WEBVTT
1
00:00:00.040 --> 00:00:02.520
Your fabric data warehouse is just a CSV graveyard. I
2
00:00:02.560 --> 00:00:04.120
know that stings, but look at how you're using it.
3
00:00:04.480 --> 00:00:08.400
Endless CSV dumps, cold tables, scheduled ETL jobs lumbering along
4
00:00:08.480 --> 00:00:11.480
like it's twenty fifteen. You bought Fabric to launch your
5
00:00:11.519 --> 00:00:13.759
data into the age of AI, and then you turned
6
00:00:13.759 --> 00:00:16.199
it into an archive. The irony is exquisite. Fabric was
7
00:00:16.199 --> 00:00:20.480
built for intelligence, real time insight, contextual reasoning, self adjusting analytics,
8
00:00:20.640 --> 00:00:23.839
yet here you are treating it like digital tupperware. Meanwhile,
9
00:00:23.920 --> 00:00:26.559
the AI layer you paid for, the data agents, the
10
00:00:26.600 --> 00:00:30.640
contextual governance, the semantic reasoning sits dormant, waiting for instructions
11
00:00:30.640 --> 00:00:32.960
that never come. So the problem isn't capacity and it's
12
00:00:33.000 --> 00:00:36.759
not data quality. It's thinking. You don't have a data problem.
13
00:00:36.799 --> 00:00:40.200
You have a conceptual one, mistaking intelligence infrastructure for storage.
14
00:00:40.399 --> 00:00:43.119
Let's fix that mental model before your CFO realizers. You've
15
00:00:43.119 --> 00:00:47.079
reinvented a network drive with better branding. The dead data problem.
16
00:00:47.280 --> 00:00:51.000
Legacy behavior dies hard. Most organizations still run nightly ETL
17
00:00:51.079 --> 00:00:55.119
jobs that sweep operational systems, flattened tables into commer separated relics,
18
00:00:55.119 --> 00:00:59.039
and upload the corpses into one lake It's comforting, predictable, measurable,
19
00:00:59.079 --> 00:01:01.200
seductively simple. But what you end up with is a
20
00:01:01.200 --> 00:01:04.560
static museum of snapshots. Each file represents how things looked
21
00:01:04.560 --> 00:01:07.239
at one moment and immediately begins to decay. There's no motion,
22
00:01:07.359 --> 00:01:10.439
no relationships, no evolving context, just files, lots of them.
23
00:01:10.560 --> 00:01:13.159
The truth that approach made sense when data lived on
24
00:01:13.239 --> 00:01:16.040
prem in constrained systems. Fabric was designed for something else,
25
00:01:16.159 --> 00:01:19.400
entirely living data, streaming data, context, away intelligence. One lake
26
00:01:19.439 --> 00:01:21.920
isn't a filing cabinet. It's supposed to be the circulatory
27
00:01:21.959 --> 00:01:25.319
system of your organization's information flow. Treating it like cold
28
00:01:25.319 --> 00:01:28.480
storage is the digital equivalent of embalming your business metrics.
29
00:01:28.799 --> 00:01:32.159
Without semantic models, your data has no language. Without relationships,
30
00:01:32.200 --> 00:01:34.959
it has no memory. A CSV from sales, a CSV
31
00:01:35.040 --> 00:01:38.280
from marketing, a CSV from finance. They can coexist peacefully
32
00:01:38.280 --> 00:01:40.200
in the same lake and still never talk to each other.
33
00:01:40.319 --> 00:01:45.200
Governance structures, missing metadata optional. Apparently, the result is isolation
34
00:01:45.319 --> 00:01:49.239
so pure that even Copilot, Microsoft's conversational AI can't interpret it.
35
00:01:49.519 --> 00:01:52.920
If you ask Copilot what were last quarters revenue drivers,
36
00:01:53.200 --> 00:01:55.079
it doesn't know where to look because you never told
37
00:01:55.120 --> 00:01:57.439
it what revenue means in your schema. Let's take a
38
00:01:57.480 --> 00:02:02.439
micro example. Suppose your sales data set contains transaction records, dates, amounts,
39
00:02:02.480 --> 00:02:04.879
product skews, and regent codes. You happily dump it into
40
00:02:04.879 --> 00:02:08.199
one lake, no semantic model, no named relationships, just raw
41
00:02:08.240 --> 00:02:12.159
table columns. Now ask Fabric's AI to identify top performing regions.
42
00:02:12.280 --> 00:02:16.000
It shrugs. It cannot contextualize region code without metadata linking
43
00:02:16.000 --> 00:02:19.560
it to geography or organizational units. To the machine, USN
44
00:02:19.599 --> 00:02:23.080
could mean North America or user segment North. Humans rely
45
00:02:23.199 --> 00:02:26.639
on inference. AI requires explicit structure. That's the gap, turning
46
00:02:26.639 --> 00:02:29.360
your warehouse into a morgue. Here's what most people miss.
47
00:02:29.479 --> 00:02:32.159
Fabric doesn't treat data at rest and data in motion
48
00:02:32.240 --> 00:02:34.960
as separate species. It assumes every data set could one
49
00:02:35.000 --> 00:02:38.479
day become an intelligent participant, queried in real time, enriched
50
00:02:38.479 --> 00:02:41.280
by context, reshaped by governance rules, and even reasoned over
51
00:02:41.319 --> 00:02:45.280
by agents. When you persist CSVS without activating those connections,
52
00:02:45.400 --> 00:02:48.800
you're ignoring Fabric's metabolic design. You chop off its nervous system.
53
00:02:49.080 --> 00:02:51.879
Compare that to data in motion in Fabric real time
54
00:02:51.919 --> 00:02:56.599
intelligence modules ingest streaming signals, IoT events, transaction logs, sensor
55
00:02:56.639 --> 00:02:58.879
pinks and feed them into live data sets that can
56
00:02:58.919 --> 00:03:02.680
trigger responses in instantly. Anomaly detection isn't run weekly, it
57
00:03:02.680 --> 00:03:06.039
happens continuously. Trend analysis doesn't wait for the quarter's end,
58
00:03:06.159 --> 00:03:08.800
It updates on every new record. This is what a
59
00:03:08.840 --> 00:03:13.240
live data looks like, constantly evaluated, contextualized by AI agents,
60
00:03:13.560 --> 00:03:16.960
and subject to governance rules in milliseconds. The difference between
61
00:03:17.000 --> 00:03:20.120
data address and data in motion is fundamental. Resting data
62
00:03:20.159 --> 00:03:23.759
answers what happened with moving data answers what's happening and
63
00:03:23.840 --> 00:03:26.639
what should we do next. If your warehouse only does
64
00:03:26.680 --> 00:03:28.960
the former, you are running a historical archive, not a
65
00:03:29.000 --> 00:03:32.360
decision engine. Fabric's purpose is to compress that timeline until
66
00:03:32.400 --> 00:03:37.000
observation and action are indistinguishable. Without AI activation, your storing fossils.
67
00:03:37.400 --> 00:03:40.400
With it, you're managing living organisms that adapt to context.
68
00:03:40.919 --> 00:03:43.400
Think of your warehouse like a body. One lake is
69
00:03:43.400 --> 00:03:46.639
the blood stream, Semantic models are the DNA, and data
70
00:03:46.639 --> 00:03:50.039
agents are the brain cells firing signals across systems. Right now,
71
00:03:50.080 --> 00:03:52.120
most of you have the bloodstream but no brain function.
72
00:03:52.319 --> 00:03:55.599
The organs exist, but nothing coordinates. And yes, it's comfortable
73
00:03:55.639 --> 00:03:59.400
that way. No surprises, no sudden automation, no rogue recommendations.
74
00:03:59.680 --> 00:04:02.879
STAT systems don't disobey, but they also don't compete in
75
00:04:02.919 --> 00:04:06.319
an environment where ninety percent of large enterprises are feeding
76
00:04:06.360 --> 00:04:09.280
their warehouses to AI agents. Leaving your data inert is
77
00:04:09.319 --> 00:04:11.879
like stocking a luxury aquarium with plastic fish because you
78
00:04:11.879 --> 00:04:15.120
prefer predictability over life. So what should be alive in
79
00:04:15.159 --> 00:04:17.920
your one leg The relationships, the context, and the intelligence
80
00:04:17.920 --> 00:04:20.920
that link your data sets into a cohesive worldview. Once
81
00:04:20.959 --> 00:04:24.040
you stop dumping raw csvs and start modeling information for
82
00:04:24.120 --> 00:04:28.600
AI consumption, Fabric starts behaving as intended, an ecosystem of living,
83
00:04:28.720 --> 00:04:31.560
thinking data instead of an ice box of obsolete numbers.
84
00:04:32.079 --> 00:04:35.680
If your ETL pipeline still ends with store CSV, congratulations,
85
00:04:35.759 --> 00:04:38.800
you've automated the world's most expensive burial process. In the
86
00:04:38.839 --> 00:04:41.399
next section will exhume those files, give them a brain,
87
00:04:41.839 --> 00:04:45.120
and show you what actually makes Fabric intelligent. The data agents,
88
00:04:45.600 --> 00:04:48.920
the missing intelligence layer, enter the part everyone skips, the
89
00:04:48.959 --> 00:04:51.920
actual intelligence layer, the thing that separates a warehouse from
90
00:04:51.920 --> 00:04:54.920
a brain. Microsoft calls them data agents, but think of
91
00:04:54.920 --> 00:04:57.240
them as neurons that finally start firing once you stop
92
00:04:57.240 --> 00:04:59.839
treating one lake like a storage locker. These agents are
93
00:04:59.839 --> 00:05:03.120
not decorative features. They are the operational cortex that Fabric
94
00:05:03.199 --> 00:05:07.079
quietly installs for you, and that most of you heroically ignore.
95
00:05:07.240 --> 00:05:10.319
Let's begin with the mistake. People obsess over dashboards. They
96
00:05:10.319 --> 00:05:13.319
think if powerbi shows a colorful line trending upward, they've
97
00:05:13.319 --> 00:05:16.759
achieved enlightenment. Meanwhile, they've left the reasoning layer, the dynamic
98
00:05:16.759 --> 00:05:20.000
element that interprets patterns and acts on them, unplugged. That's
99
00:05:20.079 --> 00:05:23.600
like buying a Tesla admiring the screen graphics and never
100
00:05:23.680 --> 00:05:27.360
pressing the accelerator. The average user believes Fabric's beauty lies
101
00:05:27.360 --> 00:05:30.920
in uniform metrics. In reality, it lies in synaptic activity
102
00:05:31.040 --> 00:05:33.879
agents that think. So what exactly are these data agents?
103
00:05:33.959 --> 00:05:37.120
They are AI powered interfaces between your warehouse and Azure's
104
00:05:37.160 --> 00:05:40.720
cognitive services, build to reason across data, not just query it.
105
00:05:40.839 --> 00:05:42.959
They live over one lake, but integrate through as your
106
00:05:43.000 --> 00:05:46.160
AI foundry, where they inherit the ability to retrieve, infer,
107
00:05:46.240 --> 00:05:50.000
and apply logic based on your organization's context. And here's
108
00:05:50.000 --> 00:05:53.199
the crucial twist. They participate in a framework called Model
109
00:05:53.240 --> 00:05:56.920
Context Protocols that allows multiple agents to share memory and
110
00:05:57.000 --> 00:06:01.079
goals so they can collaborate handoff tasks and negotiate outcomes
111
00:06:01.120 --> 00:06:04.279
like colleagues who actually read the company manual. Each agent
112
00:06:04.319 --> 00:06:07.279
can be configured to respect governance and security boundaries. They
113
00:06:07.279 --> 00:06:11.199
don't wander blindly into sensitive data because Fabric enforces policies
114
00:06:11.199 --> 00:06:14.319
through purview and role based access. This governance link gives
115
00:06:14.319 --> 00:06:18.560
them something legacy analytics never had moral restraint. Your cfo's
116
00:06:18.600 --> 00:06:23.519
financial agent cannot accidentally read HR's salary data unless expressly allowed.
117
00:06:23.560 --> 00:06:26.879
It's the difference between reasoning and rummaging. Now contrast these
118
00:06:26.920 --> 00:06:30.120
data agents with Copilot, the celebrity assistant. Everyone loves to
119
00:06:30.160 --> 00:06:34.399
talk to. Copilot sits inside teams or powerbi. It's charming, reactive,
120
00:06:34.480 --> 00:06:37.680
and somewhat shallow. It answers what you ask. Data agents,
121
00:06:37.720 --> 00:06:40.680
by comparison, are the ones who already read the quarterly forecast,
122
00:06:40.800 --> 00:06:45.480
spotted inconsistencies, and drafted recommendations before you even open the dashboard.
123
00:06:45.720 --> 00:06:49.519
Copilot is a student. Agents are auditors. One obeys, the
124
00:06:49.600 --> 00:06:52.879
other anticipates. Let's ground this in an example. Your retail
125
00:06:52.920 --> 00:06:56.600
business process is daily transactions through Fabric. Without agents, you'd
126
00:06:56.639 --> 00:07:01.879
spend fridays exporting summaries, top selling products, regents trending up anomalies.
127
00:07:01.879 --> 00:07:05.439
Over threshold with agents, the warehouse becomes sentient enough to
128
00:07:05.480 --> 00:07:08.040
notice that sales in Regent East are spiking twenty percent
129
00:07:08.079 --> 00:07:11.959
above forecast, while supply chain logs show delayed deliveries. An
130
00:07:12.000 --> 00:07:14.759
agent detects the mismatch, tags it as a fulfillment risk,
131
00:07:14.839 --> 00:07:19.959
alerts operations, and proposes redistributing inventory preemptively. Nobody asked it inferred.
132
00:07:20.000 --> 00:07:22.319
This is in science fiction. It's Fabric's real time intelligence
133
00:07:22.319 --> 00:07:25.000
merged with agentic reasoning. Pause on what that means. Your
134
00:07:25.000 --> 00:07:27.680
warehouse just performed judgment, not a query, not an alert,
135
00:07:27.680 --> 00:07:31.959
but analysis that required understanding business intent. It identified an anomaly,
136
00:07:32.199 --> 00:07:36.319
cross referenced context, and acted responsibly. That's the threshold where
137
00:07:36.439 --> 00:07:39.720
data warehouse becomes decision system. Without agents, you'd still be
138
00:07:39.720 --> 00:07:42.959
exporting powerbi visuals into slide decks, pretending you discovered the
139
00:07:43.000 --> 00:07:46.199
issue manually. Here's the weird part. Most companies have this
140
00:07:46.279 --> 00:07:49.480
capability already activated within their Fabric capacities. They just haven't
141
00:07:49.480 --> 00:07:51.720
configured it. They spent the money, got the software, and
142
00:07:51.720 --> 00:07:56.240
forgot to initialize cognition because that requires thinking architecturally, defining
143
00:07:56.279 --> 00:08:00.839
semantic relationships, establishing AI instructions, and connecting one leg endpoints
144
00:08:00.879 --> 00:08:03.759
to the reasoning infrastructure. But once you do, everything changes.
145
00:08:03.879 --> 00:08:08.199
Dashboards become side effects of intelligence rather than destinations for analysis.
146
00:08:08.560 --> 00:08:12.399
Think back to the CSV graveyard metaphor those csvs were tombstones,
147
00:08:12.439 --> 00:08:14.800
marking where all data sets went to die. Turn on
148
00:08:14.879 --> 00:08:17.600
agents and its resurrection day. The warehouse begins to breathe
149
00:08:17.800 --> 00:08:21.439
tables align themselves, attributes acquire meaning, and metrics synchronize autonomously.
150
00:08:21.519 --> 00:08:24.959
The system doesn't merely report reality, it interprets it while
151
00:08:25.000 --> 00:08:28.319
you're still drafting an email about last quarter's KPIs. Of course,
152
00:08:28.360 --> 00:08:31.360
this shift requires a mental upgrade from storage management to
153
00:08:31.439 --> 00:08:35.879
cognitive orchestration. Data agents don't wait for instructions. They follow goals.
154
00:08:36.279 --> 00:08:39.480
They use model context protocols to communicate with other Microsoft agents,
155
00:08:39.639 --> 00:08:42.320
the ones in power, Automate three sixty five, and Azure
156
00:08:42.360 --> 00:08:46.759
AI services sharing reasoning context across platforms. That's how data
157
00:08:46.759 --> 00:08:50.000
fluctuation can trigger an adaptive workflow or generate new insights
158
00:08:50.039 --> 00:08:53.480
inside Excel without human mediation, and yes, when configured poorly,
159
00:08:53.799 --> 00:08:56.759
this autonomy can look unnerving, like having interns who act
160
00:08:56.799 --> 00:08:59.799
decisively after misreading a spreadsheet. That's why governance which will
161
00:08:59.799 --> 00:09:03.799
reach soon exists, but first accept this truth. Intelligence delayed
162
00:09:03.840 --> 00:09:06.919
is advantage lost. The longer you treat fabric as cold storage,
163
00:09:06.960 --> 00:09:09.600
the more you pay for an AI platform functioning as
164
00:09:09.639 --> 00:09:12.919
a glorified backup. So stop mourning your data's potential. Wake
165
00:09:12.960 --> 00:09:15.919
the agents. Let your warehouse graduate from archive to organism.
166
00:09:16.200 --> 00:09:19.600
Because the next era of analytics isn't about asking better questions,
167
00:09:19.919 --> 00:09:23.039
It's about owning systems that answer before you can type them.
168
00:09:23.960 --> 00:09:26.399
How to resurrect your warehouse with AI. Time to bring
169
00:09:26.399 --> 00:09:29.279
the corps back to life. Resurrection starts not with code,
170
00:09:29.279 --> 00:09:32.960
but with context, because context is oxygen for data. Step
171
00:09:33.000 --> 00:09:36.120
one is infusing your warehouse with meaning. That means creating
172
00:09:36.120 --> 00:09:39.559
semantic models. These models define how your data thinks about itself.
173
00:09:39.960 --> 00:09:42.519
Sales are tied to customers, customers to regions, regions to
174
00:09:42.559 --> 00:09:45.799
revenue structures. Without them, even the most powerful AI agent
175
00:09:45.840 --> 00:09:48.840
is like a linguist handed a dictionary without syntax. In Fabric,
176
00:09:48.919 --> 00:09:52.320
you use the data modeling layer to declare these relationships explicitly,
177
00:09:52.639 --> 00:09:54.840
so your agents can reason instead of guests. Now for
178
00:09:54.919 --> 00:09:58.279
step two, actually deploying a Fabric data agent. This is
179
00:09:58.320 --> 00:10:00.240
where you give your warehouse not just a brain, but
180
00:10:00.279 --> 00:10:03.919
a personality, an operational mind that knows what to look for,
181
00:10:04.200 --> 00:10:06.840
when to alert you, and how to connect dots across
182
00:10:06.879 --> 00:10:10.840
one lake. In practice, you open Azure AI foundry, define
183
00:10:10.840 --> 00:10:14.600
a data agent and pointed at your fabric data sets. Instantly,
184
00:10:14.639 --> 00:10:17.480
it inherits access to the entire semantic layer. It's not
185
00:10:17.519 --> 00:10:20.519
a chatboard. It's a sentient indexer trained on your actual
186
00:10:20.519 --> 00:10:23.480
business structure. From now on, every table has a guardian
187
00:10:23.519 --> 00:10:27.360
angel capable of pattern recognition and inference. Step three is instruction,
188
00:10:27.600 --> 00:10:29.960
and agent without parameters is a toddler with access to
189
00:10:30.000 --> 00:10:34.120
the corporate VPN. You must provide organizations specific directives. What risk,
190
00:10:34.519 --> 00:10:38.840
revenue or priority mean, Which data sources are authoritative, Which
191
00:10:38.879 --> 00:10:42.720
systems must not be touched without human approval. Governance policies
192
00:10:42.720 --> 00:10:45.360
from purview sink here automatically, but you must define the
193
00:10:45.399 --> 00:10:48.519
logical intent, tell your agent how to behave. The clearer
194
00:10:48.559 --> 00:10:51.679
your definitions, the more coherent its reasoning. Think of it
195
00:10:51.720 --> 00:10:54.879
as drafting the company handbook for an employee who never sleeps.
196
00:10:55.440 --> 00:10:58.240
The fourth step is integration, the part that transforms clever
197
00:10:58.320 --> 00:11:02.360
prototypes into daily companions. Connect your data agent to copilot studio.
198
00:11:02.639 --> 00:11:06.120
Why because Copilot provides the natural language interface your employees
199
00:11:06.120 --> 00:11:09.360
already understand. When someone in sales types show me emerging
200
00:11:09.440 --> 00:11:12.480
churn patterns. Copilot politely forwards the request to your agent,
201
00:11:12.639 --> 00:11:16.000
which performs genuine reasoning across data sets and sends a
202
00:11:16.080 --> 00:11:20.080
human readable summary back, complete with citations and traceable lineage.
203
00:11:20.159 --> 00:11:23.799
This is intelligence served conversationally. Once this foundation is active,
204
00:11:24.120 --> 00:11:28.360
the system begins performing quiet miracles. Consider trend detection. Your
205
00:11:28.399 --> 00:11:32.720
agent continually examines transactional data, inventory levels, and forecast metrics.
206
00:11:32.799 --> 00:11:37.519
When behavior deviates from expectation, say a holiday surge developing
207
00:11:37.600 --> 00:11:40.840
earlier than predicted, it notifies marketing two weeks before the
208
00:11:40.879 --> 00:11:44.480
anomaly would have appeared in a dashboard or picture. KPI
209
00:11:44.519 --> 00:11:49.320
alerts instead of manual threshold rules. The agent recognizes trajectories
210
00:11:49.320 --> 00:11:53.559
that historically precede misses and flags them preemptively. Churn prediction,
211
00:11:53.879 --> 00:11:57.200
supply chain optimization, compliance verification. Every one of these becomes
212
00:11:57.200 --> 00:11:59.519
a living process, not a quarterly report. And here's where
213
00:11:59.519 --> 00:12:02.200
Fabrics is Dezi shines. These agents don't live in isolation.
214
00:12:02.360 --> 00:12:05.960
They communicate through model context protocols with other Microsoft services,
215
00:12:06.240 --> 00:12:09.600
creating multi agent orchestration. A Fabric data agent can identify
216
00:12:09.600 --> 00:12:12.279
a slow moving squ notify a power automate agent to
217
00:12:12.320 --> 00:12:15.519
trigger a discount workflow, sync results into dynamics through another
218
00:12:15.559 --> 00:12:18.679
as your AI agent, and finally present the outcome insight
219
00:12:18.759 --> 00:12:22.440
teams as a business alert. That sequence requires no custom scripts,
220
00:12:22.440 --> 00:12:26.399
only properly defined intentions and connections. You've just witnessed distributed
221
00:12:26.440 --> 00:12:29.840
intelligence performing genuine work. This is the real point so
222
00:12:29.879 --> 00:12:32.679
many miss. Fabric isn't a place for storing results. It's
223
00:12:32.679 --> 00:12:35.919
an operating environment for continuous reasoning. Treating it like a
224
00:12:35.919 --> 00:12:39.159
static data vault wastes the one architectural innovation that sets
225
00:12:39.159 --> 00:12:42.000
it apart. You are supposed to think in agents. Every
226
00:12:42.080 --> 00:12:45.159
data set becomes an actor, Every insight becomes an event,
227
00:12:45.480 --> 00:12:49.559
Every business process becomes an orchestrated adaptive conversation between them.
228
00:12:49.919 --> 00:12:54.000
Your job shifts from building pipelines to defining intentions. Some
229
00:12:54.080 --> 00:12:57.320
recoil at that they want comforting determinism, the assurance that
230
00:12:57.360 --> 00:12:59.919
nothing changes unless a human press is run. But inteen
231
00:13:00.000 --> 00:13:03.080
diligent systems thrive on feedback loops. When an agent refines
232
00:13:03.120 --> 00:13:06.039
a metric or automates an alert, it's not taking control,
233
00:13:06.159 --> 00:13:09.159
it's taking responsibility. This is how data finally earns its
234
00:13:09.200 --> 00:13:13.240
keep by detecting issues, making recommendations, and learning from corrections.
235
00:13:13.519 --> 00:13:16.080
If you've ever wondered why competitors move faster with the
236
00:13:16.080 --> 00:13:19.559
same data sets, it's because their warehouses aren't waiting for instructions.
237
00:13:19.679 --> 00:13:23.639
They're conversing internally, resolving micro problems before executives even hear
