1
00:00:00,120 --> 00:00:01,360
The AI everyone keeps telling
2
00:00:01,360 --> 00:00:02,640
you to learn is already hiding
3
00:00:02,640 --> 00:00:03,960
inside the tools you open every
4
00:00:03,960 --> 00:00:04,320
morning.
5
00:00:04,320 --> 00:00:06,759
That's AI builder and it lives
inside power Apps and Power
6
00:00:06,759 --> 00:00:09,359
Automate, the low code tools
you're probably already using.
7
00:00:09,359 --> 00:00:11,320
You don't need to write code,
and you don't need a data
8
00:00:11,320 --> 00:00:12,800
science team to put it to work.
9
00:00:12,800 --> 00:00:13,839
By the end of this, you'll know
10
00:00:13,839 --> 00:00:15,320
what AI builder actually is,
11
00:00:15,320 --> 00:00:16,120
what it can do with your
12
00:00:16,120 --> 00:00:17,440
documents, and how you pay for
13
00:00:17,440 --> 00:00:17,879
it.
14
00:00:17,879 --> 00:00:21,800
Welcome to the M365 FM podcast,
where we explore the people,
15
00:00:21,800 --> 00:00:24,760
ideas, and technologies shaping
the future of work.
16
00:00:24,760 --> 00:00:27,399
And by the time we're done, the
jargon will be gone and the
17
00:00:27,399 --> 00:00:30,320
whole thing will feel a lot
smaller than it sounds.
18
00:00:30,320 --> 00:00:33,359
What AI builder actually is AI
builder is a feature inside
19
00:00:33,359 --> 00:00:34,840
Microsoft Power Platform.
20
00:00:34,840 --> 00:00:37,880
It lets you create and use AI
models to take over business
21
00:00:37,880 --> 00:00:39,640
tasks you'd normally do by hand.
22
00:00:39,640 --> 00:00:42,079
The word model is the one that
scares people, so it helps to
23
00:00:42,079 --> 00:00:43,280
shrink it down to size.
24
00:00:43,280 --> 00:00:44,799
A model is a pattern finder.
25
00:00:44,799 --> 00:00:46,119
You show it a pile of examples
26
00:00:46,119 --> 00:00:47,259
and it learns what tends to go
27
00:00:47,259 --> 00:00:47,920
with what.
28
00:00:47,920 --> 00:00:49,140
Then it applies that lesson to
29
00:00:49,140 --> 00:00:50,240
something new and that is the
30
00:00:50,240 --> 00:00:50,960
whole trick.
31
00:00:50,960 --> 00:00:54,159
The reason beginners freeze up
is rarely the model itself.
32
00:00:54,159 --> 00:00:55,039
It's the story they tell
33
00:00:55,039 --> 00:00:56,420
themselves about what using AI
34
00:00:56,420 --> 00:00:57,000
requires.
35
00:00:57,000 --> 00:00:57,840
They picture a team of
36
00:00:57,840 --> 00:00:59,560
engineers, a data warehouse, and
37
00:00:59,560 --> 00:01:00,759
six months of setup before
38
00:01:00,759 --> 00:01:01,789
anything works.
39
00:01:01,789 --> 00:01:04,750
That picture is what makes AI
feel out of reach.
40
00:01:04,750 --> 00:01:08,030
AI builder is designed to skip
every one of those steps.
41
00:01:08,030 --> 00:01:10,430
Microsoft's own recipe is for
moves.
42
00:01:10,430 --> 00:01:13,069
You pick a model type, you
connect your data, the training
43
00:01:13,069 --> 00:01:15,109
happens on its own, and then you
use the result.
44
00:01:15,109 --> 00:01:16,109
There's no coding, no data
45
00:01:16,109 --> 00:01:17,609
science degree, and no secret
46
00:01:17,609 --> 00:01:18,430
handshake.
47
00:01:18,430 --> 00:01:21,310
You also don't have to go find a
separate tool to use it.
48
00:01:21,310 --> 00:01:22,430
AI builder is built straight
49
00:01:22,430 --> 00:01:23,709
into Power Apps and Power
50
00:01:23,709 --> 00:01:25,349
Automate, which are the same low
51
00:01:25,349 --> 00:01:25,989
code platforms.
52
00:01:25,989 --> 00:01:28,670
You're likely already learning
that placement matters more than
53
00:01:28,670 --> 00:01:31,489
it sounds, because it means the
AI shows up exactly where the
54
00:01:31,489 --> 00:01:32,510
work already happens.
55
00:01:32,510 --> 00:01:35,349
It sits inside the app, your
team fills out or inside the
56
00:01:35,349 --> 00:01:37,230
flow that moves your files
around at night.
57
00:01:37,230 --> 00:01:38,670
You're not learning a second
career.
58
00:01:38,670 --> 00:01:40,230
You're adding one new capability
59
00:01:40,230 --> 00:01:41,230
to something you already
60
00:01:41,230 --> 00:01:41,790
touched.
61
00:01:41,790 --> 00:01:44,069
There's a reason Microsoft built
it this way.
62
00:01:44,069 --> 00:01:46,629
Most AI headlines are about
giant models that answer
63
00:01:46,629 --> 00:01:48,069
questions in a chat box.
64
00:01:48,069 --> 00:01:50,109
Those are impressive, but
they're hard to point at one
65
00:01:50,109 --> 00:01:51,510
specific business chore.
66
00:01:51,510 --> 00:01:52,989
AI builder flips that around.
67
00:01:52,989 --> 00:01:54,909
It starts from the task instead
of the chat.
68
00:01:54,909 --> 00:01:57,269
So you begin with a job that
needs doing and then pick the
69
00:01:57,269 --> 00:01:57,989
model that does it.
70
00:01:57,989 --> 00:01:59,349
That's why a beginner can use it
71
00:01:59,349 --> 00:02:00,709
without ever training a network
72
00:02:00,709 --> 00:02:01,469
by hand.
73
00:02:01,469 --> 00:02:02,859
So keep this picture in mind.
74
00:02:02,859 --> 00:02:03,819
AI builder is AI.
75
00:02:03,819 --> 00:02:05,219
You can actually use sitting
76
00:02:05,219 --> 00:02:06,659
quietly inside tools you already
77
00:02:06,659 --> 00:02:07,060
know.
78
00:02:07,060 --> 00:02:09,939
And everything else in this
video hangs off that one idea.
79
00:02:09,939 --> 00:02:12,020
Ready made models versus models
you build.
80
00:02:12,020 --> 00:02:14,180
Once you accept that this is
real and reachable, the next
81
00:02:14,180 --> 00:02:16,379
question is which kind of model
you actually reach for.
82
00:02:16,379 --> 00:02:19,000
Microsoft gives you two paths,
and the difference comes down to
83
00:02:19,000 --> 00:02:20,419
whose data does the learning.
84
00:02:20,419 --> 00:02:22,099
The first path is pre-built
models.
85
00:02:22,099 --> 00:02:24,259
These are models Microsoft
already trained for problems
86
00:02:24,259 --> 00:02:26,740
that show up in almost every
business, so you skip straight
87
00:02:26,740 --> 00:02:29,379
to using them point one at an
invoice and it pulls out the
88
00:02:29,379 --> 00:02:30,659
totals and the dates.
89
00:02:30,659 --> 00:02:31,780
Point one at a picture of a
90
00:02:31,780 --> 00:02:32,979
receipt and it reads the line
91
00:02:32,979 --> 00:02:33,379
items.
92
00:02:33,379 --> 00:02:34,659
There are pre-built models for
93
00:02:34,659 --> 00:02:36,500
reading text inside an image for
94
00:02:36,500 --> 00:02:37,740
judging whether feedback sounds
95
00:02:37,740 --> 00:02:39,099
positive or negative, and for
96
00:02:39,099 --> 00:02:40,300
pulling names and numbers out of
97
00:02:40,300 --> 00:02:41,419
messy text.
98
00:02:41,419 --> 00:02:42,819
Others read ID documents.
99
00:02:42,819 --> 00:02:45,819
Grab the key phrases from a
paragraph, scan a business card,
100
00:02:45,819 --> 00:02:47,939
or detect what language
something is written in.
101
00:02:47,939 --> 00:02:50,259
You didn't train any of them and
you didn't have to.
102
00:02:50,259 --> 00:02:52,819
The second path is custom
models, and this is where your
103
00:02:52,819 --> 00:02:54,300
own data does the work.
104
00:02:54,300 --> 00:02:57,060
These models are trained on your
documents, your text, your
105
00:02:57,060 --> 00:03:00,219
images, and your history, which
makes them fit the way your
106
00:03:00,219 --> 00:03:02,099
business actually looks.
107
00:03:02,099 --> 00:03:04,780
That list includes document
processing, which reads your
108
00:03:04,780 --> 00:03:07,979
specific forms and category
classification, which sorts text
109
00:03:07,979 --> 00:03:09,500
into categories you define.
110
00:03:09,500 --> 00:03:11,780
It also includes entity
extraction, which finds the
111
00:03:11,780 --> 00:03:14,939
exact details you care about,
prediction, which guesses future
112
00:03:14,939 --> 00:03:17,979
outcomes from your past data,
and object detection, which
113
00:03:17,979 --> 00:03:19,740
spots your products in a photo.
114
00:03:19,740 --> 00:03:22,099
There's even a way to bring your
own model over from Azure if you
115
00:03:22,099 --> 00:03:23,020
already built one.
116
00:03:23,020 --> 00:03:25,500
So the rule of thumb is simple
if the problem is common and
117
00:03:25,500 --> 00:03:27,699
generic, like reading any
receipt that lands in your
118
00:03:27,699 --> 00:03:30,939
inbox, a pre-built model has you
covered on day one when the
119
00:03:30,939 --> 00:03:32,379
problem is uniquely yours.
120
00:03:32,379 --> 00:03:35,020
Like reading invoices from one
specific supplier.
121
00:03:35,020 --> 00:03:36,460
You train a custom model on your
122
00:03:36,460 --> 00:03:37,860
own examples, and it learns your
123
00:03:37,860 --> 00:03:38,419
quirks.
124
00:03:38,419 --> 00:03:39,860
The difference is easier to deal
with.
125
00:03:39,860 --> 00:03:40,659
An example.
126
00:03:40,659 --> 00:03:43,340
A pre-built model can read the
total of a standard invoice from
127
00:03:43,340 --> 00:03:45,479
almost any vendor because
Microsoft trained it on
128
00:03:45,479 --> 00:03:46,379
thousands of them.
129
00:03:46,379 --> 00:03:49,620
But if your invoices come in a
layout nobody else uses, with
130
00:03:49,620 --> 00:03:52,860
the account number buried inside
a table, that pre-built model
131
00:03:52,860 --> 00:03:53,500
will miss it.
132
00:03:53,500 --> 00:03:56,060
That is exactly the job you hand
to a custom model.
133
00:03:56,060 --> 00:03:58,939
Most beginners assume custom
always means better, and that
134
00:03:58,939 --> 00:03:59,939
assumption is wrong.
135
00:03:59,939 --> 00:04:01,740
Custom means more specific, and
136
00:04:01,740 --> 00:04:03,020
more specific is only better
137
00:04:03,020 --> 00:04:04,370
when your problem is specific
138
00:04:04,370 --> 00:04:04,610
to.
139
00:04:04,610 --> 00:04:06,050
One more thing before you pick.
140
00:04:06,050 --> 00:04:08,449
A few of these models are still
in preview, which means they're
141
00:04:08,449 --> 00:04:10,889
not meant for production work
yet, and some aren't available
142
00:04:10,889 --> 00:04:11,770
in every region.
143
00:04:11,770 --> 00:04:12,930
So if you go looking for a model
144
00:04:12,930 --> 00:04:13,849
and can't find it, that's
145
00:04:13,849 --> 00:04:14,849
usually why.
146
00:04:14,849 --> 00:04:16,810
Why this is not just another
chatbot.
147
00:04:16,810 --> 00:04:18,050
Before we go further, there's a
148
00:04:18,050 --> 00:04:19,649
confusion worth clearing up
149
00:04:19,649 --> 00:04:20,930
because it trips up almost every
150
00:04:20,930 --> 00:04:21,490
beginner.
151
00:04:21,490 --> 00:04:24,129
When people hear AI, they now
think of a chat window that
152
00:04:24,129 --> 00:04:26,889
answers questions and they
assume AI builder is the same
153
00:04:26,889 --> 00:04:28,410
thing wearing a Microsoft badge.
154
00:04:28,410 --> 00:04:30,610
That assumption is wrong, and
the difference is the whole
155
00:04:30,610 --> 00:04:31,449
reason it's useful.
156
00:04:31,449 --> 00:04:32,649
A chatbot is general.
157
00:04:32,649 --> 00:04:33,649
It will talk about anything,
158
00:04:33,649 --> 00:04:34,810
which is impressive, but it's
159
00:04:34,810 --> 00:04:36,449
also vague because it has no
160
00:04:36,449 --> 00:04:38,129
idea what your specific job
161
00:04:38,129 --> 00:04:39,089
looks like.
162
00:04:39,089 --> 00:04:40,730
AI builder takes the opposite
approach.
163
00:04:40,730 --> 00:04:43,209
It starts with a task like
reading an invoice or scoring a
164
00:04:43,209 --> 00:04:46,370
record, and it brings a model
that's focused on exactly that.
165
00:04:46,370 --> 00:04:48,449
You're not asking it to be
clever about everything.
166
00:04:48,449 --> 00:04:51,529
You're asking it to be reliable
about one thing over and over
167
00:04:51,529 --> 00:04:52,689
without getting bored.
168
00:04:52,689 --> 00:04:55,290
That focus is why the results
are easier to trust.
169
00:04:55,290 --> 00:04:58,050
A focused model does the same
job the same way every time,
170
00:04:58,050 --> 00:05:00,810
which is what you want when the
output feeds a business process.
171
00:05:00,810 --> 00:05:02,529
It's also why AI builder fits
172
00:05:02,529 --> 00:05:03,920
inside power apps and Power
173
00:05:03,920 --> 00:05:04,279
Automate.
174
00:05:04,279 --> 00:05:05,959
So naturally, those tools are
175
00:05:05,959 --> 00:05:07,220
already about doing a defined
176
00:05:07,220 --> 00:05:08,120
job well.
177
00:05:08,120 --> 00:05:10,399
So a model that does one thing
well slides right in.
178
00:05:10,399 --> 00:05:11,720
Now AI builder does include
179
00:05:11,720 --> 00:05:12,600
prompts, which are the more
180
00:05:12,600 --> 00:05:14,079
conversational generative side
181
00:05:14,079 --> 00:05:14,439
of things.
182
00:05:14,439 --> 00:05:15,959
So the line isn't perfectly
clean.
183
00:05:15,959 --> 00:05:18,180
When people talk about AI
builder for beginners though,
184
00:05:18,180 --> 00:05:20,920
they usually mean the models,
the focused pattern finders.
185
00:05:20,920 --> 00:05:23,879
And that's where the value shows
up fastest.
186
00:05:23,879 --> 00:05:25,160
What document processing really
187
00:05:25,160 --> 00:05:26,680
does document processing is the
188
00:05:26,680 --> 00:05:27,839
one people ask about most,
189
00:05:27,839 --> 00:05:29,040
because it attacks a job that
190
00:05:29,040 --> 00:05:30,439
quietly eats hours every single
191
00:05:30,439 --> 00:05:31,199
week.
192
00:05:31,199 --> 00:05:32,600
In plain terms, it reads and
193
00:05:32,600 --> 00:05:33,720
saves the information out of
194
00:05:33,720 --> 00:05:35,639
standard documents and invoices
195
00:05:35,639 --> 00:05:37,040
and tax forms are the classic
196
00:05:37,040 --> 00:05:37,600
example.
197
00:05:37,600 --> 00:05:40,920
Instead of a person opening a
PDF, squinting at a total, and
198
00:05:40,920 --> 00:05:43,240
retyping it into a system, the
model does that.
199
00:05:43,240 --> 00:05:43,920
Reading for you.
200
00:05:43,920 --> 00:05:45,120
The way you get there is three
201
00:05:45,120 --> 00:05:46,279
steps, and none of them are
202
00:05:46,279 --> 00:05:46,920
scary.
203
00:05:46,920 --> 00:05:48,079
First, you train the model,
204
00:05:48,079 --> 00:05:49,120
which just means you feed it
205
00:05:49,120 --> 00:05:50,199
examples and show it which
206
00:05:50,199 --> 00:05:51,439
pieces of information you want
207
00:05:51,439 --> 00:05:51,720
back.
208
00:05:51,720 --> 00:05:52,839
Then you publish it, which makes
209
00:05:52,839 --> 00:05:53,959
it available to your apps and
210
00:05:53,959 --> 00:05:54,839
your flows.
211
00:05:54,839 --> 00:05:55,600
Then you use it.
212
00:05:55,600 --> 00:05:58,399
And from that point on, the
reading takes care of itself.
213
00:05:58,399 --> 00:06:00,360
Every time a new document
arrives.
214
00:06:00,360 --> 00:06:03,160
Once it is published, the model
keeps doing that same job
215
00:06:03,160 --> 00:06:04,589
quietly in the background.
216
00:06:04,589 --> 00:06:07,310
The part that surprises almost
everyone comes next.
217
00:06:07,310 --> 00:06:09,430
You only need five documents to
get started.
218
00:06:09,430 --> 00:06:12,189
Not five thousand, and not some
labeled data set.
219
00:06:12,189 --> 00:06:14,029
You'd have to hire someone to
build first.
220
00:06:14,029 --> 00:06:17,029
Microsoft designed it so a
business user can point it at a
221
00:06:17,029 --> 00:06:19,829
handful of real documents and
still get accurate results
222
00:06:19,829 --> 00:06:21,430
tailored to their own content.
223
00:06:21,430 --> 00:06:22,910
That single detail is usually
224
00:06:22,910 --> 00:06:24,110
the moment the idea stops
225
00:06:24,110 --> 00:06:25,029
feeling like an enterprise
226
00:06:25,029 --> 00:06:26,269
project and starts feeling like
227
00:06:26,269 --> 00:06:27,389
something you could try this
228
00:06:27,389 --> 00:06:28,110
afternoon.
229
00:06:28,110 --> 00:06:30,790
Notice how this fits the pattern
from the last section.
230
00:06:30,790 --> 00:06:33,589
Document processing is a custom
model because it learns the
231
00:06:33,589 --> 00:06:34,589
shape of your documents.
232
00:06:34,589 --> 00:06:37,629
Specifically, a generic invoice
reader can't know that your
233
00:06:37,629 --> 00:06:40,110
supplier hides the due date down
in the footer.
234
00:06:40,110 --> 00:06:41,589
Yours can because you taught it.
235
00:06:41,589 --> 00:06:42,350
And that's the difference
236
00:06:42,350 --> 00:06:43,550
between a tool that sort of
237
00:06:43,550 --> 00:06:44,550
works and one that actually
238
00:06:44,550 --> 00:06:44,910
fits.
239
00:06:44,910 --> 00:06:46,389
The payoff isn't just speed
either.
240
00:06:46,389 --> 00:06:48,670
When a machine reads the
document, the data lands in your
241
00:06:48,670 --> 00:06:50,910
system clean and consistent in
the right fields.
242
00:06:50,910 --> 00:06:52,189
At the moment the document shows
243
00:06:52,189 --> 00:06:53,689
up, the errors that come from
244
00:06:53,689 --> 00:06:55,029
tired eyes and copy paste
245
00:06:55,029 --> 00:06:55,870
disappear.
246
00:06:55,870 --> 00:06:57,029
And the person who used to do
247
00:06:57,029 --> 00:06:58,149
that reading gets their
248
00:06:58,149 --> 00:06:58,910
afternoon back.
249
00:06:58,910 --> 00:06:59,750
That is the bargain.
250
00:06:59,750 --> 00:07:02,990
Document processing makes how
prediction works.
251
00:07:02,990 --> 00:07:05,750
If document processing is about
reading what already exists.
252
00:07:05,750 --> 00:07:06,990
Prediction is about guessing
253
00:07:06,990 --> 00:07:08,550
what comes next, and it works
254
00:07:08,550 --> 00:07:09,629
from a completely different
255
00:07:09,629 --> 00:07:10,350
angle.
256
00:07:10,350 --> 00:07:12,589
A prediction model studies the
patterns in your past data,
257
00:07:12,589 --> 00:07:15,389
figures out what tends to lead
to what, and then uses that on
258
00:07:15,389 --> 00:07:16,509
new records as they arrive.
259
00:07:16,509 --> 00:07:18,069
You're not feeding it documents
this time.
260
00:07:18,069 --> 00:07:19,550
You're feeding it history.
261
00:07:19,550 --> 00:07:22,029
The answers it gives come in
three shapes, and that changes
262
00:07:22,029 --> 00:07:23,110
what you can ask.
263
00:07:23,110 --> 00:07:24,670
The first is a simple yes or no,
264
00:07:24,670 --> 00:07:25,910
like whether a transaction looks
265
00:07:25,910 --> 00:07:27,350
fraudulent or whether an account
266
00:07:27,350 --> 00:07:28,709
is likely to pay its invoice on
267
00:07:28,709 --> 00:07:29,189
time.
268
00:07:29,189 --> 00:07:31,509
The second is a choice between
several outcomes, like whether a
269
00:07:31,509 --> 00:07:34,069
shipment will arrive early on
time, late or very late.
270
00:07:34,069 --> 00:07:36,470
The third is a number like how
many items you should keep in
271
00:07:36,470 --> 00:07:39,029
inventory, or how many days a
delivery will take notice.
272
00:07:39,029 --> 00:07:40,550
How ordinary those questions
are.
273
00:07:40,550 --> 00:07:42,230
None of them require a research
lab.
274
00:07:42,230 --> 00:07:44,790
They're the exact questions a
manager already asks in a
275
00:07:44,790 --> 00:07:47,230
meeting, and prediction just
gives a data backed answer
276
00:07:47,230 --> 00:07:48,149
instead of a hunch.
277
00:07:48,149 --> 00:07:49,389
It takes a guess that used to
278
00:07:49,389 --> 00:07:50,910
live in someone's head and turns
279
00:07:50,910 --> 00:07:52,329
it into something you can see
280
00:07:52,329 --> 00:07:53,670
and act on, and that is why it
281
00:07:53,670 --> 00:07:54,550
feels useful.
282
00:07:54,550 --> 00:07:57,389
Keep one honest caveat in mind,
though, because it matters.
283
00:07:57,389 --> 00:07:58,509
Prediction is a pattern.
284
00:07:58,509 --> 00:08:00,069
Guess not a crystal ball.
285
00:08:00,069 --> 00:08:01,750
It's only as good as the history
you feed it.
286
00:08:01,750 --> 00:08:03,069
And if your past data is messy
287
00:08:03,069 --> 00:08:04,589
or thin, the predictions will be
288
00:08:04,589 --> 00:08:05,420
messy too.
289
00:08:05,420 --> 00:08:08,100
So before you trust a
prediction, look at the history
290
00:08:08,100 --> 00:08:10,459
behind it, because that history
is doing most of the work.
291
00:08:10,459 --> 00:08:11,579
The model isn't inventing new
292
00:08:11,579 --> 00:08:12,600
wisdom, it's reflecting what
293
00:08:12,600 --> 00:08:13,699
already happened and projecting
294
00:08:13,699 --> 00:08:14,379
it forward.
295
00:08:14,379 --> 00:08:14,939
Used well.
296
00:08:14,939 --> 00:08:15,660
That's powerful.
297
00:08:15,660 --> 00:08:16,740
Used carelessly.
298
00:08:16,740 --> 00:08:19,980
It just automates an old bias
where you'd actually use it.
299
00:08:19,980 --> 00:08:22,620
So far, we've talked about what
AI builder is and what it does.
300
00:08:22,620 --> 00:08:24,899
And now it's worth asking the
practical question, where does
301
00:08:24,899 --> 00:08:25,980
this actually plug in?
302
00:08:25,980 --> 00:08:28,259
The answer is in the two tools
you're already using.
303
00:08:28,259 --> 00:08:29,300
And each one gives you a
304
00:08:29,300 --> 00:08:30,620
slightly different way to reach
305
00:08:30,620 --> 00:08:31,300
the models.
306
00:08:31,300 --> 00:08:34,580
In Power Automate, you add an AI
builder action to a cloud flow.
307
00:08:34,580 --> 00:08:35,539
A flow is just a set of
308
00:08:35,539 --> 00:08:36,740
instructions that runs on its
309
00:08:36,740 --> 00:08:36,940
own.
310
00:08:36,940 --> 00:08:38,820
So a new invoice lands in a
folder.
311
00:08:38,820 --> 00:08:41,419
The flow hands it to the
document processing model, and
312
00:08:41,419 --> 00:08:44,019
the extracted details get filed
into your system without a
313
00:08:44,019 --> 00:08:45,059
person touching it.
314
00:08:45,059 --> 00:08:46,740
You're not building the AI here.
315
00:08:46,740 --> 00:08:48,340
You're wiring an already trained
316
00:08:48,340 --> 00:08:50,039
model into a process you set up
317
00:08:50,039 --> 00:08:50,899
in PowerApps.
318
00:08:50,899 --> 00:08:54,100
You reach the same models
through power effects, which is
319
00:08:54,100 --> 00:08:56,379
the Excel like formula language
behind the app.
320
00:08:56,379 --> 00:08:58,580
That means a form can read a
document while the user is
321
00:08:58,580 --> 00:09:01,179
working or score a record the
moment it's saved.
322
00:09:01,179 --> 00:09:03,059
The AI isn't a separate screen
you switch to.
323
00:09:03,059 --> 00:09:04,179
It's a step inside the app,
324
00:09:04,179 --> 00:09:05,500
which is exactly why beginners
325
00:09:05,500 --> 00:09:06,690
pick it up without learning a
326
00:09:06,690 --> 00:09:07,330
new discipline.
327
00:09:07,330 --> 00:09:08,490
There's one caveat that trips
328
00:09:08,490 --> 00:09:09,809
people up, and it's worth naming
329
00:09:09,809 --> 00:09:10,490
plainly.
330
00:09:10,490 --> 00:09:12,250
Adding an AI builder action
331
00:09:12,250 --> 00:09:13,610
turns a flow into a premium
332
00:09:13,610 --> 00:09:13,889
flow.
333
00:09:13,889 --> 00:09:17,090
And if that flow lives inside an
app, it makes the app a premium.
334
00:09:17,090 --> 00:09:18,769
App premium means a paid
license.
335
00:09:18,769 --> 00:09:21,330
So this isn't a free add on to
the cheapest plan.
336
00:09:21,330 --> 00:09:22,450
It's better to know that before
337
00:09:22,450 --> 00:09:23,490
you build because the cost
338
00:09:23,490 --> 00:09:25,090
conversation starts the moment a
339
00:09:25,090 --> 00:09:26,690
model runs in something real
340
00:09:26,690 --> 00:09:27,529
that leads straight to the
341
00:09:27,529 --> 00:09:28,769
question everyone asks next,
342
00:09:28,769 --> 00:09:29,850
which is what this actually
343
00:09:29,850 --> 00:09:30,370
costs.
344
00:09:30,370 --> 00:09:31,809
That's where credits come in.
345
00:09:31,809 --> 00:09:33,769
What credits are and why they
matter.
346
00:09:33,769 --> 00:09:35,529
The reassuring part comes first
347
00:09:35,529 --> 00:09:36,889
because it's the part beginners
348
00:09:36,889 --> 00:09:37,809
never hear.
349
00:09:37,809 --> 00:09:38,970
Building, training and testing
350
00:09:38,970 --> 00:09:40,250
models in AI builder costs
351
00:09:40,250 --> 00:09:40,730
nothing.
352
00:09:40,730 --> 00:09:43,850
You can design a model, teach
it, run it against samples and
353
00:09:43,850 --> 00:09:47,049
confirm that it works all
without spending a cent only
354
00:09:47,049 --> 00:09:49,929
running a model for real inside
a live app or flow spends
355
00:09:49,929 --> 00:09:50,730
anything at all.
356
00:09:50,730 --> 00:09:51,649
So what is a credit?
357
00:09:51,649 --> 00:09:53,289
It's a prepaid unit of AI usage
358
00:09:53,289 --> 00:09:54,370
and you spend it when a model
359
00:09:54,370 --> 00:09:54,970
does work.
360
00:09:54,970 --> 00:09:56,370
Microsoft has historically sold
361
00:09:56,370 --> 00:09:57,929
an add on of about a million
362
00:09:57,929 --> 00:09:59,210
credits for roughly five hundred
363
00:09:59,210 --> 00:10:00,330
dollars a month, and some
364
00:10:00,330 --> 00:10:02,269
premium licenses also seed a few
365
00:10:02,269 --> 00:10:03,370
thousand credits to get you
366
00:10:03,370 --> 00:10:04,009
started.
367
00:10:04,009 --> 00:10:07,200
Treat those numbers as a rough
shape rather than a fixed price.
368
00:10:07,200 --> 00:10:08,559
Because licensing moves around
369
00:10:08,559 --> 00:10:09,799
and the details change over
370
00:10:09,799 --> 00:10:10,360
time.
371
00:10:10,360 --> 00:10:12,720
What burns through credits is
worth a look, too, because the
372
00:10:12,720 --> 00:10:13,639
spread is wide.
373
00:10:13,639 --> 00:10:16,600
Reading a page with basic text
recognition costs a few credits.
374
00:10:16,600 --> 00:10:18,039
Reading a page of a receipt, an
375
00:10:18,039 --> 00:10:19,559
invoice, or an ID costs a bit
376
00:10:19,559 --> 00:10:19,879
more.
377
00:10:19,879 --> 00:10:22,480
Custom document processing where
the model is tuned to your own
378
00:10:22,480 --> 00:10:25,440
forms costs the most around one
hundred credits per page.
379
00:10:25,440 --> 00:10:26,480
A couple of tools are simply
380
00:10:26,480 --> 00:10:27,679
free, and those include
381
00:10:27,679 --> 00:10:29,039
prediction and the business card
382
00:10:29,039 --> 00:10:29,480
reader.
383
00:10:29,480 --> 00:10:32,240
So the same platform can be
almost nothing or expensive,
384
00:10:32,240 --> 00:10:35,240
depending on which model you
lean on and how often it runs.
385
00:10:35,240 --> 00:10:36,320
One more shift deserves a
386
00:10:36,320 --> 00:10:37,519
sentence because you'll hear
387
00:10:37,519 --> 00:10:38,159
about it.
388
00:10:38,159 --> 00:10:41,200
Microsoft has been moving AI
builder onto copilot credits,
389
00:10:41,200 --> 00:10:43,919
the same currency that pays for
its other AI tools, and a
390
00:10:43,919 --> 00:10:46,519
feature will try your AI builder
credits first and then fall back
391
00:10:46,519 --> 00:10:47,960
to copilot credits.
392
00:10:47,960 --> 00:10:50,759
The practical upshot is that the
idea stays the same even as the
393
00:10:50,759 --> 00:10:52,240
label on the currency changes.
394
00:10:52,240 --> 00:10:53,480
And here's the calm down fact
395
00:10:53,480 --> 00:10:54,600
for anyone worried about a
396
00:10:54,600 --> 00:10:56,000
surprise bill, going over your
397
00:10:56,000 --> 00:10:57,279
monthly credits is treated as a
398
00:10:57,279 --> 00:10:58,360
grace period, and it isn't
399
00:10:58,360 --> 00:10:58,960
charged.
400
00:10:58,960 --> 00:11:00,240
Unused credits don't roll into
401
00:11:00,240 --> 00:11:01,519
the next month, so the goal is
402
00:11:01,519 --> 00:11:02,919
simply to buy roughly what you
403
00:11:02,919 --> 00:11:03,600
use.
404
00:11:03,600 --> 00:11:05,120
Because consumption resets every
405
00:11:05,120 --> 00:11:06,399
month, your planning window
406
00:11:06,399 --> 00:11:08,059
stays short, so you can adjust
407
00:11:08,059 --> 00:11:09,320
what you buy as your usage
408
00:11:09,320 --> 00:11:10,200
becomes clear.
409
00:11:10,200 --> 00:11:12,559
Nothing about this is designed
to trap you.
410
00:11:12,559 --> 00:11:14,600
Where AI builder fits in real
work.
411
00:11:14,600 --> 00:11:15,720
Now it's time to bring this down
412
00:11:15,720 --> 00:11:16,759
to earth because the concept
413
00:11:16,759 --> 00:11:18,120
only lands once you can picture
414
00:11:18,120 --> 00:11:19,000
it doing something you
415
00:11:19,000 --> 00:11:19,960
recognize.
416
00:11:19,960 --> 00:11:21,759
AI builder isn't one tool doing
one job.
417
00:11:21,759 --> 00:11:23,360
It's a set of models, and each
418
00:11:23,360 --> 00:11:25,080
one maps to a task that probably
419
00:11:25,080 --> 00:11:26,559
already lives on someone's to do
420
00:11:26,559 --> 00:11:27,080
list.
421
00:11:27,080 --> 00:11:28,179
Think about expense reports,
422
00:11:28,179 --> 00:11:29,399
receipt processing, reads the
423
00:11:29,399 --> 00:11:30,960
receipts, pulls the amounts and
424
00:11:30,960 --> 00:11:32,159
dates, and drops them into a
425
00:11:32,159 --> 00:11:33,179
report, which is the kind of
426
00:11:33,179 --> 00:11:34,399
chore nobody misses once it's
427
00:11:34,399 --> 00:11:34,879
gone.
428
00:11:34,879 --> 00:11:35,840
Think about contracts.
429
00:11:35,840 --> 00:11:37,039
Contract processing lifts out
430
00:11:37,039 --> 00:11:38,399
the key clauses and data points,
431
00:11:38,399 --> 00:11:39,840
so a legal team isn't hunting
432
00:11:39,840 --> 00:11:41,039
through forty pages for one
433
00:11:41,039 --> 00:11:41,559
date.
434
00:11:41,559 --> 00:11:43,159
Customer feedback is another
easy one.
435
00:11:43,159 --> 00:11:46,159
Category classification sorts it
into topics, and sentiment
436
00:11:46,159 --> 00:11:47,960
analysis tells you whether
people are happy.
437
00:11:47,960 --> 00:11:50,600
So you stop reading every
message just to find the trend.
438
00:11:50,600 --> 00:11:51,759
Then there's the work that used
439
00:11:51,759 --> 00:11:53,200
to need a specialist prediction
440
00:11:53,200 --> 00:11:54,480
can flag transactions that look
441
00:11:54,480 --> 00:11:55,679
fraudulent, or customers who
442
00:11:55,679 --> 00:11:56,759
might pay late and object
443
00:11:56,759 --> 00:11:58,120
detection can take inventory by
444
00:11:58,120 --> 00:11:59,440
spotting products in a photo of
445
00:11:59,440 --> 00:12:00,120
a shelf.
446
00:12:00,120 --> 00:12:02,360
Even something as small as
snapping a photo of text and
447
00:12:02,360 --> 00:12:05,399
saving it into a database is a
model away, and text recognition
448
00:12:05,399 --> 00:12:06,399
does exactly that.
449
00:12:06,399 --> 00:12:08,110
Each of these tasks has the same
shape.
450
00:12:08,110 --> 00:12:09,429
It's repetitive, it's rules,
451
00:12:09,429 --> 00:12:10,509
light, and a person can do it
452
00:12:10,509 --> 00:12:11,590
but would rather not do it a
453
00:12:11,590 --> 00:12:12,590
thousand times.
454
00:12:12,590 --> 00:12:14,470
That is where AI builder earns
its keep.
455
00:12:14,470 --> 00:12:15,669
The pattern repeats across
456
00:12:15,669 --> 00:12:16,990
teams, which is why the same
457
00:12:16,990 --> 00:12:18,230
handful of models keeps showing
458
00:12:18,230 --> 00:12:20,029
up in expense, legal, sales and
459
00:12:20,029 --> 00:12:20,950
support work.
460
00:12:20,950 --> 00:12:22,029
It isn't here to think for you.
461
00:12:22,029 --> 00:12:23,389
It's here to take the boring
462
00:12:23,389 --> 00:12:25,110
half of a task off your desk so
463
00:12:25,110 --> 00:12:26,230
you can spend your attention on
464
00:12:26,230 --> 00:12:27,350
the part that actually needs a
465
00:12:27,350 --> 00:12:27,750
human.
466
00:12:27,750 --> 00:12:29,230
And that's really the whole
point.
467
00:12:29,230 --> 00:12:32,350
AI builder is the bridge that
lets you use AI without becoming
468
00:12:32,350 --> 00:12:33,389
an AI person.
469
00:12:33,389 --> 00:12:36,269
So you keep working in the tools
you already know, while the
470
00:12:36,269 --> 00:12:39,190
intelligence shows up as one
more step in an app or a flow
471
00:12:39,190 --> 00:12:40,710
you built yourself.
472
00:12:40,710 --> 00:12:41,590
Wrapping up.
473
00:12:41,590 --> 00:12:43,110
So here's the quick version.
474
00:12:43,110 --> 00:12:46,029
AI builder puts AI inside Power
apps and Power Automate so you
475
00:12:46,029 --> 00:12:47,269
can add it without writing code.
476
00:12:47,269 --> 00:12:50,230
You choose between ready made
models and custom models trained
477
00:12:50,230 --> 00:12:51,629
on your own data.
478
00:12:51,629 --> 00:12:54,690
Document processing reads your
paperwork prediction forecasts,
479
00:12:54,690 --> 00:12:57,669
what happens next, and credits
are how you pay once a model
480
00:12:57,669 --> 00:12:58,590
runs for real.
481
00:12:58,590 --> 00:13:00,269
And here's the question I'll
leave you with.
482
00:13:00,269 --> 00:13:01,549
If you could hand just one
483
00:13:01,549 --> 00:13:03,590
repetitive task to a model this
484
00:13:03,590 --> 00:13:04,629
month, which task would you
485
00:13:04,629 --> 00:13:04,950
choose?
486
00:13:04,950 --> 00:13:06,750
And would a ready made model do
the job?
487
00:13:06,750 --> 00:13:09,309
Or would you need to teach one
your own way of working?
00:00:00,120 --> 00:00:01,360
The AI everyone keeps telling
2
00:00:01,360 --> 00:00:02,640
you to learn is already hiding
3
00:00:02,640 --> 00:00:03,960
inside the tools you open every
4
00:00:03,960 --> 00:00:04,320
morning.
5
00:00:04,320 --> 00:00:06,759
That's AI builder and it lives
inside power Apps and Power
6
00:00:06,759 --> 00:00:09,359
Automate, the low code tools
you're probably already using.
7
00:00:09,359 --> 00:00:11,320
You don't need to write code,
and you don't need a data
8
00:00:11,320 --> 00:00:12,800
science team to put it to work.
9
00:00:12,800 --> 00:00:13,839
By the end of this, you'll know
10
00:00:13,839 --> 00:00:15,320
what AI builder actually is,
11
00:00:15,320 --> 00:00:16,120
what it can do with your
12
00:00:16,120 --> 00:00:17,440
documents, and how you pay for
13
00:00:17,440 --> 00:00:17,879
it.
14
00:00:17,879 --> 00:00:21,800
Welcome to the M365 FM podcast,
where we explore the people,
15
00:00:21,800 --> 00:00:24,760
ideas, and technologies shaping
the future of work.
16
00:00:24,760 --> 00:00:27,399
And by the time we're done, the
jargon will be gone and the
17
00:00:27,399 --> 00:00:30,320
whole thing will feel a lot
smaller than it sounds.
18
00:00:30,320 --> 00:00:33,359
What AI builder actually is AI
builder is a feature inside
19
00:00:33,359 --> 00:00:34,840
Microsoft Power Platform.
20
00:00:34,840 --> 00:00:37,880
It lets you create and use AI
models to take over business
21
00:00:37,880 --> 00:00:39,640
tasks you'd normally do by hand.
22
00:00:39,640 --> 00:00:42,079
The word model is the one that
scares people, so it helps to
23
00:00:42,079 --> 00:00:43,280
shrink it down to size.
24
00:00:43,280 --> 00:00:44,799
A model is a pattern finder.
25
00:00:44,799 --> 00:00:46,119
You show it a pile of examples
26
00:00:46,119 --> 00:00:47,259
and it learns what tends to go
27
00:00:47,259 --> 00:00:47,920
with what.
28
00:00:47,920 --> 00:00:49,140
Then it applies that lesson to
29
00:00:49,140 --> 00:00:50,240
something new and that is the
30
00:00:50,240 --> 00:00:50,960
whole trick.
31
00:00:50,960 --> 00:00:54,159
The reason beginners freeze up
is rarely the model itself.
32
00:00:54,159 --> 00:00:55,039
It's the story they tell
33
00:00:55,039 --> 00:00:56,420
themselves about what using AI
34
00:00:56,420 --> 00:00:57,000
requires.
35
00:00:57,000 --> 00:00:57,840
They picture a team of
36
00:00:57,840 --> 00:00:59,560
engineers, a data warehouse, and
37
00:00:59,560 --> 00:01:00,759
six months of setup before
38
00:01:00,759 --> 00:01:01,789
anything works.
39
00:01:01,789 --> 00:01:04,750
That picture is what makes AI
feel out of reach.
40
00:01:04,750 --> 00:01:08,030
AI builder is designed to skip
every one of those steps.
41
00:01:08,030 --> 00:01:10,430
Microsoft's own recipe is for
moves.
42
00:01:10,430 --> 00:01:13,069
You pick a model type, you
connect your data, the training
43
00:01:13,069 --> 00:01:15,109
happens on its own, and then you
use the result.
44
00:01:15,109 --> 00:01:16,109
There's no coding, no data
45
00:01:16,109 --> 00:01:17,609
science degree, and no secret
46
00:01:17,609 --> 00:01:18,430
handshake.
47
00:01:18,430 --> 00:01:21,310
You also don't have to go find a
separate tool to use it.
48
00:01:21,310 --> 00:01:22,430
AI builder is built straight
49
00:01:22,430 --> 00:01:23,709
into Power Apps and Power
50
00:01:23,709 --> 00:01:25,349
Automate, which are the same low
51
00:01:25,349 --> 00:01:25,989
code platforms.
52
00:01:25,989 --> 00:01:28,670
You're likely already learning
that placement matters more than
53
00:01:28,670 --> 00:01:31,489
it sounds, because it means the
AI shows up exactly where the
54
00:01:31,489 --> 00:01:32,510
work already happens.
55
00:01:32,510 --> 00:01:35,349
It sits inside the app, your
team fills out or inside the
56
00:01:35,349 --> 00:01:37,230
flow that moves your files
around at night.
57
00:01:37,230 --> 00:01:38,670
You're not learning a second
career.
58
00:01:38,670 --> 00:01:40,230
You're adding one new capability
59
00:01:40,230 --> 00:01:41,230
to something you already
60
00:01:41,230 --> 00:01:41,790
touched.
61
00:01:41,790 --> 00:01:44,069
There's a reason Microsoft built
it this way.
62
00:01:44,069 --> 00:01:46,629
Most AI headlines are about
giant models that answer
63
00:01:46,629 --> 00:01:48,069
questions in a chat box.
64
00:01:48,069 --> 00:01:50,109
Those are impressive, but
they're hard to point at one
65
00:01:50,109 --> 00:01:51,510
specific business chore.
66
00:01:51,510 --> 00:01:52,989
AI builder flips that around.
67
00:01:52,989 --> 00:01:54,909
It starts from the task instead
of the chat.
68
00:01:54,909 --> 00:01:57,269
So you begin with a job that
needs doing and then pick the
69
00:01:57,269 --> 00:01:57,989
model that does it.
70
00:01:57,989 --> 00:01:59,349
That's why a beginner can use it
71
00:01:59,349 --> 00:02:00,709
without ever training a network
72
00:02:00,709 --> 00:02:01,469
by hand.
73
00:02:01,469 --> 00:02:02,859
So keep this picture in mind.
74
00:02:02,859 --> 00:02:03,819
AI builder is AI.
75
00:02:03,819 --> 00:02:05,219
You can actually use sitting
76
00:02:05,219 --> 00:02:06,659
quietly inside tools you already
77
00:02:06,659 --> 00:02:07,060
know.
78
00:02:07,060 --> 00:02:09,939
And everything else in this
video hangs off that one idea.
79
00:02:09,939 --> 00:02:12,020
Ready made models versus models
you build.
80
00:02:12,020 --> 00:02:14,180
Once you accept that this is
real and reachable, the next
81
00:02:14,180 --> 00:02:16,379
question is which kind of model
you actually reach for.
82
00:02:16,379 --> 00:02:19,000
Microsoft gives you two paths,
and the difference comes down to
83
00:02:19,000 --> 00:02:20,419
whose data does the learning.
84
00:02:20,419 --> 00:02:22,099
The first path is pre-built
models.
85
00:02:22,099 --> 00:02:24,259
These are models Microsoft
already trained for problems
86
00:02:24,259 --> 00:02:26,740
that show up in almost every
business, so you skip straight
87
00:02:26,740 --> 00:02:29,379
to using them point one at an
invoice and it pulls out the
88
00:02:29,379 --> 00:02:30,659
totals and the dates.
89
00:02:30,659 --> 00:02:31,780
Point one at a picture of a
90
00:02:31,780 --> 00:02:32,979
receipt and it reads the line
91
00:02:32,979 --> 00:02:33,379
items.
92
00:02:33,379 --> 00:02:34,659
There are pre-built models for
93
00:02:34,659 --> 00:02:36,500
reading text inside an image for
94
00:02:36,500 --> 00:02:37,740
judging whether feedback sounds
95
00:02:37,740 --> 00:02:39,099
positive or negative, and for
96
00:02:39,099 --> 00:02:40,300
pulling names and numbers out of
97
00:02:40,300 --> 00:02:41,419
messy text.
98
00:02:41,419 --> 00:02:42,819
Others read ID documents.
99
00:02:42,819 --> 00:02:45,819
Grab the key phrases from a
paragraph, scan a business card,
100
00:02:45,819 --> 00:02:47,939
or detect what language
something is written in.
101
00:02:47,939 --> 00:02:50,259
You didn't train any of them and
you didn't have to.
102
00:02:50,259 --> 00:02:52,819
The second path is custom
models, and this is where your
103
00:02:52,819 --> 00:02:54,300
own data does the work.
104
00:02:54,300 --> 00:02:57,060
These models are trained on your
documents, your text, your
105
00:02:57,060 --> 00:03:00,219
images, and your history, which
makes them fit the way your
106
00:03:00,219 --> 00:03:02,099
business actually looks.
107
00:03:02,099 --> 00:03:04,780
That list includes document
processing, which reads your
108
00:03:04,780 --> 00:03:07,979
specific forms and category
classification, which sorts text
109
00:03:07,979 --> 00:03:09,500
into categories you define.
110
00:03:09,500 --> 00:03:11,780
It also includes entity
extraction, which finds the
111
00:03:11,780 --> 00:03:14,939
exact details you care about,
prediction, which guesses future
112
00:03:14,939 --> 00:03:17,979
outcomes from your past data,
and object detection, which
113
00:03:17,979 --> 00:03:19,740
spots your products in a photo.
114
00:03:19,740 --> 00:03:22,099
There's even a way to bring your
own model over from Azure if you
115
00:03:22,099 --> 00:03:23,020
already built one.
116
00:03:23,020 --> 00:03:25,500
So the rule of thumb is simple
if the problem is common and
117
00:03:25,500 --> 00:03:27,699
generic, like reading any
receipt that lands in your
118
00:03:27,699 --> 00:03:30,939
inbox, a pre-built model has you
covered on day one when the
119
00:03:30,939 --> 00:03:32,379
problem is uniquely yours.
120
00:03:32,379 --> 00:03:35,020
Like reading invoices from one
specific supplier.
121
00:03:35,020 --> 00:03:36,460
You train a custom model on your
122
00:03:36,460 --> 00:03:37,860
own examples, and it learns your
123
00:03:37,860 --> 00:03:38,419
quirks.
124
00:03:38,419 --> 00:03:39,860
The difference is easier to deal
with.
125
00:03:39,860 --> 00:03:40,659
An example.
126
00:03:40,659 --> 00:03:43,340
A pre-built model can read the
total of a standard invoice from
127
00:03:43,340 --> 00:03:45,479
almost any vendor because
Microsoft trained it on
128
00:03:45,479 --> 00:03:46,379
thousands of them.
129
00:03:46,379 --> 00:03:49,620
But if your invoices come in a
layout nobody else uses, with
130
00:03:49,620 --> 00:03:52,860
the account number buried inside
a table, that pre-built model
131
00:03:52,860 --> 00:03:53,500
will miss it.
132
00:03:53,500 --> 00:03:56,060
That is exactly the job you hand
to a custom model.
133
00:03:56,060 --> 00:03:58,939
Most beginners assume custom
always means better, and that
134
00:03:58,939 --> 00:03:59,939
assumption is wrong.
135
00:03:59,939 --> 00:04:01,740
Custom means more specific, and
136
00:04:01,740 --> 00:04:03,020
more specific is only better
137
00:04:03,020 --> 00:04:04,370
when your problem is specific
138
00:04:04,370 --> 00:04:04,610
to.
139
00:04:04,610 --> 00:04:06,050
One more thing before you pick.
140
00:04:06,050 --> 00:04:08,449
A few of these models are still
in preview, which means they're
141
00:04:08,449 --> 00:04:10,889
not meant for production work
yet, and some aren't available
142
00:04:10,889 --> 00:04:11,770
in every region.
143
00:04:11,770 --> 00:04:12,930
So if you go looking for a model
144
00:04:12,930 --> 00:04:13,849
and can't find it, that's
145
00:04:13,849 --> 00:04:14,849
usually why.
146
00:04:14,849 --> 00:04:16,810
Why this is not just another
chatbot.
147
00:04:16,810 --> 00:04:18,050
Before we go further, there's a
148
00:04:18,050 --> 00:04:19,649
confusion worth clearing up
149
00:04:19,649 --> 00:04:20,930
because it trips up almost every
150
00:04:20,930 --> 00:04:21,490
beginner.
151
00:04:21,490 --> 00:04:24,129
When people hear AI, they now
think of a chat window that
152
00:04:24,129 --> 00:04:26,889
answers questions and they
assume AI builder is the same
153
00:04:26,889 --> 00:04:28,410
thing wearing a Microsoft badge.
154
00:04:28,410 --> 00:04:30,610
That assumption is wrong, and
the difference is the whole
155
00:04:30,610 --> 00:04:31,449
reason it's useful.
156
00:04:31,449 --> 00:04:32,649
A chatbot is general.
157
00:04:32,649 --> 00:04:33,649
It will talk about anything,
158
00:04:33,649 --> 00:04:34,810
which is impressive, but it's
159
00:04:34,810 --> 00:04:36,449
also vague because it has no
160
00:04:36,449 --> 00:04:38,129
idea what your specific job
161
00:04:38,129 --> 00:04:39,089
looks like.
162
00:04:39,089 --> 00:04:40,730
AI builder takes the opposite
approach.
163
00:04:40,730 --> 00:04:43,209
It starts with a task like
reading an invoice or scoring a
164
00:04:43,209 --> 00:04:46,370
record, and it brings a model
that's focused on exactly that.
165
00:04:46,370 --> 00:04:48,449
You're not asking it to be
clever about everything.
166
00:04:48,449 --> 00:04:51,529
You're asking it to be reliable
about one thing over and over
167
00:04:51,529 --> 00:04:52,689
without getting bored.
168
00:04:52,689 --> 00:04:55,290
That focus is why the results
are easier to trust.
169
00:04:55,290 --> 00:04:58,050
A focused model does the same
job the same way every time,
170
00:04:58,050 --> 00:05:00,810
which is what you want when the
output feeds a business process.
171
00:05:00,810 --> 00:05:02,529
It's also why AI builder fits
172
00:05:02,529 --> 00:05:03,920
inside power apps and Power
173
00:05:03,920 --> 00:05:04,279
Automate.
174
00:05:04,279 --> 00:05:05,959
So naturally, those tools are
175
00:05:05,959 --> 00:05:07,220
already about doing a defined
176
00:05:07,220 --> 00:05:08,120
job well.
177
00:05:08,120 --> 00:05:10,399
So a model that does one thing
well slides right in.
178
00:05:10,399 --> 00:05:11,720
Now AI builder does include
179
00:05:11,720 --> 00:05:12,600
prompts, which are the more
180
00:05:12,600 --> 00:05:14,079
conversational generative side
181
00:05:14,079 --> 00:05:14,439
of things.
182
00:05:14,439 --> 00:05:15,959
So the line isn't perfectly
clean.
183
00:05:15,959 --> 00:05:18,180
When people talk about AI
builder for beginners though,
184
00:05:18,180 --> 00:05:20,920
they usually mean the models,
the focused pattern finders.
185
00:05:20,920 --> 00:05:23,879
And that's where the value shows
up fastest.
186
00:05:23,879 --> 00:05:25,160
What document processing really
187
00:05:25,160 --> 00:05:26,680
does document processing is the
188
00:05:26,680 --> 00:05:27,839
one people ask about most,
189
00:05:27,839 --> 00:05:29,040
because it attacks a job that
190
00:05:29,040 --> 00:05:30,439
quietly eats hours every single
191
00:05:30,439 --> 00:05:31,199
week.
192
00:05:31,199 --> 00:05:32,600
In plain terms, it reads and
193
00:05:32,600 --> 00:05:33,720
saves the information out of
194
00:05:33,720 --> 00:05:35,639
standard documents and invoices
195
00:05:35,639 --> 00:05:37,040
and tax forms are the classic
196
00:05:37,040 --> 00:05:37,600
example.
197
00:05:37,600 --> 00:05:40,920
Instead of a person opening a
PDF, squinting at a total, and
198
00:05:40,920 --> 00:05:43,240
retyping it into a system, the
model does that.
199
00:05:43,240 --> 00:05:43,920
Reading for you.
200
00:05:43,920 --> 00:05:45,120
The way you get there is three
201
00:05:45,120 --> 00:05:46,279
steps, and none of them are
202
00:05:46,279 --> 00:05:46,920
scary.
203
00:05:46,920 --> 00:05:48,079
First, you train the model,
204
00:05:48,079 --> 00:05:49,120
which just means you feed it
205
00:05:49,120 --> 00:05:50,199
examples and show it which
206
00:05:50,199 --> 00:05:51,439
pieces of information you want
207
00:05:51,439 --> 00:05:51,720
back.
208
00:05:51,720 --> 00:05:52,839
Then you publish it, which makes
209
00:05:52,839 --> 00:05:53,959
it available to your apps and
210
00:05:53,959 --> 00:05:54,839
your flows.
211
00:05:54,839 --> 00:05:55,600
Then you use it.
212
00:05:55,600 --> 00:05:58,399
And from that point on, the
reading takes care of itself.
213
00:05:58,399 --> 00:06:00,360
Every time a new document
arrives.
214
00:06:00,360 --> 00:06:03,160
Once it is published, the model
keeps doing that same job
215
00:06:03,160 --> 00:06:04,589
quietly in the background.
216
00:06:04,589 --> 00:06:07,310
The part that surprises almost
everyone comes next.
217
00:06:07,310 --> 00:06:09,430
You only need five documents to
get started.
218
00:06:09,430 --> 00:06:12,189
Not five thousand, and not some
labeled data set.
219
00:06:12,189 --> 00:06:14,029
You'd have to hire someone to
build first.
220
00:06:14,029 --> 00:06:17,029
Microsoft designed it so a
business user can point it at a
221
00:06:17,029 --> 00:06:19,829
handful of real documents and
still get accurate results
222
00:06:19,829 --> 00:06:21,430
tailored to their own content.
223
00:06:21,430 --> 00:06:22,910
That single detail is usually
224
00:06:22,910 --> 00:06:24,110
the moment the idea stops
225
00:06:24,110 --> 00:06:25,029
feeling like an enterprise
226
00:06:25,029 --> 00:06:26,269
project and starts feeling like
227
00:06:26,269 --> 00:06:27,389
something you could try this
228
00:06:27,389 --> 00:06:28,110
afternoon.
229
00:06:28,110 --> 00:06:30,790
Notice how this fits the pattern
from the last section.
230
00:06:30,790 --> 00:06:33,589
Document processing is a custom
model because it learns the
231
00:06:33,589 --> 00:06:34,589
shape of your documents.
232
00:06:34,589 --> 00:06:37,629
Specifically, a generic invoice
reader can't know that your
233
00:06:37,629 --> 00:06:40,110
supplier hides the due date down
in the footer.
234
00:06:40,110 --> 00:06:41,589
Yours can because you taught it.
235
00:06:41,589 --> 00:06:42,350
And that's the difference
236
00:06:42,350 --> 00:06:43,550
between a tool that sort of
237
00:06:43,550 --> 00:06:44,550
works and one that actually
238
00:06:44,550 --> 00:06:44,910
fits.
239
00:06:44,910 --> 00:06:46,389
The payoff isn't just speed
either.
240
00:06:46,389 --> 00:06:48,670
When a machine reads the
document, the data lands in your
241
00:06:48,670 --> 00:06:50,910
system clean and consistent in
the right fields.
242
00:06:50,910 --> 00:06:52,189
At the moment the document shows
243
00:06:52,189 --> 00:06:53,689
up, the errors that come from
244
00:06:53,689 --> 00:06:55,029
tired eyes and copy paste
245
00:06:55,029 --> 00:06:55,870
disappear.
246
00:06:55,870 --> 00:06:57,029
And the person who used to do
247
00:06:57,029 --> 00:06:58,149
that reading gets their
248
00:06:58,149 --> 00:06:58,910
afternoon back.
249
00:06:58,910 --> 00:06:59,750
That is the bargain.
250
00:06:59,750 --> 00:07:02,990
Document processing makes how
prediction works.
251
00:07:02,990 --> 00:07:05,750
If document processing is about
reading what already exists.
252
00:07:05,750 --> 00:07:06,990
Prediction is about guessing
253
00:07:06,990 --> 00:07:08,550
what comes next, and it works
254
00:07:08,550 --> 00:07:09,629
from a completely different
255
00:07:09,629 --> 00:07:10,350
angle.
256
00:07:10,350 --> 00:07:12,589
A prediction model studies the
patterns in your past data,
257
00:07:12,589 --> 00:07:15,389
figures out what tends to lead
to what, and then uses that on
258
00:07:15,389 --> 00:07:16,509
new records as they arrive.
259
00:07:16,509 --> 00:07:18,069
You're not feeding it documents
this time.
260
00:07:18,069 --> 00:07:19,550
You're feeding it history.
261
00:07:19,550 --> 00:07:22,029
The answers it gives come in
three shapes, and that changes
262
00:07:22,029 --> 00:07:23,110
what you can ask.
263
00:07:23,110 --> 00:07:24,670
The first is a simple yes or no,
264
00:07:24,670 --> 00:07:25,910
like whether a transaction looks
265
00:07:25,910 --> 00:07:27,350
fraudulent or whether an account
266
00:07:27,350 --> 00:07:28,709
is likely to pay its invoice on
267
00:07:28,709 --> 00:07:29,189
time.
268
00:07:29,189 --> 00:07:31,509
The second is a choice between
several outcomes, like whether a
269
00:07:31,509 --> 00:07:34,069
shipment will arrive early on
time, late or very late.
270
00:07:34,069 --> 00:07:36,470
The third is a number like how
many items you should keep in
271
00:07:36,470 --> 00:07:39,029
inventory, or how many days a
delivery will take notice.
272
00:07:39,029 --> 00:07:40,550
How ordinary those questions
are.
273
00:07:40,550 --> 00:07:42,230
None of them require a research
lab.
274
00:07:42,230 --> 00:07:44,790
They're the exact questions a
manager already asks in a
275
00:07:44,790 --> 00:07:47,230
meeting, and prediction just
gives a data backed answer
276
00:07:47,230 --> 00:07:48,149
instead of a hunch.
277
00:07:48,149 --> 00:07:49,389
It takes a guess that used to
278
00:07:49,389 --> 00:07:50,910
live in someone's head and turns
279
00:07:50,910 --> 00:07:52,329
it into something you can see
280
00:07:52,329 --> 00:07:53,670
and act on, and that is why it
281
00:07:53,670 --> 00:07:54,550
feels useful.
282
00:07:54,550 --> 00:07:57,389
Keep one honest caveat in mind,
though, because it matters.
283
00:07:57,389 --> 00:07:58,509
Prediction is a pattern.
284
00:07:58,509 --> 00:08:00,069
Guess not a crystal ball.
285
00:08:00,069 --> 00:08:01,750
It's only as good as the history
you feed it.
286
00:08:01,750 --> 00:08:03,069
And if your past data is messy
287
00:08:03,069 --> 00:08:04,589
or thin, the predictions will be
288
00:08:04,589 --> 00:08:05,420
messy too.
289
00:08:05,420 --> 00:08:08,100
So before you trust a
prediction, look at the history
290
00:08:08,100 --> 00:08:10,459
behind it, because that history
is doing most of the work.
291
00:08:10,459 --> 00:08:11,579
The model isn't inventing new
292
00:08:11,579 --> 00:08:12,600
wisdom, it's reflecting what
293
00:08:12,600 --> 00:08:13,699
already happened and projecting
294
00:08:13,699 --> 00:08:14,379
it forward.
295
00:08:14,379 --> 00:08:14,939
Used well.
296
00:08:14,939 --> 00:08:15,660
That's powerful.
297
00:08:15,660 --> 00:08:16,740
Used carelessly.
298
00:08:16,740 --> 00:08:19,980
It just automates an old bias
where you'd actually use it.
299
00:08:19,980 --> 00:08:22,620
So far, we've talked about what
AI builder is and what it does.
300
00:08:22,620 --> 00:08:24,899
And now it's worth asking the
practical question, where does
301
00:08:24,899 --> 00:08:25,980
this actually plug in?
302
00:08:25,980 --> 00:08:28,259
The answer is in the two tools
you're already using.
303
00:08:28,259 --> 00:08:29,300
And each one gives you a
304
00:08:29,300 --> 00:08:30,620
slightly different way to reach
305
00:08:30,620 --> 00:08:31,300
the models.
306
00:08:31,300 --> 00:08:34,580
In Power Automate, you add an AI
builder action to a cloud flow.
307
00:08:34,580 --> 00:08:35,539
A flow is just a set of
308
00:08:35,539 --> 00:08:36,740
instructions that runs on its
309
00:08:36,740 --> 00:08:36,940
own.
310
00:08:36,940 --> 00:08:38,820
So a new invoice lands in a
folder.
311
00:08:38,820 --> 00:08:41,419
The flow hands it to the
document processing model, and
312
00:08:41,419 --> 00:08:44,019
the extracted details get filed
into your system without a
313
00:08:44,019 --> 00:08:45,059
person touching it.
314
00:08:45,059 --> 00:08:46,740
You're not building the AI here.
315
00:08:46,740 --> 00:08:48,340
You're wiring an already trained
316
00:08:48,340 --> 00:08:50,039
model into a process you set up
317
00:08:50,039 --> 00:08:50,899
in PowerApps.
318
00:08:50,899 --> 00:08:54,100
You reach the same models
through power effects, which is
319
00:08:54,100 --> 00:08:56,379
the Excel like formula language
behind the app.
320
00:08:56,379 --> 00:08:58,580
That means a form can read a
document while the user is
321
00:08:58,580 --> 00:09:01,179
working or score a record the
moment it's saved.
322
00:09:01,179 --> 00:09:03,059
The AI isn't a separate screen
you switch to.
323
00:09:03,059 --> 00:09:04,179
It's a step inside the app,
324
00:09:04,179 --> 00:09:05,500
which is exactly why beginners
325
00:09:05,500 --> 00:09:06,690
pick it up without learning a
326
00:09:06,690 --> 00:09:07,330
new discipline.
327
00:09:07,330 --> 00:09:08,490
There's one caveat that trips
328
00:09:08,490 --> 00:09:09,809
people up, and it's worth naming
329
00:09:09,809 --> 00:09:10,490
plainly.
330
00:09:10,490 --> 00:09:12,250
Adding an AI builder action
331
00:09:12,250 --> 00:09:13,610
turns a flow into a premium
332
00:09:13,610 --> 00:09:13,889
flow.
333
00:09:13,889 --> 00:09:17,090
And if that flow lives inside an
app, it makes the app a premium.
334
00:09:17,090 --> 00:09:18,769
App premium means a paid
license.
335
00:09:18,769 --> 00:09:21,330
So this isn't a free add on to
the cheapest plan.
336
00:09:21,330 --> 00:09:22,450
It's better to know that before
337
00:09:22,450 --> 00:09:23,490
you build because the cost
338
00:09:23,490 --> 00:09:25,090
conversation starts the moment a
339
00:09:25,090 --> 00:09:26,690
model runs in something real
340
00:09:26,690 --> 00:09:27,529
that leads straight to the
341
00:09:27,529 --> 00:09:28,769
question everyone asks next,
342
00:09:28,769 --> 00:09:29,850
which is what this actually
343
00:09:29,850 --> 00:09:30,370
costs.
344
00:09:30,370 --> 00:09:31,809
That's where credits come in.
345
00:09:31,809 --> 00:09:33,769
What credits are and why they
matter.
346
00:09:33,769 --> 00:09:35,529
The reassuring part comes first
347
00:09:35,529 --> 00:09:36,889
because it's the part beginners
348
00:09:36,889 --> 00:09:37,809
never hear.
349
00:09:37,809 --> 00:09:38,970
Building, training and testing
350
00:09:38,970 --> 00:09:40,250
models in AI builder costs
351
00:09:40,250 --> 00:09:40,730
nothing.
352
00:09:40,730 --> 00:09:43,850
You can design a model, teach
it, run it against samples and
353
00:09:43,850 --> 00:09:47,049
confirm that it works all
without spending a cent only
354
00:09:47,049 --> 00:09:49,929
running a model for real inside
a live app or flow spends
355
00:09:49,929 --> 00:09:50,730
anything at all.
356
00:09:50,730 --> 00:09:51,649
So what is a credit?
357
00:09:51,649 --> 00:09:53,289
It's a prepaid unit of AI usage
358
00:09:53,289 --> 00:09:54,370
and you spend it when a model
359
00:09:54,370 --> 00:09:54,970
does work.
360
00:09:54,970 --> 00:09:56,370
Microsoft has historically sold
361
00:09:56,370 --> 00:09:57,929
an add on of about a million
362
00:09:57,929 --> 00:09:59,210
credits for roughly five hundred
363
00:09:59,210 --> 00:10:00,330
dollars a month, and some
364
00:10:00,330 --> 00:10:02,269
premium licenses also seed a few
365
00:10:02,269 --> 00:10:03,370
thousand credits to get you
366
00:10:03,370 --> 00:10:04,009
started.
367
00:10:04,009 --> 00:10:07,200
Treat those numbers as a rough
shape rather than a fixed price.
368
00:10:07,200 --> 00:10:08,559
Because licensing moves around
369
00:10:08,559 --> 00:10:09,799
and the details change over
370
00:10:09,799 --> 00:10:10,360
time.
371
00:10:10,360 --> 00:10:12,720
What burns through credits is
worth a look, too, because the
372
00:10:12,720 --> 00:10:13,639
spread is wide.
373
00:10:13,639 --> 00:10:16,600
Reading a page with basic text
recognition costs a few credits.
374
00:10:16,600 --> 00:10:18,039
Reading a page of a receipt, an
375
00:10:18,039 --> 00:10:19,559
invoice, or an ID costs a bit
376
00:10:19,559 --> 00:10:19,879
more.
377
00:10:19,879 --> 00:10:22,480
Custom document processing where
the model is tuned to your own
378
00:10:22,480 --> 00:10:25,440
forms costs the most around one
hundred credits per page.
379
00:10:25,440 --> 00:10:26,480
A couple of tools are simply
380
00:10:26,480 --> 00:10:27,679
free, and those include
381
00:10:27,679 --> 00:10:29,039
prediction and the business card
382
00:10:29,039 --> 00:10:29,480
reader.
383
00:10:29,480 --> 00:10:32,240
So the same platform can be
almost nothing or expensive,
384
00:10:32,240 --> 00:10:35,240
depending on which model you
lean on and how often it runs.
385
00:10:35,240 --> 00:10:36,320
One more shift deserves a
386
00:10:36,320 --> 00:10:37,519
sentence because you'll hear
387
00:10:37,519 --> 00:10:38,159
about it.
388
00:10:38,159 --> 00:10:41,200
Microsoft has been moving AI
builder onto copilot credits,
389
00:10:41,200 --> 00:10:43,919
the same currency that pays for
its other AI tools, and a
390
00:10:43,919 --> 00:10:46,519
feature will try your AI builder
credits first and then fall back
391
00:10:46,519 --> 00:10:47,960
to copilot credits.
392
00:10:47,960 --> 00:10:50,759
The practical upshot is that the
idea stays the same even as the
393
00:10:50,759 --> 00:10:52,240
label on the currency changes.
394
00:10:52,240 --> 00:10:53,480
And here's the calm down fact
395
00:10:53,480 --> 00:10:54,600
for anyone worried about a
396
00:10:54,600 --> 00:10:56,000
surprise bill, going over your
397
00:10:56,000 --> 00:10:57,279
monthly credits is treated as a
398
00:10:57,279 --> 00:10:58,360
grace period, and it isn't
399
00:10:58,360 --> 00:10:58,960
charged.
400
00:10:58,960 --> 00:11:00,240
Unused credits don't roll into
401
00:11:00,240 --> 00:11:01,519
the next month, so the goal is
402
00:11:01,519 --> 00:11:02,919
simply to buy roughly what you
403
00:11:02,919 --> 00:11:03,600
use.
404
00:11:03,600 --> 00:11:05,120
Because consumption resets every
405
00:11:05,120 --> 00:11:06,399
month, your planning window
406
00:11:06,399 --> 00:11:08,059
stays short, so you can adjust
407
00:11:08,059 --> 00:11:09,320
what you buy as your usage
408
00:11:09,320 --> 00:11:10,200
becomes clear.
409
00:11:10,200 --> 00:11:12,559
Nothing about this is designed
to trap you.
410
00:11:12,559 --> 00:11:14,600
Where AI builder fits in real
work.
411
00:11:14,600 --> 00:11:15,720
Now it's time to bring this down
412
00:11:15,720 --> 00:11:16,759
to earth because the concept
413
00:11:16,759 --> 00:11:18,120
only lands once you can picture
414
00:11:18,120 --> 00:11:19,000
it doing something you
415
00:11:19,000 --> 00:11:19,960
recognize.
416
00:11:19,960 --> 00:11:21,759
AI builder isn't one tool doing
one job.
417
00:11:21,759 --> 00:11:23,360
It's a set of models, and each
418
00:11:23,360 --> 00:11:25,080
one maps to a task that probably
419
00:11:25,080 --> 00:11:26,559
already lives on someone's to do
420
00:11:26,559 --> 00:11:27,080
list.
421
00:11:27,080 --> 00:11:28,179
Think about expense reports,
422
00:11:28,179 --> 00:11:29,399
receipt processing, reads the
423
00:11:29,399 --> 00:11:30,960
receipts, pulls the amounts and
424
00:11:30,960 --> 00:11:32,159
dates, and drops them into a
425
00:11:32,159 --> 00:11:33,179
report, which is the kind of
426
00:11:33,179 --> 00:11:34,399
chore nobody misses once it's
427
00:11:34,399 --> 00:11:34,879
gone.
428
00:11:34,879 --> 00:11:35,840
Think about contracts.
429
00:11:35,840 --> 00:11:37,039
Contract processing lifts out
430
00:11:37,039 --> 00:11:38,399
the key clauses and data points,
431
00:11:38,399 --> 00:11:39,840
so a legal team isn't hunting
432
00:11:39,840 --> 00:11:41,039
through forty pages for one
433
00:11:41,039 --> 00:11:41,559
date.
434
00:11:41,559 --> 00:11:43,159
Customer feedback is another
easy one.
435
00:11:43,159 --> 00:11:46,159
Category classification sorts it
into topics, and sentiment
436
00:11:46,159 --> 00:11:47,960
analysis tells you whether
people are happy.
437
00:11:47,960 --> 00:11:50,600
So you stop reading every
message just to find the trend.
438
00:11:50,600 --> 00:11:51,759
Then there's the work that used
439
00:11:51,759 --> 00:11:53,200
to need a specialist prediction
440
00:11:53,200 --> 00:11:54,480
can flag transactions that look
441
00:11:54,480 --> 00:11:55,679
fraudulent, or customers who
442
00:11:55,679 --> 00:11:56,759
might pay late and object
443
00:11:56,759 --> 00:11:58,120
detection can take inventory by
444
00:11:58,120 --> 00:11:59,440
spotting products in a photo of
445
00:11:59,440 --> 00:12:00,120
a shelf.
446
00:12:00,120 --> 00:12:02,360
Even something as small as
snapping a photo of text and
447
00:12:02,360 --> 00:12:05,399
saving it into a database is a
model away, and text recognition
448
00:12:05,399 --> 00:12:06,399
does exactly that.
449
00:12:06,399 --> 00:12:08,110
Each of these tasks has the same
shape.
450
00:12:08,110 --> 00:12:09,429
It's repetitive, it's rules,
451
00:12:09,429 --> 00:12:10,509
light, and a person can do it
452
00:12:10,509 --> 00:12:11,590
but would rather not do it a
453
00:12:11,590 --> 00:12:12,590
thousand times.
454
00:12:12,590 --> 00:12:14,470
That is where AI builder earns
its keep.
455
00:12:14,470 --> 00:12:15,669
The pattern repeats across
456
00:12:15,669 --> 00:12:16,990
teams, which is why the same
457
00:12:16,990 --> 00:12:18,230
handful of models keeps showing
458
00:12:18,230 --> 00:12:20,029
up in expense, legal, sales and
459
00:12:20,029 --> 00:12:20,950
support work.
460
00:12:20,950 --> 00:12:22,029
It isn't here to think for you.
461
00:12:22,029 --> 00:12:23,389
It's here to take the boring
462
00:12:23,389 --> 00:12:25,110
half of a task off your desk so
463
00:12:25,110 --> 00:12:26,230
you can spend your attention on
464
00:12:26,230 --> 00:12:27,350
the part that actually needs a
465
00:12:27,350 --> 00:12:27,750
human.
466
00:12:27,750 --> 00:12:29,230
And that's really the whole
point.
467
00:12:29,230 --> 00:12:32,350
AI builder is the bridge that
lets you use AI without becoming
468
00:12:32,350 --> 00:12:33,389
an AI person.
469
00:12:33,389 --> 00:12:36,269
So you keep working in the tools
you already know, while the
470
00:12:36,269 --> 00:12:39,190
intelligence shows up as one
more step in an app or a flow
471
00:12:39,190 --> 00:12:40,710
you built yourself.
472
00:12:40,710 --> 00:12:41,590
Wrapping up.
473
00:12:41,590 --> 00:12:43,110
So here's the quick version.
474
00:12:43,110 --> 00:12:46,029
AI builder puts AI inside Power
apps and Power Automate so you
475
00:12:46,029 --> 00:12:47,269
can add it without writing code.
476
00:12:47,269 --> 00:12:50,230
You choose between ready made
models and custom models trained
477
00:12:50,230 --> 00:12:51,629
on your own data.
478
00:12:51,629 --> 00:12:54,690
Document processing reads your
paperwork prediction forecasts,
479
00:12:54,690 --> 00:12:57,669
what happens next, and credits
are how you pay once a model
480
00:12:57,669 --> 00:12:58,590
runs for real.
481
00:12:58,590 --> 00:13:00,269
And here's the question I'll
leave you with.
482
00:13:00,269 --> 00:13:01,549
If you could hand just one
483
00:13:01,549 --> 00:13:03,590
repetitive task to a model this
484
00:13:03,590 --> 00:13:04,629
month, which task would you
485
00:13:04,629 --> 00:13:04,950
choose?
486
00:13:04,950 --> 00:13:06,750
And would a ready made model do
the job?
487
00:13:06,750 --> 00:13:09,309
Or would you need to teach one
your own way of working?
