The Probability Shift: How AI is Rewriting Power Platform Design

1
00:00:00,000 --> 00:00:03,440
Your low-code automations are breaking because they are too smart for their own good.

2
00:00:03,440 --> 00:00:08,480
Rigid if then logic is a death sentence when you face messy real world data inconsistencies

3
00:00:08,480 --> 00:00:10,000
that crash your systems daily.

4
00:00:10,000 --> 00:00:13,080
The assumption that data is clean, it's the original sin of low-code design.

5
00:00:13,080 --> 00:00:17,560
We built our entire world on the if then model, but that model requires absolute certainty

6
00:00:17,560 --> 00:00:18,560
to function.

7
00:00:18,560 --> 00:00:19,560
Real world data isn't certain.

8
00:00:19,560 --> 00:00:20,560
It has typos.

9
00:00:20,560 --> 00:00:22,200
It has missing context.

10
00:00:22,200 --> 00:00:24,280
It's a chaotic stream of human error.

11
00:00:24,280 --> 00:00:29,360
When a deterministic system hits a 1% variation, it doesn't bend and it certainly doesn't adapt.

12
00:00:29,360 --> 00:00:31,560
It simply breaks.

13
00:00:31,560 --> 00:00:34,920
This creates the brittleness trap where your automation is only as strong as your cleanest

14
00:00:34,920 --> 00:00:38,640
data point in the enterprise that's a very low bar to set for success, but the problem

15
00:00:38,640 --> 00:00:41,840
isn't the data, it's the model we're using to process it.

16
00:00:41,840 --> 00:00:43,760
The death of determinism.

17
00:00:43,760 --> 00:00:46,880
Deterministic logic is a relic of the structured data era.

18
00:00:46,880 --> 00:00:51,040
It worked when everything lived in a SQL table and it worked when we forced humans to fit

19
00:00:51,040 --> 00:00:53,400
into predefined boxes.

20
00:00:53,400 --> 00:00:54,400
But that world is gone.

21
00:00:54,400 --> 00:00:59,140
Today, 80% of enterprise data is unstructured, which means it lives in emails, chats and

22
00:00:59,140 --> 00:01:00,740
voice recordings.

23
00:01:00,740 --> 00:01:06,640
Traditional power-automate flows treat apples and 4 ppl3s as two different universes.

24
00:01:06,640 --> 00:01:09,900
To a binary system, they are completely unrelated strings.

25
00:01:09,900 --> 00:01:13,500
We've reached a breaking point where the cost of manual data cleaning is now exceeding

26
00:01:13,500 --> 00:01:15,380
the value of the automation itself.

27
00:01:15,380 --> 00:01:19,340
If you have to spend two hours fixing a CSV just so a flow can run for five seconds,

28
00:01:19,340 --> 00:01:21,180
you haven't actually automated anything.

29
00:01:21,180 --> 00:01:22,340
You've just shifted the labor.

30
00:01:22,340 --> 00:01:26,380
AI isn't just a shiny new feature we're sticking onto the side of the power platform.

31
00:01:26,380 --> 00:01:30,460
It's a fundamental shift from exact match to similarity.

32
00:01:30,460 --> 00:01:32,500
The old model asks, is this A?

33
00:01:32,500 --> 00:01:35,380
The new model asks, how much does this look like A?

34
00:01:35,380 --> 00:01:37,420
Think about how you navigate your inbox.

35
00:01:37,420 --> 00:01:40,940
You don't run a rejects in your head to find an invoice, you look for patterns, you look

36
00:01:40,940 --> 00:01:46,940
for context, you recognize that NV202 and invoice number 202 are the same thing even if

37
00:01:46,940 --> 00:01:48,700
the characters don't match.

38
00:01:48,700 --> 00:01:49,780
Deterministic systems can't do that.

39
00:01:49,780 --> 00:01:50,780
They are blind to intent.

40
00:01:50,780 --> 00:01:52,100
They only see the syntax.

41
00:01:52,100 --> 00:01:56,700
And as we move toward 2026, syntax is becoming less important than semantics.

42
00:01:56,700 --> 00:02:01,220
We are seeing a massive explosion in data volume from IoT sensors, rule time streams and

43
00:02:01,220 --> 00:02:03,220
LLM generated summaries.

44
00:02:03,220 --> 00:02:07,220
If you try to manage that volume with hard coded if then branches, your logic will become

45
00:02:07,220 --> 00:02:08,220
a labyrinth.

46
00:02:08,220 --> 00:02:11,740
You'll end up with flows that have 50 nested conditions and they'll still break the

47
00:02:11,740 --> 00:02:14,100
moment of ender changes their email signature.

48
00:02:14,100 --> 00:02:18,980
The if then model assumes a perfect world where the input will always match the schema.

49
00:02:18,980 --> 00:02:21,060
But in reality, it's the opposite.

50
00:02:21,060 --> 00:02:22,420
The data is the variable.

51
00:02:22,420 --> 00:02:25,860
The logic should be the constant, but our logic has been too rigid to survive the contact

52
00:02:25,860 --> 00:02:27,060
with reality.

53
00:02:27,060 --> 00:02:30,740
We've been building glass houses in a world where it's constantly hailing and then we wonder

54
00:02:30,740 --> 00:02:32,860
why the floor is covered in shards.

55
00:02:32,860 --> 00:02:36,860
The shift to probabilistic design is about building with rubber instead of glass.

56
00:02:36,860 --> 00:02:41,780
It's about creating systems that can absorb the impact of a typo without shattering.

57
00:02:41,780 --> 00:02:46,620
We are moving away from the idea that a system must be 100% right or 100% wrong because

58
00:02:46,620 --> 00:02:48,860
in the real world nothing is 100%.

59
00:02:48,860 --> 00:02:51,820
We've been lying to ourselves about the quality of our data for decades.

60
00:02:51,820 --> 00:02:54,100
We built clean rooms for our logic.

61
00:02:54,100 --> 00:02:57,580
And then we acted surprised when the dirty data from the sales team caused a system wide

62
00:02:57,580 --> 00:02:58,580
failure.

63
00:02:58,580 --> 00:03:00,860
The death of determinism is the death of that lie.

64
00:03:00,860 --> 00:03:04,820
It's an admission that we cannot control the input, so we must modernize the processing.

65
00:03:04,820 --> 00:03:08,940
We are entering the era of the approximate enterprise where we accept that close enough

66
00:03:08,940 --> 00:03:14,780
is actually better than exactly wrong because a system that can process a 90% match is

67
00:03:14,780 --> 00:03:19,540
infinitely more valuable than a system that rejects everything except a 100% match.

68
00:03:19,540 --> 00:03:22,020
This isn't just a technical change, it's a psychological one.

69
00:03:22,020 --> 00:03:26,460
You have to get comfortable with the idea that your system might be pretty sure about

70
00:03:26,460 --> 00:03:27,460
a piece of data.

71
00:03:27,460 --> 00:03:31,460
You have to learn to design for doubt because doubt is the only honest way to handle unstructured

72
00:03:31,460 --> 00:03:32,700
information.

73
00:03:32,700 --> 00:03:36,580
This shift requires us to move from binary triggers to something much more nuanced.

74
00:03:36,580 --> 00:03:39,540
The confidence score, the architecture of uncertainty.

75
00:03:39,540 --> 00:03:44,100
If we accept that the era of binary logic is over, we have to replace it with a new framework.

76
00:03:44,100 --> 00:03:46,220
I call this the symphony orchestra model.

77
00:03:46,220 --> 00:03:50,500
In a traditional IT architecture, everyone is a soloist playing off a rigid score, and

78
00:03:50,500 --> 00:03:53,740
if one person misses a single note, the whole performance stops.

79
00:03:53,740 --> 00:03:57,660
But in a probabilistic architecture, we have deterministic conductors managing probabilistic

80
00:03:57,660 --> 00:03:58,660
musicians.

81
00:03:58,660 --> 00:04:00,020
The conductor is your power-automate flow.

82
00:04:00,020 --> 00:04:03,020
It provides the structure, it sets the tempo, and it defines the boundaries.

83
00:04:03,020 --> 00:04:04,860
The musicians are the AI models.

84
00:04:04,860 --> 00:04:07,340
They are talented, but they are expressive.

85
00:04:07,340 --> 00:04:11,020
And they don't always hit the exact same frequency every time they play.

86
00:04:11,020 --> 00:04:14,080
The conductor's job isn't to force the musicians to be robots.

87
00:04:14,080 --> 00:04:18,280
It's to listen to them and decide if the sound they're making is good enough to continue.

88
00:04:18,280 --> 00:04:22,040
This brings us to the core pillar of modern design, the confidence score.

89
00:04:22,040 --> 00:04:26,000
A confidence score is a value from 0 to 1 that quantifies doubt.

90
00:04:26,000 --> 00:04:29,360
It is the models way of saying, "I think this is a match, and here is how much I'm willing

91
00:04:29,360 --> 00:04:30,860
to bet on it."

92
00:04:30,860 --> 00:04:33,400
In the old world, we had hard-coded paths.

93
00:04:33,400 --> 00:04:36,620
If the email subject contains invoice, then save the attachment.

94
00:04:36,620 --> 00:04:38,880
In the new world, we use confidence-based rooting.

95
00:04:38,880 --> 00:04:42,040
We are making the models internal doubt visible to the system.

96
00:04:42,040 --> 00:04:45,760
We aren't treating the AI as a black box anymore, but instead we are treating it as a witness

97
00:04:45,760 --> 00:04:48,000
with a specific level of credibility.

98
00:04:48,000 --> 00:04:50,280
Here is how you actually build this in the power platform.

99
00:04:50,280 --> 00:04:54,240
When you use AI builder or Azure AI, you don't just get the extracted data.

100
00:04:54,240 --> 00:04:58,120
You get a metadata layer that reveals the probability for every single field.

101
00:04:58,120 --> 00:05:03,200
If the model extracts a total amount of $1,500, it also tells you it's 95% sure about that

102
00:05:03,200 --> 00:05:04,200
number.

103
00:05:04,200 --> 00:05:06,800
As an architect, you now have three distinct zones of action.

104
00:05:06,800 --> 00:05:08,680
First, the green zone.

105
00:05:08,680 --> 00:05:10,680
Confidence is 0.90 or higher.

106
00:05:10,680 --> 00:05:15,320
The risk of error is low, the cost of a mistake is manageable, and the result is that you automate.

107
00:05:15,320 --> 00:05:17,640
The system proceeds without human intervention.

108
00:05:17,640 --> 00:05:19,440
Second, the yellow zone.

109
00:05:19,440 --> 00:05:22,480
Confidence is between 0.60 and 0.89.

110
00:05:22,480 --> 00:05:24,520
The model is pretty sure, but it's not certain.

111
00:05:24,520 --> 00:05:28,000
Maybe the document was slightly blurry, or maybe the handwriting was messy.

112
00:05:28,000 --> 00:05:29,640
In this zone, you flag.

113
00:05:29,640 --> 00:05:32,960
The system roots the data to a human reviewer for a quick check.

114
00:05:32,960 --> 00:05:34,480
Third, the red zone.

115
00:05:34,480 --> 00:05:36,120
Confidence is below 0.60.

116
00:05:36,120 --> 00:05:37,120
The model is guessing.

117
00:05:37,120 --> 00:05:39,360
In this zone, you reject or escalate.

118
00:05:39,360 --> 00:05:42,240
You don't let that data anywhere near your production database.

119
00:05:42,240 --> 00:05:44,560
This is a transparent layer of risk management.

120
00:05:44,560 --> 00:05:47,920
It allows you to scale your automation while maintaining a safety valve.

121
00:05:47,920 --> 00:05:49,800
Think about the implications for governance.

122
00:05:49,800 --> 00:05:51,960
In a deterministic system, errors are silent.

123
00:05:51,960 --> 00:05:55,840
A flow runs, it picks the wrong number because of a typo, and it writes that wrong number

124
00:05:55,840 --> 00:05:56,840
to your ERP.

125
00:05:56,840 --> 00:05:58,840
You don't find out until the end of the quarter.

126
00:05:58,840 --> 00:06:01,040
In a probabilistic system, the error has a signature.

127
00:06:01,040 --> 00:06:04,800
The system knows it's struggling, and it tells you, hey, I processed this, but my confidence

128
00:06:04,800 --> 00:06:06,400
was only 62%.

129
00:06:06,400 --> 00:06:08,120
That is an incredibly powerful signal.

130
00:06:08,120 --> 00:06:09,640
It allows you to build a Doud dashboard.

131
00:06:09,640 --> 00:06:11,520
You can see exactly where your models are failing.

132
00:06:11,520 --> 00:06:14,720
You can see which vendors are sending documents that the AI can't read.

133
00:06:14,720 --> 00:06:18,200
And you can see which process steps are consistently hitting the yellow zone.

134
00:06:18,200 --> 00:06:19,440
You aren't just building a flow.

135
00:06:19,440 --> 00:06:21,000
You're building an intelligence loop.

136
00:06:21,000 --> 00:06:23,920
And this is where Azure AI and AI build a really shine.

137
00:06:23,920 --> 00:06:26,520
They expose these scores at the granular level.

138
00:06:26,520 --> 00:06:30,080
So you can see the confidence for a specific table cell or a specific intent.

139
00:06:30,080 --> 00:06:33,080
You can even use power effects to create dynamic thresholds.

140
00:06:33,080 --> 00:06:37,160
Maybe for a $50 expense report, a 0.70 confidence is fine.

141
00:06:37,160 --> 00:06:41,120
But for a $50,000 purchase order, you want a 0.98.

142
00:06:41,120 --> 00:06:44,520
You are finally aligning your technical logic with your business risk.

143
00:06:44,520 --> 00:06:48,120
We've been treating every data point as if it carries the same weight, but it doesn't.

144
00:06:48,120 --> 00:06:52,160
Probabilistic design allows you to treat a high stakes decision differently than a low stakes

145
00:06:52,160 --> 00:06:54,400
one, even within the same flow.

146
00:06:54,400 --> 00:06:56,880
You are no longer building for a perfect world.

147
00:06:56,880 --> 00:07:00,040
You are building for a world where you know exactly how much uncertainty you are willing

148
00:07:00,040 --> 00:07:01,320
to tolerate.

149
00:07:01,320 --> 00:07:05,600
And once you master the score, you can start looking at the logic that bridges the gaps.

150
00:07:05,600 --> 00:07:11,920
It's not only matching and semantic logic, but how do we actually handle the data that falls into that middle ground of uncertainty?

151
00:07:11,920 --> 00:07:14,440
We start by admitting that exact matching is failing.

152
00:07:14,440 --> 00:07:16,240
It's failing because it lacks context.

153
00:07:16,240 --> 00:07:20,120
In the old model, if a vendor name was off by a single period, the lookup failed.

154
00:07:20,120 --> 00:07:23,800
If you look up doesn't care about your intent, it only cares about the ASCII values.

155
00:07:23,800 --> 00:07:25,560
This is where fuzzy matching comes in.

156
00:07:25,560 --> 00:07:30,440
It's a technique that identifies and links nonidentical records by calculating a similarity score.

157
00:07:30,440 --> 00:07:33,400
And it access the bridge between IBM and IBM.

158
00:07:33,400 --> 00:07:34,960
And the impact isn't theoretical.

159
00:07:34,960 --> 00:07:39,320
The research shows that implementing fuzzy matching can improve your match rates by up to 24%.

160
00:07:39,320 --> 00:07:44,040
That's 24% of your data that would have previously been rejected or sent to a human for manual fixing.

161
00:07:44,040 --> 00:07:47,720
In a high volume environment, that represents thousands of hours of reclaimed time.

162
00:07:47,720 --> 00:07:49,240
But you have to understand the mechanics.

163
00:07:49,240 --> 00:07:51,080
There are two main ways to think about this.

164
00:07:51,080 --> 00:07:52,560
First, character-based logic.

165
00:07:52,560 --> 00:07:54,640
This is often called levenstein distance.

166
00:07:54,640 --> 00:07:57,760
It counts the number of edits needed to turn one string into another.

167
00:07:57,760 --> 00:07:59,000
It's perfect for typos.

168
00:07:59,000 --> 00:08:04,240
If someone writes "managey" a manatee instead of "management", the system sees the missing E and A.

169
00:08:04,240 --> 00:08:07,840
It calculates the distance and it sees they are nearly identical.

170
00:08:07,840 --> 00:08:09,800
Second, we have semantic logic.

171
00:08:09,800 --> 00:08:11,040
This is where things get interesting.

172
00:08:11,040 --> 00:08:12,960
Semantic logic doesn't care about the spelling.

173
00:08:12,960 --> 00:08:14,040
It cares about the meaning.

174
00:08:14,040 --> 00:08:19,640
If one system says "customer" and another says "client", a character-based match will fail every time.

175
00:08:19,640 --> 00:08:22,000
But a semantic model knows they are the same thing.

176
00:08:22,000 --> 00:08:24,920
This is the shift from matching characters to matching concepts.

177
00:08:24,920 --> 00:08:31,160
When you combine confident scores with fuzzy matching, you create a system that can navigate the messy reality of business data.

178
00:08:31,160 --> 00:08:35,000
You are no longer limited by rigid rules that break the moment a human makes a typo.

179
00:08:35,000 --> 00:08:37,440
You are building a system that understands nuance.

180
00:08:37,440 --> 00:08:41,440
And that is how you move from simple automation to true digital intelligence.

181
00:08:41,440 --> 00:08:43,680
Building self-correcting workflows.

182
00:08:43,680 --> 00:08:48,080
Even with the best fuzzy matching in the world, we have to face the reality that AI still hallucinates.

183
00:08:48,080 --> 00:08:50,280
It is just how the technology works right now.

184
00:08:50,280 --> 00:08:55,680
Current data shows that even the most advanced models have hallucination rates between 15 and 20%.

185
00:08:55,680 --> 00:08:59,720
If you are building systems for a business, you cannot ignore a one in five failure rate.

186
00:08:59,720 --> 00:09:02,520
You definitely cannot just prompt engineer your way out of it.

187
00:09:02,520 --> 00:09:06,920
You have to accept that reaching a 0% hallucination rate is not a magic trick.

188
00:09:06,920 --> 00:09:08,720
It is a deliberate architectural choice.

189
00:09:08,720 --> 00:09:13,400
This is where we move past simple automation and start building resilient self-correcting workflows.

190
00:09:13,400 --> 00:09:16,840
We do this by using a structure I call the dual path validation pattern.

191
00:09:16,840 --> 00:09:21,560
It acts as a safeguard that stops the creative imagination of the AI from messing up your actual data.

192
00:09:21,560 --> 00:09:24,080
There is one fundamental rule for the modern architect.

193
00:09:24,080 --> 00:09:26,440
Never let a large language model do the math.

194
00:09:26,440 --> 00:09:30,920
LLMs are engines for soft reasoning, which makes them great at understanding intent

195
00:09:30,920 --> 00:09:34,320
or finding a specific detail in a pile of messy text.

196
00:09:34,320 --> 00:09:37,720
However, they are statistically unreliable when it comes to basic arithmetic.

197
00:09:37,720 --> 00:09:41,560
If you ask an AI to pull three items from an invoice and calculate the total,

198
00:09:41,560 --> 00:09:43,680
you are basically asking for the system to fail.

199
00:09:43,680 --> 00:09:47,600
In the self-correcting design, you use the LLM for the soft reasoning part,

200
00:09:47,600 --> 00:09:49,760
like pulling data out of a messy PDF.

201
00:09:49,760 --> 00:09:54,640
Once you have those numbers, you pass them into a deterministic system to handle the hard constraints.

202
00:09:54,640 --> 00:09:59,200
Your power-automate flow or a basic rules engine should be the thing doing the addition.

203
00:09:59,200 --> 00:10:02,400
It should be the one checking the tax or verifying the currency conversion.

204
00:10:02,400 --> 00:10:07,120
If the LLM claims the total is $500, but your math shows it should be $480,

205
00:10:07,120 --> 00:10:09,720
the system has caught a hallucination in real time.

206
00:10:09,720 --> 00:10:12,320
Instead of crashing, the system triggers a recovery path.

207
00:10:12,320 --> 00:10:16,520
It might try a different prompt or it might flag the error for a person to look at.

208
00:10:16,520 --> 00:10:20,920
This is the difference between a flow that breaks and a flow that actually heals itself.

209
00:10:20,920 --> 00:10:23,920
We are seeing the same shift happen in robotic process automation.

210
00:10:23,920 --> 00:10:27,800
Traditional RPA is very brittle because it depends on exact coordinates on a screen.

211
00:10:27,800 --> 00:10:31,520
If a button moves just three pixels to the left, the whole bot fails.

212
00:10:31,520 --> 00:10:34,640
Self-healing RPA changes that by using probabilistic vision.

213
00:10:34,640 --> 00:10:37,720
The bot does not look for a specific pixel coordinate anymore.

214
00:10:37,720 --> 00:10:41,400
Instead, it looks for the button that looks like a save icon near the top of the page.

215
00:10:41,400 --> 00:10:45,200
When the interface changes, the bot uses context to adapt on the fly.

216
00:10:45,200 --> 00:10:49,320
It recognizes the new layout, updates its own logic and finishes the task.

217
00:10:49,320 --> 00:10:52,960
This creates a resilient loop where the system sees a potential failure

218
00:10:52,960 --> 00:10:56,840
and builds a bridge to the solution before a user even realises there was a problem.

219
00:10:56,840 --> 00:11:00,280
This shift moves our focus from success rates to recovery rates.

220
00:11:00,280 --> 00:11:03,440
We stop trying to make the AI perfect, which is impossible anyway,

221
00:11:03,440 --> 00:11:05,560
and we start making the system unshakable.

222
00:11:05,560 --> 00:11:09,120
It is about building a check and balance layer into every single transaction.

223
00:11:09,120 --> 00:11:12,840
You should treat the AI like a high-speed intern who gets distracted easily.

224
00:11:12,840 --> 00:11:15,520
You let them do the heavy lifting of reading thousands of emails,

225
00:11:15,520 --> 00:11:19,640
but you never let them sign the checks without a senior auditor verifying the signature.

226
00:11:19,640 --> 00:11:22,280
In this case, your deterministic code is that auditor.

227
00:11:22,280 --> 00:11:24,600
By separating the thinking from the calculating,

228
00:11:24,600 --> 00:11:27,880
you build a system that handles the messiness of human language

229
00:11:27,880 --> 00:11:30,040
while keeping the precision of a Swiss watch.

230
00:11:30,040 --> 00:11:34,080
This is how you move AI out of the testing phase and into the core of your business.

231
00:11:34,080 --> 00:11:37,680
You do not trust the model, you trust the architecture you built around it.

232
00:11:37,680 --> 00:11:40,880
You create a safety net so strong that the occasional stumble from the model

233
00:11:40,880 --> 00:11:42,400
does not even count as an error.

234
00:11:42,400 --> 00:11:46,040
It just becomes another part of the self-correction cycle.

235
00:11:46,040 --> 00:11:48,360
This sounds great in a lab, but in production.

236
00:11:48,360 --> 00:11:50,520
The biggest bottleneck isn't the AI.

237
00:11:50,520 --> 00:11:53,360
It's the human, the human in the loop reality.

238
00:11:53,360 --> 00:11:57,400
This sounds great in a lab, but in production, the biggest bottleneck isn't the AI.

239
00:11:57,400 --> 00:11:59,040
It's the human.

240
00:11:59,040 --> 00:12:03,440
We usually talk about automation, like it is a simple choice between a human or a machine.

241
00:12:03,440 --> 00:12:07,000
But the probability shift forces us into a messy middle ground.

242
00:12:07,000 --> 00:12:09,920
Industry benchmarks show a pretty sobering truth.

243
00:12:09,920 --> 00:12:14,520
About 50% of AI workflows still need an expert to review them before they are finished.

244
00:12:14,520 --> 00:12:17,480
This is the main reason we see a pilot to production crisis.

245
00:12:17,480 --> 00:12:20,960
If you look at the numbers, 88% of AI pilots fail within a year.

246
00:12:20,960 --> 00:12:23,120
They do not fail because the tech is broken.

247
00:12:23,120 --> 00:12:28,000
They fail because the people designing them tried to automate 100% of the task way too soon.

248
00:12:28,000 --> 00:12:30,240
They built a system that works on clean test data,

249
00:12:30,240 --> 00:12:33,680
but the logic snaps the moment it hits the friction of a real department.

250
00:12:33,680 --> 00:12:37,200
When you ignore the human element, you create a verification burden.

251
00:12:37,200 --> 00:12:41,040
If your AI saves 10 minutes of data entry, but adds 50 minutes of checking the work

252
00:12:41,040 --> 00:12:44,520
because the team does not trust it, you have actually lost productivity.

253
00:12:44,520 --> 00:12:47,120
You have built a liability instead of an asset.

254
00:12:47,120 --> 00:12:51,600
Successful designs in 2026 will treat verification work as a primary feature.

255
00:12:51,600 --> 00:12:54,320
They do not see human review as a failure of the system.

256
00:12:54,320 --> 00:12:57,160
Instead, they see it as a necessary part of the architecture.

257
00:12:57,160 --> 00:13:01,160
They build the cost of that review into the return on investment from day one.

258
00:13:01,160 --> 00:13:04,160
The secret to making this work is tiered review logic.

259
00:13:04,160 --> 00:13:06,880
This is how you scale things without losing your mind.

260
00:13:06,880 --> 00:13:09,280
High confidence transactions happen silently.

261
00:13:09,280 --> 00:13:13,960
If the model is 98% sure and the math is right, the system just moves forward.

262
00:13:13,960 --> 00:13:16,760
But low confidence cases like that yellow zone we talked about

263
00:13:16,760 --> 00:13:18,160
will trigger an escalation.

264
00:13:18,160 --> 00:13:19,840
This changes the human role entirely.

265
00:13:19,840 --> 00:13:22,800
We are moving away from data entry and toward data auditing.

266
00:13:22,800 --> 00:13:26,760
In the old model, the human was the engine because they did all the typing and clicking.

267
00:13:26,760 --> 00:13:28,360
In the new model, the human is the governor.

268
00:13:28,360 --> 00:13:33,440
They handle the exceptions and provide the nuance that a probabilistic model cannot quite understand.

269
00:13:33,440 --> 00:13:34,880
Think about how this feels for the team.

270
00:13:34,880 --> 00:13:38,440
When people know they only have to look at the 5% of cases that are actually difficult,

271
00:13:38,440 --> 00:13:40,160
they stay much more engaged.

272
00:13:40,160 --> 00:13:42,080
They are not bored by the routine anymore.

273
00:13:42,080 --> 00:13:44,520
They are challenged by the complexity of the hard cases.

274
00:13:44,520 --> 00:13:47,360
But you have to build the right interface for this to work.

275
00:13:47,360 --> 00:13:50,080
You cannot just send an email saying something went wrong.

276
00:13:50,080 --> 00:13:52,120
You need to show a side by side comparison.

277
00:13:52,120 --> 00:13:55,560
Show the person the source document, show what the AI extracted,

278
00:13:55,560 --> 00:13:58,800
and highlight the exact field where the confidence score dropped.

279
00:13:58,800 --> 00:14:03,000
If you want to scale, you have to stop acting like human review means your automation is weak.

280
00:14:03,000 --> 00:14:04,280
It is actually the opposite.

281
00:14:04,280 --> 00:14:10,160
A well-integrated human is what makes a probabilistic system strong enough to survive in the real world.

282
00:14:10,160 --> 00:14:13,680
It is the final layer of your self-correcting architecture.

283
00:14:13,680 --> 00:14:15,040
The Agentec Future.

284
00:14:15,040 --> 00:14:18,160
We are looking toward a new horizon called the Agentec Enterprise.

285
00:14:18,160 --> 00:14:21,320
This is the logical endpoint for everything we have been discussing today.

286
00:14:21,320 --> 00:14:25,160
We are moving into a world where agents are no longer just simple scripts or basic flows

287
00:14:25,160 --> 00:14:28,000
because they are becoming first-class digital workers.

288
00:14:28,000 --> 00:14:30,800
The power platform is changing right in front of us.

289
00:14:30,800 --> 00:14:34,680
It is shifting from a standard app builder into a full orchestration hub.

290
00:14:34,680 --> 00:14:38,280
In the old model, you build a specific app to solve a specific problem.

291
00:14:38,280 --> 00:14:42,720
But in the agentec model, you build an environment where agents solve those problems for you.

292
00:14:42,720 --> 00:14:45,440
These agents do not just follow a static list of steps.

293
00:14:45,440 --> 00:14:49,040
They observe their environment, evaluate their own confidence levels, and decide which

294
00:14:49,040 --> 00:14:51,040
tool is actually the right one to use next.

295
00:14:51,040 --> 00:14:54,880
By 2026, governance will look nothing like it did two years ago.

296
00:14:54,880 --> 00:14:59,160
Tools like Perview and Sentinel are moving toward probabilistic anomaly detection.

297
00:14:59,160 --> 00:15:01,680
They are not just searching for a leaked password anymore.

298
00:15:01,680 --> 00:15:05,920
Instead, they are looking for behavioral patterns that simply do not look right.

299
00:15:05,920 --> 00:15:09,680
They are applying the same similarity logic we use for data extraction to the security

300
00:15:09,680 --> 00:15:11,080
of your entire tenant.

301
00:15:11,080 --> 00:15:12,520
This shift is permanent.

302
00:15:12,520 --> 00:15:16,320
People who stick to rigid if-then-logic will be left managing broken systems and frustrated

303
00:15:16,320 --> 00:15:17,320
teams.

304
00:15:17,320 --> 00:15:20,920
They will be stuck in a cycle of constant maintenance while trying to patch up a glass house

305
00:15:20,920 --> 00:15:22,920
that was never built for this kind of weather.

306
00:15:22,920 --> 00:15:25,640
The agentec enterprise starts with a single realization.

307
00:15:25,640 --> 00:15:28,960
We do not need to control every single variable to get a reliable outcome.

308
00:15:28,960 --> 00:15:30,800
We just need to control the architecture.

309
00:15:30,800 --> 00:15:35,000
Our job is to define the thresholds and set the guardrails so the agents can work within

310
00:15:35,000 --> 00:15:36,160
those boundaries.

311
00:15:36,160 --> 00:15:38,600
In this world, your business logic becomes elastic.

312
00:15:38,600 --> 00:15:42,000
It stretches to handle a new vendor or a different document format without needing a

313
00:15:42,000 --> 00:15:44,480
developer to rewrite a single line of code.

314
00:15:44,480 --> 00:15:46,040
This is the real promise of low code.

315
00:15:46,040 --> 00:15:49,560
It is the ability to move at the speed of thought rather than the speed of syntax.

316
00:15:49,560 --> 00:15:51,200
But to get there, you have to start today.

317
00:15:51,200 --> 00:15:54,720
You have to stop building for a perfect world and start building for the messy one we actually

318
00:15:54,720 --> 00:15:55,720
live in.

319
00:15:55,720 --> 00:15:58,000
You have to embrace the doubt.

320
00:15:58,000 --> 00:15:59,000
Implementation Challenge

321
00:15:59,000 --> 00:16:00,320
Here is your homework for the week.

322
00:16:00,320 --> 00:16:02,640
I want you to find your highest failure float.

323
00:16:02,640 --> 00:16:06,160
Do not pick the most complex one, but find the one that crashes most often because of

324
00:16:06,160 --> 00:16:07,160
bad data.

325
00:16:07,160 --> 00:16:08,280
The goal is not to fix the data.

326
00:16:08,280 --> 00:16:11,920
You cannot control what your users or your vendors decide to send you.

327
00:16:11,920 --> 00:16:15,400
Instead of fighting that reality, I want you to add a confidence threshold.

328
00:16:15,400 --> 00:16:19,480
Go into your AI builder or your LLM action and expose that probability score.

329
00:16:19,480 --> 00:16:20,480
Create a simple condition.

330
00:16:20,480 --> 00:16:24,640
If the confidence score is below 85%, do not let the flow crash.

331
00:16:24,640 --> 00:16:28,680
Route that specific case to a review folder or a share point list for a manual check.

332
00:16:28,680 --> 00:16:30,640
Watch what happens to your success rate.

333
00:16:30,640 --> 00:16:34,400
You will see the stress leave your system the moment you stop pretending that every input

334
00:16:34,400 --> 00:16:35,400
will be perfect.

335
00:16:35,400 --> 00:16:39,240
If the shift in thinking changed your approach, follow me, Mercupieters, on LinkedIn, I am

336
00:16:39,240 --> 00:16:42,400
constantly sharing deep dives into these architectural patterns.

337
00:16:42,400 --> 00:16:45,400
Share this with your team, especially if your flows are breaking right now.

338
00:16:45,400 --> 00:16:48,120
We need to stop building with glass and start building with intent.

339
00:16:48,120 --> 00:16:52,040
That is the technique and it cuts your downtime in half once you start practicing it.

340
00:16:52,040 --> 00:16:56,000
If you want the advanced version that the pros use, check out this video next.

341
00:16:56,000 --> 00:16:57,760
Subscribe for more tutorials.

342
00:16:57,760 --> 00:17:01,800
Stop fighting the messy reality of your data and start designing for it instead.

343
00:17:01,800 --> 00:17:05,680
When you use confidence scores, you transform brittle flows into resilient systems that

344
00:17:05,680 --> 00:17:06,920
actually work.

345
00:17:06,920 --> 00:17:07,920
But here is the thing.

346
00:17:07,920 --> 00:17:10,120
The model only works if you build it correctly.

347
00:17:10,120 --> 00:17:14,280
If you want to see exactly how to set this up in co-pilot studio, watch this next video

348
00:17:14,280 --> 00:17:15,280
right now.

349
00:17:15,280 --> 00:17:19,280
And if this changed how you think about AI, follow me on LinkedIn or subscribe to Master