WEBVTT
1
00:00:00.171 --> 00:00:03.253
Agentic AI, you've heard the term, but what is it really?
2
00:00:03.493 --> 00:00:06.495
And how is it already changing the way that energy companies are working?
3
00:00:07.075 --> 00:00:13.239
Our guest today on Power Perspectives is AI strategist, Jeff Wilser, and he's here to answer those questions for us.
4
00:00:13.559 --> 00:00:15.020
Jeff, welcome to the show today.
5
00:00:21.913 --> 00:00:23.377
Jeff, let's start with the basics.
6
00:00:23.477 --> 00:00:27.105
When you say agentic AI workflow, what exactly does that mean?
7
00:00:27.145 --> 00:00:31.796
And how is that different than say, someone at a utility company using a chat bot to ask a question?
8
00:00:32.605 --> 00:00:32.965
Sure.
9
00:00:32.985 --> 00:00:36.068
I know you've had a lot of recent great episodes on AI.
10
00:00:36.128 --> 00:00:38.770
So I'm sure many of your listeners have some basics here.
11
00:00:38.890 --> 00:00:50.618
But for those who are new to the agent world, I think when most people think of something like ChatGPT or Chatbot, it's mostly about just asking a question and getting a response, right?
12
00:00:50.718 --> 00:00:56.662
So simple analogy, hey, I want to go to Tokyo for a vacation.
13
00:00:57.163 --> 00:00:57.523
What do you
14
00:01:00.445 --> 00:01:02.588
would actually do the things for you.
15
00:01:02.748 --> 00:01:09.476
So instead of just getting an answer, like here is a good itinerary in Tokyo, you can say, hey, I want to buy plane tickets.
16
00:01:09.696 --> 00:01:12.379
Please buy me plane tickets, book me hotels.
17
00:01:12.860 --> 00:01:18.567
So you can think of it as actually the arms and legs of doing things, not just the brain itself.
18
00:01:19.063 --> 00:01:19.363
Gotcha.
19
00:01:19.383 --> 00:01:20.643
No, that's definitely helpful.
20
00:01:20.703 --> 00:01:22.584
It's a little bit more action-based.
21
00:01:23.004 --> 00:01:33.246
And for our audience that's in the utility and energy world, what's an example of something that maybe one of their employees spends hours doing today that an AI agent could realistically handle?
22
00:01:33.786 --> 00:01:34.367
Absolutely.
23
00:01:34.547 --> 00:01:39.988
And Matt, I think a lot of folks hear about AI agents and they immediately race to
24
00:01:40.888 --> 00:01:46.452
the most provocative, scary, crazy thing that can be done.
25
00:01:46.572 --> 00:01:51.335
Like some agents who has full autonomy of a control room.
26
00:01:51.915 --> 00:02:01.702
But there is a lot of middle grounds and a lot of far less glamorous, more basic things that can be done before you get to that point.
27
00:02:02.202 --> 00:02:09.469
I think the number one thing, and this is true, so many industries and top of utilities and energy is paperwork.
28
00:02:09.969 --> 00:02:16.816
Virtually every industry, including you guys, we all have more paperwork than we want to do.
29
00:02:17.516 --> 00:02:34.165
And that ranging from compliance documents to trying to find what your rates were in some niche area from 12 years ago, there are buckets and buckets of paperwork that no one really has time to do.
30
00:02:34.745 --> 00:02:41.969
And AI agents can be built to actually sift through the paperwork and summarize and actually draft documents.
31
00:02:42.189 --> 00:02:46.151
So kind of the two-year-ago and one-year-ago AI thing was,
32
00:02:46.671 --> 00:02:57.779
was to manually, let's say, feed a PDF into something like ChatDBT or Claude or Copilot and say, hey, summarize this for me.
33
00:02:58.520 --> 00:03:11.950
Now that could be done with far more efficiency where agents can actually romp through stacks and stacks of paperwork and glean insights and help you create summary documents.
34
00:03:12.697 --> 00:03:16.278
And is the benefit just that it's able to do it that much quicker and more efficiently?
35
00:03:16.358 --> 00:03:19.699
Or do you find that the AI agent can actually do it better too?
36
00:03:20.319 --> 00:03:35.922
So in your world, think about a compliance issue that no one analyst could possibly have time to read the hundreds of precedents and documents behind this, but an agent can go through.
37
00:03:36.602 --> 00:03:46.840
And not just do it faster, but have access to this sprawling amount of paperwork to turn up insights you wouldn't otherwise have known to even look for.
38
00:03:47.262 --> 00:03:52.365
Yeah, that's definitely, I think, something that the utility executives would be eager to learn more about.
39
00:03:52.626 --> 00:03:57.629
And you mentioned that people jumped to the idea of like, oh, the AI is running the control room.
40
00:03:57.649 --> 00:03:59.790
And then there's a little bit of an aversion to that.
41
00:04:00.090 --> 00:04:05.074
But even when you're just looking at the paperwork and the regulatory aspect of it, it's still important.
42
00:04:05.194 --> 00:04:14.380
And I'm curious how much you find that there's oversight or fact checking still needed from the human side of the equation when the AI agents are doing these tasks.
43
00:04:14.832 --> 00:04:24.035
For sure, there's a cliche that I'm sure you've heard, but it's one of those cliches that happens to be true, and that's keeping a human in the loop.
44
00:04:24.395 --> 00:04:27.877
And that can happen in a couple of different levels, right?
45
00:04:27.917 --> 00:04:32.498
You could have the human spot check and look for errors.
46
00:04:33.178 --> 00:04:37.580
Obviously, you want someone to be sanity checking it before it goes anywhere sensitive.
47
00:04:38.200 --> 00:04:45.663
But one thing I think that one thing I share when I work with companies and workshops is, yes, human loop.
48
00:04:45.743 --> 00:04:53.787
But one additional thing you can do to help with control and help guardrails is actually using AI agents and AI models.
49
00:04:54.327 --> 00:04:59.833
kind of to cross-check each other in an adversarial, constructive way, right?
50
00:04:59.873 --> 00:05:13.649
So imagine that you, to keep it very simple, imagine you're working with something like Claude to create some agentic workflow and your agent says, aha, I am now going to create this compliance document.
51
00:05:14.310 --> 00:05:17.471
You can then feed it to a different system entirely.
52
00:05:17.671 --> 00:05:32.277
Let's say in ChachiPT, give it the rules of the road, give it the context and say, hey, I want you with fresh eyes and to get the nice splash cold water here, cold water, I want you to evaluate this and find blind spots.
53
00:05:32.777 --> 00:05:38.460
So by using AI system to check the other, that can help actually determine
54
00:05:39.240 --> 00:05:51.515
reduce errors and give yourself more of a more um again i wouldn't use the word certainty a human should always be in a loop but it's helpful to have ais check each other and that can reduce errors as well
55
00:05:52.713 --> 00:06:00.478
Is that one of the go to solutions that we hear about hallucinations or just, you know, faulty conclusions sometimes come from the AI?
56
00:06:00.518 --> 00:06:05.281
And, you know, the natural inclination is to have the humans be the one checking that, like you said, spot checking.
57
00:06:05.341 --> 00:06:12.686
But do you find that it's a better practice or almost time saver to have that, you know, AI checking its own work?
58
00:06:13.138 --> 00:06:13.839
For sure.
59
00:06:14.059 --> 00:06:20.083
It's, again, not to take the human out of this, but it's like one additional level of protection before the human, right?
60
00:06:20.303 --> 00:06:31.230
And it's also, it's kind of a lot of fun too when you have these two different models and you almost surprise them and say, hey, by the way, I asked Claude the same question.
61
00:06:31.830 --> 00:06:47.296
here's what it found, and you give it its answer, that will almost, is a way to spur the first model to make sure it's buttoned up, and to actually second and triple and quadruple check itself.
62
00:06:47.756 --> 00:06:59.065
I know in the utility sector, we're in an environment where these companies are seeing massive liabilities with things like wildfire management, outages that impact vulnerable customers.
63
00:07:00.026 --> 00:07:03.929
The utility leaders don't really feel like they can afford to let a mistake through.
64
00:07:04.189 --> 00:07:08.492
And this is great advice for how to capture and minimize those mistakes.
65
00:07:08.592 --> 00:07:14.256
But on the chance that there is an AI agent involved and a mistake does get through, you
66
00:07:15.017 --> 00:07:18.038
Who ultimately is held accountable when something goes wrong?
67
00:07:18.478 --> 00:07:24.900
There's one group that is definitely without question will benefit from AI, and that is lawyers.
68
00:07:25.280 --> 00:07:30.102
There are going to be so many gray area legal questions about who ultimately is accountable.
69
00:07:30.662 --> 00:07:34.663
That is a legal question beyond my pay grade as far as the actual liability.
70
00:07:35.123 --> 00:07:40.365
That said, I think the North Star, the way I view all of this for your industry,
71
00:07:41.986 --> 00:07:47.850
is something that I call the red line theory and where you are on the red line.
72
00:07:48.571 --> 00:07:59.999
And what I mean by that is what I found is the most successful way companies can use AI, whether in utilities or in publishing or in healthcare,
73
00:08:00.719 --> 00:08:14.113
is to think clearly about where is the red line of what you hold sacred that AI should not touch, that humans, to your point, Matt, ultimately must be accountable for.
74
00:08:14.634 --> 00:08:22.182
So for one example, in media and editing and journalism, Time magazine, they said, okay,
75
00:08:22.662 --> 00:08:28.927
We're not going to use AI to write AI slop to write our articles and do our research.
76
00:08:29.548 --> 00:08:31.809
That is still the work of humans.
77
00:08:32.650 --> 00:08:41.678
But the chief operating officer of Time, who I interviewed recently, what he told me was, yes, the AI is not going to replace the human writers.
78
00:08:42.258 --> 00:08:43.499
However, that's a red line.
79
00:08:43.539 --> 00:08:45.060
That's what time is sacred for.
80
00:08:45.140 --> 00:08:48.263
That is their sacred editorial judgment and credibility.
81
00:08:48.803 --> 00:08:50.405
On the other side of the red line,
82
00:08:50.985 --> 00:08:57.610
They want to use AI ambitiously, and they created AI agents to do far more things that cannot be done with time.
83
00:08:57.750 --> 00:09:07.717
Now they have AI agents and AI bots that let you interact with Time Magazine and explore its 103 year history in ways you couldn't do before.
84
00:09:07.797 --> 00:09:09.098
They added new functionality.
85
00:09:09.378 --> 00:09:19.826
The goal is not to take humans out of the loop of using judgment with respect to the wildfires and all the tricky questions you're facing.
86
00:09:20.326 --> 00:09:34.109
The goal is to give humans more time to think deeply about edge cases, about the sensitive issues, by automating the more drudgery stuff, the mundane stuff, to give humans more time to do what they are really good at.
87
00:09:34.732 --> 00:09:35.712
Yeah, that's interesting.
88
00:09:35.732 --> 00:09:45.177
It makes me think of, we ran a survey a little while back asking utility executives, what do they think AI will not replace in the utility sector?
89
00:09:45.277 --> 00:09:52.800
And the main tranche of answers came into, one, there's the physical work of repairing a line and that much was obvious.
90
00:09:52.820 --> 00:09:54.060
Oh, the robots are coming.
91
00:09:54.120 --> 00:09:56.361
I'm sure the robots are coming imminently and they'll do all of that.
92
00:09:56.381 --> 00:09:57.262
Eventually, yeah.
93
00:09:58.342 --> 00:10:06.067
But the other aspects that were cited multiple times were empathy for the customers and, you know, having kind of that human connection.
94
00:10:06.207 --> 00:10:12.271
And then the sort of judgment that leaders have to make based on all the information that AI can help collect for them.
95
00:10:12.631 --> 00:10:18.635
But ultimately, a human has to make, you know, weigh the pros and cons and the different tradeoffs and make that ultimate decision.
96
00:10:19.221 --> 00:10:20.462
Yeah, well said.
97
00:10:20.642 --> 00:10:25.066
And along those lines, think about flying and aviation.
98
00:10:26.027 --> 00:10:27.048
That's similar to your world, right?
99
00:10:27.068 --> 00:10:28.589
The stakes just as utilities.
100
00:10:29.030 --> 00:10:32.593
When you're flying a plane, human lives are quite literally at stake.
101
00:10:33.233 --> 00:10:41.140
When autopilots, when that technology developed years and decades ago,
102
00:10:41.901 --> 00:10:44.063
pilots were understandably skittish about it.
103
00:10:44.223 --> 00:10:47.146
They did not want to trust flying to a computer.
104
00:10:47.286 --> 00:10:47.966
Like, fair.
105
00:10:48.086 --> 00:10:49.548
That is an understandable reaction.
106
00:10:50.108 --> 00:10:55.573
They also worried they could be replaced by robots and by self-flying planes.
107
00:10:56.234 --> 00:11:02.259
And what happened was not that pilots were replaced or that pilots stopped using their judgment.
108
00:11:02.459 --> 00:11:05.622
What happened instead was pilots, they still had to know how to fly, but
109
00:11:06.102 --> 00:11:16.895
They still made the judgment calls, but they're actually able to make their role more about the judgments as opposed to doing more of the manual things.
110
00:11:16.915 --> 00:11:20.139
I think utilities can be looked at the same way, right?
111
00:11:20.339 --> 00:11:23.083
Ultimately, it's the human's job to apply the judgments.
112
00:11:23.663 --> 00:11:27.266
And if we can automate some systems to make that job easier, great.
113
00:11:27.766 --> 00:11:38.434
But I would also encourage utilities to think about it from the lens, not just of automating and cutting, but even growing.
114
00:11:38.794 --> 00:11:42.497
And by growing, I don't necessarily mean doubling revenue.
115
00:11:43.077 --> 00:11:48.782
But rather, what else could you do that you couldn't do beforehand?
116
00:11:48.902 --> 00:12:07.438
That's a big light bulb I see when I work with CEOs about this across industries is most companies and most organizations are thinking very narrowly about, okay, maybe I can use AI to cut this one thing or to automate this.
117
00:12:08.178 --> 00:12:13.982
And sure, automation is great, but it's also just scratching the surface of what you can do.
118
00:12:14.302 --> 00:12:28.392
I'd encourage everyone listening to your podcast to do is to take a moment and imagine, let's say theoretically, your headcount has doubled without spending an extra nickel.
119
00:12:29.492 --> 00:12:31.735
What then would you be able to do?
120
00:12:32.455 --> 00:12:34.718
What additional safeguards could you create?
121
00:12:35.358 --> 00:12:36.539
What additional services?
122
00:12:37.360 --> 00:12:40.964
Where are you able to add more value if your head got doubled?
123
00:12:41.545 --> 00:12:49.673
And then think, okay, how can AI allow us to actually get some of that new resource and capability?
124
00:12:49.693 --> 00:12:49.953
Okay.
125
00:12:50.374 --> 00:13:03.330
So I think that's one way that people in utilities can use AI, not just for like saving off a bit of cost at the margin, but to more ambitiously get more done in your organization.
126
00:13:03.830 --> 00:13:07.154
Yeah, that's a terrific thought experiment and how to approach it.
127
00:13:07.275 --> 00:13:07.475
Yeah.
128
00:13:07.815 --> 00:13:18.465
kind of going back to the treachery of the paperwork we were talking about earlier, you know, I know a longstanding challenge in the power sector is siloing, you know, information processes.
129
00:13:18.786 --> 00:13:23.250
They're always separated between different departments and the right hand doesn't always know what the left hand is doing.
130
00:13:23.730 --> 00:13:26.093
So given that part of that problem is that utilities have
131
00:13:26.733 --> 00:13:31.516
So much siloed data buried in PDFs, filings, engineering documents, old records.
132
00:13:32.036 --> 00:13:39.980
How close do you think we are to a world where, you know, through an agentic AI, an employee can essentially have a conversation with all of that institutional history?
133
00:13:40.983 --> 00:13:42.304
I think it's as close as you want to be.
134
00:13:42.704 --> 00:13:50.408
The technology exists right now, and it's not something that requires billions of dollars of investment.
135
00:13:50.588 --> 00:13:51.729
It's mostly about that.
136
00:13:51.749 --> 00:13:55.831
The irony, Matt, is I think this is one thing that's kind of lost in the conversation.
137
00:13:55.851 --> 00:13:58.853
There's so much AI hype out there and all these buzzwords.
138
00:13:59.773 --> 00:14:05.418
Even the word agentic AI can sound kind of magic wandy.
139
00:14:05.738 --> 00:14:12.444
Ultimately, an AI agent is mostly about context and context files.
140
00:14:13.104 --> 00:14:19.489
And by context files, kind of fancy way of saying really just like, it could be as simple as text documents.
141
00:14:19.510 --> 00:14:20.851
You mentioned PDFs, right?
142
00:14:21.011 --> 00:14:28.217
And so what's making or breaking AI adoption, both in the power industry and most industries at large,
143
00:14:28.897 --> 00:14:34.622
is not about are you using the most cutting edge, powerful frontier model.
144
00:14:35.243 --> 00:14:41.048
It's less about the model's power and more about how well are you feeding it context.
145
00:14:41.568 --> 00:14:48.815
What I mean by that is, as you point out, Matt, there could be PDFs living in, there's somewhere in IT, there's all kinds of PDFs.
146
00:14:48.835 --> 00:14:50.957
Somewhere in procurement, in marketing.
147
00:14:52.398 --> 00:14:59.825
Every different division has their own weird warehouse figuratively or literally of all these different files.
148
00:15:00.946 --> 00:15:13.398
Getting those files, having a smart infrastructure that is feeding the relevance files and takeaways from the files into a system that AI agents read is,
149
00:15:14.239 --> 00:15:15.220
That's the whole ballgame.
150
00:15:15.620 --> 00:15:25.047
The other area where, you know, utility leaders traditionally see knowledge spread out is where it's not written down at all, but it's stored in the minds of longstanding employees.
151
00:15:25.107 --> 00:15:31.892
And this is a common topic in the utility sector as much of the workforce is aging closer and closer to retirement.
152
00:15:32.252 --> 00:15:33.914
Are we risking losing that knowledge?
153
00:15:34.294 --> 00:15:40.899
I'm curious what you think AI agentic tools can do to capture that institutional knowledge and maybe re-reporting.
154
00:15:41.199 --> 00:15:49.366
plug in that gap where, yeah, hey, there might have been that 40, that employee who's been there for 40 years, who knows, oh, this is how this substation really behaves.
155
00:15:49.406 --> 00:15:54.450
How can we set up the agentic AI to be able to understand that as well?
156
00:15:55.000 --> 00:15:56.381
Yeah, I'm so glad you brought that up.
157
00:15:56.401 --> 00:15:57.142
That's a great one.
158
00:15:57.182 --> 00:15:58.002
I think you're right.
159
00:15:58.102 --> 00:16:01.024
It's super relevant to the industry here.
160
00:16:01.305 --> 00:16:05.548
And some might call it tribal knowledge that is a danger of being lost.
161
00:16:06.188 --> 00:16:11.952
And I think it's one often overlooked use case of AI agents.
162
00:16:12.233 --> 00:16:14.074
And one of my favorites is technology.
163
00:16:14.374 --> 00:16:21.779
interviews and actually having... Now, this could be done a couple of ways.
164
00:16:22.319 --> 00:16:36.127
You can set up an interview for almost like a series of prompts that this salty engineer about to retire can then answer, crucially, answer verbally, right?
165
00:16:36.307 --> 00:16:36.808
If it's...
166
00:16:37.388 --> 00:16:41.890
If we're asking for type response, that's not going to be nearly as effective.
167
00:16:42.070 --> 00:16:45.292
No one likes to type a 5,000-word essay, right?
168
00:16:45.872 --> 00:16:51.595
But if you ask them questions about things that, frankly, they'll probably be kind of fluttered.
169
00:16:51.635 --> 00:16:52.916
They're asked about, like, oh...
170
00:16:53.636 --> 00:17:00.142
Yeah, I recall well that ice storm of 1987, and no one's thought to ask me that in years, right?
171
00:17:00.542 --> 00:17:12.032
And have them go through and give verbally detailed responses that can create this library, this corpus of knowledge.
172
00:17:12.292 --> 00:17:14.914
And here's where agents really come into play.
173
00:17:16.035 --> 00:17:25.359
In a world before AI, if you were to, let's say, spend, and I'm making this roadmap up, but bear with me here.
174
00:17:25.819 --> 00:17:34.782
Let's say you have maybe two hours a week you devote to interviewing this, the salty veteran engineering knowledge, right?
175
00:17:34.802 --> 00:17:43.206
So over the course of a year, you have a hundred hours of them talking about all kinds of esoteric things they've seen.
176
00:17:44.126 --> 00:17:44.967
And do that at scale.
177
00:17:45.267 --> 00:17:51.612
So you've now interviewed, let's say, let's say there's thousands of hours of conversation of these folks.
178
00:17:52.213 --> 00:17:57.978
In a non-AI world, good luck getting anyone to go through that.
179
00:17:58.958 --> 00:18:06.104
It reminds me of the last scene of Raiders of the Lost Ark with the crates and crates and crates in the warehouse and all these things.
180
00:18:06.645 --> 00:18:10.088
No one can go through those crates in the warehouse
181
00:18:10.728 --> 00:18:11.409
and make sense of this.
182
00:18:11.669 --> 00:18:25.900
The agents can go through this, at this point, this massive library of interviews you've conducted with all these folks who are about to retire and make heads and tails of it.
183
00:18:26.180 --> 00:18:29.043
It could synthesize things.
184
00:18:29.423 --> 00:18:32.325
You can actually have a few different cuts of this library.
185
00:18:32.345 --> 00:18:37.770
You can say, hey, I want to create a masterclass about ice storms
186
00:18:38.390 --> 00:18:43.296
or about substations and how they reacted in this kind of emergency.
187
00:18:43.916 --> 00:18:51.245
And you can create an entire documents and dossier around these thousands of interviews.
188
00:18:51.946 --> 00:18:58.893
And then once you have that effectively, this massive database of tribal knowledge in one place,
189
00:18:59.514 --> 00:19:05.437
In the future, you could ask AI agents to search into it, to probe, to ask questions.
190
00:19:05.897 --> 00:19:18.423
A more advanced thing would that mean to create an actual, you can create effectively your own chatbot based off of this database to let you real-time ask questions to it as well.
191
00:19:18.503 --> 00:19:22.305
So I think we're going to see, that's still relatively fast,
192
00:19:22.865 --> 00:19:26.912
fringe and bleeding edge use case right now.
193
00:19:27.132 --> 00:19:34.064
In the future, my prediction, that sort of use case will go mainstream across many industries.
194
00:19:35.047 --> 00:19:36.567
Yeah, that's fascinating.
195
00:19:36.607 --> 00:19:39.708
And I definitely think that the utility sector is primed for that.
196
00:19:39.748 --> 00:19:45.770
Like I said, the aging workforce, this long institutional knowledge, it's too valuable to just let it go.
197
00:19:45.870 --> 00:19:52.072
So maybe the AI agents are coming at the perfect time to fill in that gap that's so needed.
198
00:19:52.472 --> 00:19:58.094
But I guess as we wrap up our conversation, Jeff, I want to ask the advice-based question.
199
00:19:58.314 --> 00:19:58.354
If
200
00:19:59.534 --> 00:20:06.740
somebody's listening in, they're running utility, they want to experiment with agentic AI this week, not two years from now, not those fringe cases.
201
00:20:07.140 --> 00:20:12.645
What is the first workflow you would tell them to try and how would they know whether it's actually working?
202
00:20:13.573 --> 00:20:24.059
One thing that, and if your listeners already tried this, great, move on to the next, but I think something that everyone busy in the utility industry can deal with.
203
00:20:24.080 --> 00:20:32.385
But before you think about how can we inject AI in most sensitive workflows of our power grid,
204
00:20:33.445 --> 00:20:35.926
Whether we like it or not, we all have busy inboxes.
205
00:20:36.207 --> 00:20:39.648
We all struggle to make the meetings on time.
206
00:20:39.889 --> 00:20:42.230
We all have a hard time keeping track of our to-dos.
207
00:20:43.150 --> 00:20:55.557
I would create your own chief of staff, your own AI agent who is able to triage email for you, who is able to hoover up all the projects you have.
208
00:20:56.598 --> 00:21:02.621
Most systems now let you easily connect in your email company Slack.
209
00:21:03.401 --> 00:21:05.082
to-do list systems, and so on.
210
00:21:05.723 --> 00:21:11.146
And crucially, my advice to even be more specific is to give it lots of context.
211
00:21:12.086 --> 00:21:24.754
Spend an hour walking through what's important to you, what your goals are, what are risks you're worried about, and have it then each morning,
212
00:21:26.795 --> 00:21:34.058
give you a briefing for the day and say, hey, what are the three most things I must do today?
213
00:21:34.258 --> 00:21:35.399
What's most important?
214
00:21:36.099 --> 00:21:44.103
This, I think that's a, one reason I like this starting point so much is that one, it is very de-risking, right?
215
00:21:44.123 --> 00:21:46.043
You are not putting the grid at risk.
216
00:21:46.444 --> 00:21:49.365
You are starting small with your own personal workflows.
217
00:21:50.105 --> 00:21:53.587
And two, I think most people will get a ton of value out of it.
218
00:21:53.627 --> 00:22:08.673
When I work with CEOs and show this in workshops, teams are like, oh my God, it is amazing how a little bit of agentic systems in your own workflow can handle so much busy work.
219
00:22:08.813 --> 00:22:18.717
And one more bit of very practical advice that I think listeners will benefit from is when you're working with AI,
220
00:22:19.497 --> 00:22:20.137
in this manner.
221
00:22:20.537 --> 00:22:26.899
The way to speed things up and the way to give it even more context is actually through voice, not through typing.
222
00:22:27.459 --> 00:22:35.441
Reason why is that almost everyone, certainly including me, despite typing for a living, is much faster at talking than typing.
223
00:22:35.741 --> 00:22:37.741
It's also more enjoyable to talk than to type.
224
00:22:38.361 --> 00:22:44.723
And so when we're going back to when we interviewed the salty 35-year veteran about to retire,
225
00:22:45.423 --> 00:22:52.346
they're going to be more likely to share lots of rich details if they're just casually, conversationally chatting.
226
00:22:52.526 --> 00:22:58.089
If they're typing, they cut corners, but the corners are where the good stuff lives.
227
00:22:58.909 --> 00:23:00.490
The corners are the edge cases.
228
00:23:00.830 --> 00:23:01.571
That's the nuance.
229
00:23:02.031 --> 00:23:03.932
The more nuance you give, the better.
230
00:23:04.072 --> 00:23:08.034
When you're talking, you're more likely to just share a random parenthetical.
231
00:23:08.474 --> 00:23:13.818
the AI will actually store that parenthetical and that's useful context in the future, right?
232
00:23:13.959 --> 00:23:27.610
And so, and last, not to be overly mechanical here, but when you hear voice mode, people might be tempted to actually turn on the kind of like voice mode in ChachiPT or Claude, or mainly ChachiPT.
233
00:23:28.250 --> 00:23:37.278
That's not what I mean, because that, when you go back and forth with voice, almost like Siri, and we all know how bad Siri is, you kind of only get the lightest,
234
00:23:37.938 --> 00:23:40.600
most surface level version of the AI.
235
00:23:40.720 --> 00:23:44.363
It's not the more powerful thinking versions of the model.
236
00:23:44.963 --> 00:23:48.946
Instead, what I mean is you can use a dictation service.
237
00:23:49.347 --> 00:24:00.015
And for those who are still not sure exactly where to start, literally after you finish listening, Matt, to your great podcast, they can go in and say, hey, chatbot of choice.
238
00:24:00.755 --> 00:24:04.338
I want to create my own chief of staff AI agents.
239
00:24:04.898 --> 00:24:05.939
Show me how to do it.
240
00:24:06.400 --> 00:24:07.621
They will show you how to do it.
241
00:24:08.121 --> 00:24:15.047
And the last bonus thing they can do too is say, because this is moving so fast, AI is evolving so fast.
242
00:24:15.847 --> 00:24:24.010
they should say in this prompt, also go and do some latest research to make sure you're up to date on what's possible.
243
00:24:24.030 --> 00:24:39.195
Because the irony is, Matt, AI is evolving so fast that a lot of these chatbots don't even know what they're capable of because the training date is cut off and things have happened in the last two months since training data.
244
00:24:39.635 --> 00:24:43.456
So that's where I would start and you will get your eugenic system before you know it.
245
00:24:44.076 --> 00:24:46.840
Jeff, again, this is a topic that's been on the top of everyone's mind.
246
00:24:46.860 --> 00:24:51.445
You mentioned we've been covering A on the podcast a lot, but you definitely gave us some new insights to chew on.
247
00:24:51.505 --> 00:24:53.087
And we thank you for joining us today.
248
00:24:53.408 --> 00:24:53.868
Thanks so much.
249
00:24:53.888 --> 00:24:54.449
Had a great time.
250
00:24:55.170 --> 00:24:59.375
And as always, I also want to thank our audience for being with us today on Power Perspectives.
251
00:24:59.555 --> 00:25:00.617
And we'll catch you next time.
1
00:00:00.171 --> 00:00:03.253
Agentic AI, you've heard the term, but what is it really?
2
00:00:03.493 --> 00:00:06.495
And how is it already changing the way that energy companies are working?
3
00:00:07.075 --> 00:00:13.239
Our guest today on Power Perspectives is AI strategist, Jeff Wilser, and he's here to answer those questions for us.
4
00:00:13.559 --> 00:00:15.020
Jeff, welcome to the show today.
5
00:00:21.913 --> 00:00:23.377
Jeff, let's start with the basics.
6
00:00:23.477 --> 00:00:27.105
When you say agentic AI workflow, what exactly does that mean?
7
00:00:27.145 --> 00:00:31.796
And how is that different than say, someone at a utility company using a chat bot to ask a question?
8
00:00:32.605 --> 00:00:32.965
Sure.
9
00:00:32.985 --> 00:00:36.068
I know you've had a lot of recent great episodes on AI.
10
00:00:36.128 --> 00:00:38.770
So I'm sure many of your listeners have some basics here.
11
00:00:38.890 --> 00:00:50.618
But for those who are new to the agent world, I think when most people think of something like ChatGPT or Chatbot, it's mostly about just asking a question and getting a response, right?
12
00:00:50.718 --> 00:00:56.662
So simple analogy, hey, I want to go to Tokyo for a vacation.
13
00:00:57.163 --> 00:00:57.523
What do you
14
00:01:00.445 --> 00:01:02.588
would actually do the things for you.
15
00:01:02.748 --> 00:01:09.476
So instead of just getting an answer, like here is a good itinerary in Tokyo, you can say, hey, I want to buy plane tickets.
16
00:01:09.696 --> 00:01:12.379
Please buy me plane tickets, book me hotels.
17
00:01:12.860 --> 00:01:18.567
So you can think of it as actually the arms and legs of doing things, not just the brain itself.
18
00:01:19.063 --> 00:01:19.363
Gotcha.
19
00:01:19.383 --> 00:01:20.643
No, that's definitely helpful.
20
00:01:20.703 --> 00:01:22.584
It's a little bit more action-based.
21
00:01:23.004 --> 00:01:33.246
And for our audience that's in the utility and energy world, what's an example of something that maybe one of their employees spends hours doing today that an AI agent could realistically handle?
22
00:01:33.786 --> 00:01:34.367
Absolutely.
23
00:01:34.547 --> 00:01:39.988
And Matt, I think a lot of folks hear about AI agents and they immediately race to
24
00:01:40.888 --> 00:01:46.452
the most provocative, scary, crazy thing that can be done.
25
00:01:46.572 --> 00:01:51.335
Like some agents who has full autonomy of a control room.
26
00:01:51.915 --> 00:02:01.702
But there is a lot of middle grounds and a lot of far less glamorous, more basic things that can be done before you get to that point.
27
00:02:02.202 --> 00:02:09.469
I think the number one thing, and this is true, so many industries and top of utilities and energy is paperwork.
28
00:02:09.969 --> 00:02:16.816
Virtually every industry, including you guys, we all have more paperwork than we want to do.
29
00:02:17.516 --> 00:02:34.165
And that ranging from compliance documents to trying to find what your rates were in some niche area from 12 years ago, there are buckets and buckets of paperwork that no one really has time to do.
30
00:02:34.745 --> 00:02:41.969
And AI agents can be built to actually sift through the paperwork and summarize and actually draft documents.
31
00:02:42.189 --> 00:02:46.151
So kind of the two-year-ago and one-year-ago AI thing was,
32
00:02:46.671 --> 00:02:57.779
was to manually, let's say, feed a PDF into something like ChatDBT or Claude or Copilot and say, hey, summarize this for me.
33
00:02:58.520 --> 00:03:11.950
Now that could be done with far more efficiency where agents can actually romp through stacks and stacks of paperwork and glean insights and help you create summary documents.
34
00:03:12.697 --> 00:03:16.278
And is the benefit just that it's able to do it that much quicker and more efficiently?
35
00:03:16.358 --> 00:03:19.699
Or do you find that the AI agent can actually do it better too?
36
00:03:20.319 --> 00:03:35.922
So in your world, think about a compliance issue that no one analyst could possibly have time to read the hundreds of precedents and documents behind this, but an agent can go through.
37
00:03:36.602 --> 00:03:46.840
And not just do it faster, but have access to this sprawling amount of paperwork to turn up insights you wouldn't otherwise have known to even look for.
38
00:03:47.262 --> 00:03:52.365
Yeah, that's definitely, I think, something that the utility executives would be eager to learn more about.
39
00:03:52.626 --> 00:03:57.629
And you mentioned that people jumped to the idea of like, oh, the AI is running the control room.
40
00:03:57.649 --> 00:03:59.790
And then there's a little bit of an aversion to that.
41
00:04:00.090 --> 00:04:05.074
But even when you're just looking at the paperwork and the regulatory aspect of it, it's still important.
42
00:04:05.194 --> 00:04:14.380
And I'm curious how much you find that there's oversight or fact checking still needed from the human side of the equation when the AI agents are doing these tasks.
43
00:04:14.832 --> 00:04:24.035
For sure, there's a cliche that I'm sure you've heard, but it's one of those cliches that happens to be true, and that's keeping a human in the loop.
44
00:04:24.395 --> 00:04:27.877
And that can happen in a couple of different levels, right?
45
00:04:27.917 --> 00:04:32.498
You could have the human spot check and look for errors.
46
00:04:33.178 --> 00:04:37.580
Obviously, you want someone to be sanity checking it before it goes anywhere sensitive.
47
00:04:38.200 --> 00:04:45.663
But one thing I think that one thing I share when I work with companies and workshops is, yes, human loop.
48
00:04:45.743 --> 00:04:53.787
But one additional thing you can do to help with control and help guardrails is actually using AI agents and AI models.
49
00:04:54.327 --> 00:04:59.833
kind of to cross-check each other in an adversarial, constructive way, right?
50
00:04:59.873 --> 00:05:13.649
So imagine that you, to keep it very simple, imagine you're working with something like Claude to create some agentic workflow and your agent says, aha, I am now going to create this compliance document.
51
00:05:14.310 --> 00:05:17.471
You can then feed it to a different system entirely.
52
00:05:17.671 --> 00:05:32.277
Let's say in ChachiPT, give it the rules of the road, give it the context and say, hey, I want you with fresh eyes and to get the nice splash cold water here, cold water, I want you to evaluate this and find blind spots.
53
00:05:32.777 --> 00:05:38.460
So by using AI system to check the other, that can help actually determine
54
00:05:39.240 --> 00:05:51.515
reduce errors and give yourself more of a more um again i wouldn't use the word certainty a human should always be in a loop but it's helpful to have ais check each other and that can reduce errors as well
55
00:05:52.713 --> 00:06:00.478
Is that one of the go to solutions that we hear about hallucinations or just, you know, faulty conclusions sometimes come from the AI?
56
00:06:00.518 --> 00:06:05.281
And, you know, the natural inclination is to have the humans be the one checking that, like you said, spot checking.
57
00:06:05.341 --> 00:06:12.686
But do you find that it's a better practice or almost time saver to have that, you know, AI checking its own work?
58
00:06:13.138 --> 00:06:13.839
For sure.
59
00:06:14.059 --> 00:06:20.083
It's, again, not to take the human out of this, but it's like one additional level of protection before the human, right?
60
00:06:20.303 --> 00:06:31.230
And it's also, it's kind of a lot of fun too when you have these two different models and you almost surprise them and say, hey, by the way, I asked Claude the same question.
61
00:06:31.830 --> 00:06:47.296
here's what it found, and you give it its answer, that will almost, is a way to spur the first model to make sure it's buttoned up, and to actually second and triple and quadruple check itself.
62
00:06:47.756 --> 00:06:59.065
I know in the utility sector, we're in an environment where these companies are seeing massive liabilities with things like wildfire management, outages that impact vulnerable customers.
63
00:07:00.026 --> 00:07:03.929
The utility leaders don't really feel like they can afford to let a mistake through.
64
00:07:04.189 --> 00:07:08.492
And this is great advice for how to capture and minimize those mistakes.
65
00:07:08.592 --> 00:07:14.256
But on the chance that there is an AI agent involved and a mistake does get through, you
66
00:07:15.017 --> 00:07:18.038
Who ultimately is held accountable when something goes wrong?
67
00:07:18.478 --> 00:07:24.900
There's one group that is definitely without question will benefit from AI, and that is lawyers.
68
00:07:25.280 --> 00:07:30.102
There are going to be so many gray area legal questions about who ultimately is accountable.
69
00:07:30.662 --> 00:07:34.663
That is a legal question beyond my pay grade as far as the actual liability.
70
00:07:35.123 --> 00:07:40.365
That said, I think the North Star, the way I view all of this for your industry,
71
00:07:41.986 --> 00:07:47.850
is something that I call the red line theory and where you are on the red line.
72
00:07:48.571 --> 00:07:59.999
And what I mean by that is what I found is the most successful way companies can use AI, whether in utilities or in publishing or in healthcare,
73
00:08:00.719 --> 00:08:14.113
is to think clearly about where is the red line of what you hold sacred that AI should not touch, that humans, to your point, Matt, ultimately must be accountable for.
74
00:08:14.634 --> 00:08:22.182
So for one example, in media and editing and journalism, Time magazine, they said, okay,
75
00:08:22.662 --> 00:08:28.927
We're not going to use AI to write AI slop to write our articles and do our research.
76
00:08:29.548 --> 00:08:31.809
That is still the work of humans.
77
00:08:32.650 --> 00:08:41.678
But the chief operating officer of Time, who I interviewed recently, what he told me was, yes, the AI is not going to replace the human writers.
78
00:08:42.258 --> 00:08:43.499
However, that's a red line.
79
00:08:43.539 --> 00:08:45.060
That's what time is sacred for.
80
00:08:45.140 --> 00:08:48.263
That is their sacred editorial judgment and credibility.
81
00:08:48.803 --> 00:08:50.405
On the other side of the red line,
82
00:08:50.985 --> 00:08:57.610
They want to use AI ambitiously, and they created AI agents to do far more things that cannot be done with time.
83
00:08:57.750 --> 00:09:07.717
Now they have AI agents and AI bots that let you interact with Time Magazine and explore its 103 year history in ways you couldn't do before.
84
00:09:07.797 --> 00:09:09.098
They added new functionality.
85
00:09:09.378 --> 00:09:19.826
The goal is not to take humans out of the loop of using judgment with respect to the wildfires and all the tricky questions you're facing.
86
00:09:20.326 --> 00:09:34.109
The goal is to give humans more time to think deeply about edge cases, about the sensitive issues, by automating the more drudgery stuff, the mundane stuff, to give humans more time to do what they are really good at.
87
00:09:34.732 --> 00:09:35.712
Yeah, that's interesting.
88
00:09:35.732 --> 00:09:45.177
It makes me think of, we ran a survey a little while back asking utility executives, what do they think AI will not replace in the utility sector?
89
00:09:45.277 --> 00:09:52.800
And the main tranche of answers came into, one, there's the physical work of repairing a line and that much was obvious.
90
00:09:52.820 --> 00:09:54.060
Oh, the robots are coming.
91
00:09:54.120 --> 00:09:56.361
I'm sure the robots are coming imminently and they'll do all of that.
92
00:09:56.381 --> 00:09:57.262
Eventually, yeah.
93
00:09:58.342 --> 00:10:06.067
But the other aspects that were cited multiple times were empathy for the customers and, you know, having kind of that human connection.
94
00:10:06.207 --> 00:10:12.271
And then the sort of judgment that leaders have to make based on all the information that AI can help collect for them.
95
00:10:12.631 --> 00:10:18.635
But ultimately, a human has to make, you know, weigh the pros and cons and the different tradeoffs and make that ultimate decision.
96
00:10:19.221 --> 00:10:20.462
Yeah, well said.
97
00:10:20.642 --> 00:10:25.066
And along those lines, think about flying and aviation.
98
00:10:26.027 --> 00:10:27.048
That's similar to your world, right?
99
00:10:27.068 --> 00:10:28.589
The stakes just as utilities.
100
00:10:29.030 --> 00:10:32.593
When you're flying a plane, human lives are quite literally at stake.
101
00:10:33.233 --> 00:10:41.140
When autopilots, when that technology developed years and decades ago,
102
00:10:41.901 --> 00:10:44.063
pilots were understandably skittish about it.
103
00:10:44.223 --> 00:10:47.146
They did not want to trust flying to a computer.
104
00:10:47.286 --> 00:10:47.966
Like, fair.
105
00:10:48.086 --> 00:10:49.548
That is an understandable reaction.
106
00:10:50.108 --> 00:10:55.573
They also worried they could be replaced by robots and by self-flying planes.
107
00:10:56.234 --> 00:11:02.259
And what happened was not that pilots were replaced or that pilots stopped using their judgment.
108
00:11:02.459 --> 00:11:05.622
What happened instead was pilots, they still had to know how to fly, but
109
00:11:06.102 --> 00:11:16.895
They still made the judgment calls, but they're actually able to make their role more about the judgments as opposed to doing more of the manual things.
110
00:11:16.915 --> 00:11:20.139
I think utilities can be looked at the same way, right?
111
00:11:20.339 --> 00:11:23.083
Ultimately, it's the human's job to apply the judgments.
112
00:11:23.663 --> 00:11:27.266
And if we can automate some systems to make that job easier, great.
113
00:11:27.766 --> 00:11:38.434
But I would also encourage utilities to think about it from the lens, not just of automating and cutting, but even growing.
114
00:11:38.794 --> 00:11:42.497
And by growing, I don't necessarily mean doubling revenue.
115
00:11:43.077 --> 00:11:48.782
But rather, what else could you do that you couldn't do beforehand?
116
00:11:48.902 --> 00:12:07.438
That's a big light bulb I see when I work with CEOs about this across industries is most companies and most organizations are thinking very narrowly about, okay, maybe I can use AI to cut this one thing or to automate this.
117
00:12:08.178 --> 00:12:13.982
And sure, automation is great, but it's also just scratching the surface of what you can do.
118
00:12:14.302 --> 00:12:28.392
I'd encourage everyone listening to your podcast to do is to take a moment and imagine, let's say theoretically, your headcount has doubled without spending an extra nickel.
119
00:12:29.492 --> 00:12:31.735
What then would you be able to do?
120
00:12:32.455 --> 00:12:34.718
What additional safeguards could you create?
121
00:12:35.358 --> 00:12:36.539
What additional services?
122
00:12:37.360 --> 00:12:40.964
Where are you able to add more value if your head got doubled?
123
00:12:41.545 --> 00:12:49.673
And then think, okay, how can AI allow us to actually get some of that new resource and capability?
124
00:12:49.693 --> 00:12:49.953
Okay.
125
00:12:50.374 --> 00:13:03.330
So I think that's one way that people in utilities can use AI, not just for like saving off a bit of cost at the margin, but to more ambitiously get more done in your organization.
126
00:13:03.830 --> 00:13:07.154
Yeah, that's a terrific thought experiment and how to approach it.
127
00:13:07.275 --> 00:13:07.475
Yeah.
128
00:13:07.815 --> 00:13:18.465
kind of going back to the treachery of the paperwork we were talking about earlier, you know, I know a longstanding challenge in the power sector is siloing, you know, information processes.
129
00:13:18.786 --> 00:13:23.250
They're always separated between different departments and the right hand doesn't always know what the left hand is doing.
130
00:13:23.730 --> 00:13:26.093
So given that part of that problem is that utilities have
131
00:13:26.733 --> 00:13:31.516
So much siloed data buried in PDFs, filings, engineering documents, old records.
132
00:13:32.036 --> 00:13:39.980
How close do you think we are to a world where, you know, through an agentic AI, an employee can essentially have a conversation with all of that institutional history?
133
00:13:40.983 --> 00:13:42.304
I think it's as close as you want to be.
134
00:13:42.704 --> 00:13:50.408
The technology exists right now, and it's not something that requires billions of dollars of investment.
135
00:13:50.588 --> 00:13:51.729
It's mostly about that.
136
00:13:51.749 --> 00:13:55.831
The irony, Matt, is I think this is one thing that's kind of lost in the conversation.
137
00:13:55.851 --> 00:13:58.853
There's so much AI hype out there and all these buzzwords.
138
00:13:59.773 --> 00:14:05.418
Even the word agentic AI can sound kind of magic wandy.
139
00:14:05.738 --> 00:14:12.444
Ultimately, an AI agent is mostly about context and context files.
140
00:14:13.104 --> 00:14:19.489
And by context files, kind of fancy way of saying really just like, it could be as simple as text documents.
141
00:14:19.510 --> 00:14:20.851
You mentioned PDFs, right?
142
00:14:21.011 --> 00:14:28.217
And so what's making or breaking AI adoption, both in the power industry and most industries at large,
143
00:14:28.897 --> 00:14:34.622
is not about are you using the most cutting edge, powerful frontier model.
144
00:14:35.243 --> 00:14:41.048
It's less about the model's power and more about how well are you feeding it context.
145
00:14:41.568 --> 00:14:48.815
What I mean by that is, as you point out, Matt, there could be PDFs living in, there's somewhere in IT, there's all kinds of PDFs.
146
00:14:48.835 --> 00:14:50.957
Somewhere in procurement, in marketing.
147
00:14:52.398 --> 00:14:59.825
Every different division has their own weird warehouse figuratively or literally of all these different files.
148
00:15:00.946 --> 00:15:13.398
Getting those files, having a smart infrastructure that is feeding the relevance files and takeaways from the files into a system that AI agents read is,
149
00:15:14.239 --> 00:15:15.220
That's the whole ballgame.
150
00:15:15.620 --> 00:15:25.047
The other area where, you know, utility leaders traditionally see knowledge spread out is where it's not written down at all, but it's stored in the minds of longstanding employees.
151
00:15:25.107 --> 00:15:31.892
And this is a common topic in the utility sector as much of the workforce is aging closer and closer to retirement.
152
00:15:32.252 --> 00:15:33.914
Are we risking losing that knowledge?
153
00:15:34.294 --> 00:15:40.899
I'm curious what you think AI agentic tools can do to capture that institutional knowledge and maybe re-reporting.
154
00:15:41.199 --> 00:15:49.366
plug in that gap where, yeah, hey, there might have been that 40, that employee who's been there for 40 years, who knows, oh, this is how this substation really behaves.
155
00:15:49.406 --> 00:15:54.450
How can we set up the agentic AI to be able to understand that as well?
156
00:15:55.000 --> 00:15:56.381
Yeah, I'm so glad you brought that up.
157
00:15:56.401 --> 00:15:57.142
That's a great one.
158
00:15:57.182 --> 00:15:58.002
I think you're right.
159
00:15:58.102 --> 00:16:01.024
It's super relevant to the industry here.
160
00:16:01.305 --> 00:16:05.548
And some might call it tribal knowledge that is a danger of being lost.
161
00:16:06.188 --> 00:16:11.952
And I think it's one often overlooked use case of AI agents.
162
00:16:12.233 --> 00:16:14.074
And one of my favorites is technology.
163
00:16:14.374 --> 00:16:21.779
interviews and actually having... Now, this could be done a couple of ways.
164
00:16:22.319 --> 00:16:36.127
You can set up an interview for almost like a series of prompts that this salty engineer about to retire can then answer, crucially, answer verbally, right?
165
00:16:36.307 --> 00:16:36.808
If it's...
166
00:16:37.388 --> 00:16:41.890
If we're asking for type response, that's not going to be nearly as effective.
167
00:16:42.070 --> 00:16:45.292
No one likes to type a 5,000-word essay, right?
168
00:16:45.872 --> 00:16:51.595
But if you ask them questions about things that, frankly, they'll probably be kind of fluttered.
169
00:16:51.635 --> 00:16:52.916
They're asked about, like, oh...
170
00:16:53.636 --> 00:17:00.142
Yeah, I recall well that ice storm of 1987, and no one's thought to ask me that in years, right?
171
00:17:00.542 --> 00:17:12.032
And have them go through and give verbally detailed responses that can create this library, this corpus of knowledge.
172
00:17:12.292 --> 00:17:14.914
And here's where agents really come into play.
173
00:17:16.035 --> 00:17:25.359
In a world before AI, if you were to, let's say, spend, and I'm making this roadmap up, but bear with me here.
174
00:17:25.819 --> 00:17:34.782
Let's say you have maybe two hours a week you devote to interviewing this, the salty veteran engineering knowledge, right?
175
00:17:34.802 --> 00:17:43.206
So over the course of a year, you have a hundred hours of them talking about all kinds of esoteric things they've seen.
176
00:17:44.126 --> 00:17:44.967
And do that at scale.
177
00:17:45.267 --> 00:17:51.612
So you've now interviewed, let's say, let's say there's thousands of hours of conversation of these folks.
178
00:17:52.213 --> 00:17:57.978
In a non-AI world, good luck getting anyone to go through that.
179
00:17:58.958 --> 00:18:06.104
It reminds me of the last scene of Raiders of the Lost Ark with the crates and crates and crates in the warehouse and all these things.
180
00:18:06.645 --> 00:18:10.088
No one can go through those crates in the warehouse
181
00:18:10.728 --> 00:18:11.409
and make sense of this.
182
00:18:11.669 --> 00:18:25.900
The agents can go through this, at this point, this massive library of interviews you've conducted with all these folks who are about to retire and make heads and tails of it.
183
00:18:26.180 --> 00:18:29.043
It could synthesize things.
184
00:18:29.423 --> 00:18:32.325
You can actually have a few different cuts of this library.
185
00:18:32.345 --> 00:18:37.770
You can say, hey, I want to create a masterclass about ice storms
186
00:18:38.390 --> 00:18:43.296
or about substations and how they reacted in this kind of emergency.
187
00:18:43.916 --> 00:18:51.245
And you can create an entire documents and dossier around these thousands of interviews.
188
00:18:51.946 --> 00:18:58.893
And then once you have that effectively, this massive database of tribal knowledge in one place,
189
00:18:59.514 --> 00:19:05.437
In the future, you could ask AI agents to search into it, to probe, to ask questions.
190
00:19:05.897 --> 00:19:18.423
A more advanced thing would that mean to create an actual, you can create effectively your own chatbot based off of this database to let you real-time ask questions to it as well.
191
00:19:18.503 --> 00:19:22.305
So I think we're going to see, that's still relatively fast,
192
00:19:22.865 --> 00:19:26.912
fringe and bleeding edge use case right now.
193
00:19:27.132 --> 00:19:34.064
In the future, my prediction, that sort of use case will go mainstream across many industries.
194
00:19:35.047 --> 00:19:36.567
Yeah, that's fascinating.
195
00:19:36.607 --> 00:19:39.708
And I definitely think that the utility sector is primed for that.
196
00:19:39.748 --> 00:19:45.770
Like I said, the aging workforce, this long institutional knowledge, it's too valuable to just let it go.
197
00:19:45.870 --> 00:19:52.072
So maybe the AI agents are coming at the perfect time to fill in that gap that's so needed.
198
00:19:52.472 --> 00:19:58.094
But I guess as we wrap up our conversation, Jeff, I want to ask the advice-based question.
199
00:19:58.314 --> 00:19:58.354
If
200
00:19:59.534 --> 00:20:06.740
somebody's listening in, they're running utility, they want to experiment with agentic AI this week, not two years from now, not those fringe cases.
201
00:20:07.140 --> 00:20:12.645
What is the first workflow you would tell them to try and how would they know whether it's actually working?
202
00:20:13.573 --> 00:20:24.059
One thing that, and if your listeners already tried this, great, move on to the next, but I think something that everyone busy in the utility industry can deal with.
203
00:20:24.080 --> 00:20:32.385
But before you think about how can we inject AI in most sensitive workflows of our power grid,
204
00:20:33.445 --> 00:20:35.926
Whether we like it or not, we all have busy inboxes.
205
00:20:36.207 --> 00:20:39.648
We all struggle to make the meetings on time.
206
00:20:39.889 --> 00:20:42.230
We all have a hard time keeping track of our to-dos.
207
00:20:43.150 --> 00:20:55.557
I would create your own chief of staff, your own AI agent who is able to triage email for you, who is able to hoover up all the projects you have.
208
00:20:56.598 --> 00:21:02.621
Most systems now let you easily connect in your email company Slack.
209
00:21:03.401 --> 00:21:05.082
to-do list systems, and so on.
210
00:21:05.723 --> 00:21:11.146
And crucially, my advice to even be more specific is to give it lots of context.
211
00:21:12.086 --> 00:21:24.754
Spend an hour walking through what's important to you, what your goals are, what are risks you're worried about, and have it then each morning,
212
00:21:26.795 --> 00:21:34.058
give you a briefing for the day and say, hey, what are the three most things I must do today?
213
00:21:34.258 --> 00:21:35.399
What's most important?
214
00:21:36.099 --> 00:21:44.103
This, I think that's a, one reason I like this starting point so much is that one, it is very de-risking, right?
215
00:21:44.123 --> 00:21:46.043
You are not putting the grid at risk.
216
00:21:46.444 --> 00:21:49.365
You are starting small with your own personal workflows.
217
00:21:50.105 --> 00:21:53.587
And two, I think most people will get a ton of value out of it.
218
00:21:53.627 --> 00:22:08.673
When I work with CEOs and show this in workshops, teams are like, oh my God, it is amazing how a little bit of agentic systems in your own workflow can handle so much busy work.
219
00:22:08.813 --> 00:22:18.717
And one more bit of very practical advice that I think listeners will benefit from is when you're working with AI,
220
00:22:19.497 --> 00:22:20.137
in this manner.
221
00:22:20.537 --> 00:22:26.899
The way to speed things up and the way to give it even more context is actually through voice, not through typing.
222
00:22:27.459 --> 00:22:35.441
Reason why is that almost everyone, certainly including me, despite typing for a living, is much faster at talking than typing.
223
00:22:35.741 --> 00:22:37.741
It's also more enjoyable to talk than to type.
224
00:22:38.361 --> 00:22:44.723
And so when we're going back to when we interviewed the salty 35-year veteran about to retire,
225
00:22:45.423 --> 00:22:52.346
they're going to be more likely to share lots of rich details if they're just casually, conversationally chatting.
226
00:22:52.526 --> 00:22:58.089
If they're typing, they cut corners, but the corners are where the good stuff lives.
227
00:22:58.909 --> 00:23:00.490
The corners are the edge cases.
228
00:23:00.830 --> 00:23:01.571
That's the nuance.
229
00:23:02.031 --> 00:23:03.932
The more nuance you give, the better.
230
00:23:04.072 --> 00:23:08.034
When you're talking, you're more likely to just share a random parenthetical.
231
00:23:08.474 --> 00:23:13.818
the AI will actually store that parenthetical and that's useful context in the future, right?
232
00:23:13.959 --> 00:23:27.610
And so, and last, not to be overly mechanical here, but when you hear voice mode, people might be tempted to actually turn on the kind of like voice mode in ChachiPT or Claude, or mainly ChachiPT.
233
00:23:28.250 --> 00:23:37.278
That's not what I mean, because that, when you go back and forth with voice, almost like Siri, and we all know how bad Siri is, you kind of only get the lightest,
234
00:23:37.938 --> 00:23:40.600
most surface level version of the AI.
235
00:23:40.720 --> 00:23:44.363
It's not the more powerful thinking versions of the model.
236
00:23:44.963 --> 00:23:48.946
Instead, what I mean is you can use a dictation service.
237
00:23:49.347 --> 00:24:00.015
And for those who are still not sure exactly where to start, literally after you finish listening, Matt, to your great podcast, they can go in and say, hey, chatbot of choice.
238
00:24:00.755 --> 00:24:04.338
I want to create my own chief of staff AI agents.
239
00:24:04.898 --> 00:24:05.939
Show me how to do it.
240
00:24:06.400 --> 00:24:07.621
They will show you how to do it.
241
00:24:08.121 --> 00:24:15.047
And the last bonus thing they can do too is say, because this is moving so fast, AI is evolving so fast.
242
00:24:15.847 --> 00:24:24.010
they should say in this prompt, also go and do some latest research to make sure you're up to date on what's possible.
243
00:24:24.030 --> 00:24:39.195
Because the irony is, Matt, AI is evolving so fast that a lot of these chatbots don't even know what they're capable of because the training date is cut off and things have happened in the last two months since training data.
244
00:24:39.635 --> 00:24:43.456
So that's where I would start and you will get your eugenic system before you know it.
245
00:24:44.076 --> 00:24:46.840
Jeff, again, this is a topic that's been on the top of everyone's mind.
246
00:24:46.860 --> 00:24:51.445
You mentioned we've been covering A on the podcast a lot, but you definitely gave us some new insights to chew on.
247
00:24:51.505 --> 00:24:53.087
And we thank you for joining us today.
248
00:24:53.408 --> 00:24:53.868
Thanks so much.
249
00:24:53.888 --> 00:24:54.449
Had a great time.
250
00:24:55.170 --> 00:24:59.375
And as always, I also want to thank our audience for being with us today on Power Perspectives.
251
00:24:59.555 --> 00:25:00.617
And we'll catch you next time.
