Skip to content
Exploring Machine Learning, AI, and Data Science

Why Mike Dymott Slows His AI Agents Down on Purpose

Welcome to another episode of Data Driven, where we dive into the evolving landscape of artificial intelligence, data science, and the critical importance of data engineering.

In this episode, the conversation focused on the journey from traditional CRM systems to AI-powered business solutions, emphasizing why solid data and well-designed processes are more crucial than ever. One concept discussed was the orchestration between humans and AI agents, highlighting the need for seamless handoffs and the risks of letting agents loose on bad data. The discussion explored lessons learned from digital transformation projects, the pitfalls of simply automating broken processes, and the real-world impact of AI on customer experiences.

A key theme that emerged was the shift towards industry-specific and distributed AI models, along with the increasing importance of resilience, security, and responsible implementation. Tune in for practical insights, war stories from the field, and actionable advice on building trustworthy, effective AI-driven organizations.

Links

Time Stamps

00:00 Leading ISAM with data focus

06:44 Understanding Problems in Hackathons

10:25 Challenges in Business Transformations

11:45 Discussing digital transformation challenges

14:59 Handling Failures in Complex Systems

20:24 Testing with production-like data

24:11 Blame and responsibility in AI errors

25:13 Responsibility in AI implementation

28:07 Security and data center trends

32:24 Repurposing Cold War data centers

37:44 Discussing CEO risk and resilience

39:46 Bill Gates’ ongoing influence at Microsoft

42:15 AI in Motion podcast details

Transcript
Speaker:

In fact, pacing your agents and slowing them down is what we've done in

Speaker:

some of our implementations. You actually introduce

Speaker:

a lag, even though the agent can do it fast,

Speaker:

a deliberate lag. Really? So how do

Speaker:

customers react to that? Do they, are they— Well, actually

Speaker:

positively, because if it goes too fast, they know AI is

Speaker:

involved. Then they don't feel special. It's not

Speaker:

the economy this time, it's the data. Mike Dimond of

Speaker:

ISM on why giving AI agents bad data and bad

Speaker:

process is a recipe for disaster. This is Data Driven.

Speaker:

Hello, and welcome back to Data Driven, the podcast where we explore the emerging industry

Speaker:

of artificial intelligence, data science, and of course,

Speaker:

data engineering, because without data engineering, none of this is really possible.

Speaker:

However, my favoritest data engineer in the world is unable to make it today, so

Speaker:

it's just yours truly here, coming back from KubeCon

Speaker:

2026. And I'm talking to Mike Dymott.

Speaker:

Hopefully I pronounced that right. Yeah, that's right. How are you doing today? Awesome.

Speaker:

So welcome to the show. And you work for,

Speaker:

you're in the UK, but you work for a local company, as in local

Speaker:

to me, in Maryland. So tell us about your work at ISM.

Speaker:

Yeah. So ISM is about a 40-year-old business.

Speaker:

It began life in the early days of

Speaker:

CRM and led by Barton Goldenberg, who was a

Speaker:

very dear colleague and later friend. And

Speaker:

again, he came, I basically, I was a client over a

Speaker:

stretch of about 25 years, beginning and end of those 25

Speaker:

years, I was a client and always a friend, I'd say.

Speaker:

Barton was a true visionary. He moved along with the times

Speaker:

and is greatly missed. Interesting. So

Speaker:

how does your work at ISM intersect with AI? and their

Speaker:

impact on business systems? So when

Speaker:

I took over in May with ISAM, largely in the

Speaker:

leadership role, I actually said to the colleagues

Speaker:

at ISAM, without AI added to CRM, they didn't

Speaker:

really have a consulting business anymore. So I see

Speaker:

it as really being an extension of CRM

Speaker:

because we always looked at business process in ISAM.

Speaker:

People, process, technology, and I've added data to the method that

Speaker:

Barton wrote in many of his

Speaker:

writings, presentations, books. It was people, process,

Speaker:

technology, but you guys, you've said it already,

Speaker:

it's all about the data. And in one of my recent

Speaker:

communications, I used the phrase nicked from

Speaker:

the Clinton era, it's not the economy this

Speaker:

time, stupid, but data. That's funny. I like

Speaker:

that. That needs to be on a t-shirt or a bumper sticker.

Speaker:

Yeah. So we're very much of the CRM

Speaker:

era with a number of other things that Barton introduced to his business

Speaker:

over time. But AI, business

Speaker:

process, yes, focused on customers, but

Speaker:

very much vendor independent. We'd like to stay that way. Over

Speaker:

time, we may well have some partnerships or are looking at a few,

Speaker:

but always Vendor independent. There's a

Speaker:

certain power in that level of neutrality, right? Because you can pretty

Speaker:

much walk into any building and any customer and not really have

Speaker:

a particular, I wouldn't say not have a particular bias, but you're

Speaker:

not, you're pretty neutral in terms of you're about becoming

Speaker:

part of the solution rather than a branded solution, so to speak.

Speaker:

What are companies most excited about AI right now? I

Speaker:

would assume it's agents. Yeah,

Speaker:

agentic as opposed to just simply chatting or using

Speaker:

it. Too many people are still using it probably like it's

Speaker:

Google. Right. But yeah, very much the,

Speaker:

for me, it's all about the handoffs. So you may well

Speaker:

have a process that begins with a human, moves to an

Speaker:

agent, comes back to human. That whole orchestration is

Speaker:

really fundamental. I think orchestration is gonna become more and more

Speaker:

important. Yeah. When you say orchestration, orchestration has a

Speaker:

lot of different contexts, but I assume you meant orchestration between

Speaker:

agents and humans. It's the, if you like,

Speaker:

you could describe it as a dance. So it's a dance between

Speaker:

the human, the customer, the user, and the agent. And that

Speaker:

flow has to be really smooth and you have

Speaker:

to understand the handoffs. You have to understand your process.

Speaker:

And if your data's rubbish, you are really gonna have a bad time.

Speaker:

Makes a lot of sense. And it's very much about,

Speaker:

this is why at ISM we focus a lot on

Speaker:

looking at process. You could, as in the old

Speaker:

days with CRM, take a bad process and all you're gonna do is do it

Speaker:

faster. Yeah. So that was

Speaker:

bad then, you know, on a focus then.

Speaker:

It's even more important now because it's on steroids.

Speaker:

You can't get away with just designing your current process

Speaker:

into the agentic era. You have got to start from basics

Speaker:

and understand the journey. No, I think that's a good way to put it

Speaker:

because it's one thing to put together a very impressive AI

Speaker:

demo, but it's quite another to

Speaker:

make it, put it into production and have it add real value to the

Speaker:

business. What have you seen customers learn that

Speaker:

experience, right? Because I'm pretty sure that there's been a lot of cases where

Speaker:

people will vibe code something and then be like, oh, let's deploy this. And then

Speaker:

they realize that it's not so straightforward.

Speaker:

Are you talking to Mr. Vibe Code? I've never really been a commercial

Speaker:

developer. I only ever did it at university and at school.

Speaker:

But I'm quite hooked on it

Speaker:

myself. It's very good for prototyping. You still need to have a senior

Speaker:

developer, certainly a senior one, take a look at your code and

Speaker:

certainly check out vulnerabilities, but it does have its place.

Speaker:

And you can drive quite a lot of imagination and

Speaker:

innovation by vibing because I mean, I've seen it myself

Speaker:

and I've had some very good results. But one of the key things

Speaker:

you hit upon a very, you can tell, a passionate topic of mine.

Speaker:

It works well when you understand the problem. If you

Speaker:

don't, I was at an event recently and the people

Speaker:

in the room, quite a lot of experienced IT people in the room, having a

Speaker:

go at a hackathon with a tool. And

Speaker:

quite a few of them didn't have any real results to show after an afternoon

Speaker:

of vibing. And yet others were

Speaker:

producing some really impressive output. And I

Speaker:

think the difference is how you're describing it and whether you actually understand the

Speaker:

problem. But it does have its place and it's going to

Speaker:

increase. Let's not get— we won't get away from it. In fact,

Speaker:

what I haven't said about my background is I worked in corporate business

Speaker:

for many years and managed development teams over most of those

Speaker:

years. And it's fair to say that we were being pushed

Speaker:

very hard towards the end of my time in the corporate scene to use AI

Speaker:

in our development of product. Very, very hard

Speaker:

push. And the whole token maxing

Speaker:

question then comes to mind, although I think that's diminishing

Speaker:

quite a bit now. I think people have got budgets now

Speaker:

rather than just a free-for-all. But yeah,

Speaker:

being measured on, are you actually using AI to

Speaker:

increase your productivity? I think that's where developers are.

Speaker:

And it's interesting because there's still developers that are fighting it, but I

Speaker:

don't see it as necessarily a competitor. I see it as

Speaker:

a way to help accelerate my thinking, right? I

Speaker:

think you'd kind of, we kind of danced around this, is like, it has

Speaker:

to get back to the solutions, right? The fundamentals of why we're in IT,

Speaker:

right? Like why businesses, you know, employ

Speaker:

people in IT. It's about the solution. I will borrow

Speaker:

your Clintonism, right? Or it's like, it's about the solution, stupid, right? Like—

Speaker:

Absolutely is. It's either going to add value

Speaker:

or not. And if it doesn't add value, then it

Speaker:

could look cool, but at the end of the day, it's not worth paying for

Speaker:

it, whether it's human or AI-based or some

Speaker:

combination. Well, I quite often laugh when I go to a lot

Speaker:

of conferences, as you'd probably expect, and I sit there

Speaker:

and I listen to the presentations and I think, I've heard this

Speaker:

before, or why are they making out this is so

Speaker:

different and radical? Everybody's saying, oh, AI data is really

Speaker:

important. It's like, it was important during the CRM era. It was

Speaker:

important during the big SAP era where everybody bought

Speaker:

SAP for their ERP. It's always been about the data.

Speaker:

Why is that such a surprise now? Yeah, no,

Speaker:

I think you're right because one of the things that always comes up in any

Speaker:

sales organization is CRM hygiene, right? Like Salesforce hygiene,

Speaker:

right? You have to make sure the data's in there correctly and all of that.

Speaker:

Because you're right, like if the data's bad, then anything that is derived from the

Speaker:

data, which is pretty much everything, is going to be

Speaker:

compromised. And you're right, it always has been about the data, right? I

Speaker:

think that is a fundamental

Speaker:

fact that I think people, I don't know when

Speaker:

people are gonna learn that one. Some of your IT departments,

Speaker:their old name in the:Speaker:

department. That's right. And,

Speaker:

you know, back then though, like data processing was a much heavier

Speaker:

lift. Like you were really talking about punch cards and stuff like that and really

Speaker:

processing it. What do you, how do you think we ended up getting

Speaker:

lost in that? Right? Because you're right. If it went from data processing to, you

Speaker:

know, from information science or management of information science, what do you

Speaker:

think we went wrong as an industry? I

Speaker:

think there's no one point in which it went wrong.

Speaker:

It's the, in my experience, having been

Speaker:

largely in corporate businesses and a lot of the time on transformation

Speaker:

programs, all too often you sit in the room in the discovery

Speaker:

session. So you might well have one of the big

Speaker:

consulting firms in, and you bring them in

Speaker:

and they'll do a discovery. It might be 2 weeks, it might be a week,

Speaker:

and you'll sit there and you'll talk about, this is once in a

Speaker:

20-year experience of transforming our business.

Speaker:

Let's take the opportunity. And really run with it.

Speaker:

And everybody gets excited. They walk out the room at the end

Speaker:

of this discovery, they might write it up, and then everybody sits in the room

Speaker:

the following week without the consultants and say, oh,

Speaker:

but we don't do it that way here. And the thing I alluded to

Speaker:

earlier about taking a bad practice, you just repeated it.

Speaker:

And you take all the data from the old systems that you want to remove.

Speaker:

I have seen it all too often. Yeah. Right. So it's

Speaker:

largely political and people and change

Speaker:

management that's letting you down. You're not actually

Speaker:

transforming in the end. All you're doing is moving

Speaker:

data from one place to another, a bad process from one place to another.

Speaker:

And I think that also carried over to the digital transformation

Speaker:

era too, because I've sat in a lot of digital transformation meetings.

Speaker:

And just thinking like we're now to the point

Speaker:

where when I hear digital transformation, right, get a little sick on myself,

Speaker:

but it doesn't have to be that way. Like what, how much of it do

Speaker:

you think is, how much do you think of it is that if you

Speaker:

facilitate one of the big consulting firms

Speaker:

facilitates it? I don't think they do the handoff really well, right? Because when they're

Speaker:

not in the room, the conversations tend to, I've been in the

Speaker:

room when somebody from one of the big firms was there

Speaker:

and it's, some of it's hand waving, right?

Speaker:

Some of it's real. I don't want to trash the big consulting firms, but I

Speaker:

just don't feel like they do a good handoff to

Speaker:

the normal people in the room, the people who are stuck with it. I think

Speaker:

the handoff, I think their ideas are valid, but I think their handoff

Speaker:

process could use a little bit of tweaking. Yeah.

Speaker:

And I also wouldn't say, and I'm just saying because I've worked with

Speaker:

some of the big ones, but I think it would be true of even small

Speaker:

consulting firms going in. It's the behavior of the organization

Speaker:

that in its reaction to the transformation agenda, the

Speaker:

consultants, whoever they are, big or small, don't necessarily

Speaker:

understand the business. So while we spend a lot of time in ISM

Speaker:

looking at people, process, technology, and the data before we really

Speaker:

say, okay, these are our recommendations. And quite often in ISM,

Speaker:

we're really a governance layer. And other partners will come in and

Speaker:

actually do execution, not all the time, it varies. And

Speaker:

we quite like being that independent arbiter. And I think you do need

Speaker:

that challenge in these programs. The other thing is

Speaker:

actually the development itself. If you've got third-party

Speaker:

developers, if they are, shall we say,

Speaker:

kept at arm's length from the core team in the partner,

Speaker:

and that could be offshore and not necessarily offshore, so to speak.

Speaker:

Right. I don't actually like the phrase offshore. In another country,

Speaker:

not where you're based. They don't always get the full picture.

Speaker:

So you will end up with verbose code. You have seen it too often.

Speaker:

Inefficiencies, misunderstandings,

Speaker:

inefficiencies, and things creep in. And that

Speaker:

again is where the transformation doesn't always happen.

Speaker:

I've been in the same building and I felt like I was at arm's length,

Speaker:

right? I felt like I was offshore. You know, even if I'm in

Speaker:

the same building, I've had a couple of consulting gigs like that where it's kind

Speaker:

of like, you have to talk to this per— I'm like, why can't I just

Speaker:

talk to the actual person I'm trying to solve for? And it gets down to,

Speaker:

I think you're right, people, process, and data, right? And probably in

Speaker:

that order, right? Because people control the process, the process determines

Speaker:

the quality or lack thereof of the data. So like, there's definitely

Speaker:

something to that. And in our methodology

Speaker:

at ISM, which I've modified from Barton's CRM work,

Speaker:

is through the lens of users and customers. And the

Speaker:

mix, as Barton used to call it, he had a kind of

Speaker:

percentage mix. I think it varies from company to

Speaker:

company, but if you get one of those pieces wrong,

Speaker:

you are gonna have some failures. Well, and then also too,

Speaker:

how do you address failures? Because failures in complex systems are inevitable,

Speaker:

right? And how do you, how do you, I think the people and process

Speaker:

are gonna be large determining factors in how the

Speaker:

failure is managed, mitigated, and resolved. And if you don't

Speaker:

have that together, right, it becomes a blame game. And I've seen

Speaker:

more than one SAP implementation go sideways in my career.

Speaker:

And now, fortunately, I was not on

Speaker:

the ERP or CRM team, but I can kind of watch from a

Speaker:

safe minimum distance over the hill. Sipping

Speaker:

my coffee like, wow, that's a dumpster fire. And I think

Speaker:

in a lot of ways that actually scared me away from ERP work, right?

Speaker:

Where I, when it works well and it's

Speaker:

deployed well, it's awesome. When it's not

Speaker:

deployed well, I mean, this one particular, this

Speaker:

was in the '90s and like the production line had to be shut down.

Speaker:

Like all the whole supply chain had just become I'm trying to keep the

Speaker:

language rating here pretty civil, but it got pretty borked pretty fast.

Speaker:

And like millions of dollars a day were being like just set on fire.

Speaker:

Right. And I was just, you know, a bunch of us at the time I

Speaker:

was on the web development team, which, you know, early

Speaker:

e-commerce days, we're kind of like watching it from a distance saying, I'm glad we're

Speaker:

not there. But yeah. I think that gets

Speaker:

me thinking though, right? You know, if CRM data

Speaker:

is incomplete or wrong or, or whatever. It just

Speaker:

seems like giving an agent access to it and

Speaker:

agency to do things based on that bad data, that seems

Speaker:

like a recipe for disaster. That seems like you have a dumpster, you have a

Speaker:

can of matches and a big can of gasoline or petrol in your hand.

Speaker:

Yes. And if you really want to be agentic and successfully

Speaker:

agentic-wise, then maybe this is the one time we will

Speaker:

actually will sort our data out. In advance,

Speaker:

because we will actually, we'll just, instead of

Speaker:

driving on a, sorry, English language, an A-road, you're driving on a

Speaker:

motorway. All of a sudden you're on a speed track.

Speaker:

Right. Not even a motorway. You're gonna get to these

Speaker:

mistakes a lot quicker. You've made the same mistakes with people with bad data, but

Speaker:

you might have had time to realize it before it became a real issue. And

Speaker:

certainly you wouldn't get the volume. But now you're on

Speaker:

an absolute race to keep up with your

Speaker:

agents. In fact, pacing your agents and slowing them down is

Speaker:

what we've done in some of our implementations. You actually

Speaker:

introduce a lag, even though the agent can do

Speaker:

it fast, a deliberate lag.

Speaker:

Really? So how do customers react to that? Do they— are they— Well,

Speaker:

actually positively, because if it goes too fast, They know

Speaker:

AI's involved, then they don't feel special.

Speaker:

Interesting. Do you mean they don't feel special if it goes too fast? They know

Speaker:

it's an agent. Nobody's looked at my, um, I've issued a

Speaker:

complaint, whatever it happens to be, get a response. It's already

Speaker:

reasonably personalized, but it came within a minute.

Speaker:

Nobody looked. Oh, okay. Okay. It's kind of like, you know, the professor

Speaker:

giving out a task and then like before you leave, give me a, write a,

Speaker:

write a 300, you know, write an essay on topic X. And

Speaker:

then before you even leave the room, like you already have it done. Like clearly

Speaker:

AI is in the process. I see what you mean. So kind of introducing a

Speaker:

lag for human oversight, or is it just perception?

Speaker:

It could be both. Okay. And the other thing I've done,

Speaker:

actually, I've just written an article and I'm writing a

Speaker:

follow-on article. And this is about, you know, once your agents

Speaker:

live, what do you do? So here's a sort of

Speaker:

a trailer for it. It's

Speaker:

actually, when you implement, you don't necessarily

Speaker:

implement Big Bang. Right. So you start potentially,

Speaker:

your agent may well be, inverted commas, live, but

Speaker:

it's not actually customer-facing. It's not even

Speaker:

executing the action, but it's saying what it would do.

Speaker:

If it could. So you've got that kind of check as a

Speaker:

kind of proof. This is once you've gone live,

Speaker:

because obviously production systems and sandboxes aren't

Speaker:

necessarily the same thing. And this is where there's another topic of mine.

Speaker:

But once you've gone beyond that initial test

Speaker:

in live, is to pick a demographic or pick

Speaker:

a subset of customers, a proportion, and

Speaker:

you increase that proportion over days or weeks,

Speaker:

but also once it's properly active, is to

Speaker:

always have a log of those actions it took.

Speaker:

So you've got that auditability and that

Speaker:

traceability built in from the start. That's what we've done.

Speaker:

Interesting. What sorts of things should businesses consider before they

Speaker:

Give, obviously there's probably read access to CRM to figure

Speaker:

out what's happening, but write access seems to be another level of commitment,

Speaker:

trust to it. What are the important considerations before

Speaker:

you give an agent write access? Is it

Speaker:

inserting the human in the loop? Is it inserting a pause? What do you think?

Speaker:

Well, it's back to what I just said, really. So

Speaker:

you'll, I alluded to testing in the

Speaker:

previous conversation just now, and

Speaker:

What I'm saying is, you,

Speaker:

if in test, you might not have production-size data.

Speaker:

I'm actually becoming a huge fan of the idea of

Speaker:

anonymized, but truly replicated

Speaker:

production-like volume and

Speaker:

variance. If you can't test with

Speaker:

production anomalies in data, because all these edge

Speaker:

cases you won't think of in testing. Yeah. Unless you've got a

Speaker:

really very special user involved in testing.

Speaker:

And I have been in environments where we've had some really very effective user teams,

Speaker:

so it's not impossible. So you really, you're

Speaker:

going to go into production and you're going to hit

Speaker:

some very strange scenarios, especially in a customer

Speaker:

service environment, in a complex business like a financial

Speaker:

institution, for example, especially consumer-facing.

Speaker:

So that's why I like the approach

Speaker:

of small groups graduating upwards.

Speaker:

So you have got jump-off points before you've had a

Speaker:

really major news story about how you upset all your customers.

Speaker:

Better to take it gradually. There was an airline actually in the

Speaker:

UK recently that had an incredibly

Speaker:

successful AI implementation. They

Speaker:

went gradual and they ramped it up fast because they were getting such good

Speaker:

results. That was recently

Speaker:

presented at some— I can't say who they are, but it's pretty obvious who they

Speaker:

are. Very successful. Interesting. And that was

Speaker:

Salesforce, actually. That was Salesforce. So how good are the

Speaker:

Salesforce internal tools? Because I think there's a lot of temptation to

Speaker:

rely on the internal tools as they are. Obviously,

Speaker:

the built-in Salesforce tools are probably pretty good, but they're probably not gonna work for

Speaker:

everyone. You know, I would always roll this back and I

Speaker:

suppose this takes us full circle to the beginning of our

Speaker:

conversation today. And that's, is the solution really AI?

Speaker:

Yes, I'm big on AI and ISM

Speaker:

is very much agent-first now

Speaker:

and native in this space. But, and even

Speaker:

our business is running with agents. But

Speaker:

the first question is not, I want

Speaker:

to, rather than should say this, you don't come with a solution, you

Speaker:

come with the problem. And the solution may well be

Speaker:

classic technology. And there's a number of instances

Speaker:

in my corporate experience, very recent as well, where

Speaker:

it was agent force or classic sales force. And we

Speaker:

concluded actually for our use case, we didn't need

Speaker:

anything agentic. It was classic. Right. So

Speaker:

always that check and balance of, is it really

Speaker:

agentic or is it just automation?

Speaker:

Don't be afraid to say it isn't AI. But

Speaker:

from the top, I always use the analogy of the

Speaker:

exec, go and play golf. And they play golf with the other

Speaker:

execs of other companies, and they're all boasting about how they're using

Speaker:

AI. So he goes back, your CEO or

Speaker:

CTO goes back to the team and says, we're gonna do AI.

Speaker:

Just, I don't want any other solutions. It's

Speaker:

gotta be AI. And I want everything to have AI at the end.

Speaker:

That's the wrong approach. Start with the problem. And that has

Speaker:

always been the case. Even Visual Studio used to have,

Speaker:

what were the files for Visual Studio? Solution files, right? It's all about the solution.

Speaker:

It's about solving a problem. Hold on one second. All right.

Speaker:

Sorry about that. So where,

Speaker:

when something does go wrong, who bears the brunt of

Speaker:

when there was a failure in the data or the AI? Where does that

Speaker:

blame go? Right. And I don't mean blame, you know,

Speaker:

AIs are going to make mistakes and sometimes honest mistakes,

Speaker:

sometimes just on bad data or just we're asking them to do more than what

Speaker:

they're really capable yet. My next question for you is,

Speaker:

a good example would be, I think it was Air Canada had basically through the

Speaker:

chatbot had given a customer the wrong answer. And

Speaker:

then the court basically said, no, no, no, Air Canada has to, you know, that

Speaker:

AI agent was considered, you know, a personnel

Speaker:

making a promise for that? Like, what does that

Speaker:

mean for applying AI to

Speaker:

these types of solutions? Well, I think you've got to have a very

Speaker:

serious attitude to what you're implementing. And

Speaker:

ultimately, you can't just say, oh, it's AI, it wasn't our fault.

Speaker:

If you're Air Canada or somebody else, it actually

Speaker:

stops with the supplier, if it's a supplier,

Speaker:

consumer relationship or business-to-business, it's got to be the

Speaker:

supplier that has to take

Speaker:

responsibility for the system that they are running. And

Speaker:

so what if there are agents? But back to my

Speaker:

sequel article I've just authored this morning,

Speaker:

morning in the UK, is to go back and

Speaker:

observe even after go-live. It's not great, we've

Speaker:

implemented it, see you in 10 years' time when we're going to amend this agent.

Speaker:

Especially because the models are changing all the time. And I mean, there's an interesting

Speaker:

debate that's really getting quite hot now, large

Speaker:

language models versus small, i.e.,

Speaker:

domain-specific models, and even your own.

Speaker:

I think that could well be a trend that's going to kick off in '27.

Speaker:

So let's take a deeper dive. What do you think that trend is going to

Speaker:

look like? So I think you're going to

Speaker:

see more industry, probably we'll start

Speaker:

with verticals. Okay. So you will find, I

Speaker:

think, rather than, I mean, let's, I'll reword it. Who

Speaker:

needs the sum total of the knowledge of the world

Speaker:

to run a financial services business or

Speaker:

serve a customer, be it an

Speaker:

account holder? Who needs all that knowledge? You don't. You

Speaker:

only need knowledge relating to your business and being

Speaker:

able to answer those questions in natural language because that's what people want.

Speaker:

You don't need everything. So I think it's gonna start, I mean, we

Speaker:

saw it in CRM, you saw it in SAP,

Speaker:

getting industry-specific about this

Speaker:

now. And I think that's coming in '27. You're gonna see

Speaker:

more and more media industry, financial services,

Speaker:

you name it, healthcare, typical

Speaker:

major streams of vertical models.

Speaker:

So there are a far greater tuning

Speaker:

to the business process at hand and indeed the business at

Speaker:

hand. I think you're gonna see a lot more of that. Interesting.

Speaker:

It's always, you know, as we get to that time of year where people start

Speaker:

making predictions about next year, I'm always curious where people think we're gonna go,

Speaker:

particularly after a wild Particularly after the wild week that we had in

Speaker:

tech last week where, you know,

Speaker:

Jensen Huang kind of scandal, then Outlook went down and

Speaker:

there's a lot of debate about what caused that. Was it AI agents going rogue

Speaker:

or, you know, it's

Speaker:

not, was not a boring week at all. Yeah. And security

Speaker:

will be an ever-increasing story. And I

Speaker:

think you're gonna see it move away from the

Speaker:

race to the better and better frontier models are more and

Speaker:

more about prevention of, you know,

Speaker:

mitigation of risk and being ahead of the

Speaker:

vulnerabilities. I think that's another theme that is going to be

Speaker:

next year. And I think close on the heels of that

Speaker:

big topic, I think is probably,

Speaker:

yeah, especially as I read the US politics as an outsider to

Speaker:

the US, is the whole data center question.

Speaker:

Will we keep building data centers? I

Speaker:

think there'll be a saturation point for

Speaker:

data centers in the sense that they're being built now. I think there'll

Speaker:

be more distribution. I think there'll be a

Speaker:

little bit like we've built the mainframe, now we're gonna build the

Speaker:

mid-range system. Yep. And later on, we're gonna be running it locally.

Speaker:

And it's going to be distributed across those

Speaker:

different size dimensions for different purposes.

Speaker:

I think you're going to see more and more of that. I mean, watch Apple.

Speaker:

Apple is going to be very impressive moving forward in the AI

Speaker:

space. 100%. I think they have real potential. I

Speaker:

do think that the data center, build

Speaker:

it as a multi-acre, 100-acre, type thing, I think is

Speaker:

just those days are coming to an end because you have not

Speaker:

just community pushback, but just, I think it was,

Speaker:

they were trying to do the Walmart thing where you get economies of scale

Speaker:

and economies of scale work as long as you get economies of scale.

Speaker:

And with increasing pushback and on communities,

Speaker:

from communities where these data centers go, I don't think all of it is

Speaker:

necessarily, grassroots. I think some of it is funded by state actors

Speaker:

and people with ill intentions. But I also

Speaker:

know like clearing out 200, 300 acres of farmland

Speaker:

is not a good look, nor is that really sustainable in any

Speaker:

meaningful way. And in any case, I mean, you take the UK, it's

Speaker:

even worse in some ways because we have a planning system that

Speaker:

is very slow. We can't build a data center quick.

Speaker:

There's a very interesting company. This is no

Speaker:

endorsement, but it's called Chainergy. And they have, I

Speaker:

think, something like 1,000 sites reserved to

Speaker:

build small, very small, I wouldn't even call them a

Speaker:

data center, a little data hub, if you like. And it's

Speaker:

on a farm, in a farm shed, a fairly big

Speaker:

shed. It's an agricultural building, which in the UK is not

Speaker:

necessarily the that huge. And

Speaker:

you've got compute in there, and right

Speaker:

by it is a biomass generator. So the electricity is on

Speaker:

site, it's enclosed, there's not lots of water,

Speaker:

it just looks like an agricultural building. It could have been a

Speaker:

former agricultural building, might not even be brand new.

Speaker:

And that level of distribution,

Speaker:

and it does have grid connection for resilience.

Speaker:

You distribute that, lots of fault tolerance. I think it's

Speaker:

a pretty impressive model. And I think in California,

Speaker:

there's a builder who is building something

Speaker:

like 16 GPUs inside that basically

Speaker:

looks like a heat pump on the side of the house. And that's part of

Speaker:

their building regime. That's highly distributed.

Speaker:

No, I think that's really going to be I think that's really,

Speaker:

I think the future. There's a number of neo-clouds.

Speaker:

Again, I'm not endorsing any one of them, but one of the

Speaker:

cottage industries that's popping up is the old AT&T buildings or

Speaker:

Verizon or whatever the local, the central offices. A

Speaker:

lot of companies are buying out their old central offices

Speaker:

and putting server racks in them. Yeah. Right? And these are the big

Speaker:

buildings you see in small towns throughout the US. Yeah. Where

Speaker:

they have these, these big Verizon in the East Coast is Verizon.

Speaker:

And they're built, you know, they were built during the Cold War, many of them.

Speaker:

So like they have pretty rigid structures and things like that. There's not really a

Speaker:

lot of windows in them. They were the data centers of their day,

Speaker:

right? And I think that there's a, and the telcos can't

Speaker:

get rid of them fast enough. So a lot of the neo clouds are building

Speaker:

it. And there was actually one company that I

Speaker:

think it's called Zero Latency, Zero Lat.

Speaker:

And, you know, they're putting inference kind of in these, they're

Speaker:

buying up the old equipment, they're buying up the old buildings and

Speaker:

renting it back to the telcos as well as putting in, you know,

Speaker:

GPUs for fast local inference, which I think is brilliant, right? Like the

Speaker:

one-size-fits-all massive Walmart-type situation that we

Speaker:

have, big box store approach to data centers, I think is

Speaker:

I think those are coming to an end. There was a very interesting

Speaker:

NVIDIA announcement in their— it was at Frankfurt

Speaker:

last month, where it's basically a

Speaker:

container. So your compute arrives in a

Speaker:

container. Well, why don't you just have a few? Again, it just

Speaker:

looks like a container in a shipyard, or it

Speaker:

looks like you make it look a little bit more like an agriculture. Maybe the

Speaker:

chainage thing could just be dropped at the farm.

Speaker:

Right, right. Or put underground or something like that. And you can get

Speaker:

geothermal heating type of cooling and things like that. There's a lot more, I think

Speaker:

there's a lot more room for creativity for— And what about reusing the heat that's

Speaker:

generated? Yeah. In Norway, I think, don't Norway or

Speaker:

Finland, they reuse the heat to heat up the streets and the

Speaker:

sidewalks, which is brilliant because, you know, these things in my own

Speaker:

home lab, which has not been moved to this location yet, You know, my office

Speaker:

is in the basement and in the winter, like I

Speaker:

literally heat my home office through my servers. It's funny

Speaker:

you say that because under my desk I've got stuff and I've got 2 screens,

Speaker:

well, 3 screens here and I can feel the heat if I close the

Speaker:

door, basically heating my office. Yep.

Speaker:

Yep. I mean, that's what I do. And even in the summer, because not so

Speaker:

much in this house, but in the old house to get the top floors,

Speaker:

kind of at a decent temperature, the basement always gets cold. So for me,

Speaker:

I happen to like cold, so it's not an issue for me. So like if

Speaker:

I, so, but it's basically taking advantage of a

Speaker:

quirk of the HVAC and, you know, making

Speaker:

the best of it, right? So I think there's a lot more room for

Speaker:

creativity in that space. And I think that, I

Speaker:

think we have a lot more ways to go. And then Jensen was not that

Speaker:

long ago also promoting the idea of distributed data centers where They

Speaker:

basically bolt onto your external heating unit

Speaker:

and kind of, you know, work that way. I think there's a lot of very

Speaker:

practical engineering and security concerns around that. But

Speaker:

again, that's— I don't think that's unsolvable. And if you start to think

Speaker:

about more of the catastrophizing side of

Speaker:

things, is it not better with a good degree of

Speaker:

tolerance and distribution? Then nobody can

Speaker:

take you, take one of your data centers out, take 4 or 5 of them

Speaker:

out, but it makes no difference. I would prefer that. 100%. I mean,

Speaker:

and we saw that in Qatar where like, this is not a theoretical anymore.

Speaker:

Like there are, you know, you could have a drone attack. Like I live, well,

Speaker:

you know, Bethesda, so, you know, you're not that far from Loudoun County, Virginia,

Speaker:

where every building, I think, I swear if it's not a,

Speaker:

if it's not a supermarket or a townhouse, it's a data

Speaker:

center. Yeah. And we used

Speaker:

to live in that area maybe about 20 years ago. And

Speaker:

like, you know, it seems like every bit of square foot of space that's left

Speaker:

has been built out to a data center. And, you know, it's not,

Speaker:

you know, it's not that hard to imagine that data centers are now legitimate,

Speaker:

you know, military targets, right? They are the factories of the

Speaker:

21st century. So no, you're right. Like that type of resilience

Speaker:

and lack of reliance, I mean, I think we—

Speaker:

Microsoft was out towards the end of last week.

Speaker:

We can't be reliant on— and even the Jensen thing,

Speaker:

and long may he have a good retirement unless he comes

Speaker:

back, which is the rumor now, but it is a sad loss.

Speaker:

But on the other hand, making space for new

Speaker:

people to rise up. Right. We can't be reliant on

Speaker:

one manufacturer. We can't be reliant on one person. in one

Speaker:

manufacturer. Let's spread the love, so to speak.

Speaker:

I'm always available, NVIDIA, if you want to give me a call.

Speaker:

But just, you know, there are so many people with talent

Speaker:

and it's, you don't want to be a ceiling on the talent that's coming up

Speaker:

from underneath. Let's make room for a few more players.

Speaker:

No, I think that's a good point. Like Jensen was the single point of failure

Speaker:

at NVIDIA. Well, we don't know how that's going to play out, but He built

Speaker:

that company out of the force of his personality, which is admirable for

Speaker:

sure. But also apparently they had no succession plan,

Speaker:

right? Yeah. I really, I cannot believe it. I think every

Speaker:

board of a public company has to, that should be part of it, right? Because

Speaker:

who knows, like somebody could drop dead on a hard drive. He did a lot

Speaker:

of travel and it could have happened in his travel. Right. He could have

Speaker:

gotten sick. He could have gotten, he could have, had food

Speaker:

poisoning. I mean, any number of things could happen to him because he's a human

Speaker:

being, right? Like, and having

Speaker:

one guy I used to work with, he called it the beer truck factor. And

Speaker:

it was basically what happens if you get hit by a beer truck?

Speaker:

What's the resilience plan around that? I think CEOs, particularly if

Speaker:

they're $4 or $5 trillion worth of money,

Speaker:

need to have that level of that. I mean, that should be I'm not a

Speaker:

fan of like adding more regulation, but certainly, you know, having that kind of

Speaker:

risk should be part of it. I think we're going to find

Speaker:

in NVIDIA, there are a lot of talented people in there

Speaker:

who we haven't seen. Right. And I

Speaker:

think that'll be a great— I mean, Salesforce is a really good example. You'll hear

Speaker:

Marc Benioff a reasonable amount, but it's never

Speaker:

always Marc. And there are others and the others

Speaker:

do rotate a bit. So there are, you know,

Speaker:

there are talented people, and quite often you see them repeatedly,

Speaker:

but there's a good spread of that talent across the globe. I think

Speaker:

we need to see that in NVIDIA. Even if Jensen comes back,

Speaker:

let's see more of the others. Right. Actually, the

Speaker:

Google I/O earlier this year, I think it was about April

Speaker:

time, that was brilliant for that. There was loads of handoff

Speaker:

back and forth. You know, it's not one person,

Speaker:

but there is a key person at the top still, and they're still well known

Speaker:

and everything. Let's see

Speaker:

more of the talent in NVIDIA. And yeah, let's see Jensen come back in

Speaker:

some way. I guess once everything's resolved with whatever

Speaker:

did or didn't happen, you know, Right. A fistful of

Speaker:

knuckles for being naughty and let's move on and use

Speaker:

his creativity and leadership in a slightly different way

Speaker:

moving forward. You know, split some of the business running away from the

Speaker:

technical, you know, maybe it's a dual role CEO where it's

Speaker:

technical CEO and financial CEO. I always like what

Speaker:

Microsoft did after the Justice Department kind of forced Bill

Speaker:

Gates out. He wasn't completely gone, but he was,

Speaker:

he kind of had a transition role for, I don't know, a number of years

Speaker:

where he was chairman of the board. And even to this day,

Speaker:

Bill's opinion matters at Microsoft. because he

Speaker:

probably has a lot of stock options that he, you know, stock that's attached to

Speaker:

it. I don't think, but nothing happens, nothing

Speaker:

serious happens without his blessing or approval

Speaker:

even to this day. Right. So, or at least when I last worked there a

Speaker:

couple of years ago, you know, he was still seen as a trust advisor. I

Speaker:

think that would be a reasonable approach for Jensen

Speaker:

because, and again, like I think even if

Speaker:

worst case scenario, he does get divorced and all that, he'll still walk away

Speaker:

with somewhere between $70 and $80 billion. And

Speaker:

even in the Bay Area where real estate is expensive, he'll be okay. He

Speaker:

won't be back at Denny's anytime soon. And hey,

Speaker:

maybe it's a good buying opportunity for Nvidia. And exactly.

Speaker:

So if you've got a few spare dollars underneath the mattress,

Speaker:

this isn't financial advice, maybe now

Speaker:

Telling. I would, I would assume, yeah, I would assume like, you know, it definitely

Speaker:

looks good. But no, you're right. I mean, I think monopolies are, are bad,

Speaker:

not just because of consumer choice being restricted, but

Speaker:

also resiliency comes at a cost, right? I think this is

Speaker:

an opportunity for the AMDs of the world, the other chipmakers of the world to

Speaker:

step up and realize we can have a vibrant kind of

Speaker:

ecosystem around this. Yeah. And it

Speaker:

may well be, it may well be defined a bit by size and

Speaker:

whether you're edge AI or you're data center AI.

Speaker:

You know, let's, I think there is

Speaker:

room for creativity and innovation amongst a greater

Speaker:

number. So with that in mind, just wanted to get some

Speaker:

final thoughts in here. What can one thing listeners do this week

Speaker:

to prepare their organization to have these agents

Speaker:

be more trustworthy? I think it's always go back to

Speaker:

basics. AI is just like every other

Speaker:

hype cycle and it's just running very fast and it's

Speaker:

impossible keeping up. I would say

Speaker:

concentrate on your foundations. And where can folks

Speaker:

follow, find out more about you, follow your work and listen to

Speaker:

AI in Motion? Yes. So we're on our AI

Speaker:

in Motion podcast every couple of weeks with a a great

Speaker:

colleague, Tim Baharan, well known in Silicon Valley.

Speaker:

You can find that on isnguide.com. We are

Speaker:

on Spotify and Apple, and maybe a few other platforms I don't know about,

Speaker:

don't pay much attention to all of them, but definitely on those 2.

Speaker:

Follow us on LinkedIn as well. But yeah, we hope people

Speaker:

will engage with us and enjoy what we have to offer. Excellent.

Speaker:

Thank you very much. And We'll let the outro roll.

Speaker:

That's Mike Dimott of ISM with one thing to do this week:

Speaker:

concentrate on your foundations. Find Mike on the AI in Motion

Speaker:

podcast, out every couple of weeks on Spotify and Apple, and

Speaker:

follow ISM on LinkedIn. Thanks for listening to Data Driven.

About the author, Frank

Frank La Vigne is a software engineer and UX geek who saw the light about Data Science at an internal Microsoft Data Science Summit in 2016. Now, he wants to share his passion for the Data Arts with the world.

He blogs regularly at FranksWorld.com and has a YouTube channel called Frank's World TV. (www.FranksWorld.TV). Frank has extensive experience in web and application development. He is also an expert in mobile and tablet engineering. You can find him on Twitter at @tableteer.