The Secret to Putting SQL Data in Copilot Studio

1
00:00:00,000 –> 00:00:02,900
Your co-pilot is fluent, confident, and utterly clueless.

2
00:00:02,900 –> 00:00:04,620
It greets your employees like an expert,

3
00:00:04,620 –> 00:00:07,140
yet it’s blind to the existence of your customers,

4
00:00:07,140 –> 00:00:08,500
invoices, or inventory.

5
00:00:08,500 –> 00:00:09,580
You think it knows your business?

6
00:00:09,580 –> 00:00:10,240
It doesn’t.

7
00:00:10,240 –> 00:00:11,380
It knows Wikipedia.

8
00:00:11,380 –> 00:00:13,420
Inside your network,

9
00:00:13,420 –> 00:00:15,980
SQL Server holds your company’s actual memories.

10
00:00:15,980 –> 00:00:17,100
The sales you’ve made,

11
00:00:17,100 –> 00:00:18,300
the people you’ve invoiced,

12
00:00:18,300 –> 00:00:20,100
the chaos of human data.

13
00:00:20,100 –> 00:00:21,740
But co-pilot studio sits outside

14
00:00:21,740 –> 00:00:23,340
that fortress smiling through the glass,

15
00:00:23,340 –> 00:00:24,580
pretending it understands.

16
00:00:24,580 –> 00:00:25,820
The irony is beautiful,

17
00:00:25,820 –> 00:00:28,100
a so-called intelligent assistant

18
00:00:28,100 –> 00:00:29,500
that can’t see the data

19
00:00:29,500 –> 00:00:30,620
that built your business.

20
00:00:30,620 –> 00:00:33,700
The bridge it needs is the power platform data gateway,

21
00:00:33,700 –> 00:00:35,820
your secure tunnel through the firewall

22
00:00:35,820 –> 00:00:37,420
that lets co-pilot observe SQL

23
00:00:37,420 –> 00:00:39,580
in real time without ever exposing it.

24
00:00:39,580 –> 00:00:40,660
By the end of this session,

25
00:00:40,660 –> 00:00:43,300
you’ll wire that bridge, query life tables,

26
00:00:43,300 –> 00:00:45,300
and even teach co-pilot to write back.

27
00:00:45,300 –> 00:00:48,100
No magic, just architecture executed properly.

28
00:00:48,100 –> 00:00:51,380
Why co-pilots fail without context?

29
00:00:51,380 –> 00:00:53,740
A co-pilot disconnected from your structured data

30
00:00:53,740 –> 00:00:55,900
is little more than a verbose for Chuntella.

31
00:00:55,900 –> 00:00:57,940
It generates words that sound authoritative,

32
00:00:57,940 –> 00:01:00,700
but are entirely divorced from operational truth.

33
00:01:00,700 –> 00:01:02,700
Ask it about this quarter’s customer churn,

34
00:01:02,700 –> 00:01:03,900
and it’ll estimate,

35
00:01:03,900 –> 00:01:05,500
ask it who owed you money last month,

36
00:01:05,500 –> 00:01:07,900
and it’ll hallucinate confidence while inventing numbers.

37
00:01:07,900 –> 00:01:09,620
That’s what happens when large language models

38
00:01:09,620 –> 00:01:11,380
are forced to perform without grounding.

39
00:01:11,380 –> 00:01:13,620
They produce statistically likely nonsense.

40
00:01:13,620 –> 00:01:15,460
Enterprises perpetuate this blindness

41
00:01:15,460 –> 00:01:17,300
by keeping their AI in the cloud,

42
00:01:17,300 –> 00:01:19,100
but their data in the basement.

43
00:01:19,100 –> 00:01:21,780
Security teams erect beautiful firewalls,

44
00:01:21,780 –> 00:01:24,300
compliance officers for bid-inbound connections,

45
00:01:24,300 –> 00:01:25,540
and the poor co-pilot,

46
00:01:25,540 –> 00:01:27,460
stuck in its public sandbox,

47
00:01:27,460 –> 00:01:30,620
sifts through generic training data and calls it knowledge.

48
00:01:30,620 –> 00:01:32,060
It’s as if you hired a consultant

49
00:01:32,060 –> 00:01:33,660
who’s read every business book ever written

50
00:01:33,660 –> 00:01:35,420
but has never seen your balance sheet.

51
00:01:35,420 –> 00:01:37,660
Inside your walls, SQL Server remains

52
00:01:37,660 –> 00:01:40,140
the spinal cord of real business function.

53
00:01:40,140 –> 00:01:41,580
Every order, every update,

54
00:01:41,580 –> 00:01:44,340
every mis-keyed customer address pulses through it.

55
00:01:44,340 –> 00:01:46,380
It isn’t glamorous, but it’s reliable.

56
00:01:46,380 –> 00:01:49,060
The relational glue that binds your ERP, CRM,

57
00:01:49,060 –> 00:01:52,180
and those Excel spreadsheets labelled final V27.

58
00:01:52,180 –> 00:01:55,580
May he see time, but I’m going to be a detail.

59
00:01:55,580 –> 00:01:57,820
Without access to that structured intelligence,

60
00:01:57,820 –> 00:02:00,940
an AI agent has the literacy of a genius child reading

61
00:02:00,940 –> 00:02:02,260
random encyclopedias.

62
00:02:02,260 –> 00:02:04,020
It knows language, not meaning.

63
00:02:04,020 –> 00:02:05,420
The wall exists for good reason.

64
00:02:05,420 –> 00:02:08,860
Directly exposing SQL data to the cloud is corporate self-harm.

65
00:02:08,860 –> 00:02:11,860
Firewalls, network zones, and authentication boundaries

66
00:02:11,860 –> 00:02:14,060
exist precisely because someone once tried

67
00:02:14,060 –> 00:02:15,300
just opening a port

68
00:02:15,300 –> 00:02:18,060
and spent the next quarter explaining the breach.

69
00:02:18,060 –> 00:02:20,540
Compliance frameworks require data residency

70
00:02:20,540 –> 00:02:23,220
and auditors demand logs that show precisely who touched

71
00:02:23,220 –> 00:02:24,060
which record.

72
00:02:24,060 –> 00:02:25,540
Yes, the wall must stay.

73
00:02:25,540 –> 00:02:27,460
Yet isolation isn’t the answer either.

74
00:02:27,460 –> 00:02:29,460
The ideal is hybrid parity,

75
00:02:29,460 –> 00:02:31,740
keeping on-prem control while granting the cloud

76
00:02:31,740 –> 00:02:33,020
intelligent visibility.

77
00:02:33,020 –> 00:02:35,380
That balance transforms AI from a pilot trick

78
00:02:35,380 –> 00:02:36,860
into a dependable analyst.

79
00:02:36,860 –> 00:02:39,540
Picture a system where your co-pilot reads customer orders

80
00:02:39,540 –> 00:02:40,940
the instant they’re updated,

81
00:02:40,940 –> 00:02:43,820
where it summarizes invoices without exporting CSVs

82
00:02:43,820 –> 00:02:46,140
and where every query is authenticated, encrypted,

83
00:02:46,140 –> 00:02:47,140
and auditable.

84
00:02:47,140 –> 00:02:48,460
That’s hybrid done correctly.

85
00:02:48,460 –> 00:02:50,860
Understanding this split, the genius trapped outside

86
00:02:50,860 –> 00:02:53,180
and the data locked inside is the first step

87
00:02:53,180 –> 00:02:55,500
toward appreciating the architectural slide of hand

88
00:02:55,500 –> 00:02:56,580
that solves it.

89
00:02:56,580 –> 00:02:58,900
Before we talk about data, think in biology.

90
00:02:58,900 –> 00:03:01,860
The body operates because the spinal cord connects brain

91
00:03:01,860 –> 00:03:05,180
to muscle without exposing nerves to daylight.

92
00:03:05,180 –> 00:03:08,860
In technology, the power platform data gateway does precisely that.

93
00:03:08,860 –> 00:03:09,860
It’s not just a tunnel.

94
00:03:09,860 –> 00:03:12,140
It’s a disciplined neural bridge that keeps both

95
00:03:12,140 –> 00:03:14,220
hemispheres synchronized and secure.

96
00:03:14,220 –> 00:03:16,780
Once you understand that, everything about hybrid AI

97
00:03:16,780 –> 00:03:18,420
begins to click.

98
00:03:18,420 –> 00:03:21,780
Enter the data gateway, the spine of hybrid AI.

99
00:03:21,780 –> 00:03:23,780
Let’s start with the correction of language.

100
00:03:23,780 –> 00:03:26,460
People call the power platform data gateway middleware.

101
00:03:26,460 –> 00:03:27,820
That word is an insult.

102
00:03:27,820 –> 00:03:31,100
Middleware is what you use when two systems refuse to cooperate.

103
00:03:31,100 –> 00:03:33,540
The gateway isn’t a translator, it’s a spinal column.

104
00:03:33,540 –> 00:03:35,300
It links the cloud’s analytical brain

105
00:03:35,300 –> 00:03:38,460
with the reflex-driven body of your on-prem SQL server.

106
00:03:38,460 –> 00:03:40,380
Those two hemispheres must communicate constantly

107
00:03:40,380 –> 00:03:41,380
but never recklessly.

108
00:03:41,380 –> 00:03:43,140
The data gateway handles that conversation

109
00:03:43,140 –> 00:03:44,500
with surgical precision.

110
00:03:44,500 –> 00:03:46,700
Here’s how it thinks, nothing from the cloud ever knocks

111
00:03:46,700 –> 00:03:47,460
on your firewall.

112
00:03:47,460 –> 00:03:48,940
The gateway maintains sovereignty

113
00:03:48,940 –> 00:03:51,060
by initiating every conversation outward,

114
00:03:51,060 –> 00:03:54,340
picture it like an employee who only makes phone calls.

115
00:03:54,340 –> 00:03:56,900
They never accept incoming ones.

116
00:03:56,900 –> 00:03:58,300
The cloud sends no invitation.

117
00:03:58,300 –> 00:04:00,860
Your gateway dials the number, encrypts the session,

118
00:04:00,860 –> 00:04:03,580
verifies the credentials, and keeps the channel alive

119
00:04:03,580 –> 00:04:06,100
just long enough for safe command and response.

120
00:04:06,100 –> 00:04:07,660
From a security auditor’s perspective,

121
00:04:07,660 –> 00:04:09,900
that one architectural decision outbound only

122
00:04:09,900 –> 00:04:12,140
is the difference between compliance and chaos.

123
00:04:12,140 –> 00:04:14,700
Now, installing it is almost disappointingly simple.

124
00:04:14,700 –> 00:04:17,300
You download the on-premises data gateway client,

125
00:04:17,300 –> 00:04:20,020
sign in with your organization’s power platform account,

126
00:04:20,020 –> 00:04:23,100
and register it under a unique gateway cluster name.

127
00:04:23,100 –> 00:04:24,460
Behind that modest interface,

128
00:04:24,460 –> 00:04:26,620
lives serious engineering.

129
00:04:26,620 –> 00:04:29,100
Connection strings sealed in the Windows credential store,

130
00:04:29,100 –> 00:04:31,020
symmetric keys for data encryption,

131
00:04:31,020 –> 00:04:32,900
and a lightweight Windows service dedicated

132
00:04:32,900 –> 00:04:35,580
to maintaining secure communication with Azure.

133
00:04:35,580 –> 00:04:36,900
The moment registration completes,

134
00:04:36,900 –> 00:04:38,820
your local server quietly joins the roster

135
00:04:38,820 –> 00:04:41,820
of trusted hybrid nodes recognized by the power platform.

136
00:04:41,820 –> 00:04:44,740
Gateway clusters are the unsung heroes of enterprise resilience.

137
00:04:44,740 –> 00:04:46,700
You can deploy more than one instance

138
00:04:46,700 –> 00:04:50,060
on separate machines, each functioning as a backup route.

139
00:04:50,060 –> 00:04:51,500
Should one node stop responding,

140
00:04:51,500 –> 00:04:53,980
maybe a maintenance reboot or a hardware hiccup,

141
00:04:53,980 –> 00:04:55,900
the others continue routing traffic.

142
00:04:55,900 –> 00:04:57,820
Power platform services automatically

143
00:04:57,820 –> 00:04:59,980
balance connections between available members.

144
00:04:59,980 –> 00:05:03,700
The result, high availability without ever exposing an open port.

145
00:05:03,700 –> 00:05:05,700
Microsoft designed it so reliability never

146
00:05:05,700 –> 00:05:07,420
trades places with recklessness.

147
00:05:07,420 –> 00:05:09,100
And here’s the bonus most overlooked.

148
00:05:09,100 –> 00:05:10,780
One gateway serves them all.

149
00:05:10,780 –> 00:05:12,700
The same installation that enables your co-pilot

150
00:05:12,700 –> 00:05:15,740
to query local SQL also powers reports in Power BI,

151
00:05:15,740 –> 00:05:18,620
apps in Power Apps and flows in Power Automate.

152
00:05:18,620 –> 00:05:21,740
In other words, every hybrid connection in the power platform,

153
00:05:21,740 –> 00:05:24,460
ecosystem shares that identical spinal path.

154
00:05:24,460 –> 00:05:26,260
Each signal runs up and down the same nerve

155
00:05:26,260 –> 00:05:28,460
and none of them bypass security policy.

156
00:05:28,460 –> 00:05:30,860
That shared backbone eliminates redundant connectors

157
00:05:30,860 –> 00:05:32,180
and network clutter.

158
00:05:32,180 –> 00:05:34,700
One discipline bridge instead of four chaotic tunnels.

159
00:05:34,700 –> 00:05:37,740
Let’s preempt the paranoia that flares in every security review.

160
00:05:37,740 –> 00:05:39,860
No, the gateway does not upload your database.

161
00:05:39,860 –> 00:05:42,380
It doesn’t clone, mirror, or replicate anything.

162
00:05:42,380 –> 00:05:44,580
All it does is execute queries on your behalf

163
00:05:44,580 –> 00:05:47,460
and return the results just as if a well-trained employee

164
00:05:47,460 –> 00:05:51,140
ran a stored procedure and copied the outcome into a secure message.

165
00:05:51,140 –> 00:05:52,580
The session keys roll frequently.

166
00:05:52,580 –> 00:05:55,540
The payloads are encrypted end to end using TLS.

167
00:05:55,540 –> 00:05:58,100
And authentication goes through as your active directory

168
00:05:58,100 –> 00:05:59,860
or the credentials you explicitly supply.

169
00:05:59,860 –> 00:06:01,900
There is no ghost copy, no hidden cache,

170
00:06:01,900 –> 00:06:04,340
no covert synchronization hiding under your desk.

171
00:06:04,340 –> 00:06:06,900
For regulatory environments that live in audit logs,

172
00:06:06,900 –> 00:06:08,780
the gateway also generates telemetry.

173
00:06:08,780 –> 00:06:11,380
Every call, every result set, every authentication

174
00:06:11,380 –> 00:06:14,780
and handshake can be tracked through power platform monitoring tools.

175
00:06:14,780 –> 00:06:17,220
That means you can prove to compliance line by line

176
00:06:17,220 –> 00:06:20,260
that data never left your trusted boundary unencrypted.

177
00:06:20,260 –> 00:06:21,780
The effect is paradoxical.

178
00:06:21,780 –> 00:06:24,820
Opening the wall actually strengthens your evidence of control.

179
00:06:24,820 –> 00:06:26,700
Auditors love diagrams with gateways

180
00:06:26,700 –> 00:06:28,340
because suddenly the arrows in the network

181
00:06:28,340 –> 00:06:30,420
map point the correct way outbound.

182
00:06:30,420 –> 00:06:33,540
So to recap in biological terms, SQL server is the muscle.

183
00:06:33,540 –> 00:06:35,380
Copilot studio is the frontal cortex.

184
00:06:35,380 –> 00:06:38,340
The data gateway is the mylinated nerve fiber connecting the two,

185
00:06:38,340 –> 00:06:39,740
a high way of electrical activity

186
00:06:39,740 –> 00:06:42,300
wrapped in layers of encryption instead of tissue.

187
00:06:42,300 –> 00:06:44,100
Without it, the cloud brain sends commands

188
00:06:44,100 –> 00:06:45,740
that never reach the limbs.

189
00:06:45,740 –> 00:06:48,620
With it queries, updates and context flow symmetrically,

190
00:06:48,620 –> 00:06:51,420
both directions, without violating the skin of your perimeter.

191
00:06:51,420 –> 00:06:54,100
Once that spine exists, we can attach the brain.

192
00:06:54,100 –> 00:06:56,820
Copilot studio will soon learn to read your SQL tables

193
00:06:56,820 –> 00:06:59,660
as knowledge sources, constructing natural language questions

194
00:06:59,660 –> 00:07:02,100
that translate into precise T-School commands.

195
00:07:02,100 –> 00:07:04,540
The gateway stands guard, translating intent

196
00:07:04,540 –> 00:07:07,540
into execution and returning verified results.

197
00:07:07,540 –> 00:07:10,020
What happens next when the copilot finally understands

198
00:07:10,020 –> 00:07:12,100
the contents of those tables in real time

199
00:07:12,100 –> 00:07:14,980
is where the promise of hybrid AI stops being a buzzword

200
00:07:14,980 –> 00:07:16,820
and becomes a functioning nervous system.

201
00:07:16,820 –> 00:07:18,740
And yes, that’s our next step.

202
00:07:18,740 –> 00:07:22,380
Teaching copilot to read SQL, adding knowledge sources.

203
00:07:22,380 –> 00:07:25,340
A copilot without data is like an intern with enthusiasm

204
00:07:25,340 –> 00:07:26,340
and no memory.

205
00:07:26,340 –> 00:07:28,540
It smiles, nods and answers confidently

206
00:07:28,540 –> 00:07:30,100
while secretly improvising.

207
00:07:30,100 –> 00:07:32,260
The first lesson in hybrid AI literacy

208
00:07:32,260 –> 00:07:35,060
is giving that intern access to the company’s archives,

209
00:07:35,060 –> 00:07:36,820
carefully, securely and on your terms.

210
00:07:36,820 –> 00:07:39,580
That’s where knowledge sources in copilot studio come in.

211
00:07:39,580 –> 00:07:42,340
What you’re about to build isn’t a simple connection string.

212
00:07:42,340 –> 00:07:43,700
It’s cognition.

213
00:07:43,700 –> 00:07:45,900
We begin with a blank agent in copilot studio.

214
00:07:45,900 –> 00:07:47,860
It’s empty, no knowledge, no tools,

215
00:07:47,860 –> 00:07:50,020
just linguistic talent waiting for context.

216
00:07:50,020 –> 00:07:51,580
The moment you click at knowledge,

217
00:07:51,580 –> 00:07:53,740
you shift from wordplay to data access,

218
00:07:53,740 –> 00:07:55,740
choose Azure S-Cole as the source

219
00:07:55,740 –> 00:07:59,420
and hear the data gateway performs its first act of diplomacy.

220
00:07:59,420 –> 00:08:00,900
Because you already registered it,

221
00:08:00,900 –> 00:08:04,220
your local SQL instance quietly appears in the connection list.

222
00:08:04,220 –> 00:08:07,100
It’s that same gateway sitting inside your network,

223
00:08:07,100 –> 00:08:10,060
initiating outbound trust to power platform.

224
00:08:10,060 –> 00:08:12,500
You select it, authenticate and point to the database

225
00:08:12,500 –> 00:08:14,500
holding your operational truth.

226
00:08:14,500 –> 00:08:16,620
Authentication matters more than most realize.

227
00:08:16,620 –> 00:08:19,220
SQL authentication uses dedicated database credentials,

228
00:08:19,220 –> 00:08:20,100
simple but local.

229
00:08:20,100 –> 00:08:23,380
Windows authentication leverages existing active directory trust.

230
00:08:23,380 –> 00:08:26,540
Perfect when your gateway machine already belongs to the domain.

231
00:08:26,540 –> 00:08:28,020
Then there’s the Azure hybrid approach

232
00:08:28,020 –> 00:08:29,740
where Azure AD acts as broker

233
00:08:29,740 –> 00:08:31,740
between cloud identity and local permissions.

234
00:08:31,740 –> 00:08:33,660
Each option satisfies different combinations

235
00:08:33,660 –> 00:08:35,660
of corporate paranoia and practical need.

236
00:08:35,660 –> 00:08:38,900
The point is that copilot never sees the password directly.

237
00:08:38,900 –> 00:08:40,820
The gateway handles credential storage

238
00:08:40,820 –> 00:08:42,140
through encrypted reference

239
00:08:42,140 –> 00:08:45,460
as if it were the company’s sealed envelope policy.

240
00:08:45,460 –> 00:08:47,900
Once authenticated copilot studio politely asks

241
00:08:47,900 –> 00:08:49,180
what you’d like it to know.

242
00:08:49,180 –> 00:08:52,380
Each table or view you select defines a boundary of knowledge.

243
00:08:52,380 –> 00:08:55,780
Choose carefully, feed it messy schema and you’ll train confusion,

244
00:08:55,780 –> 00:08:57,580
feed it normalized, well-named views

245
00:08:57,580 –> 00:08:59,740
and it will respond like a seasoned analyst.

246
00:08:59,740 –> 00:09:01,660
Think of schema design as diction.

247
00:09:01,660 –> 00:09:05,220
Clear column names become vocabulary copilot can use

248
00:09:05,220 –> 00:09:08,700
while cryptic abbreviations turn sentences in coherent.

249
00:09:08,700 –> 00:09:10,540
The model doesn’t understand joints.

250
00:09:10,540 –> 00:09:13,500
It infers relationships from the structure you expose.

251
00:09:13,500 –> 00:09:15,980
That’s why many architects create read optimized views,

252
00:09:15,980 –> 00:09:18,300
condensed precise representations of the truth,

253
00:09:18,300 –> 00:09:20,860
pre-joint and scrubbed of sensitive columns.

254
00:09:20,860 –> 00:09:23,340
After linking tables, copilot studio indexes

255
00:09:23,340 –> 00:09:25,020
their metadata through the gateway.

256
00:09:25,020 –> 00:09:26,860
It doesn’t duplicate your data.

257
00:09:26,860 –> 00:09:30,100
Instead it prepares schemas for dynamic querying.

258
00:09:30,100 –> 00:09:31,700
When you ask a question say,

259
00:09:31,700 –> 00:09:33,860
what’s Greenfield Corpse recent order total?

260
00:09:33,860 –> 00:09:36,140
Copilot generates an internal SQL statement

261
00:09:36,140 –> 00:09:37,580
referencing those views.

262
00:09:37,580 –> 00:09:40,180
The gateway executes it locally, pulls back results

263
00:09:40,180 –> 00:09:42,780
and sends a sanitized JSON payload to the model.

264
00:09:42,780 –> 00:09:45,460
The model then reformats that output into natural speech.

265
00:09:45,460 –> 00:09:47,380
To you, it looks like language magic.

266
00:09:47,380 –> 00:09:49,980
To the network administrator, it’s a single outbound call,

267
00:09:49,980 –> 00:09:52,260
wrapped in TLS, logged and closed.

268
00:09:52,260 –> 00:09:55,420
Context persistence is where things feel eerily human.

269
00:09:55,420 –> 00:09:57,620
Ask about Greenfield Corpse latest order

270
00:09:57,620 –> 00:10:00,660
then immediately follow up with what items were included.

271
00:10:00,660 –> 00:10:02,340
Copilot doesn’t lose track of the subject

272
00:10:02,340 –> 00:10:04,540
because conversation history and query context

273
00:10:04,540 –> 00:10:06,140
write the same secure path.

274
00:10:06,140 –> 00:10:07,820
It remembers the customer referenced,

275
00:10:07,820 –> 00:10:10,820
constructs a second SQL query filtered by that ID

276
00:10:10,820 –> 00:10:12,660
and delivers the itemized list

277
00:10:12,660 –> 00:10:14,620
still without pre-storing anything.

278
00:10:14,620 –> 00:10:17,220
Essentially, copilot behaves like an attentive analyst

279
00:10:17,220 –> 00:10:18,980
who keeps the prior spreadsheet open

280
00:10:18,980 –> 00:10:20,300
while answering the next question.

281
00:10:20,300 –> 00:10:22,940
Because every query travels live through the gateway,

282
00:10:22,940 –> 00:10:25,420
responses reflect the current state of SQL

283
00:10:25,420 –> 00:10:26,860
at the exact moment you ask.

284
00:10:26,860 –> 00:10:30,340
Modify a record in SQL management studio and re-ask.

285
00:10:30,340 –> 00:10:31,900
The answer updates instantly.

286
00:10:31,900 –> 00:10:34,940
That’s not caching its genuine real-time data retrieval.

287
00:10:34,940 –> 00:10:36,780
This immediacy closes the classical lag

288
00:10:36,780 –> 00:10:38,780
between analytics and operations.

289
00:10:38,780 –> 00:10:41,300
Your copilot stops being a storyteller about old data

290
00:10:41,300 –> 00:10:43,260
and becomes a reporter for the present tense.

291
00:10:43,260 –> 00:10:44,380
Common mistakes?

292
00:10:44,380 –> 00:10:46,340
Over permissive access tops the list.

293
00:10:46,340 –> 00:10:48,380
Always restrict the connection to the few tables

294
00:10:48,380 –> 00:10:49,980
copilot actually needs.

295
00:10:49,980 –> 00:10:52,300
And avoid giant unfiltered resultsets.

296
00:10:52,300 –> 00:10:54,140
Language models aren’t designed to summarize

297
00:10:54,140 –> 00:10:55,420
millions of rows at once.

298
00:10:55,420 –> 00:10:58,820
Instead, scope the knowledge through concise relevant views.

299
00:10:58,820 –> 00:11:00,940
Another pitfall is forgetting data types.

300
00:11:00,940 –> 00:11:03,060
Copilot interprets the schema literally.

301
00:11:03,060 –> 00:11:05,420
If you store numeric identifiers as strings,

302
00:11:05,420 –> 00:11:06,740
expect confusion.

303
00:11:06,740 –> 00:11:08,660
The more disciplined your database design,

304
00:11:08,660 –> 00:11:10,700
the more articulate your copilot becomes.

305
00:11:10,700 –> 00:11:11,740
So what have we accomplished?

306
00:11:11,740 –> 00:11:13,340
We’ve given the intern eyesight.

307
00:11:13,340 –> 00:11:15,460
Copilot can now read live company data

308
00:11:15,460 –> 00:11:18,340
with perfect recall and zero exfiltration risk.

309
00:11:18,340 –> 00:11:19,700
It answers customer queries

310
00:11:19,700 –> 00:11:21,780
by translating natural language into SQL

311
00:11:21,780 –> 00:11:24,180
executing in milliseconds through your gateway.

312
00:11:24,180 –> 00:11:26,540
And while that’s impressive, an AI that reads your ledger

313
00:11:26,540 –> 00:11:28,460
like a novel, the real transformation

314
00:11:28,460 –> 00:11:30,180
happens when it learns to act.

315
00:11:30,180 –> 00:11:31,980
Reading data makes it informative.

316
00:11:31,980 –> 00:11:33,580
Writing data makes it valuable.

317
00:11:33,580 –> 00:11:35,340
In the next stage, we give it hands.

318
00:11:35,340 –> 00:11:37,340
With SQL actions and controlled writebacks

319
00:11:37,340 –> 00:11:40,060
that eager intern upgrades to a trusted employee capable

320
00:11:40,060 –> 00:11:42,820
of updating reality, not merely describing it.

321
00:11:42,820 –> 00:11:45,660
Giving copilot hands, SQL actions and writebacks,

322
00:11:45,660 –> 00:11:48,220
up to this point, your copilot has been the perfect data

323
00:11:48,220 –> 00:11:51,580
analyst, curious, articulate, but fundamentally harmless.

324
00:11:51,580 –> 00:11:53,740
It observes your SQL server like a museum visitor

325
00:11:53,740 –> 00:11:55,180
behind rope barriers.

326
00:11:55,180 –> 00:11:56,500
Now we remove the glass.

327
00:11:56,500 –> 00:11:59,460
The time has come for copilot to act on the world it understands

328
00:11:59,460 –> 00:12:02,420
to insert, update and maintain records through SQL

329
00:12:02,420 –> 00:12:03,860
rather than merely describe them.

330
00:12:03,860 –> 00:12:06,140
This is the moment copilot graduates from librarian

331
00:12:06,140 –> 00:12:08,100
to employee in copilot studio

332
00:12:08,100 –> 00:12:10,140
that transformation begins in the tool section,

333
00:12:10,140 –> 00:12:11,860
sometimes labeled actions.

334
00:12:11,860 –> 00:12:14,420
Here you define what the AI is allowed to do.

335
00:12:14,420 –> 00:12:16,940
Each action is a contract between human administrators

336
00:12:16,940 –> 00:12:18,020
and machine intention.

337
00:12:18,020 –> 00:12:20,700
You expose certain functions, describe them clearly

338
00:12:20,700 –> 00:12:23,180
and let the model decide when they’re appropriate.

339
00:12:23,180 –> 00:12:25,740
Conceptually these are APIs with etiquette.

340
00:12:25,740 –> 00:12:29,340
Without them, copilot speaks with them copilot performs.

341
00:12:29,340 –> 00:12:32,380
Start by adding a new action and choosing the SQL connector.

342
00:12:32,380 –> 00:12:34,940
The options mimic the verbs of database life.

343
00:12:34,940 –> 00:12:38,180
Insert, update, delete, execute, stored procedure.

344
00:12:38,180 –> 00:12:40,820
Let’s select insert row because creation is the purest form

345
00:12:40,820 –> 00:12:41,580
of proof.

346
00:12:41,580 –> 00:12:43,420
The interface prompts you to pick a connection,

347
00:12:43,420 –> 00:12:45,900
the same one we configured earlier through the data gateway.

348
00:12:45,900 –> 00:12:47,020
That continuity matters.

349
00:12:47,020 –> 00:12:49,500
It means your write operations travel along

350
00:12:49,500 –> 00:12:51,460
the same encrypted nerve as your queries.

351
00:12:51,460 –> 00:12:54,180
No extra tunnel, no unmonitored path.

352
00:12:54,180 –> 00:12:56,020
Authentication context is preserved

353
00:12:56,020 –> 00:12:57,780
and governance remains intact.

354
00:12:57,780 –> 00:13:00,300
Next, you identify where this action should operate.

355
00:13:00,300 –> 00:13:03,100
Choose your database, then your table, perhaps customers.

356
00:13:03,100 –> 00:13:06,420
The moment you selected, copilot studio introspects the schema

357
00:13:06,420 –> 00:13:08,940
and lists the columns as input parameters.

358
00:13:08,940 –> 00:13:11,060
These become the fields copilot must supply

359
00:13:11,060 –> 00:13:13,180
before executing the SQL command.

360
00:13:13,180 –> 00:13:15,460
Think of each parameter as a missing puzzle piece

361
00:13:15,460 –> 00:13:18,220
the language model has to find through conversation.

362
00:13:18,220 –> 00:13:19,780
The art lies in labeling.

363
00:13:19,780 –> 00:13:22,260
Don’t leave parameter names as cryptic identifiers

364
00:13:22,260 –> 00:13:24,260
like Cust ID or PHNUM.

365
00:13:24,260 –> 00:13:27,820
Rename them to natural prompts, customer ID, phone number,

366
00:13:27,820 –> 00:13:29,100
email address.

367
00:13:29,100 –> 00:13:31,500
In the models world, clarity is destiny.

368
00:13:31,500 –> 00:13:34,340
You can also provide concise descriptions for each field,

369
00:13:34,340 –> 00:13:37,620
unique numeric ID for the customer, primary contact email,

370
00:13:37,620 –> 00:13:38,460
and so forth.

371
00:13:38,460 –> 00:13:40,740
These hints guide copilot’s slot-filling logic

372
00:13:40,740 –> 00:13:42,140
when it lacks information.

373
00:13:42,140 –> 00:13:46,580
For example, if a user says add a new client named Dubert 365,

374
00:13:46,580 –> 00:13:49,180
the model sees it has a name but no phone or address.

375
00:13:49,180 –> 00:13:52,980
It asks politely, what’s their phone number and business address?

376
00:13:52,980 –> 00:13:54,340
That follow-up isn’t scripted.

377
00:13:54,340 –> 00:13:57,460
It’s inference born from your parameter metadata.

378
00:13:57,460 –> 00:13:59,820
Once copilot gathers all required inputs,

379
00:13:59,820 –> 00:14:03,540
the gateway executes the SQL command silently, just as before.

380
00:14:03,540 –> 00:14:05,140
Outbound encrypted logged.

381
00:14:05,140 –> 00:14:08,980
Within seconds, the new record materializes inside SQL server.

382
00:14:08,980 –> 00:14:11,220
The experience to the user feels magical.

383
00:14:11,220 –> 00:14:14,060
One conversational request creates tangible data

384
00:14:14,060 –> 00:14:17,060
in an on-prem system without any browser, plug-in,

385
00:14:17,060 –> 00:14:19,820
or direct database exposure.

386
00:14:19,820 –> 00:14:21,420
The firewall remains unsolid.

387
00:14:21,420 –> 00:14:23,020
The network admin remains calm.

388
00:14:23,020 –> 00:14:24,540
Validation is critical here.

389
00:14:24,540 –> 00:14:26,460
The connector respects SQL constraints,

390
00:14:26,460 –> 00:14:28,660
primary keys, data types, and triggers,

391
00:14:28,660 –> 00:14:31,260
but it’s wise to implement additional sanity checks.

392
00:14:31,260 –> 00:14:33,860
You can include conditional flows in copilot studio

393
00:14:33,860 –> 00:14:35,260
to confirm before committing like,

394
00:14:35,260 –> 00:14:37,380
“Are you sure you want to create this customer?”

395
00:14:37,380 –> 00:14:39,860
Each confirmation step not only prevents accidents

396
00:14:39,860 –> 00:14:42,500
but also provides a clear paper trail for auditors.

397
00:14:42,500 –> 00:14:45,660
Remember, governing AI means supervising enthusiasm.

398
00:14:45,660 –> 00:14:47,300
Now, about safety.

399
00:14:47,300 –> 00:14:49,740
Many organizations sensibly divide knowledge

400
00:14:49,740 –> 00:14:51,140
and action credentials.

401
00:14:51,140 –> 00:14:54,340
Reading might use a service account with select rights only

402
00:14:54,340 –> 00:14:57,580
while writing requires an elevated connector approved by IT.

403
00:14:57,580 –> 00:15:00,420
Copilot Studio allows you to maintain separate connections

404
00:15:00,420 –> 00:15:03,460
for these layers all under the same gateway infrastructure.

405
00:15:03,460 –> 00:15:06,420
This separation of duties ensures that even if a configuration

406
00:15:06,420 –> 00:15:08,980
misfires, no rogue agent gains right access

407
00:15:08,980 –> 00:15:10,660
beyond its intended scope.

408
00:15:10,660 –> 00:15:13,500
Observe how elegantly the gateway handles dual purpose.

409
00:15:13,500 –> 00:15:16,460
It translates natural language into T-Suckel-Bot directions

410
00:15:16,460 –> 00:15:18,820
yet keeps authentication centralized.

411
00:15:18,820 –> 00:15:21,860
The administrator doesn’t manage dozens of API keys.

412
00:15:21,860 –> 00:15:24,060
The gateway proxy manages trust once

413
00:15:24,060 –> 00:15:25,900
and replicates it responsibly.

414
00:15:25,900 –> 00:15:28,020
Compliance officers rejoice because every writeback

415
00:15:28,020 –> 00:15:30,220
is timestamped, traceable, and reversible.

416
00:15:30,220 –> 00:15:32,940
You can open Power Platform Telemetry and see precisely

417
00:15:32,940 –> 00:15:37,060
which user invoked which action against which table at what time.

418
00:15:37,060 –> 00:15:40,020
That’s not automation gone wild, that’s automation domesticated.

419
00:15:40,020 –> 00:15:41,500
Let’s return to the demo example.

420
00:15:41,500 –> 00:15:42,780
You instruct copilot.

421
00:15:42,780 –> 00:15:44,820
Create a new customer record.

422
00:15:44,820 –> 00:15:47,700
It interprets the intent, checks available tools,

423
00:15:47,700 –> 00:15:50,220
and finds your create new customer record action.

424
00:15:50,220 –> 00:15:53,060
Missing parameters trigger questions until complete.

425
00:15:53,060 –> 00:15:56,260
When it finally executes SQL Server gains an 11th customer.

426
00:15:56,260 –> 00:15:58,940
Refresh the table in management studio and there it is.

427
00:15:58,940 –> 00:16:01,300
Proof that conversation translated into commerce.

428
00:16:01,300 –> 00:16:04,860
Your AI didn’t just summarize reality, it altered it responsibly.

429
00:16:04,860 –> 00:16:06,740
That’s the essence of giving Copilot hands.

430
00:16:06,740 –> 00:16:08,860
By exposing a controlled set of SQL actions

431
00:16:08,860 –> 00:16:11,100
through the data gateway, you empower intelligence

432
00:16:11,100 –> 00:16:12,980
to participate in daily operations

433
00:16:12,980 –> 00:16:16,060
while retaining the guardrails of enterprise data governance.

434
00:16:16,060 –> 00:16:18,140
Each action is a carefully fenced off power,

435
00:16:18,140 –> 00:16:20,780
bounded capability rather than unlimited access.

436
00:16:20,780 –> 00:16:22,820
When configured well, your copilot becomes

437
00:16:22,820 –> 00:16:26,380
both informative and operational, capable of performing transactions,

438
00:16:26,380 –> 00:16:30,100
logging every keystroke, and learning proper workplace discipline.

439
00:16:30,100 –> 00:16:31,220
Congratulations.

440
00:16:31,220 –> 00:16:33,580
You’ve just hired your first digital employee

441
00:16:33,580 –> 00:16:36,100
and built its desk inside SQL Server.

442
00:16:36,100 –> 00:16:39,020
Designing the hybrid brain, architecture, and scaling.

443
00:16:39,020 –> 00:16:40,420
What you have now is more than a demo.

444
00:16:40,420 –> 00:16:42,740
It’s a nervous system, but every nervous system

445
00:16:42,740 –> 00:16:46,300
eventually meets reality, lag, failure, and scale.

446
00:16:46,300 –> 00:16:47,780
This section is for the architects,

447
00:16:47,780 –> 00:16:49,500
the people who must explain to leadership

448
00:16:49,500 –> 00:16:52,420
why the copilot doesn’t melt under enterprise load

449
00:16:52,420 –> 00:16:55,540
and why hybrid doesn’t secretly mean fragile.

450
00:16:55,540 –> 00:16:58,700
Think of the hybrid brain as four organs in one organism.

451
00:16:58,700 –> 00:17:01,580
The data source, SQL Server, is the memory cortex,

452
00:17:01,580 –> 00:17:04,140
storing knowledge in perfect tabular patterns.

453
00:17:04,140 –> 00:17:06,100
The gateway layer is the spinal cord,

454
00:17:06,100 –> 00:17:08,900
transmitting signals both ways while filtering anything unfit

455
00:17:08,900 –> 00:17:09,540
for travel.

456
00:17:09,540 –> 00:17:12,420
The cloud services, power platform, and copilot studio

457
00:17:12,420 –> 00:17:14,740
are the prefrontal cortex, interpreting language,

458
00:17:14,740 –> 00:17:17,140
applying reasoning, managing context.

459
00:17:17,140 –> 00:17:20,020
Finally, the front ends, teams, web chat, mobile,

460
00:17:20,020 –> 00:17:21,860
are the mouth and hands where humans actually

461
00:17:21,860 –> 00:17:22,980
interact with the machine.

462
00:17:22,980 –> 00:17:24,940
Keep those roles distinct when one tries

463
00:17:24,940 –> 00:17:27,060
to perform another’s function, technical back pain,

464
00:17:27,060 –> 00:17:27,580
ensues.

465
00:17:27,580 –> 00:17:29,500
Resilience begins with redundancy.

466
00:17:29,500 –> 00:17:31,620
Deploy multiple gateways on separate servers

467
00:17:31,620 –> 00:17:32,620
to form a cluster.

468
00:17:32,620 –> 00:17:35,340
They share one identity, one connection reference,

469
00:17:35,340 –> 00:17:37,140
but balance the work among themselves.

470
00:17:37,140 –> 00:17:39,020
If a single machine crashes or someone

471
00:17:39,020 –> 00:17:42,460
casually reboots it during patch week, the others carry on.

472
00:17:42,460 –> 00:17:43,700
The copilot notices nothing.

473
00:17:43,700 –> 00:17:45,900
The power platform automatically roots connections

474
00:17:45,900 –> 00:17:47,420
to the available node.

475
00:17:47,420 –> 00:17:49,820
No manual intervention, no downtime.

476
00:17:49,820 –> 00:17:52,780
For auditors, the cluster is a comforting diagram.

477
00:17:52,780 –> 00:17:55,100
Two arrows, instead of one failure point.

478
00:17:55,100 –> 00:17:56,740
Next comes load management.

479
00:17:56,740 –> 00:17:59,300
Queries generated by copilot are unpredictable.

480
00:17:59,300 –> 00:18:01,900
Short text requests, one minute, large analytical joins

481
00:18:01,900 –> 00:18:02,660
the next.

482
00:18:02,660 –> 00:18:04,620
A well-designed schema prevents those spur

483
00:18:04,620 –> 00:18:06,660
of the moment join explosions.

484
00:18:06,660 –> 00:18:08,980
Use read optimized views, indexed keys

485
00:18:08,980 –> 00:18:10,300
and row-level filters.

486
00:18:10,300 –> 00:18:12,940
The data gateway executes SQL on your local network,

487
00:18:12,940 –> 00:18:15,260
so it inherits whatever indexes you’ve built.

488
00:18:15,260 –> 00:18:17,860
Optimal indexing isn’t an academic suggestion.

489
00:18:17,860 –> 00:18:20,020
It’s the reason copilot answers in seconds,

490
00:18:20,020 –> 00:18:21,460
rather than sulking in timeout.

491
00:18:21,460 –> 00:18:23,660
Then there’s auditability, the bureaucratic soul

492
00:18:23,660 –> 00:18:24,820
of the hybrid brain.

493
00:18:24,820 –> 00:18:27,580
Every tool execution, every query, every authentication

494
00:18:27,580 –> 00:18:30,460
request, surfaces in power platform telemetry.

495
00:18:30,460 –> 00:18:31,220
Use it.

496
00:18:31,220 –> 00:18:33,700
Export logs to log analytics or Sentinel

497
00:18:33,700 –> 00:18:36,900
apply filters by user or time and demonstrate compliance

498
00:18:36,900 –> 00:18:37,820
numerically.

499
00:18:37,820 –> 00:18:40,580
When your security officer asks, who updated the customer

500
00:18:40,580 –> 00:18:42,060
table last Thursday?

501
00:18:42,060 –> 00:18:43,620
You can answer with painful precision.

502
00:18:43,620 –> 00:18:45,780
Nothing convinces governance like timestamps.

503
00:18:45,780 –> 00:18:48,820
Edge cases deserve mention because they are inevitable.

504
00:18:48,820 –> 00:18:50,580
Legacy authentication still lurks.

505
00:18:50,580 –> 00:18:52,740
Some environments run ancient SQL authentication

506
00:18:52,740 –> 00:18:55,940
where the password policy remembers the Bronze Age.

507
00:18:55,940 –> 00:18:57,540
Use the gateway’s credential store

508
00:18:57,540 –> 00:19:00,980
to hide that embarrassment and rotate keys regularly.

509
00:19:00,980 –> 00:19:04,060
Large data models can overwhelm copilot’s language interface,

510
00:19:04,060 –> 00:19:05,780
so summarizing through stored procedures

511
00:19:05,780 –> 00:19:08,860
is safer than letting it interpret million row JSONs.

512
00:19:08,860 –> 00:19:12,020
Dynamic schemas, tables that change weekly require automated

513
00:19:12,020 –> 00:19:13,420
metadata refresh.

514
00:19:13,420 –> 00:19:15,180
Schedule those connections to re-index

515
00:19:15,180 –> 00:19:18,700
nightly so your copilot doesn’t wake up confused Monday morning.

516
00:19:18,700 –> 00:19:21,660
Security philosophy underpins everything.

517
00:19:21,660 –> 00:19:23,380
The goal is not migration.

518
00:19:23,380 –> 00:19:26,540
Moving your crown jewel data to someone else’s cloud

519
00:19:26,540 –> 00:19:27,780
isn’t modernization.

520
00:19:27,780 –> 00:19:28,940
It’s surrender.

521
00:19:28,940 –> 00:19:31,340
The goal is synchronization without exposure.

522
00:19:31,340 –> 00:19:33,860
The gateway permits motion without relocation.

523
00:19:33,860 –> 00:19:36,100
Data stays in the jurisdiction auditors can visit

524
00:19:36,100 –> 00:19:39,060
while intelligence flows freely to the tools employees actually use.

525
00:19:39,060 –> 00:19:40,940
It’s the only equilibrium between control

526
00:19:40,940 –> 00:19:42,740
and productivity that scales.

527
00:19:42,740 –> 00:19:44,940
From a design standpoint, document the path.

528
00:19:44,940 –> 00:19:49,340
SQL server, memory, data gateway, spine, power platform

529
00:19:49,340 –> 00:19:52,220
cloud, brain, teams, or web, face.

530
00:19:52,220 –> 00:19:53,980
One continuous signal fully encrypted

531
00:19:53,980 –> 00:19:55,460
auditable at every hop.

532
00:19:55,460 –> 00:19:57,140
Once you internalize that pattern,

533
00:19:57,140 –> 00:19:59,900
replicating it for other systems becomes trivial.

534
00:19:59,900 –> 00:20:02,300
Change cycle for Oracle or a local API

535
00:20:02,300 –> 00:20:03,860
and the structure remains identical.

536
00:20:03,860 –> 00:20:04,780
Congratulations.

537
00:20:04,780 –> 00:20:07,780
You’ve just drawn the blueprint for hybrid AI itself.

538
00:20:07,780 –> 00:20:08,740
The real secret.

539
00:20:08,740 –> 00:20:11,900
So what’s the real secret to putting SQL data in copilot studio?

540
00:20:11,900 –> 00:20:13,820
It isn’t a command or a hidden switch.

541
00:20:13,820 –> 00:20:17,460
It’s architecture, respecting boundaries while designing pathways.

542
00:20:17,460 –> 00:20:19,780
Knowledge without connectivity is useless.

543
00:20:19,780 –> 00:20:22,380
Connectivity without control is dangerous.

544
00:20:22,380 –> 00:20:24,260
The data gateway resolves that paradox

545
00:20:24,260 –> 00:20:26,140
by letting intelligence cross the firewall

546
00:20:26,140 –> 00:20:27,300
without ever breaching it.

547
00:20:27,300 –> 00:20:30,020
With SQL as memory and copilot studio as reasoning,

548
00:20:30,020 –> 00:20:33,020
your organization finally owns a complete digital brain,

549
00:20:33,020 –> 00:20:35,460
capable of quoting invoices, adding customers,

550
00:20:35,460 –> 00:20:38,260
and learning while remaining inside policy.

551
00:20:38,260 –> 00:20:40,380
Real time hybrid intelligence isn’t law,

552
00:20:40,380 –> 00:20:42,580
it’s a symptom of wiring done properly.

553
00:20:42,580 –> 00:20:44,460
If this concept saved you another night

554
00:20:44,460 –> 00:20:47,860
of exporting CSVs, repay the favor, subscribe.

555
00:20:47,860 –> 00:20:49,700
Because next we extend this architecture

556
00:20:49,700 –> 00:20:52,500
to legacy APIs and flat file dinosaur systems,

557
00:20:52,500 –> 00:20:54,780
teaching copilot to communicate with everything else

558
00:20:54,780 –> 00:20:56,540
still haunting your server rack.

559
00:20:56,540 –> 00:20:58,260
The future of AI isn’t another model,

560
00:20:58,260 –> 00:20:59,420
it’s proper wiring.