Cutting Through Data Complexity: Where Humans Stay in Control

Duncan Cooper: Well, lovely to see everybody, virtually, obviously. And kind of my quick
introduction, I’m Duncan Cooper, run my own consulting firm now, which is based around data
strategy and data leadership. Previously Chief Data Officer at Northern Trust, senior data
leadership positions at BMW. and HSBC and then a long time in fintechs, although I’m that old,
we called them software companies, not fintechs, you know, we were that uncool, and
consulting. So did a lot of work in various consulting firms, including a fun time building the risk
system behind quantitative easing for the Bank of England.
Alpesh Doshi: Ohh, I didn’t know that.
Duncan Cooper: It was a it was a fun time.
Alpesh Doshi: And, you know, Duncan sort of plays it down what he’s achieved, right? You
know, I like Duncan because one of the things I do on my podcast, which is called the Innovation
Briefing, where this will also appear, is I like to interview people who disrupt industries, but also
disrupt companies and how they work. One of the biggest things around most firms is that if you
don’t have people like Duncan in them, change never happens. And unfortunately, you know,
it’s a lot of it’s nothing to do with technology or business knowledge and nothing. It’s about
politics. It’s about stakeholder management. It’s about communication. And it’s about vision and
planning and vision and trying to communicate that across a business. And Duncan is fantastic at
that. Yeah, I know, but you are. I mean, that’s the thing, right? If you see the other interviews
I’ve done, that’s the kind of people that I always interview. And it’s really fantastic to get that
kind of person. And now, as I said at the beginning before Duncan came on, you know, AI ready
data is It actually isn’t a hot topic. People haven’t started talking about it. People talking about
AI all the time, but they don’t think, okay, well, where’s the data coming from? And I’m sure all
of you work in organizations of any size, data is one of the biggest problems and it hasn’t
changed in 15 years. You know, when this big data concept came around in 2011, The reality
was it was early and people started spending large amounts of money. We’ve had phases from
data warehouses to lakes, data lakes to lake houses. Now we’re in a new era of data and the
way you, and I said this earlier, but the way you have to prepare data to go into AI models is a
different set of patterns and technologies and processes. And, and therefore, you know what we
should start off is… The conversation is really about AI ready data. So let’s start. I mean, I guess

let’s define what AI ready data is. So, you know, Duncan, from your perspective, what would
you, you know, given that you’re a data veteran, what does AI ready data mean to you?
Duncan Cooper: That’s a really interesting one because it’s probably not what people think.
There we are. So I’m almost assuming people think AI-ready data is clean and structured and
organized and all these things that historically we’ve been told that’s what makes good data. So
that governance function that wants to put all of the controls around it. And for me, that is very
good data, but you’ve got to spend a huge amount of time and effort putting that in place.
Realistically, AI ready data doesn’t need to be that structured. What it does need, though, is
clear explainability about what it is. So when we talk about AI-ready data, I immediately start
looking at what is the metadata around this. So I want governance in terms of catalogs, technical
catalog, business glossary, I want definitions. I want understanding that I can then point to an
LLM. And when it sees a field that says duration, which of the 25 different ways of calculating
duration does it actually mean? And what can I use that for? And how structured is that data in
terms of further calculation? That really supports AI, because… LLMs are driven by the context
of the data. And most organizations’ data is structured, but not contextualized. So that for me is
where I would tend to start the thing that people don’t look at.
Alpesh Doshi: I think, I mean, one of the other pieces I’d love to get your view on is with this AI,
you know, there’s two aspects of the AI for me. One is using AI and agents and agentic systems,
not chat bots, but proper agentic systems to fix the data itself. As I’m sure all of us know, right?
Data still isn’t fixed. We’re 15 years into the big data and now obviously the word what we call it
is changed. But with that, that far in, and actually data still isn’t fixed in most organizations. And
if it has been, most of the data that’s fixed is focused on reporting. So collect loads of data, run it
through a lakehouse or something else, and produce a bunch of reports. So if you imagine you,
one of the things for me is, How do you actually use agents and identity systems to solve the
problem of fixing data in quotes to get it ready? And then the other side is then once what we
just talked about just now, which is… How do we use that data in actually solving a business
problem or making a decision or something like that? So given that you are a veteran CDO and
you run teams that actually do the first thing, I would love to get your perspective on, okay,
look, here’s how this is how my data function works. How would you perceive the fact that you
can now apply this AI stuff and agent stuff to actually solve the data problem that frankly hasn’t
been solved in 20 years?
Duncan Cooper: Very much so. And look, I love applying AI to stop me doing tasks that are
boring or repetitive or not delivering value. And I would say that when I’ve applied AI capabilities
over the top of historical or legacy data, It solves a lot, but not everything that you need. Now, a
little example of this, we looked at mainframe schemas in an organization I was at, where they
had applications that were over 30 years old running on AS400s. And if you’re familiar with the
joys of AS400s, most column names for the data are limited to 7 characters length. And so
there’s not a lot of clarity that you can put into a seven character column name on a 30 year old
system. But when we took that data out and then reflected it and applied over the top of it, just
some very basic capabilities, nothing special, just using things like Claude and going, look at the

data and then try and guess what these fields are. And it didn’t do too bad. You know, this is,
yeah, 80% and this was about a year and a half, two years ago. So this isn’t recent with cutting
edge tools. What that tells me is this stuff is getting better. So you start to use AI to explain your
data to your AI. That is a great start. Can you then start putting rules and quality checks over the
top of it? I think this is where it gets really interesting because AI is non-deterministic by its very
nature. So it’s looking and extrapolating and analyzing and pulling things out that may be data
quality issues. much better than a very formulaic deterministic system. So there, there’s some
really cool things that you can do, but do that as your data comes in to enable data going out.
Alpesh Doshi: So, I’d have to slightly disagree with you on this non-determinism question,
because I get it regularly and I get this hallucination thing regularly as well, right? If you build the
right system in the right way, and a system, you know, an AI system isn’t just a model and a not
a model and a and a prompt, right? There’s a lot of other stuff that goes around it. The
determinism factor is significantly higher. The non-deterministic element of LLMs is significantly
lower. And I think, you know, I get this hallucination thing, we can’t trust it. And I agree, look,
you don’t want to trust it yet. You know, you have to do what are called evals. You have to test
the system and test it over and over again. You’re building swarms, you test them. before you’re
going to release it to live, but also still have a human in the loop or human over the loop. We’re
going to talk about that in a bit. So, you know, I would love to get your perspective. You know,
what, you know, when you did that example, what did it do wrong? And what did that prevent it
from going live? Because all this specific thing is, oh , we can’t put this live because this, the
specific area use case you talked about, and obviously wider, that it actually we can’t trust it.
Duncan Cooper: So I would immediately caution conflating non-deterministic with repeatability.
So it was non-deterministic, but it was repeatable. So you’re now getting for the same inputs,
you will get the same output. And that’s where I look at the repeatability. But the non-
determinism is that if you put the same inputs in, but then vary outside parameters. So for
example, a share price, if you put in the share price as 10, and the previous share price was 9,
that’s probably all right. If you put in 25 and the previous share price was 9, that’s an issue.
That’s a massive difference. We want to look at it. But if you are putting in the share price and at
the same time you put in a benchmark that’s showing that benchmark has increased by 100%
the previous day, your LLM is looking at not just the rule, but it’s taking into account the wider
context of the index. And then it’s sitting there going, okay, well, that actually looks right. Now,
it goes back to that human in the loop bit. You’re never freestyling data quality and saying, go
for it. But you’re getting it to pull out less false positives, I suppose, is the best way of describing
it. You’ve got a smarter calculation. because it is non-deterministic but repeatable, that allows
you to be smarter on the outbound.
Alpesh Doshi: So did you do did you do any kind of eval? Do you know what eval is? So you
model eval and there’s a bunch of different metrics and measures and inputs you can put in to
do proper eval to understand what the model is going to do, what it’s supposed to do. Did you
do any eval as part of the process? Well, let’s see. Let’s hold on. No, it’s gone. Is it your?

Duncan Cooper: Right, sorry about this.
Alpesh Doshi: Now, I don’t know if you got the question. Okay, I’m trying to remember what the
question was. Oh, yes, the question that I put was, was run eval. So when you did the work
around pulling the solution together,
Duncan Cooper: No, unfortunately I missed it, so…
Alpesh Doshi: How do you do what’s called LLM eval and agent eval and the system eval to see
whether you mentioned, you know, same input, same outputs. How did you do the eval process
to show that would work? Did you have a formal process around that?
Duncan Cooper: So, I’m not the best person to talk in detail about the actual, but the way we
the way we did it was with a large amount of training data.
Alpesh Doshi: Oh no, it’s fine.
Duncan Cooper: So, and it’s not just the training data of the thing you’re checking, but the
training data of things that are associated with it. So when you’re checking security prices that
you’re getting in on a daily basis, you want to look at the macro movements of the markets at
the same time from a benchmark or so FTSE 100 or DJ 500 moving, you want to understand how
that impacts an underlying stock price. And so you could vary the individual thing you’re
checking and then the benchmarks in your test data and you’re firing lots of stuff at it to
basically build up a model.
Alpesh Doshi: I mean, I think that’s one of the things for me is, you know, these systems are
very different to traditional software systems. And being able to do eval to prove that they work,
I mean, in a regulated environment, it’s absolutely essential, right? You know, you can’t just sort
of, if you’re doing an e-commerce project, an e-commerce business and you’re just buying stuff,
find the transaction and the actual payment is really important, but everything else, if it breaks,
is not going to… do anything, it might cause irritation to the user. So yeah, cool. And then things
like, you know, the traditional elements around data, around lineage, you know, on definition
and quality. What about that? I mean, how did you, what do you think about that in terms of
using AI to solve, for example, the data quality problem? but also using AI to actually build the
lineage. So one of the things I’m sure you know, lineage is quite an important piece. I get this
question regularly. Well, how do we do lineage? Well, you can do it automatically. It will
generate lineage for you. We can’t do that. Well, yeah, you can. And I’d love to get your
Duncan Cooper: So this is an interesting case because this is, it’s almost like a byproduct of what
we were doing anyway. We never thought about this as an AI capability, but it actually became.
So if you think you’re looking at something like a data mesh construct and you’re applying data
contracts. So if you’re applying data contracts, you’re immediately producing data that you’re
saying what you can use it for, who can use it, where it can go. But it’s a product. So anybody
can take that data as long as they are within the constrictions of the data contract. So you’ve

now got data products, and each of these data products has a data contract. When a consumer
is consuming that data product, they are complying with the data contract, so they’re
consuming that as well. So you now have a catalog, a business glossary, a data product, and a
data contract. All of this is producing data about the data being consumed. At that point, we
were persisting that information because we wanted to know, actually, who were our
customers? Who was using it? What are they doing? Where are they going? Now, if you do all of
that, you suddenly got a huge amount of data that can start to describe lineage. We weren’t
thinking in terms of going that direction when we started. We were thinking about get the data
about the data being used, but it came about on the outside.
Alpesh Doshi: And did you, did you capture it somewhere in a, in a, you know, there’s a bunch
of tooling out there you could capture it in. Did you actually formalize and capture lineage as an
example?
Duncan Cooper: So we would just persist in that as data dumping back into the mesh. So we
were using Snowflake, we were using Databricks, you’ve got all of the tools. Building it up
afterwards or post hoc was actually, you could do quite a lot because you’ve actually got the raw
data.
Alpesh Doshi: Yeah, yeah, interesting. Yeah, so one of the things I talk about, which I, we’ve not
talked about this before, is what I call an agentic data mesh. and an agentic data product. Yeah,
so those two things are, you know, going back to the Medallion architecture, we know you’re
building data, you know, have domain, data domains, you build data contracts, which you
describe, you build data products. So if you imagine what that looks like today, it’s not that
intelligent. It’s fairly static, right? When you build a data product, it’s a bunch of tables. It’s a
bunch of tables. And so when you call the data product and you mention the data contract, it’s
how do you access this data? What does it do? And you have a data product that you access.
One of the things I talk about is a gentic data product, which actually is live. So inside the data
product, sorry, around the data product is an agent who understands and can reason about the
data it knows about in this particular data product. So for example, Let’s say we’re talking about
the customer domain, we’re talking about a specific customer data product. If you wrap it,
rather than having a bunch of tables, you have an agent wrapping around it. So when somebody
wants to use it, they can ask a question and say more than just give me the tables, you can
actually reason and say, well, I want to know something about the customer. So it’s intelligent.
It’s an intelligent. So for me, this is given that I know about data meshes like you, and we talked
about, we’ve been talking about this for how many years, right? So this is the kind of things, and
to be honest, so I haven’t talked to you about this, but this is the kind of stuff that we’re working
on. So you’re building agentic data products which do significantly more.
Duncan Cooper: Yeah, got you, yeah.
Alpesh Doshi: than the traditional data mesh and data product definitions. And actually
transform a lot of the work that you needed to do. For example, right now, you’d have to spend

huge amounts of time defining data products, defining data contracts, trying to make them
work. It’s a bloody nightmare, right? You could have catalogs, you have huge registries of huge
numbers of data products. And if somebody wants something slightly different, You create
another data product and then the contract, right? It’s a bloody nightmare to manage. So if you
imagine creating an agentic data products, you have to do any of that. You define three or four
and it reasons about it. You ask it questions and it’ll give you what you want and it will do it
dynamically.
Duncan Cooper: So you could go down to the concept of creating your own data steward, data
owner, data custodian, as agents over a product and agents over a domain, which solves huge
amounts.
Alpesh Doshi: Yeah, yeah, yeah, and then you can describe it, yeah, exactly. So, but the thing is,
as well, it goes to the, you know, when you look at downstream, one of the things is you want
to, let’s say, doing customer domain, you want to ask a specific question, you’re solving a
specific problem. You talked about it earlier about metadata. There’s a lot of metadata that
comes out. So one of the things for me is models are really great, right? If you give well-defined
agents with inputs and outputs, you can say, I want to do this task. If you set it up properly, and
that’s hard, I’m not going to say that’s easy. You can just ask it a question, say, I want to retrieve
this kind of data for this kind of stuff for a customer. And today to do that, you’d have to set up a
whole process and load of work. In the new world, you just ask the question because it knows
about the agentic data products and you have metadata around it. You just ask it a question and
it will say, oh, I need that data product and that agentic data product and that agentic product
and I’ll give you the answer so it can work it out for you.
Duncan Cooper: It could be really interesting to dive down into this, probably not here,
otherwise I’ll go into it. But that concept, yeah.
Alpesh Doshi: Yeah, yeah, but this is so that’s what I’m working on. That’s what we’re working
on is that is exactly doing that.
Duncan Cooper: So if you think metadata is actually explainability, but also knowledge of the
data, it’s business knowledge.
Alpesh Doshi: It’s fundamental, yeah, so, yeah, and so for me, so if you, if you think about that
idea, if you think about the kind of things the teams that you’ve run, and this goes across
governance, you go across quality, you know, if you extrapolate, I mean… You know how much
time money would have saved if you did it?
Duncan Cooper: It’s the age-old. Every organisation looks at data as a cost centre and looks at it
as a control function. And if you can automate that out, you can start to change the mindset to a
little bit more around data as a revenue generator, data reserve. as something that really gives
you a competitive advantage, which is exactly where it should be. But it takes some time and if

you can accelerate that time to doing the foundations, to getting you there, you’re certainly in a
very interesting place.
Alpesh Doshi: There you go. So we’ve got two more segments, and we talk a little bit about both
the segments. One of them is keeping humans in the loop and above the loop. So why don’t we
spend a little bit of time on that? You already talked about the human in the loop stuff that you
did before. You know, maybe you can, and everybody is, as I’m sure you know, everybody’s
Most people aren’t convinced yet that you can run stuff and let it run on its own. Engineers, like,
you know, if you talk to Anthropic and a bunch of other codecs and a bunch of other guys, they
just let systems build for 24, 48 hours on their own without any interference, right? I try and do
that. But you have to be very well defined. So if you want to do that, and so you don’t need a
human in the loop, but the key question is if you imagine the data function, the data functions
you’ve run. What is the change in operating model around the fact that the human in the, so
you’ve got agentic systems being built. What do you foresee in terms of the changing operating
model and the human in the loop and the human over the loop oversight of the activities? So
nobody’s done this yet, so it’s okay. I’m asking because there isn’t an answer, but given that you
run enough CDO teams and you know, you run them, you run these teams, right? And obviously
this is a visionary discussion. This is not like holding you to anything, right?
Duncan Cooper: Yeah, there’s no right answer. Yeah, of course. Yeah, no, don’t hold me to a
count on this. When you look at data governance in any organisation, it tends to be
bureaucratic, slow, and manual. Because of the maturity of the function, usually. You know, you
don’t have the tools, you don’t have the… the foundational concepts that we’ve talked about,
like metadata, like catalogs, that drive the automation further down the line. The other side is
you also have every organization producing their own policies, procedures, and standards that
they want to apply across the organization. you know, very theoretical, very aspirational, no
understanding of the underlying business. And then you have a negotiation conversation of how
to get it working. Keeping humans in the loop, but making them focus on the things that’s
valuable as opposed to counting widgets or crunching processes. That’s where you’ll start to get
that value. And if you can go to that pattern, So you have defined policies, procedures,
standards across your org. You have clearly defined metadata. You have available pagentic data
mesh, as we’ve just discovered. You’re suddenly in a position where your entire governance
function should be just dealing with the exception. should just be applying the very specific
knowledge that they have around the business and adding value. It means that you should be
able to do a lot more important stuff without having to do the grunt work and the manual stuff
to keep the lights on. I do not know any organisation that is close to that at the moment.
Alpesh Doshi: No one is, but this is why I end up having to…
Duncan Cooper: But you could be.
Alpesh Doshi: And that’s why I end up spending a lot of time, and honestly, it’s actually a
reasonable leap for people, right? So if you imagine, if you imagine the functions across the

capabilities and functions you run, you know, pick one, pick one and say, okay, look, how will the
operating model change from large amounts of manual stuff with spreadsheets and a bunch of
systems to… hey, I’ve got agentic data mesh, I’ve got agentic data products. They’re going to run
in the background. They’re going to do a lot of the work the individual do. They’ll do, remember
the models that do, they do reasoning. They do, they’re not just like input output. They’re
actually in quotes, think about an answer. And they’ll honestly, I’m sure you’ve interacted with
these systems like Claude and anti-gravity. They have a decent dialogue with you and they’re
like, yesterday, just as an example, it told me, oh, I deleted a file of accident. I’m really sorry. I
deleted this file. It was like being apologetic. You’re like, bloody hell, this is like amazing. You’re
telling me how sorry, you know. It is great, but the thing for me is that people, the biggest
change that I’ve seen in the last year is the advent of reasoning models and the improvement
iteratively over the last year on these reasoning models. And I think if you imagine reasoning
around data, A lot of the work is rules-based. A lot of the work is human judgment. I need to do
X or Y. So if you imagine, and you know this better than me, you know, you’ve got a bunch of
teams, you’ve got a bunch of people sitting there doing that work. If you imagine having these
kind of capabilities, which really, you know, pick a capability that you’d have in a CDO function
and think,
Duncan Cooper: If you want to watch it, yeah.
Alpesh Doshi: And I put you on the spot, you know, there’s no right answer. I’ve just put you on
the spot to give you a, what would you, what would you say? What would you say in terms of
reasoning models and a human over the loop running these a capability that today is run all by
humans, basically?
Duncan Cooper: Yeah, I mean, the most obvious one applying to the very large banks is BCBS
239 as a regulation tied directly to data lineage, data explainability and traceability through the
org of what you’re using that data for. It’s, I think, it’s been around since about 2016 as a
regulation. It applies to the GSIPs, globally significant banking institutions. It’s roughly 30 of
them. I would be surprised if more than five have fully completed a BCBS 239. And that will be
very manual. There’ll be lots of people, process, tools, all munched over the top of them. Huge
cost of operation. We’ve talked a lot about foundational. products here, data products, data
contracts. How that operating model works now at the moment, I think the figures I saw were
most banks are spending between 50 and 200 million a year for BCBS 239 compliance. So a lot
of money, lot of cash being spent on a very repeatable regular basis. that has absolutely zero
value to the bottom line of the organization. It’s A regulation, you’ve got to comply with it, but it
doesn’t make you smarter, better or quicker than your competitors. So change that operating
model. Can you put the components in place within your organization? that would then allow
you to move to a position where the operating model for 239 compliance is hugely automated,
self-reporting agents all the way through the workflow, pushing anomalies up. You still need
that human at the top to go, okay, I’ve identified and I’ve seen this.

Alpesh Doshi: So, so let me throw you a little bit of jargon. I don’t like to say jargon out, but as
we’re so one of the bits that we talk about is what’s called a gentic analytics and reporting. So,
moving from traditional BI, which is frankly what this is about, to how do you how do you?
Agentic systems do the end-to-end. So you have agents sitting there, they’re all different agents.
So for example, for a BCBS 239 as the example, there’s a bunch of rules and a bunch of things
you have to apply, like lineage and a bunch of other things. If you had agents sitting in the
background tracking it all, going, oh, you’re not complying with this reg, go and change it and it
will go out and figure out, oh, , I’ve got to change it and report it. So for me, that is a massive
reasoning step. It’s very well defined. BCB, okay, I don’t know, you can tell me whether it’s well
defined or not.
Duncan Cooper: Oh, it’s well defined. It’s not pleasant, but it’s well defined.
Alpesh Doshi: Okay, so if it’s well defined, that’s actually better for a model. You mentioned it
earlier around inputs and outputs. You mentioned it, you know, determinism, it is to do with the
inputs and the outputs, right? If you do the same inputs, you’re going to get some outputs. So
that’s interesting.
Duncan Cooper: Yeah, of course, yeah.
Alpesh Doshi: You know, for me, these are the kind of things that… I, you know, if you can, so
you’re an intelligent guy. When I start talking about these things, you go, bloody hell, yeah, I
never thought of that. And that for me is one of the things that what I’m trying to do, you know,
what I try and do, and I think for me is this, plant seeds into people’s heads to say, oh, think of
this differently. If you could do that. So the question that I have for you is how, and I get this
question every day, everybody tells me I’m going to lose my job. Everyone is fearful of their jobs.
So if we then extrapolate from what we just talked about around what the human in the loop
and the human over the loop is going to look like, give you a sort of view on actually what would
that, what would you think that would be? How is the person’s role in a particular organization
going to change with this new world that’s emerging? What would they be doing as a job?
Tomorrow, and it may be completely different. I think it will be completely different, but I’d love
to get your view of what that what you think it would be.
Duncan Cooper: The short answer is yes, it’s completely different. The parallel that I like using in
this space, and I’ve been talking with friends and some very smart people about it, is the
comparison with the Industrial Revolution. So End of the 18th century, the 1780s, 1790s, we
have the invention of the steam engine that allows us to automate pumps that allowed us to
empty mines of water so that we can do better mining. That then led to steam trains, it led to
centralized steam factories that allowed power. to generate spinning looms that produce
clothes cheaper, produce metal goods cheaper, and revolutionized economies around the
world, starting in the UK. That took about 100 years from start to finish to go from first invention
of the steam engine all the way through to electric powered factories in the late Victorian age,
1890s. And during that time, you are moving a society from agrarian pastural working, working

in fields, herding sheep, cutting crops, very, very manual labour. And you’re suddenly moving
them to very industrialised supply chains. And over those 100 years, that’s 100 years, call it
three generations. People’s minds changed and their ways they worked and the ways they
interacted. Society itself changed massively over those 100 years. Now, that was easy because
you’ve got 3 generations to do it. If you compare AI to the Industrial Revolution, which is not
unfair, it is a massive disruptive seismic shock to the way that we worked. That is being
implemented in the space of 10 years, 20 years max. So, you know, if you go back to COVID,
those halcyon days, pre-COVID, 2019, AI was something that only very specialist people worked
in. It was very niche, it wasn’t popular and nobody could really see the use for it day-to-day.
We’re only seven years later and it’s disrupting the way that people work and organise. That
whiplash effect of the speed that we’re changing, we’re not doing it in three generations, we’re
doing it in the career of 1 generation, just one part of a career of 1 generation. So, it moves
quick. People will have to learn massively new skills. That disrupts the operating model, it
disrupts the way you work. Now, I’m hard to believe I am 52. You know, these youthful good
looks do confuse people, but I’m 52. I grew up with the BBC microcomputer and programming. I
was writing code when I was three. It becomes second nature to me. Yeah, I love it. I’ve always
worked with it. But I am abnormal for a mid 50s person. And that knowledge and appetite and
buying into things like AI changing the world for them now is going to be a challenge. But my
daughters, who are 13 and 15, it’s second nature to them. So If we look at how roles are going
to change, new skills, new knowledge, you need to be continuously upgrading your skill set. And
if you don’t or can’t or can’t be interested in it, then yeah, you’re probably going to have some
challenges in the future world. But If you’re smart and open to it and apply it and try it and do it,
then you’ll find it’s a multiplier to your knowledge. And as mid 50s, we know quite a lot of
things. So it just makes us even better.
Alpesh Doshi: Right, we’ve got, I’ve opened it up for questions. So if you go to the Q&A, if you
have any questions, please ask them. We’ve got a few minutes to talk about governance and I
know you have a specific view on it. So let, I’ll just let you tell us what you, what your view is.
Duncan Cooper: So governance, governance, I tend to describe much like reconciliation, an utter
waste of time. And I don’t mean that in that you shouldn’t do it. I mean that it adds no value to
you. Reconciliation is making sure that the thing you expected to happen has happened.
Governance is applying the rules that you want your organization to work in and ensuring that
people are compliant with it. Now, we live in a society where we have rules and we are not
followed around by individual policemen every single day asking us to report and self-report
that we are complying with all the rules. We know what we need to do within society and we
know that if we break those rules, we’ll get caught or we’ll get told off. Some of those rules are
petty, like jumping a traffic light. Some of them are very important, like not breaking in and
stealing things. Governance is that same thing. But how do we change governance so it’s an
enabler to people as opposed to a restriction? And this goes back to the rip the operating model
up, because the operating model for governance is set up for manual processes and the way we
worked 20, 30 years ago. It’s not fit for purpose. Governance in the AI world needs to

completely change, and it will change with the operating model, but this is going to be a long
and hard process.
Alpesh Doshi: Yeah, I think, I mean, this is going back to what we talked about around agentic
systems, swarms of agents, automation. For me, governance has to be embedded as part of the
process and then frankly. One of the ways to think about it is there’ll be a suite of agents that do
governance as part of the process automatically, right? It’ll do it because you can’t frankly do it
afterwards. If it’s running a BCBS 239 report, you have humans doing the reporting on BCBS or
the aspects of BCBS 239 as we talked about. In the new world, it’s doing it all on its own. And
unless you put some kind of system in between to manage, do the governance, and you can still
have human oversight, you know, human over the loop oversight, you’re screwed, right?
Because you’re not, you can’t wait until after the systems has done what it’s doing. What did it
do? Because today what happens is, I’m sure you know, do a bit of process, have a team that
does the governance and reporting, so you report them and create their report on a weekly
basis and go, oh, that didn’t work. And that’s how it works, right? So I think, and this is the other
thing for me, if you think about, I don’t even thought about this. I don’t see people even thinking
about the fact that AI is going to change how governance is done. And in fact, what governance
can do, you know, because a lot of the things you want to do in governance, you can’t, because
it’s too much effort. It might take too long. So the kind of reporting, if you talk to anybody,
anybody of governor, head of governance, say, I want to do this, this and this, like, well, I
haven’t got time for that. I can do this first thing, but the other two things are going to take
weeks. So they can’t do it. They go, okay, fine, don’t worry. We only have to do the minimum
that we need to do, and that’s where we’re fine. We’re compliant and the story. But I think the
new world is, actually, you know, we can do all three. We can automate the whole thing and we
can get more business value, going back to what it said around the value of data.
Duncan Cooper: Yeah, I mean, when it comes down to it, governance is a people and process
problem generating data. The trouble is the way that it generates that data, project statuses,
processes, policy compliance, is a really manual way of doing things. It’s very human intensive.
But it’s just the same as the problems we’re already solving further up the line in places like 239
and various business operations like cutting a nav or calculating your security masters change
over the past day. It’s just a people process and data problem. So look at it as a people process
and data problem that we’re solving already. Agentic AI, your agentic PMO. I would love it.
Alpesh Doshi: So that’s your future team, right? Is you and 20 agents doing the job of 50 people.
Duncan Cooper: Someone that has to deal with Jira on a day-to-day basis, so I don’t have to. I
love it.
Alpesh Doshi: Hey, Jira, that was a whole conversation on Jira, I can tell you. But so we have a
question and it says, I’m going to check my glasses and I can’t bloody read it. It says, could you
please comment on KYA, know your agent, as there appears no trust layer or identity protocol
for agents performing on behalf of humans?

Duncan Cooper: It’s a good one, isn’t it? It’s a good one because it is a problem. How do you
define an agent and how do you control that agent? We know how to define and control our
people and we give them appraisals and we tell them off if they’re bad and give them pay rises.
We are ill suited To doing so with agents, but the concepts are the same.
Alpesh Doshi: So, so the kind of things that we’re working on is, so I was with a large financial
institution and the Chief Data Officer, and one of the things that he’s looking for is what’s called
an agentic control plane. One of the elements of the Gentic control plane is an agent registry
and an agent identity. So if we take the world that you’re in, you’re going to have 50, 100, 200,
1000 agents all running. Firstly, you need to know what the hell they do. They’ve got to be
logged somewhere so you can find them. They need metadata around to explain what they do.
You need to identify them because you need to know their uniqueness and what they can do.
You need to have an identity of the agent. And what I’ve seen so far is there are no systems that
do that yet. There are a few emerging than you, but the traditional providers all systems have
been developed, IAM or, you know, that kind of thing, for people. They’re not developed for
agents, not developed for the lifecycle of agents and managing identity for an agent, right? So I
think what I’m seeing is… People are starting to talk about it, but it’s not at a place where…
They’re going, it’s a serious problem because everybody’s using chatbots, let’s be honest. Right
now, they’ve got conversation layer, they’ve got pick your model underneath and they’ve got an
interface and you type into it. Well, that’s it. There’s no like 50 agents sitting around gonna do
jobs automatically and you need to know where they are, who they are, what they’re doing,
what permissions they have. Right, so, so for me, it’s um… It’s a new thing that I think people are
talking about. We’ve got another question. Here we go. What are your thoughts on how should
organisations rethink governance when agents become decision makers rather than just
decision support tools?
Duncan Cooper: It’s a very good question, but I don’t see agents becoming detailed decision
makers immediately. And this goes back to our human in the loop kind of comment. I think
there are industries where they will, and you touched on things like anthropic and their swarms
of engineers working autonomously. But we work in financial services and it’s slow and it’s
methodical and it’s risk averse. I think it’s going to be a long, long time before any bank is
comfortable saying, yeah, this is now the decision maker, this agentic platform. When it does, I
completely agree that there needs to be a lot of thinking about how we manage that and how
we record those decisions. But I don’t see it happening that quickly. Could be contentious.
Alpesh Doshi: So the, I mean, one of the other things that we’re working on is doing that is
tracing decisions. So having reasoning and decisions, so what you do is recording every decision
the model makes. So you have reasoning steps and you have a decision that’s made, you’re
capturing all the reasoning steps and the decision itself. So, I mean, these are the kind of tools
that people need, right? You know, and these are large, complex systems, right? If you’re going
to have, if you’re going to have a model and an agent throwing off lots of telemetry, which it
does, you know, some of the things I can show you, I haven’t shown you, but some of the things

I can show you, it’s throwing off just one agent running for 5 minutes, because throwing off
hundreds of messages.
Duncan Cooper: So the flip side to this is if you look at most organisations, how do they manage
decisions at the moment without AI? And quite frankly, it’s usually pretty badly. You know, this
is not stuff that’s tracked or traced or controlled in any particular formulaic way. So your point
is, as soon as you start developing that model, of structure and repeatability in controlling that
process and controlling those decisions. You’re actually doing it better than the humans are
anyway, so it’s a good thing.
Alpesh Doshi: And we’ve got a few minutes left. As agents and Gen. AI is becoming more
capable of planning and executing complex workflows, what unique human capabilities do you
believe become even more valuable in few domains, specifically in software engineering?
Duncan Cooper: The gaps that I see Jen I having is around the soft human skills of ethics and
empathy, which is hard, it’s hard to capture in itself, and knowledge or historical or institutional
knowledge. Most organisations are again poor at recording their institutional knowledge and
tracing and tracking why you’ve done things. Go into any big bank and their business processes
are 20, 30, 40 years of scar tissue, of checks and balances put in there because we had a mistake
once and we don’t want it to happen again. And now, nobody knows why or what that is, so…
That knowledge, as soon as you can start to capture knowledge better and control knowledge of
workflows, that will allow AI to become much better at that process flow. So you’re moving out
of the engineering world and you’re moving into the operations world. Will it ever get to the
front office and will you start allowing machines to run your portfolios? You’re seeing it already,
people running portfolios. That’s easy until it goes wrong. And I think when it goes wrong will be
the interesting point.
Alpesh Doshi: No, it’s, I think, so, you know, given you’ve, you know, this stuff has been around
for 3 1/2 years, right, really, you know, end of 22 is when these models came around. But
actually for me, the bigger, the bigger sort of turning point is actually when reasoning models
really sort of came along, which was about a year and about April last year when the first
versions of the reasoning models came out. For me, that was really the turning point rather than
the fact that LLMs came along, they could you could summarize documents and that kind of
thing. And if you look at the adoption of that. It’s so basic right now, so basic, I just don’t see it.
You know, given your work career, what you’ve been around and what you’ve seen, you
explained a historic version, you know, give me a view of what you think, you know, where are
the sort of phases? What is the sort of… indicators of adoption and change going to happen over
the next whatever period you tell me, you know, is it going to be one year, is it going to be two
years, is it going to be five years? And some characteristics of what changes that are going to
happen over that period. Don’t have to give everything. It’s just for me, it’s trying to explain to
people, look, in a year’s time, we’re going to see these things. In 2 years’ time, we’re going to
see these things. and just trying to get vindications to people. what they can expect.

Duncan Cooper: Any support function in an organization, so things like HR and finance and IT
support, a big indicator for me will be the automation and headcount reduction in those spaces.
That’s not to say they don’t add value, but it’s manual. So, for me… Those sort of areas,
suddenly getting AI agents running off of the top of ServiceNow or an HR system that’s handling
hiring and firing. These tools there, when you start seeing movement in that space and
reduction in headcount, reduction in cost, that will be a very interesting trigger. When you start
them separate, stepping into non-revenue generating functions within the organization. So we
touched on PMO and automated governance. When you see platforms coming in and handling
business tasks then, touched on fund accounting or nav creation. You’re now seeing big seismic
movements. These are all within the constrained world or landscape of existing players in the
financial institution space. So another big thing for me will be a new entrant coming in. Can
someone come in and create, I don’t know, an asset servicer in a box? Because what an asset
servicer does isn’t particularly complex. It’s just very driven by regulations and regulatory setup.
But you could automate. everything that an asset servicer does nowadays. Would it scale?
Would it be as controlled? Would it be as governed? Probably not, because it’s the first of
anything. But if you can get it to the point where you’re starting to adapt and improve it and
scale it up, you don’t have the legacy infrastructure. of an asset servicer that’s been around for
100 years and is doing the same thing over and over again. Because they’re doing the same
thing they were doing in the 1950s. They’ve automated communication a little bit. They’ve
automated fund accounting a little bit. But this is essentially the same business process that
hasn’t changed. since the mid 50s. Can you come in with a new business process, new
technology, new processing and start from scratch, which is really weird because every other
industry has had these new entrants coming in with no legacy or no history and scaling really
quickly. The only place we haven’t really seen it is in traditional banking.
Alpesh Doshi: And do you think, I mean, one of the challenges is, you know, most of these firms,
we’re talking about financial services here in general and the kind of things you worked on, you
know, outsource asset finance, asset servicing collections and other things, fund administration
and custody and all those things. You know, most of these firms, have large, long-term
contracts, multi-year contracts. So, you know, my assumption is that they’re not going to take it
lying down. If some, if a player comes along and says, you know, you’re, you know, we’ll, you
know, you’re spending, pick your number 50 million today, we’ll do, we’ll do it for 10 million a
year, you know, Do you think that the senior leaders in these firms are ready to take the leap?
Because that’s the problem, right? Let’s take the lead.
Duncan Cooper: It’s as blunt as that. No, I don’t think they are. I think they want the cost
savings. They want the financial impacts, but I’m not sure they’re ready to go through the
heartache and pain that is reorganizing your business around a completely different governance
and operating model. But more importantly, it’s the layers below them. the people who are
invested in the organisation and the status quo, who are going to be sitting there going, I don’t
want things to change. I don’t want to change my job or be more complex. I want things to
continue as they are. So I don’t think change is going to come from inside financial services. I

think it’s going to be forced Into them from outside. And that barrier to that barrier to entry to
being a custodian is very high from a regulatory burden. But if that’s the only thing stopping you,
at some point someone will think, yeah, it’s actually worth the risk.
Alpesh Doshi: Yeah, it’s true. Well, I think we’ve run out of time. Thanks a lot for agreeing to
participate. Hopefully we’ve given the audience some insights and new ideas to think about and
new thoughts and old thoughts. You know, going back to the industrial revolution and how the
world worked then and what it’s going to do now. That kind of analogy is always useful, right?
So for people to get it. So thank you very much, Duncan, for participating. And we’ll speak again,
I’m sure. And thank you for the audience for participating. And you’ll get a summary of what we
talked about on an e-mail. And look out for more webinars from us soon.
Duncan Cooper: Alpesh Doshi 1:00:08 Thank you everybody for joining. Thanks a lot everyone.