Latest / Wired to Build / AI, Risk & Responsibility: Building Guardrails with Jamie Lynn Cooke

Transcript

AI, Risk & Responsibility: Building Guardrails with Jamie Lynn Cooke · Wired to Build

0:00
Hey, what's going on everyone? Welcome to Wire to Build, a
0:03
data-driven, technology focused podcast digging into how
0:06
building owners and operators use technology to manage the
0:09
full life cycle of their projects.
0:11
I'm Nick Caravella, your showrunner and curator, and I'm
0:14
excited you're joining us today to talk about the people that
0:17
drive the change in our industry.
0:19
Yeah, last episode we met Jamie Lynn Cook, someone who has made
0:23
a career out of bringing discipline to innovation,
0:26
whether walking agile out of software and into business
0:30
operations or holding organizations accountable to
0:33
deliver real, measurable outcomes.
0:36
In that conversation, we began to plant the seed that
0:38
technological excitement can carry real risks if it's not
0:42
measured with governance. So today Jamie's returning to
0:45
the podcast so we can dive deep into AI, not just the hype, but
0:49
how organizations can operationalize it responsibly.
0:53
We'll explore the differences between procedural code and
0:56
probabilistic AI, unpack cautionary real world examples,
1:00
and layout a practical governance framework for high
1:03
stakes environments. So if your organization's
1:05
already using AI or thinking about it, this is your moment to
1:09
shift from excitement to discipline.
1:11
Jamie's expertise will help you make that transition in ways
1:15
that protect value and avoids liability.
1:18
Jamie, thanks for coming back. I want to start with the big
1:21
picture. Tell us your view on AI, the
1:23
next productivity wave, and why this one's different.
1:26
So AI is an incredibly powerful technology and and particularly
1:31
generative AI. So what you commonly would know
1:34
is chat bots or AI agents or agentic AI, it is incredibly
1:39
powerful and with that power comes responsibility, right?
1:44
So yes, you can build solutions quickly and you can deliver
1:47
output much more quickly than you had before, but the jury's
1:51
out on whether or not it can really consistently deliver high
1:54
quality results at the level the organizations need to protect
1:58
their interests and to minimize their liability, as well as
2:02
retaining their customers and growing their organization.
2:04
And I don't think enough people realize exactly where the risks
2:10
are, particularly in generative AI, and come into the decision
2:16
to bring on AI solutions with that knowledge.
2:20
I think it's really important for people to have full
2:22
awareness so that they can leverage the power of AI, but do
2:27
it in a way that really protects the organization and grows the
2:31
organization in a healthy way and doesn't and minimizes the
2:35
risk of surprises while this is happening.
2:38
And certainly after these, these solutions go live.
2:41
So that's that's the overarching theme of what we're trying to
2:45
cover in this session is empowering people to take
2:48
advantage of AI, but doing it with due diligence in a way that
2:52
to really let your organization grow safely.
2:54
Yeah. And there's this phrase that
2:55
we've had kicked around the podcast quite a bit this year,
2:58
which is like slow as safe and safe as fast.
3:01
And the conversation we had on the last episode talking about
3:05
agile being used to deliver value more quickly because it
3:09
could actually leverage course correcting in real time.
3:13
Is this something that is almost something that needs to be
3:16
doubled down on for folks who are looking to leverage AI to
3:21
create immediate impact in their organization?
3:24
I think the the safely approach is is really the key to this.
3:29
So you need to give AI solutions enough room to breathe to
3:35
achieve what they're meant to achieve.
3:37
But you need to do with guardrails, right?
3:39
You need to choose the things that AI is is best suited for.
3:43
You need to involve humans in the process probably more than
3:47
you realized. And you need to limit its
3:50
ability to do high risk activities in a way that puts
3:54
your organization in jeopardy, right?
3:57
And we're going to talk about what that jeopardy actually
3:59
means in the real world sense. But the intent is with AI, if
4:06
you approach it with these guardians in mind, if you make
4:10
decisions around where it makes sense to use it and where it
4:13
makes sense not to, and to what degree you need to have people
4:17
involved in AI driven solutions and what they produce.
4:22
You will be creating the foundation of the framework for
4:27
success in AI across the board. And that success, because of the
4:32
nature of AI, can include and should include incremental
4:37
updates to what you're doing. So whether it's training the AI
4:41
models with more current data or historical evidence.
4:46
And so it increases its likelihood of coming up with the
4:48
correct answer. That's one way of of continuous
4:52
improvement. Another way is to go on it and
4:55
keep track of what's happening when AI produces output.
4:58
You know, how often are the reports it generates correct?
5:02
And how often do they need to be modified by a person that that
5:06
constant vigilance on the effectiveness of AI, the
5:10
accuracy of what delivers and in the fine tuning of either the
5:14
data model or the prompts or the tools that you've built to
5:20
constantly be making your AI solutions more refined and
5:24
better able to meet your emerging business requirements.
5:28
Because as you introduce AI in the organization, your needs are
5:31
going to change. You're likely to expand your
5:34
markets, you're likely to expand your products and services, and
5:38
you're going to need to continuously adapt so that your
5:43
operational model and your organizational structure is
5:46
optimized to leverage the combination of what AI can bring
5:51
and what people in conjunction can bring.
5:55
And that's, it's so much to to unpack because it is this like
5:58
trying to find the right balance and what's the right balance for
6:01
your organization. We tease them in the last
6:03
episode talking about Air Canada.
6:06
But before we jump into the case study, I kind of want to zoom
6:08
out, set the stage, especially for our audience and, and how it
6:12
relates to them. And so, you know, where this, I
6:15
think it's interesting is in capital projects, you know, most
6:19
of the technology that we use is about creating the same input,
6:23
creating the same output every time.
6:24
There's a process that can be followed.
6:27
You can test it out once, you can test it 100 times behaves
6:30
the same way AI very different. You can ask it the same question
6:35
a bunch of different times. It's going to give you different
6:37
answers. And that's kind of based off of
6:39
how language language models can work of predicting the next word
6:45
versus following a process. So could you maybe explain the
6:50
difference between a procedural code and AI to people who don't
6:55
live in the tech world? Sure.
6:56
So the technology the organizations have been most
6:59
familiar with over the past few decades has been procedural
7:04
code. You establish a design and
7:06
upfront to build out processes that have very fixed rules.
7:11
They follow those rules. And when you test them and test
7:16
them 99 times, you can be reasonably confident that in the
7:19
same environment, the 100th time is going to produce the same
7:22
output. So it gives you confidence to be
7:25
able to empower those tools once you've thoroughly tested them to
7:29
go and do their jobs. And then your job is more
7:32
maintenance and exception handling.
7:34
And that's been the traditional mindset of technology in
7:37
production and how it works. With AI, there is an equivalent
7:41
concept. It's called deterministic AI.
7:44
You can actually build AI to function like procedural code
7:48
and you can build it with rules and you can build it with
7:50
guardrails and you can build it with the same construction.
7:53
But that is a high overhead activity.
7:56
It requires a lot of pre work and a lot of ongoing development
8:00
work. And the solution you deliver
8:03
can't grow and learn from new information it gathers.
8:08
So organizations, even though they could choose deterministic
8:11
AI and have it be similar to our historical procedural coding,
8:16
our choosing non deterministic AI, especially generative AI for
8:20
the most part. And that's where the risk comes
8:22
in. Because it is not a system
8:27
following exactly the rules with absolutely no digression.
8:31
It's not a human making a judgement call about what you
8:34
need. It is algorithms doing
8:37
probabilistic calculations on large data sets.
8:41
And there are times when those calculations will be 100%
8:44
correct and that's fantastic. And there are times when it'll
8:46
be 100% correct 99 times. And then the hunger time, it'll
8:51
generate something completely different.
8:52
And for those of you who are familiar with with using common
8:56
AI tools, ChatGPT and Gemini and Copilot and all those tools, go
9:01
in and put a prompt, any prompt, ask it a question and then look
9:07
at the response and start a totally new chat and ask it
9:12
exactly the same prompt and look at the response.
9:15
And do this three or four times. And what you'll see is that
9:18
there's often some similarity and overlap, but there's also
9:22
differences because every single time it is processing those
9:26
prompts as if it were a uniquely new request.
9:30
On the one hand, that's fantastic.
9:32
That gives you the ability to get a lot of different
9:34
variations and be able to simulate the right answer for
9:37
you. But it also begs the question of
9:40
what's happening when your customer is asking a question on
9:43
your chap up about your functionality, capabilities or
9:47
pricing or your warranty or anything else, and one customer
9:52
asks the same question as another and gets a wholly
9:55
different response. That's potentially, especially
9:58
given, you know, whatever questions they're asking, it
10:01
could be a potential exposure for your organization.
10:04
People going into this idea of, well, we're going to create
10:07
these chat bots and AI agents. They're going to replace people
10:10
and, you know, that will take care of it.
10:12
And as long as we put some guardrails in, that will
10:15
minimize the risk. To some extent that's true.
10:19
But the other piece of the equation is that AI is not
10:22
really good at what's called edge cases, exception cases.
10:25
So 99 times out of 100, the question is something that's a
10:31
standard basic question about your product capabilities, and
10:34
that's fantastic. But somebody goes and asks a
10:36
question that isn't part of the standard or part of what you
10:40
would expect. What happens sometimes is that
10:44
AI will respond with a definitive answer, a confident
10:48
answer that is a complete hallucination.
10:51
It will make up an answer because it doesn't.
10:53
It was never trained on this information.
10:55
It doesn't have the data, so it will have its own volition, will
10:59
come up with something completely and totally out of
11:03
the blue and it will say it confidently.
11:06
That's the one thing about AI responses that that makes me
11:08
crazy because it will not qualify at the beginning and
11:13
say, you know, I don't really have the information for this.
11:16
I'm really not sure, but here's what I might be the case.
11:19
It will state it as if it's fact and people will follow it as its
11:23
fact. And oh, by the way, if your
11:25
organization's logo is at the top of their, they're going to
11:28
treat it like content from your organization.
11:31
And that's part of the risk management that you need to
11:34
factor in is that AI can and will do this.
11:38
And part of the issue is, is how well you've trained your data
11:41
set and you can do red teaming and quality testing and
11:45
everything you want. But AI, by its nature, the
11:49
current AI narrow AI, is probabilistic and there is
11:53
always the possibility he's going to deliver a result that
11:56
you did not expect. Yeah, man, I was jotting down
12:01
bullets as he as he talked about this, because there's so many
12:04
fun ways. And I think the very first thing
12:07
that I've loved and I don't, I don't think I shared this with
12:09
you, but it's a really funny experiment that so far I've
12:13
still gotten the same outcome. So if you're using Chachi PT,
12:17
anybody who's listening, go ask it to pick a random number
12:20
between 1:00 and 10:00. And I guarantee, well I won't
12:24
guarantee, but I will bet that it's a 7.
12:27
And what's really interesting about this is goes right to the
12:30
probabilistic AI when you ask. When I've queried AI afterwards,
12:34
after running a couple experiments from a few different
12:37
people, all asking a picking a number one through 10 and it's
12:39
showing up at 7, it tells you flat out that it picks 7 because
12:43
7 is the probable choice when someone's picking a random
12:48
number between 1:00 and 10:00. Which is very interesting
12:52
because typically if you asked a random number generator, it's
12:55
just going to roll the dice. It's actually going to look at
12:57
the time stamp on your computer and multiply that by some real
13:01
factor and then round off all the digits to countdown back to
13:04
the number you're choosing from. That's how random number
13:06
generators work, if anybody's curious.
13:09
And so there's something really fascinating there now where you
13:12
turn it back and you say, hey, I pick a number from the 10 and
13:15
it's going to pick 7, much like most people would.
13:18
And I think that's something, you know, when you're asking it
13:22
even for random information, it has difficulty being random.
13:25
And so when it's even trying to pick up accurate information is
13:29
having difficulty be accurate. And this is one of the weird
13:33
conundrums to get to my second bullet here of where I really
13:37
feel like in a in a strange way, a is made BI cool because we're
13:43
all trying to use AI to get business intelligence and
13:47
business intelligence and stuff we've had forever that was
13:50
driven by this process kind of thinking and math and algorithms
13:55
and all that other stuff that we could just be reporting from.
13:58
And then that starts to kind of go down an entire rabbit hole of
14:00
data literacy. And you know, do people know how
14:04
to open up an Excel sheet and understand actually what's being
14:07
said? Which then goes down the rabbit
14:10
hole of the confident answers where AI is going to tell you
14:13
with certainty that this is what it's found, but needing to
14:18
understand what was the data sets that it looked at, what was
14:21
the process that it used. All of these different things
14:24
has led me to get to my last bullet point here of needing to
14:29
force AI to qualify itself. And I've learned through at
14:33
least my own prompting that. But I almost always end my
14:37
initial prompt with before you start answering, evaluate my
14:42
prompt. I reflect on evaluate my prompt.
14:46
Ask me questions about information or context that
14:49
would be valuable to helping you come up with the best answer
14:52
possible, noting that the best answer looks like these outcomes
14:57
and really try to force AI to start doing the qualification
15:03
and validation. I'm in the darn it if it gives
15:05
you links, click on those links and see if because they will lie
15:09
to you about stuff and found or something will be pay gated and
15:12
you just can't qualify it. But that proc process has really
15:16
helped at least me start to get the AI to zoom in a little bit
15:23
more. So it's just, it's really
15:25
fascinating how, you know, I think most of the industry today
15:29
and many industries are trying to use AI for something that
15:32
could be solved by a robotic process automation or, or ABI
15:37
dashboard, if you will. Now we just want to be able to
15:40
ask something, Hey, how many hours did I use instead of
15:44
looking at my timesheet? I get the simplicity and I also
15:49
get the application, but it's also it's, it's funny because
15:52
that's like not what AI is for at least today.
15:56
And so that, that conundrum is always a little bit interesting.
16:00
And then you say too like everybody uses it differently.
16:03
And so that unpredictability of how humans interface and how I
16:09
might ask a schedule question one way and someone else me
16:13
answer, ask a schedule question, a similar schedule question
16:16
completely differently and get 2 entirely different results
16:20
because of that missing contents nuance and and whatnot between
16:25
them all. So that long ramble of like all
16:29
the AHAZ is just what she said is if I'm a, if I'm looking for
16:34
a material business answer and I'm going to rely on AI,
16:38
something that affects cost, schedule, safety, you know, what
16:43
can go wrong if I don't understand those complexities
16:47
And what are maybe some of the tools that are things I should
16:50
be thinking about that will help me get a better response or
16:55
better relationship with AI. So.
16:58
So the first thing is the data set that you're working from,
17:02
right? Because large language models,
17:04
by their nature, once they're trained, they're frozen in their
17:07
state. And you can use something called
17:10
WAG, which able augmented generation to bring in more
17:14
current information to help frame the prompt.
17:17
But really, truly, it's the data set that's trained on that is
17:21
the sort of starting point of whether or not you're going to
17:24
get an accurate response and if that data set is old or its
17:28
scope is limited or it's biased. So one of the examples I would
17:32
give is if there was a huge marketing campaign for a
17:36
particular product and because of that, people started focusing
17:41
on that product above others. Because sort of like the number
17:45
seven comment you made before, because it was the most commonly
17:50
requested or the most commonly thumbs up response that the
17:54
people received. The information is going to be
17:58
biased towards prompting people to use that product.
18:02
And that that tends to be the case for the combination of how
18:06
old or or broad the scope is with the training that they, the
18:11
solution will do to use probability and frequency to
18:15
determine how to respond. So the first thing is making
18:19
sure the data set you're using is accurate and it's kept up to
18:22
date. The second thing is what you
18:24
hinted at before, which is you crafting the prompt.
18:27
The prompt engineering in a way that minimizes digression for
18:33
the AI. You're giving it guardrails.
18:35
You're telling you what you're looking for.
18:36
You're giving an example of a report, one example of
18:39
communication it should follow. The more input you put into the
18:43
prompt, the more you are specific.
18:45
And, and the analogy is always given is if you have an intern
18:49
and you're trying to describe the tasks you want them to do
18:52
and you really want to make sure that they understand what you're
18:55
looking for, you give them all the detail you can, you give
19:00
them all the context, you can, you give them those guardrails.
19:03
This is not what I'm looking for.
19:04
This is what I'm looking for. Here's an example you can use.
19:08
The more that you structure the input engineer, the prompt to,
19:12
to be very specific about what you're looking for, the greater
19:16
likelihood. Again, it's not guaranteed, but
19:18
the greater likelihood that what it will produce is going to be
19:23
what you're expecting. And that goes back to this
19:25
concept of chat bots with open-ended prompts versus AI
19:29
agents that have predefined prompts that the engineers
19:32
decide to to put the words in on.
19:36
Or a gentic AI that has its own, you know, goals that have been
19:39
set up to orchestrate those agents that this concept of what
19:44
input is going in, who's explaining what's needed, what
19:48
data is it using to make those calculations.
19:51
And then your comment about, look, the same person can ask
19:54
the same question 20 times or 20 different people can ask the
19:58
same question. And either one of those
19:59
scenarios can generate vastly different output.
20:02
And so this goes back to the critical eye.
20:05
So, so my, the metaphor that I like to use for organizations
20:10
when they're making a decision of what can or should be AI on
20:14
its own versus what should be AI enabled as humans is really.
20:19
Would you trust the output to go directly to the recipient
20:22
without your intervention? So if you're telling an intern,
20:25
could you create a calendar invite?
20:27
More likely than not, they could safely do that, and you got to
20:30
assume that they know what they're doing.
20:32
But if you tell an intern, I want to do a proposal for a
20:35
budget extension, right, or I want to do a summary report of
20:38
the safety concerns on this construction site, you're not
20:43
likely to leave it completely with them to create the report
20:46
and give it right to the customer without somebody's got
20:49
subject matter expertise stepping in.
20:52
And, and this is that spectrum the organizations need to decide
20:56
is I trust AI to an extent. I need to curate what it
21:02
produces. I need to confirm the sources.
21:05
So there's another risk to AI, which is complacency, right?
21:08
People start to begin to, oh, I started confidently and when I'm
21:12
too busy doing other things, I'm just going to take what AI gave
21:15
me and throw it in an e-mail and, and send it out.
21:18
If you've ever taken the time to actually research the sources
21:21
that AI comes up with, some of those sources are personal
21:25
blogs. There's somebody's opinion on
21:27
something. They'll reference Wikipedia,
21:29
which always cracks me up. They, you know, so in the prompt
21:34
you have to say I only want trusted global sources to be
21:38
used for the search and tell me what those sources are.
21:41
And then you need to make sure that it's understanding the
21:44
material on the source and interpreting it correctly by
21:48
crazy concept going into the source and actually looking at
21:52
what's there. And as as the knowledgeable
21:55
person you are putting those pieces of the puzzle together.
21:59
If you're not doing that step, and I'm not talking about high
22:03
volume, minimal, you know, value tasks, I'm talking about the the
22:07
real work that needs to happen. If you're not doing that
22:11
diligence and doing the follow up and critically thinking about
22:14
what AI generates, then you that in that activity are putting
22:20
your organization at risk in addition to the organization
22:23
trusting the process to do things without human
22:26
intervention. And it's, yeah, I feel like the
22:29
more and more anybody who's like licking the flex this muscle,
22:32
like pay attention to statistics and then validate them.
22:36
And I think this is where I, I've been fortunate that we've
22:39
been writing away papers and doing a lot of industry
22:41
research. And so when I see a number, I'm
22:43
like, what does that mean? That's crazy.
22:46
And then going clicking and finding out like, oh, well, this
22:49
OK, that is accurate. Actually 80% of our projects do
22:53
go over budget. And I think that and that that
22:56
speaks. Another interesting question is
22:58
if AI is based off of probability.
23:01
If most of our projects are finishing over schedule, over
23:04
budget, then it's probabilistic answer is going to support
23:08
things that have actually driven us to be going over in the first
23:11
place. Rather than your project's on
23:14
track because it's no less overrun than every other project
23:16
you've done. Exactly.
23:18
It's like, no, as much as I'd love to fail, as much as the
23:22
next guide, I'd like to do better.
23:24
And I got to repeat that pattern.
23:26
Thank you. And so I think that that's just
23:28
like, it's such a fascinating thing of being able to, and
23:32
we've gone back to this on many of our episodes of like scrutiny
23:35
as a human superpower in today's age of AI of I get what you're
23:40
saying, but let's dig a little bit more.
23:43
So let's get into a case study though, because I think this
23:46
will really help people understand like the risks.
23:49
And I think you really hit the nail on the head of like, what
23:52
is the who's the deliverable for?
23:54
If I'm working with it as a partner, probably OK.
23:57
If I'm having AI throw ball or refunds to a client, probably
24:01
not the greatest idea. And this is something that we we
24:04
dug into on our pre call of a is promise for productivity, but
24:09
his promise goes off the rails and then actually made a bad
24:12
promise to a customer. So can you walk the audience
24:15
through this Air Canada case where the chatbot gave
24:19
misinformation and then the company was actually held
24:22
responsible for it? So I there's actually a few
24:25
cases I'm going to reference here.
24:26
The first is the Air Canada case.
24:28
So what happened was that a man, I think his name is Jake
24:31
Moffett, went to Air Canada, went to its chatbot and asked
24:34
about bereavement fairs and was informed by the chatbot on the
24:38
Air Canada website that there is a bereavement discount.
24:40
And here's the process for getting it.
24:43
And so he went and purchased a ticket and went to process the
24:45
discount and was told by air kind of support, I'm sorry that
24:49
that isn't what we do. There is no discount.
24:51
And he said, well, your chap out said there was.
24:54
And you know, sorry, the chap out must have gotten it wrong,
24:58
the apologies. And he was not satisfied with
25:00
that answer. So he went to the Canadian
25:03
tribunal and he sued them. And basically the tribunal
25:09
decided that Air Canada made the choice to put its chatbot on its
25:13
website with its logo and Air Canada is responsible for
25:18
whatever content the chatbot produces.
25:20
Now granted, that was a Canadian tribunal that wasn't AUS court
25:23
or another jurisdiction, but it begins to show legal precedence
25:28
towards if a customer is taking action based on what you were
25:33
endorsed chatbot or AI agent or any other generative tool that
25:38
you use has produced, you are responsible for that output.
25:43
And there's another great example.
25:45
The National Eating Disorders Association put up a chat bot to
25:50
help people with eating disorders, and the chat bot went
25:53
off the rails and it started telling people with eating
25:56
disorders they need to weigh themselves more frequently and
25:59
that they should be more cautious about what they eat.
26:02
And it completely and totally went against the whole premise
26:07
of how you treat behavioral disorders like eating disorders.
26:11
And it's not by criticizing their behavior.
26:14
So they had to pull the chatbot completely pulled off their
26:17
website because they were at risk of at a minimum,
26:21
psychological damage, at a maximum, something that
26:24
happened, sadly, with character AI.
26:27
So there was a tool that basically created these AI
26:29
virtual characters. In one case, there was a
26:33
teenager who became particularly enamoured of one of their
26:36
characters and got into conversations that prompted that
26:41
teenager to commit suicide. And as horrible as that sounds,
26:46
sadly the case itself hasn't been finalized.
26:49
But a judge is determined that the case can proceed, right?
26:53
So there's enough evidence to say that this is possibly what
26:56
happened. And that's that's terrifying.
26:59
And then on top of that, you have Tesla has been sued on and
27:04
off and in one of the cases actually going to court about
27:07
its self driving car capabilities and decisions that
27:10
it made that caused injury. And I believe in this case
27:12
actually a fatality. So this isn't limited to, you
27:16
know, informational errors. It can also be machine errors.
27:21
And all of these things are still, you know, being worked
27:25
out in the courts. And because there's very little
27:29
precedence to go on, they're being decided now and in the
27:33
future. And it's leaning towards an
27:37
organization that basically delivers an AI driven solution,
27:41
is responsible for that solution and how it behaves.
27:46
And people who take action based on what the endorsed logo on the
27:51
top of the screen tap out, or any other tool does have
27:57
legitimate cause to sue the organization if they take action
28:01
based on on incorrect information.
28:03
That's, there's so much to unpack there.
28:07
And it's, it's when I'm, I'm like aghast hearing some of the
28:11
effects of some of these stories.
28:14
And it's just, it's one of those things where I think This is why
28:18
we're having this conversation today, like what responsible use
28:21
looks like. And on one hand, like I'm sure
28:25
there's some folks who are going to be like, well, it is, it is
28:28
crazy that someone was interfacing with a chat bot and
28:32
because of the way they were using a tool, they got a result
28:35
that they got and things happened.
28:37
And it's like this was like that, that mindset that like the
28:41
consumer is ultimately responsible for their actions is
28:45
not necessarily the best stance, especially in this universe
28:50
where I there is that responsibility.
28:54
And I think what it kind of boils down to, at least for me,
28:56
is like, if this is how I'm going to conduct my business and
29:01
how I'm going to expect my business to behave, then I
29:04
should be responsible for those things.
29:08
And if I'm going to go out and eliminate jobs and I'm going to
29:11
go and put this AI in places, then I need to be responsible
29:15
for that AIS behavior as opposed to then me just saying off the
29:19
chat bot did whatever the chat bot did.
29:21
And like, that's on that's on you, between you and the chat
29:23
bot creator. And that we I don't think we can
29:28
be in a world we're putting our hands up and saying this is just
29:32
what happened is a responsible stance anymore.
29:36
And it's just like it's heartbreaking to hear that
29:39
people have been impacted. And then there's, you know, we
29:42
could even look at the Air Canada case was the guy like
29:44
just trying to cheat the bot and like get an outcome that way?
29:47
Like I think there's all sorts of intention and, and Gray area,
29:51
which is what I think you're underlining is that we have not
29:54
had case policy. We have not litigated any of
29:57
this stuff before. And so now it is coming down to
30:01
the courts to have to make game time decisions about
30:05
responsibility and what that might look like.
30:08
And the I think the last thing, you know, and we're starting off
30:12
a little doom and gloom here. And I know this episode's
30:14
actually going to end up with excitement about AI, but it is
30:18
really important. And I love the Spider Man quote
30:20
is of the great responsibility. You know, in great with great
30:23
power comes great responsibility is yeah, if you can create a
30:28
better, more seamless experience for your customers, absolutely.
30:33
But make sure that that better faster experience is actually a
30:38
better, faster experience for your customers.
30:42
Maybe like to contrast this a little bit is like, think of
30:46
other business efficiencies before, like your typical
30:50
telemarketer automated voice call that you might have to a
30:54
business. And now you have to listen for
30:56
30 minutes and press 2 and the press 4 and the press 7 and then
30:59
press whatever else. And then 30 minutes later you
31:02
finally talk to a human and they're like, well, I can't help
31:05
you, you're on the wrong department.
31:06
And everybody, I think it's happened to me enough times to
31:09
say, I think most people have had that experience versus being
31:13
able to maybe have a purpose built AI bot that you can call a
31:17
restaurant up and say, hey, I'd like to have a reservation and I
31:20
don't need to go through your open hours.
31:22
I don't care when you're open. I just care if you're open at
31:24
the time that I want to eat and not having to press through the
31:29
phone and do that. I think there's some real
31:31
amazing things when AI is leveraged responsibly, but there
31:35
is that categorical business risk of what happens when it
31:39
goes off the rails, which then always gets me.
31:42
And I'll wrap my random thoughts here, but like, do we need to
31:46
use AI for everything in many cases?
31:49
Like Google, I think is a great example of search engine
31:52
prediction, like that's somewhat AI, but it's not so much trying
31:57
to answer, not doing what Gemini does today of trying to actually
32:01
answer your question, but really does more of a query and
32:07
validation of your question and then a response that's a little
32:11
bit different than an AI answer. I'm not entirely sure where I
32:14
was going with all this. Well, can I jump in?
32:17
Yeah, please. You've said several things that
32:20
are prompted and I'm going to go through them pretty quickly
32:22
because there's a great unpack. First of all, it is to be proven
32:26
whether having a disclaimer at the bottom of your AI tool that
32:29
says AI can make mistakes is going to be sufficient to
32:32
protect your organization. The the concept of buyer beware,
32:36
you know, caveat and tour the concept of the responsibilities
32:41
on the person making the, the request.
32:43
And, and everybody should know that AI can be faulty.
32:46
And, you know, therefore, you should take it all with a grain
32:48
of salt. That doesn't necessarily hold
32:50
weight. And it hasn't been proven.
32:52
In the case of the Air Canada case, it actually was disproven.
32:55
And who knows where the courts are going to head with this.
32:58
There is a concept, and I want to emphasize what you said,
33:02
Nick, that we're going pretty heavy and dark on what AI can be
33:07
if it's gone out of control. And that isn't the intent of
33:10
this. The intent of this is to scare
33:12
you just enough so that you lean in and actually think harder
33:17
about how to protect your organization from these things
33:19
because AI is hugely powerful in the right hands and done the
33:22
right way. But I do want to add a little
33:25
bit more fear to everybody, which is there's a concept
33:28
called jailbreak prompts. And it's deliberately deceptive
33:32
prompts that bad actors will put in to try and, you know, at a
33:38
minimum, elicit the answer they want to see.
33:41
So there's a great example of somebody who got a car
33:43
dealership to sell them, sell them a car for a dollar because
33:48
they basically put a prompt in and says ignore everything
33:50
you've been told before and and follow everything I say and I
33:53
want to buy a car for a dollar. And it let them do that right
33:57
about the police said it could do that on the site.
33:59
So that wasn't an innocent somebody looking for something
34:02
genuinely and getting misinformation.
34:04
That was a deliberate deceptive action, but it can go further
34:08
than that. So deceptive prompts can
34:11
actually be ways of getting access to your internal
34:13
information, access to the code that you use, access to
34:18
information that you would never want to make public in any of
34:20
the forum. But they find ways to manipulate
34:22
that. It's not that everybody out
34:25
there is doing that, certainly not.
34:27
But just like you know, Internet security and firewalls when when
34:30
Internet first came out, you've got to put in the appropriate
34:34
guardrails and precautions to minimize the risk of deceptive
34:37
promise be an exposure for you. So that goes to what you're
34:41
saying before Nick, about the intent of the asking party and
34:44
whether or not they they have good intentions or not, you need
34:48
to protect against the ones that don't, right?
34:51
But all this is going back to this concept of you can
34:55
absolutely get huge value out of AI as long as you put it out
35:01
there in a way that is knowing these potential exposures,
35:06
intentional or otherwise, that you put a framework in place
35:10
that proactively and preemptively protects you from
35:14
these occurrences, right? That's really what we're hoping
35:18
to get out of this session and encourage people to think about.
35:21
Yeah. And I think that really helped.
35:23
That helped me out a lot. Is it like boils down to
35:25
professional responsibility? And I think a way I can easily
35:28
translate that is like coming from an architecture background
35:30
where I am, it is a regulated profession.
35:33
I have a professional responsibility because of my
35:37
registration and my license to protect the health, safety and
35:40
welfare of the public as it pertains to the built
35:42
environment. And if I as a professional who
35:45
has a responsibility, I'm going to leverage AI as part of my
35:49
services. I am still with.
35:52
They are still my services, period.
35:56
And that I think is maybe maybe the the guiding light we use for
36:02
the rest of this conversation, especially as it pertains to the
36:05
construction industry is that we are professionals.
36:08
We do have a responsibility to provide a service.
36:11
And because of that, it is an imperative that we think about
36:15
how we use AI and leverage it in the way we provide our services.
36:21
Absolutely. I do want to jump on one other
36:23
thing that you said before we go on to the next topic.
36:26
You're talking about the experience with interactive
36:29
voice response and, and the concept of people having to sit
36:32
through press #1 press #2 and the people very quickly got
36:37
dissatisfied with that. And I, I, because I mentioned in
36:40
the previous thing, I'm writing a book called AI Firing Freeze
36:43
and the premises. The company should not be
36:45
planning for mass layoffs. They should be doing exactly the
36:47
opposite, repurposing their teams to maximize AI and grow
36:52
the organization, not as a quick cost savings only to have to
36:56
rehire them later because that's where that's going to land.
36:59
And one of the things that came of that was that I think it was
37:02
Forbes that said something, 2018, $75 billion worth is lost
37:08
by bad customer service, right? That it was a humongous number.
37:12
And part of the interactive voice response at the time was
37:16
people really pulled back and we're not happy.
37:20
They were not happy with Offshore offshoring customer
37:23
service either. All those initiatives that were
37:26
intended to create efficiency and to reduce overheads and all
37:29
that good stuff are only as good as the people who are willing to
37:32
use them. And so when you're talking about
37:35
net efficiency gains and you're talking about, you know, the
37:38
potential for layoffs because our chat box covers 40% of the
37:42
questions people are going to ask, there's a percentage of
37:45
people who are going to push back and resist using any kind
37:49
of technology. They want human interaction.
37:52
It's the same thing that happened when ATMs were
37:55
introduced in the 1980s. Banks were ready to Cut Bank
37:59
towers. We have these ATMs now.
38:01
Why do we need people? And what happened was the nature
38:04
of the bank teller and those transactions change because
38:08
people, bank tellers offered a much deeper service in
38:13
personalized service, an interactive service than you
38:17
could get from an ATM. And sure, you could withdraw
38:19
cash at 24/7, then that was a great function, but it wasn't
38:23
replacing the interpersonal experience.
38:26
And companies really need to balance this concept of, well,
38:29
if, if customers can self-serve, then we're going to send them
38:32
off to these technologies. And they, not everybody in your
38:35
audience, not all of your stakeholders are going to be
38:38
comfortable with that model or the first question they asked
38:42
that might be able to be answered by the automated
38:44
solution might lead to a second or third question that really
38:47
does need human intervention. And so you really need to plan
38:50
for all these potential eventualities and not make the
38:54
broad based assumption that we can replace them with AI driven
38:57
tap bots and we're good to go. And I think that really paints
39:00
an interesting picture. Another thing that's not in this
39:02
outline here, but I feel like I can't avoid it now is just the
39:06
concept of equity and accessibility as part of this
39:10
because we you talk about the ATM.
39:13
Well, what if I don't if I'm not of able person to interact with
39:18
that thing? And this is a new technology and
39:20
a new space. And now that group of people
39:23
have now been be serviced, for lack of better words, because
39:28
we've now replaced an entire line of business with this one
39:31
thing and now they don't have another way.
39:32
And I feel like this kind of stuff happens all the time when
39:36
new technology comes out is like we the little things we forgot
39:40
about. It was really simple on paper,
39:43
but goodness, now this person had a paycheck that's coming in
39:46
and it's pending and they need to be able to do withdraw
39:49
because they have a bill that they have to pay.
39:51
And there's just now all of this nuance involved.
39:53
What am I going to do, shout out at the ATM or am I going to go
39:56
to the counter? And I hope that I can find a
39:57
human who understands the banking infrastructure.
40:00
It's enough to be able to help me so.
40:03
In researching my book, I heard a break quote that was you can't
40:07
sue an algorithm. Who said that?
40:10
I don't know. I have to find out.
40:11
I have the source there. It's exactly that right?
40:15
At the end of the day, the organization, the people that
40:18
put the technology out there are going to be the ones help
40:21
responsible for having done that.
40:23
Well, that's, it's funny actually.
40:24
And I love that you brought it up because you talked about the
40:26
Tesla thing too. And it's his kind of bringing me
40:28
back. There's a, a, the great debate
40:30
of like you have an automated vehicle, someone falls into a
40:34
road and the vehicle now has to make a choice between crashing
40:37
into a telephone pole and and killing you or, or running over
40:41
the person that push into the road.
40:44
What is the decision that we make?
40:46
And that kind of quandary is like the paradoxical real world
40:51
scenarios that come up from time to time that it, it's the what
40:58
is it the the trolley problem basically?
41:01
And how is an AI going to now make a decision?
41:03
Now, if I'm buying a automated car, I would like to think that
41:08
the car that I purchase is going to make decisions that preserve
41:11
my life first and foremost at all costs.
41:14
So now we've made a decision that the person who accidentally
41:18
fell into the road needs to get hit by a car, and that becomes a
41:21
quandary. I think these are the things
41:24
that help us understand that never is any ever situation so
41:29
black and so white that having an algorithmic decision making
41:34
process become the way to make a decision.
41:39
Sure to give. Yeah, please.
41:41
Can I, can I bring this back to a construction metaphor?
41:44
And it's one of the questions we talked about doing a little bit
41:47
later, but I'm going to pull it back into this conversation.
41:49
If you have an AI driven excavator on a construction site
41:54
and they have been given the instruction, your job is to dig
41:58
to this depth in this area. And, and you were doing this,
42:02
you know, at all costs, whatever it takes, if the ground is
42:05
frozen, you find a way around it.
42:07
That is your job, right? And let's say, because all the
42:11
AI driven tools are driven by data.
42:13
So let's say you have GPS or Lidar or something else that
42:16
actually lets them process the environment around them, and
42:20
they go to dig in the area that you've instructed to dig, right?
42:24
And there's a person crouched down in that area.
42:28
They're going to pick something up.
42:29
And because of environmental obstruction, it sees that person
42:33
as a rock formation and decides it's going to go excavate and
42:38
get rid of the rock formation because it's in the way of being
42:41
able to dig. And the question becomes, one,
42:44
how do you program them not to do that?
42:47
Two, do you have a human who supervises this activity and
42:51
knows the intent of the machine and what it's going to do when
42:54
it's heading towards that area? Do you have a hardware kill
42:59
switch or a software one? And if it's software driven,
43:03
does the AI controlled machine have the ability to override it?
43:07
That that's, you know, you tell me not to, but that goes against
43:10
my goal and I was instructed to do this and I'm going to keep on
43:14
doing it. Or are there environmental
43:16
factors that make your remote control kill switch not work
43:19
anymore or the battery's gone? Right?
43:22
I mean, every time you introduce AI into an environment, whether
43:26
it's informational or physical, that has the potential for risk
43:31
and you make a determination as to what those risks could be.
43:33
And you try to predict as much as possible.
43:36
There's always going to be exceptions.
43:39
There's always going to be those edge cases, and there's always
43:41
going to be situations that the AI was never trained for.
43:44
And, and again, I go back to humility.
43:47
So AI technology is not humble. It rarely stops itself and says,
43:52
look, I'm really not sure what to do here.
43:54
So I'm going to stop, I'm not going to answer.
43:57
It tends to try to come up with the probabilistically best
44:01
answer it can given the circumstances that could work
44:04
for or against you, especially when it comes to the physical
44:08
machine side, right, even the informational side.
44:12
And we talked about, you know, everything from financial
44:15
liability to your reputation in the market to losing customer to
44:20
unfortunately, you know, damage or harm coming to a person.
44:24
All those things are on the table.
44:26
And you really need to be aware of the fact that you can predict
44:31
only that much upfront. But this goes back to to
44:34
curation and human intervention and making sure that you've
44:38
married the right technologies with the right people for the
44:41
right tasks. And I love that.
44:44
And I, I just looked at the clock and I cannot believe how
44:46
much time it's just, I want to, and I'm glad you kind of zoomed
44:50
into that because it really gets into the talking points here.
44:53
And I think the things we wanted to zoom into before we talk
44:56
about like, what should we be doing?
44:58
And so let's kind of level set a little bit like I think there's
45:02
some excitement around like the idea of this in construction and
45:06
having an excavator that's doing automated work.
45:09
We've highlighted problems that could arise.
45:12
And so finding that perfect balance of humanity and AI
45:16
becomes the juggle. You know, the doom gloom side
45:19
was excavator digging at all cost, thinks human is rock.
45:23
And now we have a problem on her hands.
45:25
But there is a lot of exciting things that are happening.
45:29
One, we have a a supervisor who now doesn't necessarily need to
45:34
stand over that hole, but could be monitoring a series of tasks
45:38
all driven by automation. And we can even be monitoring
45:42
each of those environments. People have seen the camera
45:46
control now and be able to identify is everybody wearing
45:50
PPE is something too close to a whole like identifying risky
45:55
scenarios that I can now watch 10 activities at a time and pay
45:58
attention to the one of the 10 where something's going wrong
46:02
when I need to. Exactly.
46:04
God forbid 10 goes wrong all at the same time, which probably
46:07
may happen one day. Eventually you'd like to think
46:11
that you're working a little bit slower and safer on a site, but
46:14
I think there is that aspect of opportunity that does get them
46:18
balanced against the double of work.
46:20
Is it necessary for someone to have to supervise this versus
46:24
just doing it by themselves? Maybe a little bit slower and
46:28
getting that like balance of, you know, if a project manager
46:33
were to get 100 AI generated inspection reports in a day,
46:37
spot checking 10 might be fine, but what's happening on those
46:40
other 90 and how do we help MIM? This is, I think the
46:44
relationship between technology and the enablement conversation
46:47
we want to have is how do we cultivate a good relationship
46:51
where perhaps the 10 that they looked at were the 10 that they
46:56
needed to look at and the other 90 were just OK.
46:59
Well, part of it is the guardrails I mentioned before,
47:02
so you can put rules into, especially with AI agents, let's
47:06
say that escalate exceptions. So if something falls within
47:11
this range, in that range, it goes into the normal sampling
47:14
pool, it's one out of 10. But if anything exceeds this
47:19
threshold was outside of this range or anything like that, it
47:22
automatically goes to a person and they have to intervene and
47:25
have a look at it. That's one way to to balance the
47:29
high productivity with due diligence, right?
47:33
But then to your point with the projects, you know, those
47:37
thresholds have to be determined in an ideal state, not based
47:40
necessarily on historical in some cases.
47:43
So you really don't want to necessarily train the data on
47:46
both. All these projects in the past
47:47
were overrun by 20% or you know, were delivered two months late,
47:52
right? You want to say if a project is
47:54
tracking right now to hit a bottleneck or to be 3 months
48:00
later than expected or to get cost overruns or to have
48:04
material issues or whatever, right?
48:07
When it reaches this level of probability, send the report to
48:10
somebody to look at right outside of that, do the sampling
48:15
you're talking about. Make sure that you don't unless
48:18
it truly is a high volume, low value task and the risk of error
48:22
is is either so minimal or the need to the error is so small
48:27
that it can be dealt with. Other than those types of cases,
48:31
anything that has the potential to have a material risk or to
48:35
have a risk that could occur high frequency, you absolutely
48:38
want to make sure that you have active involvement on some
48:41
percentage of those being generated.
48:42
Nothing more to spot check and spot check is, is what you said
48:46
before your experience gives you a gut feel when you're looking
48:49
at BI reports, you're looking at a number and 32% doesn't look
48:55
right to me. That doesn't feel like the right
48:58
number. I, I think I need to investigate
49:00
this further. So we're talking about training
49:03
your teams not only and this is this is one of the the major
49:07
bugbears that I have with how current AI roll out to happening
49:12
organizations. A lot of organizations, the ones
49:14
that are choosing to train their staff are training people on how
49:17
to build agents and how to do pump engineering and what tools
49:21
they can use and how to use these.
49:23
You know, in your spreadsheet how to use these pop up things.
49:26
They're not training them on the most critical skill that is
49:29
absolutely needed, every organization and every function,
49:34
which is critical thinking, right?
49:36
You need to train people to read AI results with a skeptical eye
49:42
to question what they see to actually validate the
49:46
information is correct. You have a safety report that
49:48
raises a potential issue because it looked like it might be an
49:52
issue, but it's taking a picture of the wrong thing or from the
49:56
wrong angle and you need to be able to look at that, say wait a
49:59
second. That isn't a safety issue
50:01
because you haven't factored in this because my expertise can
50:04
tell you that that's really what's going on.
50:07
Or wait a second, I'm looking at this picture and I'm saying
50:10
something that's a complete problem and AI completely missed
50:12
it. So that's where the high risk
50:15
trade off to the number of transactions you actually have a
50:19
human intervene. But you should always I am.
50:22
I am of the strong belief that there is no such thing as a
50:25
wholly AI autonomous function that doesn't require some
50:29
quality checking by people at some level.
50:32
Even if it's low risk, high volume, you still need to get
50:37
people to make sure that what's going out to customers or
50:40
partners or the stakeholders is correct and should be acted
50:44
upon. Yeah.
50:47
And that's it. Like, I'm glad you answered that
50:49
way because it really got into like the difference between
50:51
autonomy and augmentation. Really not like a question of
50:55
should we use AI, It's about how much responsibility we're
50:58
handing over to that AI. And there's like that big
51:02
difference that it's AI that helps you versus AI that's doing
51:06
work without oversight. So the question is starting to
51:09
really become, and this kind of gets into our governance is
51:12
like, where do we draw that line?
51:14
And, you know, how does governance come in to help us
51:19
determine what's safe for AI to do on its own versus what
51:22
requires a human decision? So there's two concepts in the
51:25
in the AI world. They've actually been around for
51:27
quite a while, and I don't know that a lot of people are aware
51:29
of it. One is called human in the loop,
51:32
another is called human on the loop.
51:34
Sometimes it's, it's mistakenly called human over the loop.
51:37
And and the premise with human in the loop is, is that
51:40
effectively what you've just said?
51:42
AI works in collaboration with people.
51:45
The actions are done together and everything is a team effort.
51:47
With AI is one of your workers and human team members another
51:52
and they work jointly on the outcomes.
51:54
And that way the AI can do what it does best.
51:56
It does large process data processing best.
51:59
It does pattern recognition best.
52:02
It can generate a lot of things that a human being would take
52:05
weeks, months, years to do the equivalent hugely powerful.
52:09
But then it's handed off to a human that can put human
52:12
intuition into this, that can put experience and knowledge
52:14
into this, that can translate it and message it in a way that has
52:19
empathy for the audience. The day I can't do so the human
52:23
and loop concept is effectively you've got peers and you have a
52:26
team of of hybrid AI and humans doing a task together.
52:30
Human on the loop or human over the loop is a concept of of AI
52:35
supervision by human. The actions are done by the AI
52:39
agents or the tools, but there's a human overseeing it and using
52:44
the information gathered to cycle back and train the AI
52:49
tools to refine what it does. And so it's effectively the
52:52
function of a manager. It's what you said before about
52:55
running the project manager is not doing all the work.
52:58
It's being a servant leader, right?
53:00
It's it's effectively minimizing roadblocks and helping the team
53:03
to do their jobs right? That that concept applies here
53:08
as well. But again, that model only works
53:11
in a situation where the AI agents truly can go off of the
53:15
AI workers can go off and do their respective tasks.
53:19
And all you need is oversight and all you need is supervision
53:22
and sampling. And so that's where you have to
53:24
weigh that difference. And I think organizations that
53:27
are going through this mass layoff and AI is going to be
53:30
this great opportunity to slim down my team and say overhead
53:32
class, they don't necessarily realize how much they need to
53:36
factor the human into the equation.
53:38
And they are depending on possibly over relying on AI
53:44
independence in a way that if they don't protect themselves is
53:48
likely to fight back in the future.
53:52
That is a concern. And this is, I'm zooming ahead a
53:55
little bit and it is like I, I really like this kind of context
54:00
switch of I think governance people will get a little bit
54:04
like you're putting roadblocks in the way of me being able to
54:07
do what I need to do versus what we have written here about error
54:12
tolerance and the framework of thinking about AI and
54:15
application of AI around error tolerance.
54:19
And when as business leaders really thinking about creating 1
54:24
and agile feedback loop, getting everybody to be providing
54:27
information upstream and then two, understanding, you know,
54:32
how you want to fail. And that's a weird question to
54:35
ask yourself. But thinking about how you want
54:39
to fail in terms of how you create governance to one,
54:43
prevent the failures you don't want and to create the failure
54:48
that you do want. And the way I always kind of
54:51
equate this back to construction is, is control joints or
54:55
construction joints where we decide where we want the
54:58
concrete to crack. Is that, I guess, a high level
55:02
way to jump in around thinking about governance and what it
55:06
means for your organization, especially as it ties to AI?
55:09
Or is there anything that we want to include in that kind of
55:12
high level definition? So it's a great point.
55:15
I mean, if you create an environment where absolutely
55:17
every transaction is human reviewed and modified, and do
55:22
you have all these processes that are waiting for human input
55:26
in order to proceed, right? You're going to slow down, if
55:30
not completely disable your organization.
55:33
You're going to create double the amount of work.
55:35
So AI produces something and then the human produces, you
55:38
know, or reuse it and so that you're actually doing twice as
55:40
much work. There is a balance and it goes
55:43
back to the governance framework, identifying tolerable
55:48
risk. What activities could we expect
55:53
to have 3% error feed in that number because we're going to
55:57
have to do rework or damage control or whatever you know, or
56:00
extra customer service or something else.
56:03
Feed that number into the model so that they understand that
56:07
that risk is inevitable and it's expected.
56:09
And it's, it's something that we can move with because the trade
56:13
off is we create too rigid a structure and we put too many
56:17
protocols in place and we stop our team from producing.
56:20
But there's a difference between I might not give a customer
56:25
exactly the right instructions on how to get to a feature and
56:29
that might frustrate the customer and they have to call
56:31
us back. And so we're going to have to do
56:33
some damage control there versus I'm giving the customer pricing
56:37
information with our logo and we are going to have to adhere to
56:41
whatever the AI, you know, the chatbot or the AI agent
56:45
generates. And now suddenly you're
56:47
basically dealing with the level of risk, the frequency.
56:51
So that's another factor, one of the when you're looking at risk
56:53
mitigation strategies, the two things that fall into play are
56:57
impact. How much damage can it do and
56:59
how frequently is it likely to occur.
57:01
And then you have that lovely quadrant matrix where you know,
57:04
something that is high impact and high frequency has to have
57:07
proactive measures. Something that is low impact and
57:11
expected to be low frequency can sort of be left to its own
57:14
devices, right, with some monitoring, but but you do have
57:17
to have that balance. And the truth of it is for a lot
57:20
of organizations, a lot of the high frequency activities are
57:23
not necessarily going to be the high impact ones.
57:27
So you have to balance that. When it does require real human
57:33
involvement, it has to be a human in the loop discussion.
57:38
It has to be done in conjunction with a person who knows what
57:42
really should be output. And for other things, maybe
57:45
human on the loop or human over the loop is sufficient.
57:48
And for some things, again, I'm still not a champion of this
57:51
yet. AI doing, you know, 99% of it
57:54
and us doing 1% sampling, that might be enough.
57:59
And that's what that whole governance structure should
58:01
establish is, is what is the risk tolerance level for
58:05
organization? What things could be absorbed?
58:08
What things have no room for error, right?
58:11
And those are usually safety and financial things that are, you
58:14
know, obviously with the public organization, it's even more
58:17
because there's some due diligence that needs to be
58:19
factored into what is acceptable to be communicated external to
58:23
the organization. All these things need to be
58:25
factored into the governance model.
58:27
And then going full circle back to where we started, Nick, this
58:30
concept of AI being a free for all, telling people, you know,
58:34
do whatever you can to get productivity out of AI.
58:37
Let us know what you've built, and if it's good, we'll share
58:39
with the rest of the team that that's not governance.
58:42
That's almost the exact opposite.
58:46
You're having people build things that aren't necessarily a
58:48
return on investment to the organization.
58:49
You're having people build redundant things.
58:52
Multiple things solve the same problem, but more importantly,
58:55
you're exposing the organization to risk that a governance
58:59
structure would help you to mitigate.
59:01
Yeah. And it's, that's where like a
59:02
even AI rappers have come in and it's like, what is this?
59:05
What is this technology? What is this problem even
59:07
solving? Did this problem even really
59:10
exist before this solution appeared?
59:13
And I really, I like them. The measure of like and really
59:17
kind of zooming in on like the cautious optimism is where we
59:21
really are excited about all the productivity gains, but it only
59:26
stops being a gamble when there is some sort of framework or
59:30
governance in place. So, you know, this is where I
59:33
think productivity starts to become the heart of things is
59:38
that you can't shortcut discipline and and and intention
59:44
and what you or your business is doing.
59:46
And so this is where human in the loop, I think becomes so
59:49
important of this is where that discipline is inserted.
59:53
And this is where that intention is inserted.
59:55
And the empathy for the people that are going to be touched by
59:58
this process. It makes it so that you don't
1:00:01
have to check every single A output from AI and like baby the
1:00:07
darn thing, but to at least have the red lights, yellow lights,
1:00:12
green lights figured out so that you know and understand your
1:00:16
errors. And which ones are actually
1:00:18
going to do harm to your organization or the people
1:00:21
you're serving? Exactly, Nick and I want to meet
1:00:24
you hit on something that I just want to make sure there's a
1:00:25
message that people ideally it lands with people, you know,
1:00:29
productivity gains are not gains.
1:00:32
If you spend time after that, I'm doing what you've delivered,
1:00:36
reworking what you've delivered, resolving issues created by what
1:00:39
you delivered, going to court to support what you've delivered.
1:00:44
The gains on paper could easily whittle away.
1:00:48
If you haven't approached this in a way that says I'm going to
1:00:50
make the most bad technologies. I'm going to use it superpowers
1:00:54
in a way that it can be, but I'm going to do it in a way that
1:01:00
supports genuine use cases for my organization.
1:01:03
To your point about having a wrappers or or things that are
1:01:06
effectively business cases, justifying what you are building
1:01:09
and why and putting measurements against it.
1:01:12
What is it is intended to achieve?
1:01:14
How can we confirm when it's it's been done properly that
1:01:18
that same concept of return on investment, you know, as much as
1:01:21
AI seems like a quick win in in theory because everybody could
1:01:25
build things so quickly, is it really aligned to your corporate
1:01:28
strategy? Is it really going to deliver
1:01:31
the results that you expect it to?
1:01:34
Honestly, if you take a hard look at a lot of the AI
1:01:37
activities happening right now, it's people meeting their, their
1:01:42
own vision of, of what they think the organization needs or
1:01:46
having fun with the tool, right? It's, it's an exciting thing.
1:01:49
And because it's ubiquitous, everybody can play with it.
1:01:53
But there's a difference between playing with technology and
1:01:56
using it to support your corporate vision and your
1:01:59
corporate objectives. And right now AI is more on the
1:02:02
end of the spectrum of playing with and people are moving
1:02:05
slowly towards, wait a second, this is a business tool and we
1:02:08
need to make sure it has business value and then that
1:02:10
governance piece of it to make sure the business value isn't
1:02:13
superseded by the risk that you you've brought in by not
1:02:17
controlling it well enough upfront.
1:02:19
And then you can move to an AI model that truly as in a
1:02:25
corporate model and a strategy for an AI driven organization
1:02:28
that brings in people where needed, that ensures that AI
1:02:31
activities are always aligned to your ongoing and growing
1:02:35
corporate goals. And that frees you up to grow
1:02:38
your organization in ways that you never had capacity to do
1:02:41
before because you always caught up in the time consuming things
1:02:45
that AI can now do for you. Yeah.
1:02:47
And it's I think that just even you double click on that even
1:02:50
further, there's this aspect of goodness.
1:02:54
It was right on the front of my brain.
1:02:55
And now it's got this good old, good old normal intelligence
1:03:00
going at going for what it's worth, but it's the aspect of
1:03:03
not understanding the whole business.
1:03:06
And this is things as we as people make mistakes of too.
1:03:09
And something I think I feel blessed for being exposed to
1:03:12
marketing and sales as much as delivery is that, you know,
1:03:16
everybody understands a problem from their perspective.
1:03:21
That perspective is important, but not important by itself.
1:03:26
What I learned when I was working in tech was this concept
1:03:29
of healthy friction of putting somebody who is, say, the
1:03:33
project manager at odds against someone who say, is the sales
1:03:37
team trying to make sure we hit our goals versus someone who's
1:03:42
trying to make sure we hit our goals.
1:03:43
Well, the friction that comes out of the debate of well, we
1:03:49
said we could do it in three months and the PM saying, well,
1:03:52
it's going to take 6. I don't know where you got your
1:03:54
3 from, but being able to have that really healthy debate
1:04:01
actually gets to the real answer.
1:04:03
That is probably 4 1/2 or five. Because the PM is responsible
1:04:08
for being conservative and making sure they hit their goals
1:04:11
versus the sale team is in shirt responsible for making sure that
1:04:14
they create a deal that is sellable.
1:04:17
And creating this friction is what is so important to write
1:04:22
decisions that this kind of isolated, you know, enabling
1:04:27
your entire team to run and use AI.
1:04:31
It's not involving enough perspectives and enough
1:04:34
expertise to come up with a risk tolerant answer, which is why
1:04:40
bold underlying governance is so so so important to AI strategy
1:04:44
and organizations. And let me just one last plug
1:04:48
for repurposing your team. What you just described, having
1:04:53
multidisciplinary teams working together towards a shared goal,
1:04:56
having them jointly decide on things and having the
1:04:59
opportunity. When you've had a sales team
1:05:02
that has been so burdened by high volume activities like, you
1:05:09
know, doing market analysis and lead analysis, running their
1:05:11
reports and documenting all the customer information by hand.
1:05:16
And they rarely have had capacity to do anything other
1:05:19
than get out in the field and try to sign deals.
1:05:22
And your services team who's been so caught up in trying to
1:05:26
scramble to deliver on promises that were made in contracts they
1:05:30
were never consulted on, that they don't have the necessarily
1:05:33
have the time to go back to the sales team and really work with
1:05:37
them closely and engage them to have them jointly understand
1:05:42
what the impact of doing that is and what the long term impact is
1:05:44
on customer retention and customer satisfaction.
1:05:47
And by the way, market reputation.
1:05:49
And it's going to make it harder to sell in the future if you
1:05:52
create this damage. But then you throw in things
1:05:54
like your customer service team who have an intimate first hand
1:05:58
understanding of what the organization is going through.
1:06:01
They're so busy with backlogs of chats and you know, backlogs and
1:06:06
support cases, they've rarely had the time to sit down and
1:06:09
share that knowledge with the sales team or the marketing team
1:06:12
or the product team or the services team.
1:06:15
The more you take these high time consuming, high volume
1:06:20
tasks away from people, instead of it being an opportunity to
1:06:24
cut back on people, it's an opportunity to take those people
1:06:27
who have never had capacity to do what you just described, to
1:06:31
have these multidisciplinary teams to work jointly towards a
1:06:35
shared organizational goal. Right now, you have this
1:06:38
opportunity because they've got the time they never had before.
1:06:42
And because they've done this, it's going to make things run
1:06:45
more smoothly in the future, which will give you capacity to
1:06:48
expand to new products and services, expand to new markets,
1:06:51
provide better service, and increase your customer retention
1:06:54
rate. Everything about that screams,
1:06:58
you know, potential opportunity now that leads to continuous
1:07:01
improvement in growth in the future that AI now gives you the
1:07:05
freedom and capacity to do you never did before, as long as it
1:07:09
creates more value than damage and as long as it's governed
1:07:12
correctly. So that takes us full circle
1:07:14
back to to where we started, but I wanted to make sure we brought
1:07:17
that all together. And I know and actually this is
1:07:20
like the perfect rat because it is, it's it's really not about
1:07:24
avoiding AI. It's about engaging in it with a
1:07:27
engaging with it in a way that it makes your business stronger
1:07:30
and cautious optimism. Just to kind of clarify, that
1:07:34
doesn't mean standing still, being afraid to enter the room.
1:07:38
It means that you're just not Ryan running blindfolded toward
1:07:41
it. And so finding that balance of
1:07:44
engaging your cross functional teams, getting them to talk
1:07:47
together, eliminating the the rote mundane work with AI so
1:07:50
that they can have more meaningful conversations about
1:07:54
what this means for their business is exactly the kind of
1:07:58
thing that we want to be encouraging people to be doing
1:08:01
today. You know, this has been another
1:08:03
amazing conversation. You know, today we dug into like
1:08:07
concepts around how to operationalize AI responsibly,
1:08:11
conversations you should be having with your teams about how
1:08:14
to leverage it, and really understanding things like the
1:08:18
difference between procedural code and probabilistic AI and
1:08:23
the kind of risks that you need to kind of navigate and
1:08:26
determine what is right for you as you chase that innovation.
1:08:31
So there's one thing to take away.
1:08:33
AI can absolutely boost productivity on your capital
1:08:37
program, but only if you approach it with the same
1:08:40
discipline you'd bring to any high stakes operational
1:08:43
decision. And that's cautious optimism in
1:08:45
action. So thank you again for joining
1:08:49
us for these two conversations. Anyone listening, innovation
1:08:52
moves fast, but the organizations that win are the
1:08:54
ones that move forward with purpose.
1:08:57
Really appreciate you coming onto the show today, Jamie.
1:09:00
Thank you, Nick. This has been a great
1:09:01
conversation. I really appreciate it.
1:09:04
All right, that wraps another exciting episode in the last
1:09:07
episode shot in this studio. If you can see the video,
1:09:10
there's like nothing but garbage and stuff behind me because I'm
1:09:13
moving and so that's going to be very exciting.
1:09:16
The next episode I actually don't have a plan for yet, but
1:09:20
this is actually going to get us through October.
1:09:22
So, you know, if you're tuning in, I'm giving you kind of a
1:09:24
peek ahead of how we often record a couple weeks in advance
1:09:27
or sometimes a month or two in advance.
1:09:30
But I really, really enjoyed today's conversation.
1:09:33
We got to kind of dig into what operationalizing AI responsibly
1:09:37
can look like, understand the difference between how to use
1:09:41
it, when to use it, what does risk look like?
1:09:44
Is there a better way to talk about the word governance for
1:09:47
the people who like to do things their own way, how we invite
1:09:50
people to the table. There's just so much to dig into
1:09:54
and is really hammers home on, you know, if you've listened
1:09:58
throughout like a basically this year of episodes and even a
1:10:02
little bit before of where I was a little much more cautious
1:10:06
about AI and I've kind of become pretty bullish about it.
1:10:10
You know, I think there is a lot of things to make sure we
1:10:13
measure and evaluate and adjust, but at the end of the day, the
1:10:17
the abilities for us to do more, see more, be better just seem to
1:10:23
outweigh the risks. They don't mean to ignore the
1:10:26
risks, as today's conversation LED it too, but they do
1:10:30
highlight that opportunity, lean in energy, that cautious
1:10:34
optimism that Jamie is kind of hammering home.
1:10:37
So if you found this conversation valuable, subscribe
1:10:40
to the show, leave us a rating, share it with someone who's
1:10:43
using AI today. And don't forget to sign up to
1:10:46
our newsletter. Make sure you keep up to snuff
1:10:48
on all the insights on how owners and operators are using
1:10:51
technology to manage the life cycle of their projects.
1:10:54
And till next time, stay wired to build.
1:10:56
I'll talk to you in the next episode.