02:03,680 --> 01:02:05,760
than pure printed text extraction.

1494
01:02:05,760 --> 01:02:07,680
Cross-document entity resolution

1495
01:02:07,680 --> 01:02:10,640
is the advanced challenge that separates basic extraction

1496
01:02:10,640 --> 01:02:12,560
from true document intelligence.

1497
01:02:12,560 --> 01:02:14,320
Suppose you have a master service agreement

1498
01:02:14,320 --> 01:02:16,240
and five associated statements of work.

1499
01:02:16,240 --> 01:02:19,600
The MSA contains the governing law and the liability cap.

1500
01:02:19,600 --> 01:02:23,280
Each SOW contains the project scope, timeline, and budget.

1501
01:02:23,280 --> 01:02:26,720
Your taxonomy needs to capture the relationship between these documents.

1502
01:02:26,720 --> 01:02:29,600
A standalone extraction of the SOW budget is useful,

1503
01:02:29,600 --> 01:02:32,880
but an extraction that links the SOW to its parent MSA

1504
01:02:32,880 --> 01:02:36,240
and inherits the governing law from the MSA is far more powerful.

1505
01:02:36,240 --> 01:02:38,320
This requires not just document level extraction

1506
01:02:38,320 --> 01:02:39,840
but document relationship modeling,

1507
01:02:39,840 --> 01:02:42,240
which is beyond the current native capabilities of syntax

1508
01:02:42,240 --> 01:02:46,160
and requires custom orchestration through power, automate, or as your logic apps.

1509
01:02:46,160 --> 01:02:49,440
Language and jurisdiction variations add another layer of complexity.

1510
01:02:49,440 --> 01:02:51,520
A global organization might process contracts

1511
01:02:51,520 --> 01:02:53,920
in English, German, Japanese, and Portuguese.

1512
01:02:53,920 --> 01:02:56,000
The pre-built models handle some of these languages,

1513
01:02:56,000 --> 01:02:58,560
but custom models need language-specific training.

1514
01:02:58,560 --> 01:03:01,760
A termination clause in English uses different linguistic patterns

1515
01:03:01,760 --> 01:03:03,040
than a clause in German.

1516
01:03:03,040 --> 01:03:04,960
You cannot train a model on English contracts

1517
01:03:04,960 --> 01:03:07,920
and expect it to extract accurately from Japanese contracts.

1518
01:03:07,920 --> 01:03:09,920
You need separate models for each language

1519
01:03:09,920 --> 01:03:12,320
or you need to use as your document intelligences

1520
01:03:12,320 --> 01:03:14,880
multilingual capabilities as a pre-processing layer

1521
01:03:14,880 --> 01:03:16,480
before syntax extraction.

1522
01:03:16,480 --> 01:03:19,760
The metadata schema, however, can be shared across languages.

1523
01:03:19,760 --> 01:03:21,600
The contract type column means the same thing

1524
01:03:21,600 --> 01:03:24,080
whether the document is in English or Japanese.

1525
01:03:24,080 --> 01:03:26,720
This is why taxonomy design must be language-agnostic

1526
01:03:26,720 --> 01:03:29,040
even when model training is language-specific.

1527
01:03:29,040 --> 01:03:31,040
Testing protocols for production readiness

1528
01:03:31,040 --> 01:03:33,120
should include a formal validation suite.

1529
01:03:33,120 --> 01:03:36,560
Select 100 documents that represent your full production mix,

1530
01:03:36,560 --> 01:03:38,400
including clean digital files,

1531
01:03:38,400 --> 01:03:42,080
scan paper, multi-page contracts, tables, signatures,

1532
01:03:42,080 --> 01:03:43,280
and edge cases.

1533
01:03:43,280 --> 01:03:44,880
Run these documents through the model

1534
01:03:44,880 --> 01:03:48,160
and compare the extractions against a human-verified ground truth.

1535
01:03:48,160 --> 01:03:51,360
Calculate precision, recall, and F1 score for each field.

1536
01:03:51,360 --> 01:03:53,440
Publish the results to the governance council.

1537
01:03:53,440 --> 01:03:56,400
If accuracy is below your threshold, do not deploy.

1538
01:03:56,400 --> 01:03:59,600
Retrain and retest until the numbers meet your standard.

1539
01:03:59,600 --> 01:04:03,360
This discipline separates experiments from production systems.

1540
01:04:03,360 --> 01:04:05,920
Deploying to live libraries, autofill, and skills.

1541
01:04:05,920 --> 01:04:07,040
Once the model is trained,

1542
01:04:07,040 --> 01:04:08,800
you need to connect it to live workflows.

1543
01:04:08,800 --> 01:04:10,800
In the classic syntax model, you publish the model

1544
01:04:10,800 --> 01:04:12,800
and bind it to a SharePoint library.

1545
01:04:12,800 --> 01:04:15,120
Every time a document is uploaded to that library,

1546
01:04:15,120 --> 01:04:16,800
the model runs automatically.

1547
01:04:16,800 --> 01:04:19,600
It classifies the document type, extracts the fields,

1548
01:04:19,600 --> 01:04:21,360
and populates the library columns.

1549
01:04:21,360 --> 01:04:22,880
The process happens in seconds,

1550
01:04:22,880 --> 01:04:26,000
and the user sees the metadata appear as if by magic.

1551
01:04:26,000 --> 01:04:28,000
That binding mechanism still works.

1552
01:04:28,000 --> 01:04:29,680
And for multi-library deployments,

1553
01:04:29,680 --> 01:04:31,520
it is still the most scalable approach.

1554
01:04:31,520 --> 01:04:33,680
In the AI era, the experience is different.

1555
01:04:33,680 --> 01:04:35,200
Users open a document library,

1556
01:04:35,200 --> 01:04:36,720
select create autofill columns,

1557
01:04:36,720 --> 01:04:38,720
and describe what they want in natural language.

1558
01:04:38,720 --> 01:04:41,840
Co-pilot processes the documents and suggests column values.

1559
01:04:41,840 --> 01:04:44,480
There are no model galleries, no complex binding,

1560
01:04:44,480 --> 01:04:46,800
no extractors to configure, just a prompt.

1561
01:04:46,800 --> 01:04:49,040
The trade-off is flexibility for simplicity.

1562
01:04:49,040 --> 01:04:52,640
You lose the ability to reuse the configuration across libraries.

1563
01:04:52,640 --> 01:04:54,320
You lose batch testing visibility,

1564
01:04:54,320 --> 01:04:56,240
but you gain speed and accessibility.

1565
01:04:56,240 --> 01:04:59,280
Real-time classification means that the moment a file hits the library,

1566
01:04:59,280 --> 01:05:00,880
the metadata is generated.

1567
01:05:00,880 --> 01:05:04,320
For high-volume libraries processing hundreds of documents per day,

1568
01:05:04,320 --> 01:05:06,320
this automatic processing is critical.

1569
01:05:06,320 --> 01:05:08,800
Documents can also be processed via API,

1570
01:05:08,800 --> 01:05:11,440
which means that files uploaded through third-party systems,

1571
01:05:11,440 --> 01:05:13,280
email attachments, or migration tools,

1572
01:05:13,280 --> 01:05:15,440
can still trigger the same extraction pipeline.

1573
01:05:15,440 --> 01:05:17,040
The metadata schema is consistent

1574
01:05:17,040 --> 01:05:19,040
regardless of how the document arrived.

1575
01:05:19,040 --> 01:05:21,440
Multimodal inputs are increasingly common.

1576
01:05:21,440 --> 01:05:24,160
Your contract library might contain digital word files,

1577
01:05:24,160 --> 01:05:26,320
scanned PDFs, image attachments,

1578
01:05:26,320 --> 01:05:28,880
and even photographs of signed documents taken on a phone.

1579
01:05:28,880 --> 01:05:32,640
The OCR layer handles the image-to-text conversion

1580
01:05:32,640 --> 01:05:35,200
and the extraction model processes the resulting text.

1581
01:05:35,200 --> 01:05:36,800
The user experience is the same.

1582
01:05:36,800 --> 01:05:40,000
The underlying pipeline adapts to the input type automatically.

1583
01:05:40,000 --> 01:05:43,040
What matters is that the extraction quality depends on OCR accuracy

1584
01:05:43,040 --> 01:05:44,160
for scanned inputs,

1585
01:05:44,160 --> 01:05:46,800
so monitor your scan quality as part of the deployment,

1586
01:05:46,800 --> 01:05:49,760
connecting extracted metadata to views and filters,

1587
01:05:49,760 --> 01:05:52,960
transforms the library from a file dump into a business dashboard,

1588
01:05:52,960 --> 01:05:56,160
create views that show contracts expiring in the next 90 days,

1589
01:05:56,160 --> 01:05:58,720
create views that filter by governing low jurisdiction,

1590
01:05:58,720 --> 01:06:01,120
create views that flag auto-renewal contracts

1591
01:06:01,120 --> 01:06:02,640
approaching their notice deadline.

1592
01:06:02,640 --> 01:06:04,640
These views turn metadata into action.

1593
01:06:04,640 --> 01:06:06,080
A legal team can open SharePoint

1594
01:06:06,080 --> 01:06:07,760
and see exactly what needs attention today

1595
01:06:07,760 --> 01:06:09,600
without running a single manual search.

1596
01:06:09,600 --> 01:06:11,040
Skills take this a step further

1597
01:06:11,040 --> 01:06:13,600
by automating workflows across the entire site.

1598
01:06:13,600 --> 01:06:14,640
A skill might say,

1599
01:06:14,640 --> 01:06:16,960
for every contract uploaded to the legal library,

1600
01:06:16,960 --> 01:06:19,360
extract the counterparty name and expiration date,

1601
01:06:19,360 --> 01:06:22,320
create a calendar event for 90 days before expiration

1602
01:06:22,320 --> 01:06:23,760
and notify the contract manager.

1603
01:06:23,760 --> 01:06:26,400
The skill is written in natural language,

1604
01:06:26,400 --> 01:06:28,720
stored as markdown in the agent assets library

1605
01:06:28,720 --> 01:06:30,320
and executed by the site agent.

1606
01:06:30,320 --> 01:06:32,320
It chains extraction metadata update,

1607
01:06:32,320 --> 01:06:34,080
list operation and notification

1608
01:06:34,080 --> 01:06:36,000
into a single automated workflow

1609
01:06:36,000 --> 01:06:37,760
that runs without human intervention.

1610
01:06:37,760 --> 01:06:40,800
But production deployment requires caution.

1611
01:06:40,800 --> 01:06:42,400
Skills have no version control.

1612
01:06:42,400 --> 01:06:44,800
Edit the markdown, save it, and it's live.

1613
01:06:44,800 --> 01:06:47,120
There is no rollback, there is no sandbox for testing.

1614
01:06:47,120 --> 01:06:48,560
You write the skill, submit it,

1615
01:06:48,560 --> 01:06:50,000
and it runs on real data.

1616
01:06:50,000 --> 01:06:50,960
One syntax error,

1617
01:06:50,960 --> 01:06:52,960
and you've modified thousands of records

1618
01:06:52,960 --> 01:06:54,960
and by default, everyone can create a skill.

1619
01:06:54,960 --> 01:06:56,400
Before you enable skills,

1620
01:06:56,400 --> 01:06:57,520
in a production site,

1621
01:06:57,520 --> 01:06:59,840
restrict creation to designated owners.

1622
01:06:59,840 --> 01:07:02,240
Before you deploy a skill to a high-stakes library,

1623
01:07:02,240 --> 01:07:03,840
test it on a copy of the data.

1624
01:07:03,840 --> 01:07:05,920
The speed of skills is a double-edged sword.

1625
01:07:05,920 --> 01:07:07,840
Fast execution means fast mistakes.

1626
01:07:07,840 --> 01:07:09,360
You also need to monitor costs.

1627
01:07:09,360 --> 01:07:10,640
Azure Payers, you go meters,

1628
01:07:10,640 --> 01:07:12,000
track every page processed,

1629
01:07:12,000 --> 01:07:13,120
every translation character,

1630
01:07:13,120 --> 01:07:14,480
every signature request.

1631
01:07:14,480 --> 01:07:16,160
The meters are granular and transparent,

1632
01:07:16,160 --> 01:07:18,080
but they accumulate quickly at scale.

1633
01:07:18,080 --> 01:07:20,800
Set up Azure Cost Management alerts before you go live.

1634
01:07:20,800 --> 01:07:22,240
Process a pilot batch,

1635
01:07:22,240 --> 01:07:23,840
measure the meter cost per document

1636
01:07:23,840 --> 01:07:25,840
and extrapolate to your full volume.

1637
01:07:25,840 --> 01:07:28,480
A surprise Azure Bill is not the kind of production feedback

1638
01:07:28,480 --> 01:07:30,320
your finance team wants.

1639
01:07:30,320 --> 01:07:32,640
Search configuration is an often overlooked benefit

1640
01:07:32,640 --> 01:07:34,240
of populated metadata.

1641
01:07:34,240 --> 01:07:36,480
SharePoint Search uses crawled properties

1642
01:07:36,480 --> 01:07:38,560
and manage properties to index content.

1643
01:07:38,560 --> 01:07:41,200
When your model populates a managed metadata column,

1644
01:07:41,200 --> 01:07:42,800
that value becomes a managed property

1645
01:07:42,800 --> 01:07:45,040
that Search can filter, refine, and query.

1646
01:07:45,040 --> 01:07:46,800
Users can search for contract type,

1647
01:07:46,800 --> 01:07:49,840
MSA, and find every master service agreement in the tenant.

1648
01:07:49,840 --> 01:07:52,960
They can search for expiration date now to now nine days

1649
01:07:52,960 --> 01:07:55,760
and find every contract expiring in the next 90 days.

1650
01:07:55,760 --> 01:07:57,280
These are not advanced queries.

1651
01:07:57,280 --> 01:07:58,800
They are standard search syntax

1652
01:07:58,800 --> 01:08:01,120
that becomes powerful when metadata exists.

1653
01:08:01,120 --> 01:08:02,960
Without metadata, Search is limited

1654
01:08:02,960 --> 01:08:05,840
to full text keyword matching, which returns noise.

1655
01:08:05,840 --> 01:08:08,880
With metadata, Search becomes precision retrieval.

1656
01:08:08,880 --> 01:08:11,200
The user experience of automated metadata population

1657
01:08:11,200 --> 01:08:12,240
also drives adoption.

1658
01:08:12,240 --> 01:08:13,680
When a user uploads a contract

1659
01:08:13,680 --> 01:08:15,280
and sees the expiration date,

1660
01:08:15,280 --> 01:08:17,920
counterparty name, and contract type appear automatically

1661
01:08:17,920 --> 01:08:19,120
in the library columns,

1662
01:08:19,120 --> 01:08:21,040
they understand the value immediately.

1663
01:08:21,040 --> 01:08:23,680
When they see empty columns that require manual tagging,

1664
01:08:23,680 --> 01:08:24,960
they skip the step.

1665
01:08:24,960 --> 01:08:27,440
The model's accuracy directly affects user trust.

1666
01:08:27,440 --> 01:08:30,640
If the model populates columns correctly nine times out of 10,

1667
01:08:30,640 --> 01:08:32,560
users learn to trust the 10th.

1668
01:08:32,560 --> 01:08:34,160
If it fails five times out of 10,

1669
01:08:34,160 --> 01:08:35,680
users learn to ignore it.

1670
01:08:35,680 --> 01:08:37,520
Accuracy is not just a technical metric.

1671
01:08:37,520 --> 01:08:38,800
It is an adoption metric.

1672
01:08:38,800 --> 01:08:42,320
Migration scenarios are another place where deployment planning matters.

1673
01:08:42,320 --> 01:08:44,560
If you are migrating documents from a file share

1674
01:08:44,560 --> 01:08:47,680
or a legacy document management system into SharePoint,

1675
01:08:47,680 --> 01:08:49,920
you have a one time opportunity to apply metadata

1676
01:08:49,920 --> 01:08:51,040
during the migration.

1677
01:08:51,040 --> 01:08:52,960
Tools like SharePoint Migration Manager

1678
01:08:52,960 --> 01:08:55,520
can map source properties to destination columns.

1679
01:08:55,520 --> 01:08:57,920
But most source systems have no structured metadata.

1680
01:08:57,920 --> 01:08:59,600
They have folder paths and file names.

1681
01:08:59,600 --> 01:09:01,600
You can use the folder path as a hint,

1682
01:09:01,600 --> 01:09:02,960
mapping the file share of a contract

1683
01:09:02,960 --> 01:09:05,600
due to 24 bars to a contract year column.

1684
01:09:05,600 --> 01:09:08,000
But the real value comes from running the syntax model

1685
01:09:08,000 --> 01:09:10,240
on the migrated documents after they land in SharePoint.

1686
01:09:10,240 --> 01:09:12,000
Plan the migration in two waves.

1687
01:09:12,000 --> 01:09:13,360
First move the files.

1688
01:09:13,360 --> 01:09:14,880
Second process them with the model.

1689
01:09:14,880 --> 01:09:16,240
The model will populate columns

1690
01:09:16,240 --> 01:09:17,840
that the migration could not.

1691
01:09:17,840 --> 01:09:19,840
Disaster recovery and business continuity

1692
01:09:19,840 --> 01:09:21,440
also depend on metadata.

1693
01:09:21,440 --> 01:09:24,400
If a site is accidentally deleted and restored from backup,

1694
01:09:24,400 --> 01:09:27,200
the documents return, but the model bindings might not.

1695
01:09:27,200 --> 01:09:29,760
If the model was bound to a library that no longer exists,

1696
01:09:29,760 --> 01:09:32,080
the restored library might not have the same binding,

1697
01:09:32,080 --> 01:09:34,160
document your model library relationships,

1698
01:09:34,160 --> 01:09:36,800
include them in your backup and recovery runbooks.

1699
01:09:36,800 --> 01:09:38,960
Test restoration in a sandbox environment,

1700
01:09:38,960 --> 01:09:41,520
a model that worked perfectly before a disaster

1701
01:09:41,520 --> 01:09:43,520
and failed silently after recovery

1702
01:09:43,520 --> 01:09:45,280
is not a resilient system.

1703
01:09:45,280 --> 01:09:47,680
Metadata governance includes disaster planning,

1704
01:09:47,680 --> 01:09:49,760
not just day-to-day operations.

1705
01:09:49,760 --> 01:09:51,920
From metadata to governed action.

1706
01:09:51,920 --> 01:09:53,840
Extracted metadata is just data

1707
01:09:53,840 --> 01:09:55,600
until you enforce something with it.

1708
01:09:55,600 --> 01:09:56,960
The real transformation happens

1709
01:09:56,960 --> 01:09:58,480
when you connect those library columns

1710
01:09:58,480 --> 01:10:00,960
to policies, workflows, and compliance controls.

1711
01:10:00,960 --> 01:10:02,800
Retention labels in Microsoft PerView

1712
01:10:02,800 --> 01:10:04,240
are the first place to connect.

1713
01:10:04,240 --> 01:10:06,160
When your model extracts an expiration date

1714
01:10:06,160 --> 01:10:07,840
and writes it to a SharePoint column,

1715
01:10:07,840 --> 01:10:09,600
you can configure a retention policy

1716
01:10:09,600 --> 01:10:11,600
that triggers six months before that date.

1717
01:10:11,600 --> 01:10:13,200
The document is flagged for review,

1718
01:10:13,200 --> 01:10:14,480
the owner is notified,

1719
01:10:14,480 --> 01:10:16,400
and the organization makes a conscious decision

1720
01:10:16,400 --> 01:10:18,480
to renew, renegotiate or dispose.

1721
01:10:18,480 --> 01:10:21,200
Without that connection, contracts expire silently,

1722
01:10:21,200 --> 01:10:22,560
auto-renew by default,

1723
01:10:22,560 --> 01:10:25,200
or sit in storage long past their useful life.

1724
01:10:25,200 --> 01:10:27,680
Retention without metadata is a blunt instrument.

1725
01:10:27,680 --> 01:10:30,640
Retention with extracted metadata is a precision tool.

1726
01:10:30,640 --> 01:10:32,560
Sensitivity labels work the same way.

1727
01:10:32,560 --> 01:10:34,240
When your model classifies a document

1728
01:10:34,240 --> 01:10:35,600
as containing personal data,

1729
01:10:35,600 --> 01:10:37,360
financial records, or intellectual property

1730
01:10:37,360 --> 01:10:39,760
that classification can trigger a sensitivity label

1731
01:10:39,760 --> 01:10:41,680
that restricts external sharing,

1732
01:10:41,680 --> 01:10:44,320
encrypts the file, or applies a watermark.

1733
01:10:44,320 --> 01:10:46,560
The classification happens automatically at upload,

1734
01:10:46,560 --> 01:10:48,160
the protection follows immediately.

1735
01:10:48,160 --> 01:10:50,560
Users don't need to remember which documents are sensitive,

1736
01:10:50,560 --> 01:10:51,840
the model remembers for them.

1737
01:10:51,840 --> 01:10:54,080
Data loss prevention policies benefit enormously

1738
01:10:54,080 --> 01:10:55,760
from accurate classification.

1739
01:10:55,760 --> 01:10:58,480
DLP rules detect and protect sensitive information

1740
01:10:58,480 --> 01:11:00,880
across Microsoft 365 locations.

1741
01:11:00,880 --> 01:11:03,680
They use predefined or custom sensitive information types

1742
01:11:03,680 --> 01:11:04,880
and can take actions,

1743
01:11:04,880 --> 01:11:06,560
such as blocking external sharing,

1744
01:11:06,560 --> 01:11:08,400
sending alerts, or requiring justification

1745
01:11:08,400 --> 01:11:09,600
for policy overrides.

1746
01:11:09,600 --> 01:11:12,800
But DLP is only as accurate as the classification feeding it.

1747
01:11:12,800 --> 01:11:15,360
When documents carry reliable metadata generated

1748
01:11:15,360 --> 01:11:18,640
by trained models, DLP policies trigger on the right documents

1749
01:11:18,640 --> 01:11:20,800
and ignore the false positives.

1750
01:11:20,800 --> 01:11:22,480
When metadata is missing or wrong,

1751
01:11:22,480 --> 01:11:25,200
DLP becomes a nuisance that uses bypass.

1752
01:11:25,200 --> 01:11:27,520
Legal hold scenarios demonstrate the full value

1753
01:11:27,520 --> 01:11:28,960
of govern metadata.

1754
01:11:28,960 --> 01:11:30,720
When litigation is anticipated,

1755
01:11:30,720 --> 01:11:32,720
legal teams need to identify, preserve,

1756
01:11:32,720 --> 01:11:34,560
and produce all relevant documents.

1757
01:11:34,560 --> 01:11:36,320
Without metadata that means manual review

1758
01:11:36,320 --> 01:11:37,760
of every file in every site.

1759
01:11:37,760 --> 01:11:40,560
With metadata, it means filtering by document type,

1760
01:11:40,560 --> 01:11:43,360
date range, counterparty, and custodian.

1761
01:11:43,360 --> 01:11:46,400
A well-classified SharePoint environment reduces e-discovery

1762
01:11:46,400 --> 01:11:49,520
from weeks of manual labor to hours of targeted export.

1763
01:11:49,520 --> 01:11:52,000
The chain of custody is documented automatically.

1764
01:11:52,000 --> 01:11:53,920
The metadata integrity is preserved.

1765
01:11:53,920 --> 01:11:55,280
The defensibility is stronger.

1766
01:11:55,280 --> 01:11:58,160
Dashboard visibility closes the loop for business stakeholders.

1767
01:11:58,160 --> 01:12:01,440
Power BI can connect directly to SharePoint lists and libraries,

1768
01:12:01,440 --> 01:12:03,920
visualizing the extracted metadata in real time.

1769
01:12:03,920 --> 01:12:06,240
Executive C, contract renewal pipelines,

1770
01:12:06,240 --> 01:12:07,360
compliance officers,

1771
01:12:07,360 --> 01:12:09,680
C retention status across departments.

1772
01:12:09,680 --> 01:12:12,320
Security teams see sensitivity label coverage.

1773
01:12:12,320 --> 01:12:14,320
The documents that were once invisible

1774
01:12:14,320 --> 01:12:16,880
become the data source for strategic decisions.

1775
01:12:16,880 --> 01:12:18,640
The metadata gap that was a liability

1776
01:12:18,640 --> 01:12:20,320
becomes a competitive advantage.

1777
01:12:20,320 --> 01:12:21,920
And the digital trail is complete.

1778
01:12:21,920 --> 01:12:23,680
Every extraction, every column update,

1779
01:12:23,680 --> 01:12:24,960
every policy application,

1780
01:12:24,960 --> 01:12:27,920
and every user access is logged in SharePoint audit records.

1781
01:12:27,920 --> 01:12:30,720
When an auditor asks how you knew which documents to retain,

1782
01:12:30,720 --> 01:12:33,200
you show them the model that extracted the expiration date

1783
01:12:33,200 --> 01:12:34,960
and the policy that applied the label.

1784
01:12:34,960 --> 01:12:37,680
When a regulator asks how you protected sensitive data,

1785
01:12:37,680 --> 01:12:39,520
you show them the classification model

1786
01:12:39,520 --> 01:12:41,520
and the sensitivity label automation.

1787
01:12:41,520 --> 01:12:43,520
The metadata doesn't just organize your content,

1788
01:12:43,520 --> 01:12:44,880
it proves your governance.

1789
01:12:44,880 --> 01:12:46,880
Workflow automation triggered by metadata

1790
01:12:46,880 --> 01:12:49,520
is where the operational value becomes tangible.

1791
01:12:49,520 --> 01:12:51,280
When a model extracts a contract value

1792
01:12:51,280 --> 01:12:52,960
and writes it to a SharePoint column,

1793
01:12:52,960 --> 01:12:55,120
a power automate flow can evaluate that value

1794
01:12:55,120 --> 01:12:56,720
against approval thresholds.

1795
01:12:56,720 --> 01:12:59,360
A contract under $10,000 might route

1796
01:12:59,360 --> 01:13:01,120
to a department manager for approval.

1797
01:13:01,120 --> 01:13:04,480
A contract over $100,000 might route to legal, finance,

1798
01:13:04,480 --> 01:13:06,080
and the CFO in sequence.

1799
01:13:06,080 --> 01:13:08,400
The routing logic is driven by the extracted metadata,

1800
01:13:08,400 --> 01:13:10,160
not by manual form submissions.

1801
01:13:10,160 --> 01:13:12,560
The metadata becomes the workflow engine.

1802
01:13:12,560 --> 01:13:15,440
Document lifecycle management also improves dramatically.

1803
01:13:15,440 --> 01:13:18,480
Most organizations have a loose concept of document lifecycle,

1804
01:13:18,480 --> 01:13:20,720
but they lack the mechanism to enforce it.

1805
01:13:20,720 --> 01:13:23,040
A contract might have six lifecycle stages,

1806
01:13:23,040 --> 01:13:26,800
draft, under review, approved, executed, active, and expired.

1807
01:13:26,800 --> 01:13:28,480
Each stage has different permissions,

1808
01:13:28,480 --> 01:13:31,440
different retention rules, and different notification requirements.

1809
01:13:31,440 --> 01:13:33,360
With metadata-driven lifecycle management,

1810
01:13:33,360 --> 01:13:35,040
the stage is a SharePoint column

1811
01:13:35,040 --> 01:13:37,520
populated by the model or by workflow logic.

1812
01:13:37,520 --> 01:13:39,920
When the model detects signatures on the final page,

1813
01:13:39,920 --> 01:13:41,840
it updates the stage to execute it.

1814
01:13:41,840 --> 01:13:43,440
When the expiration date passes,

1815
01:13:43,440 --> 01:13:45,520
the workflow updates the stage to expired

1816
01:13:45,520 --> 01:13:47,040
and triggers a retention review.

1817
01:13:47,040 --> 01:13:49,680
The document moves through its lifecycle automatically,

1818
01:13:49,680 --> 01:13:51,680
guided by the metadata it carries.

1819
01:13:51,680 --> 01:13:53,760
External collaboration scenarios benefit

1820
01:13:53,760 --> 01:13:55,760
from metadata-driven access control.

1821
01:13:55,760 --> 01:13:57,840
When you share a document with an external law firm

1822
01:13:57,840 --> 01:14:00,640
or audit partner, you need to know what you are sharing.

1823
01:14:00,640 --> 01:14:03,680
A document tagged with client, confidential, and legal privilege

1824
01:14:03,680 --> 01:14:05,200
requires different sharing controls

1825
01:14:05,200 --> 01:14:08,080
than a document tagged with public reference material.

1826
01:14:08,080 --> 01:14:12,000
Metadata-driven sharing policies can enforce these distinctions automatically.

1827
01:14:12,000 --> 01:14:14,400
The model classifies the document at upload.

1828
01:14:14,400 --> 01:14:16,960
The sensitivity label applies the sharing restriction.

1829
01:14:16,960 --> 01:14:19,920
The external recipient receives only what they are supposed to receive

1830
01:14:19,920 --> 01:14:21,840
and the organization maintains an audit trail

1831
01:14:21,840 --> 01:14:24,560
of what was shared, why, and when.

1832
01:14:24,560 --> 01:14:27,520
Version control and document history also improve with metadata.

1833
01:14:27,520 --> 01:14:29,760
When a contract goes through five revision cycles,

1834
01:14:29,760 --> 01:14:31,440
the version history shows what changed,

1835
01:14:31,440 --> 01:14:32,880
but it doesn't explain why.

1836
01:14:32,880 --> 01:14:35,120
A metadata column for approval stage can show

1837
01:14:35,120 --> 01:14:37,520
that version two was the legal review draft,

1838
01:14:37,520 --> 01:14:39,520
version three included finance feedback,

1839
01:14:39,520 --> 01:14:42,720
and version four was the final executive approval copy.

1840
01:14:42,720 --> 01:14:44,480
The version history tells you when.

1841
01:14:44,480 --> 01:14:46,000
The metadata tells you why.

1842
01:14:46,000 --> 01:14:48,240
Together, they create a complete narrative

1843
01:14:48,240 --> 01:14:51,040
of the document lifecycle that satisfies auditors,

1844
01:14:51,040 --> 01:14:53,040
protects the organization in disputes,

1845
01:14:53,040 --> 01:14:56,080
and educates new team members about how decisions were made.

1846
01:14:56,080 --> 01:14:59,840
Training and compliance reporting also rely on metadata.

1847
01:14:59,840 --> 01:15:02,400
Regulatory frameworks often require evidence

1848
01:15:02,400 --> 01:15:04,800
that employees have read and acknowledged policies.

1849
01:15:04,800 --> 01:15:06,640
A policy document tagged with acknowledgement

1850
01:15:06,640 --> 01:15:09,520
required an employee handbook can trigger a training workflow

1851
01:15:09,520 --> 01:15:11,360
that assigns the document to every new hire

1852
01:15:11,360 --> 01:15:12,800
and tracks their acknowledgement.

1853
01:15:12,800 --> 01:15:14,720
The metadata drives the assignment list.

1854
01:15:14,720 --> 01:15:16,800
The workflow tracks the completion status,

1855
01:15:16,800 --> 01:15:20,080
the audit report shows who has acknowledged which policies and when.

1856
01:15:20,080 --> 01:15:22,560
Without metadata, this becomes a manual mailing list

1857
01:15:22,560 --> 01:15:24,400
and a spreadsheet that nobody updates.

1858
01:15:24,400 --> 01:15:26,080
The cumulative effect of these connections

1859
01:15:26,080 --> 01:15:28,320
is that SharePoint stops being a file repository

1860
01:15:28,320 --> 01:15:30,080
and starts being a business system.

1861
01:15:30,080 --> 01:15:31,680
Documents are not passive storage.

1862
01:15:31,680 --> 01:15:33,760
They are active records that trigger workflows

1863
01:15:33,760 --> 01:15:36,880
in force policies, generate reports and drive decisions.

1864
01:15:36,880 --> 01:15:39,360
The metadata is what makes this transition possible

1865
01:15:39,360 --> 01:15:41,760
and the model is what generates the metadata at scale.

1866
01:15:41,760 --> 01:15:44,400
Scaling AI transformation across the enterprise.

1867
01:15:44,400 --> 01:15:46,000
One library is proof of concept.

1868
01:15:46,000 --> 01:15:47,680
The real return comes when you scale,

1869
01:15:47,680 --> 01:15:50,720
but scaling requires more than copying the model to another site.

1870
01:15:50,720 --> 01:15:53,280
It requires a strategy, a governance framework

1871
01:15:53,280 --> 01:15:54,800
and a change management plan.

1872
01:15:54,800 --> 01:15:56,640
Start by measuring what matters.

1873
01:15:56,640 --> 01:15:57,680
Before you expand,

1874
01:15:57,680 --> 01:15:59,840
establish baseline metrics for your pilot.

1875
01:15:59,840 --> 01:16:01,280
How many documents were processed?

1876
01:16:01,280 --> 01:16:03,520
What was the average extraction accuracy?

1877
01:16:03,520 --> 01:16:06,000
How many hours did staff spend on manual data entry

1878
01:16:06,000 --> 01:16:07,840
before the model and how many after?

1879
01:16:07,840 --> 01:16:11,120
How many compliance incidents were prevented by automated retention?

1880
01:16:11,120 --> 01:16:13,280
These numbers build the business case for expansion.

1881
01:16:13,280 --> 01:16:16,400
Without them, you're asking for budget based on faith.

1882
01:16:16,400 --> 01:16:19,920
The ROI calculation combines time-saved, error reduction

1883
01:16:19,920 --> 01:16:21,280
and risk avoidance.

1884
01:16:21,280 --> 01:16:24,000
Industry benchmarks suggest that automated document processing

1885
01:16:24,000 --> 01:16:28,000
can reduce processing time by 70% and error rates by 80%.

1886
01:16:28,000 --> 01:16:30,000
But these are assumptions, not guarantees.

1887
01:16:30,000 --> 01:16:32,240
Your actual results depend on document quality,

1888
01:16:32,240 --> 01:16:34,640
model accuracy and process maturity.

1889
01:16:34,640 --> 01:16:35,920
Measure your own results.

1890
01:16:35,920 --> 01:16:38,480
Use the pilot data to build a custom ROI model

1891
01:16:38,480 --> 01:16:40,480
that reflects your organization's reality,

1892
01:16:40,480 --> 01:16:42,400
not a vendor's best case scenario.

1893
01:16:42,400 --> 01:16:45,120
Expanding from legal to finance, HR and operations

1894
01:16:45,120 --> 01:16:46,160
follows a pattern.

1895
01:16:46,160 --> 01:16:48,400
Identify the next document type with high volume,

1896
01:16:48,400 --> 01:16:50,640
high manual effort and high business risk.

1897
01:16:50,640 --> 01:16:52,960
Finance might process invoices and purchase orders,

1898
01:16:52,960 --> 01:16:56,400
HR might process employee agreements and benefits forms,

1899
01:16:56,400 --> 01:16:58,480
operations might process safety reports

1900
01:16:58,480 --> 01:17:00,400
and vendor certifications.

1901
01:17:00,400 --> 01:17:02,720
Each domain has its own document patterns,

1902
01:17:02,720 --> 01:17:05,600
its own taxonomy and its own compliance requirements.

1903
01:17:05,600 --> 01:17:08,160
Don't assume the contract model transfers directly.

1904
01:17:08,160 --> 01:17:10,000
Train a new model for each document type

1905
01:17:10,000 --> 01:17:11,920
using domain experts from that department.

1906
01:17:11,920 --> 01:17:14,800
Change management is the most underestimated part of scaling.

1907
01:17:14,800 --> 01:17:16,080
The technology is the easy part.

1908
01:17:16,080 --> 01:17:17,760
The hard part is getting people to trust it.

1909
01:17:17,760 --> 01:17:20,080
Contract managers need to see that the models' extractions

1910
01:17:20,080 --> 01:17:23,440
are accurate before they stop manually checking every field.

1911
01:17:23,440 --> 01:17:26,400
Legal teams need to understand that the retention policy is defensible

1912
01:17:26,400 --> 01:17:28,320
before they rely on it for e-discovery.

1913
01:17:28,320 --> 01:17:31,760
IT teams need to know that the model won't generate surprise

1914
01:17:31,760 --> 01:17:34,320
Azure Bills before they let it run unattended.

1915
01:17:34,320 --> 01:17:36,000
Trust is built through transparency,

1916
01:17:36,000 --> 01:17:37,520
show people the confidence scores,

1917
01:17:37,520 --> 01:17:39,200
show them the correction workflow,

1918
01:17:39,200 --> 01:17:41,280
let them override the model when it's wrong

1919
01:17:41,280 --> 01:17:43,840
and feed those corrections back into training.

1920
01:17:43,840 --> 01:17:45,760
Governance councils become critical at scale.

1921
01:17:45,760 --> 01:17:48,560
A single model in one library can be managed by one owner.

1922
01:17:48,560 --> 01:17:51,440
50 models across 10 departments need coordination.

1923
01:17:51,440 --> 01:17:53,440
Establish across functional governance council

1924
01:17:53,440 --> 01:17:54,960
with representatives from IT,

1925
01:17:54,960 --> 01:17:56,720
legal compliance and business units.

1926
01:17:56,720 --> 01:17:58,640
Define who owns the term sets,

1927
01:17:58,640 --> 01:18:00,000
who approves new models,

1928
01:18:00,000 --> 01:18:02,960
who monitors accuracy and who handles exceptions.

1929
01:18:02,960 --> 01:18:04,240
Without this structure,

1930
01:18:04,240 --> 01:18:06,320
models proliferate inconsistently,

1931
01:18:06,320 --> 01:18:07,760
taxonomies diverge,

1932
01:18:07,760 --> 01:18:10,720
and the metadata gap reappears at a larger scale.

1933
01:18:10,720 --> 01:18:11,920
The roadmap is simple.

1934
01:18:11,920 --> 01:18:13,200
Audit your current workflows

1935
01:18:13,200 --> 01:18:15,520
and identify the highest impact document type.

1936
01:18:15,520 --> 01:18:17,840
Pile it a model for that type in a single library.

1937
01:18:17,840 --> 01:18:20,000
Measure accuracy, cost and time savings,

1938
01:18:20,000 --> 01:18:21,440
iterate on the training data

1939
01:18:21,440 --> 01:18:23,120
until the model is production ready.

1940
01:18:23,120 --> 01:18:24,640
Connect the extracted metadata

1941
01:18:24,640 --> 01:18:27,520
to retention, sensitivity and DLP policies,

1942
01:18:27,520 --> 01:18:28,960
measure compliance outcomes,

1943
01:18:28,960 --> 01:18:30,800
then expand to the next document type

1944
01:18:30,800 --> 01:18:31,760
and the next,

1945
01:18:31,760 --> 01:18:34,320
until the entire document lifecycle is governed.

1946
01:18:34,320 --> 01:18:35,920
This is not a big bank transformation,

1947
01:18:35,920 --> 01:18:37,280
it is a series of deliberate,

1948
01:18:37,280 --> 01:18:39,200
measured steps that compound over time.

1949
01:18:39,200 --> 01:18:41,680
Department specific expansion requires different planning

1950
01:18:41,680 --> 01:18:42,720
for each domain.

1951
01:18:42,720 --> 01:18:44,880
Finance departments typically start with invoices

1952
01:18:44,880 --> 01:18:46,000
and purchase orders

1953
01:18:46,000 --> 01:18:47,920
because these documents have high volume,

1954
01:18:47,920 --> 01:18:48,960
consistent structure,

1955
01:18:48,960 --> 01:18:50,400
and direct cost impact.

1956
01:18:50,400 --> 01:18:52,080
The ROI is easy to calculate,

1957
01:18:52,080 --> 01:18:53,840
multiply the number of invoices per month

1958
01:18:53,840 --> 01:18:55,600
by the minutes saved per invoice

1959
01:18:55,600 --> 01:18:57,440
by the fully loaded cost per minute.

1960
01:18:57,440 --> 01:18:58,960
The result is a monthly savings number

1961
01:18:58,960 --> 01:19:00,720
that justifies the model investment.

1962
01:19:00,720 --> 01:19:02,960
HR departments often start with employee agreements

1963
01:19:02,960 --> 01:19:04,320
and benefits enrollment forms

1964
01:19:04,320 --> 01:19:06,240
because these carry compliance risk.

1965
01:19:06,240 --> 01:19:07,680
A misfiled i9 form

1966
01:19:07,680 --> 01:19:09,440
or an unsigned non-competed agreement

1967
01:19:09,440 --> 01:19:11,760
creates legal exposure that is easy to quantify.

1968
01:19:11,760 --> 01:19:13,200
Operations departments might start

1969
01:19:13,200 --> 01:19:14,480
with safety incident reports

1970
01:19:14,480 --> 01:19:15,760
or vendor qualification documents

1971
01:19:15,760 --> 01:19:17,360
because these tie directly to insurance

1972
01:19:17,360 --> 01:19:18,640
and audit requirements.

1973
01:19:18,640 --> 01:19:20,240
Each department has its own pain point,

1974
01:19:20,240 --> 01:19:21,440
its own document type

1975
01:19:21,440 --> 01:19:22,960
and its own success metric.

1976
01:19:22,960 --> 01:19:24,640
Don't impose a single template.

1977
01:19:24,640 --> 01:19:26,240
Let each department solve its own

1978
01:19:26,240 --> 01:19:28,240
highest priority problem first.

1979
01:19:28,240 --> 01:19:30,320
Technology adoption curves vary by department

1980
01:19:30,320 --> 01:19:31,120
and by role.

1981
01:19:31,120 --> 01:19:32,800
Early adopters in IT and legal

1982
01:19:32,800 --> 01:19:34,800
might embrace the model immediately.

1983
01:19:34,800 --> 01:19:36,560
Skeptics in finance and operations

1984
01:19:36,560 --> 01:19:38,400
might resist until they see proof.

1985
01:19:38,400 --> 01:19:39,600
Plan for this variation.

1986
01:19:39,600 --> 01:19:41,520
Identify your champions in each department,

1987
01:19:41,520 --> 01:19:43,840
give them early access to the pilot results,

1988
01:19:43,840 --> 01:19:45,680
let them become internal advocates.

1989
01:19:45,680 --> 01:19:47,440
When a contract manager tells her colleagues

1990
01:19:47,440 --> 01:19:49,600
that the model saved her four hours last week,

1991
01:19:49,600 --> 01:19:50,960
that testimony is more powerful

1992
01:19:50,960 --> 01:19:52,720
than any business case presentation.

1993
01:19:52,720 --> 01:19:54,960
Peer influence drives adoption more effectively

1994
01:19:54,960 --> 01:19:56,400
than top-down mandates.

1995
01:19:56,400 --> 01:19:58,560
Training programs need to cover three audiences.

1996
01:19:58,560 --> 01:20:00,720
End users need to know how to upload documents

1997
01:20:00,720 --> 01:20:02,720
and how to verify automated extractions.

1998
01:20:02,720 --> 01:20:05,280
They need to understand what the confidence score means

1999
01:20:05,280 --> 01:20:06,960
and when to override the model.

2000
01:20:06,960 --> 01:20:08,720
Site owners need to know how to configure

2001
01:20:08,720 --> 01:20:11,200
libraries, bind models and create views.

2002
01:20:11,200 --> 01:20:13,200
They need to understand taxonomy basics

2003
01:20:13,200 --> 01:20:15,760
so they don't break the model by changing column names.

2004
01:20:15,760 --> 01:20:18,080
Administrators need to know how to monitor costs,

2005
01:20:18,080 --> 01:20:20,800
manage model versions, and enforce governance rules.

2006
01:20:20,800 --> 01:20:23,120
Each audience needs a different depth of training

2007
01:20:23,120 --> 01:20:24,480
and each needs hands-on practice

2008
01:20:24,480 --> 01:20:26,640
with real documents from their own department.

2009
01:20:26,640 --> 01:20:28,800
Generic training on sample documents fails

2010
01:20:28,800 --> 01:20:31,440
because it doesn't reflect the messiness of real-world content.

2011
01:20:31,440 --> 01:20:34,160
Executive sponsorship is the single biggest predictor

2012
01:20:34,160 --> 01:20:35,520
of scaling success.

2013
01:20:35,520 --> 01:20:36,880
Without a senior leader who believes

2014
01:20:36,880 --> 01:20:38,800
in the value of metadata governance,

2015
01:20:38,800 --> 01:20:40,800
the project becomes an IT experiment

2016
01:20:40,800 --> 01:20:43,120
that dies when the initial budget runs out.

2017
01:20:43,120 --> 01:20:44,480
With executive sponsorship,

2018
01:20:44,480 --> 01:20:46,320
the project becomes a strategic initiative

2019
01:20:46,320 --> 01:20:48,560
that survives reorganizations, budget cuts,

2020
01:20:48,560 --> 01:20:49,760
and staff turnover.

2021
01:20:49,760 --> 01:20:51,520
The sponsor does not need to understand

2022
01:20:51,520 --> 01:20:53,600
transformer models or Azure meters.

2023
01:20:53,600 --> 01:20:55,600
They need to understand that unmanaged documents

2024
01:20:55,600 --> 01:20:58,080
are a liability, that AI can fix it,

2025
01:20:58,080 --> 01:21:01,440
and that the fix requires sustained investment in taxonomy,

2026
01:21:01,440 --> 01:21:03,120
training, and governance.

2027
01:21:03,120 --> 01:21:05,200
Their job is to protect the program long enough

2028
01:21:05,200 --> 01:21:07,840
for the compounding benefits to become visible.

2029
01:21:07,840 --> 01:21:10,640
Metrics and reporting sustain the program over time.

2030
01:21:10,640 --> 01:21:12,880
After the pilot establish a governance dashboard

2031
01:21:12,880 --> 01:21:15,360
that tracks the metrics that matter to your organization.

2032
01:21:15,360 --> 01:21:17,120
Document volume processed per month.

2033
01:21:17,120 --> 01:21:19,200
Average extraction accuracy by field,

2034
01:21:19,200 --> 01:21:20,880
cost per document processed.

2035
01:21:20,880 --> 01:21:22,320
Hours saved per department.

2036
01:21:22,320 --> 01:21:23,920
Compliance incidents prevented,

2037
01:21:23,920 --> 01:21:25,360
audit findings resolved,

2038
01:21:25,360 --> 01:21:26,880
share this dashboard monthly

2039
01:21:26,880 --> 01:21:28,160
with the governance council

2040
01:21:28,160 --> 01:21:30,000
and quarterly with executive leadership.

2041
01:21:30,000 --> 01:21:32,080
When the numbers improve, celebrate the progress.

2042
01:21:32,080 --> 01:21:33,200
When the numbers stall,

2043
01:21:33,200 --> 01:21:34,560
diagnose the problem.

2044
01:21:34,560 --> 01:21:36,080
Visibility creates accountability

2045
01:21:36,080 --> 01:21:37,760
and accountability creates improvement.

2046
01:21:37,760 --> 01:21:39,520
The organizations that succeed treat this

2047
01:21:39,520 --> 01:21:41,040
not as a technology deployment

2048
01:21:41,040 --> 01:21:42,320
but as a capability build.

2049
01:21:42,320 --> 01:21:43,440
Technology changes.

2050
01:21:43,440 --> 01:21:46,080
Models get renamed, interfaces evolve.

2051
01:21:46,080 --> 01:21:47,360
The capability that endures

2052
01:21:47,360 --> 01:21:49,840
is the organization's ability to classify its content,

2053
01:21:49,840 --> 01:21:51,040
govern its metadata,

2054
01:21:51,040 --> 01:21:53,120
and automate its document processes.

2055
01:21:53,120 --> 01:21:54,160
Build that capability

2056
01:21:54,160 --> 01:21:56,240
and you will survive every rebrand.

2057
01:21:56,240 --> 01:21:57,840
Build only a point solution

2058
01:21:57,840 --> 01:21:58,960
and you will be rebuilding

2059
01:21:58,960 --> 01:22:00,720
when the next rename arrives.

2060
01:22:00,720 --> 01:22:02,000
Finally, consider the platform

2061
01:22:02,000 --> 01:22:03,520
around your SharePoint deployment.

2062
01:22:03,520 --> 01:22:06,560
Microsoft 365 is not a collection of separate products.

2063
01:22:06,560 --> 01:22:07,840
It is an integrated platform

2064
01:22:07,840 --> 01:22:10,160
where SharePoint teams, outlook, one drive,

2065
01:22:10,160 --> 01:22:10,960
power platform,

2066
01:22:10,960 --> 01:22:13,760
and per view, share data, permissions, and metadata.

2067
01:22:13,760 --> 01:22:15,920
When you fix the metadata gap in SharePoint,

2068
01:22:15,920 --> 01:22:18,080
you improve the signal for co-pilot in Teams,

2069
01:22:18,080 --> 01:22:19,440
for Power BI in Outlook,

2070
01:22:19,440 --> 01:22:21,600
and for Per View across the entire tenant.

2071
01:22:21,600 --> 01:22:24,400
The metadata you extract from a contract in SharePoint today

2072
01:22:24,400 --> 01:22:27,840
becomes the data that drives a power app's approval workflow tomorrow,

2073
01:22:27,840 --> 01:22:30,160
a Power BI risk dashboard next quarter,

2074
01:22:30,160 --> 01:22:32,640
and a co-pilot answer six months from now.

2075
01:22:32,640 --> 01:22:34,320
The investment in taxonomy and governance

2076
01:22:34,320 --> 01:22:37,520
pays dividends across every workload in Microsoft 365.

2077
01:22:37,520 --> 01:22:40,320
That is why the metadata gap is not just a SharePoint problem.

2078
01:22:40,320 --> 01:22:41,520
It is a platform problem

2079
01:22:41,520 --> 01:22:43,440
and fixing it is not just an IT project,

2080
01:22:43,440 --> 01:22:44,880
it is a business transformation.

2081
01:22:44,880 --> 01:22:46,640
And remember, nothing you built is wasted,

2082
01:22:46,640 --> 01:22:48,480
but nothing you built is finished either.

2083
01:22:48,480 --> 01:22:49,680
The syntax error has ended,

2084
01:22:49,680 --> 01:22:52,720
but its capabilities live on as part of AI in SharePoint.

2085
01:22:52,720 --> 01:22:54,240
Your existing model still bill,

2086
01:22:54,240 --> 01:22:55,760
still run, and still extract,

2087
01:22:55,760 --> 01:22:57,280
but their behavior has changed.

2088
01:22:57,280 --> 01:22:59,040
The user interface is conversational.

2089
01:22:59,040 --> 01:23:00,800
The configuration is intent-based.

2090
01:23:00,800 --> 01:23:03,760
The governance is distributed across agents and skills.

2091
01:23:03,760 --> 01:23:05,600
Treat this as a refresh, not a swap,

2092
01:23:05,600 --> 01:23:06,880
validate before production,

2093
01:23:06,880 --> 01:23:08,560
plan for continuous iteration,

2094
01:23:08,560 --> 01:23:09,840
and build your architecture

2095
01:23:09,840 --> 01:23:11,600
so that it survives the next rename

2096
01:23:11,600 --> 01:23:12,720
because there will be one.

2097
01:23:13,920 --> 01:23:16,160
The metadata gap is not a technology problem.

2098
01:23:16,160 --> 01:23:18,160
It is a structural problem that technology can fix

2099
01:23:18,160 --> 01:23:19,760
if you build the foundation first.

2100
01:23:19,760 --> 01:23:21,280
Custom AI document models

2101
01:23:21,280 --> 01:23:23,360
turn your SharePoint liability into a governed,

2102
01:23:23,360 --> 01:23:24,880
searchable, co-pilot-ready asset.

2103
01:23:24,880 --> 01:23:27,600
But the model is only as good as the taxonomy it feeds,

2104
01:23:27,600 --> 01:23:29,920
and the taxonomy is only as good as the governance

2105
01:23:29,920 --> 01:23:30,720
that maintains it.

2106
01:23:30,720 --> 01:23:32,400
If this changed how you think about SharePoint,

2107
01:23:32,400 --> 01:23:33,680
follow me on LinkedIn.

2108
01:23:33,680 --> 01:23:35,120
And if you want the next detailed look

2109
01:23:35,120 --> 01:23:36,560
at Power Automate Orchestration

2110
01:23:36,560 --> 01:23:37,920
for Extracted Metadata,

2111
01:23:37,920 --> 01:23:38,960
that's coming next.