Delivering Enterprise Value with AI Agents in a Regulated Industry

This is the Data Science Conversations podcast with Damien Deighan and Dr. Philip Diesinger.

We feature cutting-edge data science and AI research from the world’s leading academic minds and industry practitioners so you can expand your knowledge and grow your career. This podcast is sponsored by Data Science Talent, the data science recruitment experts.

Welcome to the Data Science Conversations podcast. My name is Damien Deighan and I’m here with my co-host Philip Diesinger. How’s it going, Philip? Great. Thanks, Damien. Looking forward to this one.

Good. Yes, very much so, because today we’re going to be talking about how to deliver enterprise value with AI agents in a regulatory environment. So our special guest on today’s show is Andrew Mohammed. the current director of data and AI at OVO Energy in Bristol.

By way of background, Andrew began his now 20-year long corporate career by working at Centrica, then EDF Energy, where he worked in a number of divergent roles across the business, spanning commercial, strategy, risk hedging, and even sales and marketing. From there, he moved into a senior data science management role in financial services with Lloyds Bank. And in 2022, he joined OVO Energy as head of data science, where he now is also in charge of everything AI. So in terms of Andrew’s academic background, he graduated with a math degree from the University of Birmingham before completing a master’s in quant finance at the University of Reading. Andrew, we’re delighted to have you on the podcast. Welcome along. Hi, Damian. Hi, Philip. Pleasure to be here. So if we just start with your career journey so far, Andrew, you know, despite your maths and quant background, very earlier in your career, you ended up in more business type roles and then you made it into data science leadership roles.

Can you maybe talk about that and how that has shaped your overall approach to your current role?

Absolutely. The journey most people go on is what they learn about themselves during their career. I think what’s been true to me is that I I really get energized by problem solving, real world problems for customers or for businesses. I just happen to be strong in maths and coding. And over the years, that’s just kind of grown as the technology has grown to converge into the AI space. So I started my career off in commodity trading, in forward procurement of energy contracts and optimizing them. I was very fortunate to be, you know, mentored by a kind of colleague there who was, you know, proficient in coding. One of the things I kind of learned from that was the sheer amount of routine, you know, you do day to day, kind of check the news, check the prices, but the real value laying like the anomalies, you know, market events that were bubbling up or shifting prices or changing behavior, you know, from a key industrial customer. So the more you start to automate the day to day, then we could look for the real. pockets of value.

Those are the strengths I’ve thrived upon, really. So we’ve learned commodity trading and then have worked in two regulated environments, you know, working on business problems across EDF, whether it’s long-term, 50-year-out strategy and planning to running commercial parts of the business, but always relying on automation. And then that kind of grew. Financial services and now energy, we’re kind of leading teams that can bring that automation in a far more intelligent and heavier place. I’d say what I’ve taken more from the academic world is just the sheer rigor that you can take into any business situation. Always have a hypothesis, but be ready to be proven wrong. And then where’s your evidence? It’s kind of through doing a maths degree, through doing a finance degree, and whichever room I’ve kind of walked in in business. It’s just the models have got smarter, the data’s got richer, and we’re in the age of AI now. Great.

And your current role, you’ve very quickly built a team from 10 to 12 people up to, I think, 80 or 90. Do you want to quickly talk about how you structure things and what you’ve learned about that process? 

Yes, it’s been a privilege to kind of lead a large function now as director of data and AI. And it’s always been helpful to think of leadership in its different levels and how you adapt personally. So I always like to think of it in terms of people, systems, value. But when you’re working in a team of 10, you’re very much part of the team generating the ideas. It’s about how you solicit and harness the ideas from the group as a working member to branch for the people side. When you start to build into a team of, you know, 20 to 40 people. As we kind of did at OVO, we realized there was some real value to be gained from the proof of concepts we were kind of shipping in the AI space. We started to scale and grow the team. You then have to think about the systems, you know, as a technology leader you can bring in.

By systems, I kind of mean best practice, how you upskill people. You can’t do this individually or team by team. So systems attest the quality. How do we know we’re shipping the right things? What’s a good level of accuracy? What are the principles you want to instill in communities of practice?

And then as that becomes even bigger, you know, into 80 or 90 people, your role then very much becomes about the value that you can add. You know, are you allocating directly in those teams the best use of value that those teams feel that they’re working on the biggest problems of the company? And therefore, you then start to focus on how do we measure the value and the kind of outputs. And for me, that was like one of the biggest. challenges you know in leadership transition to go for someone who’s highly energized by problem solving working as part of the team on the solution as the team gets bigger you can no longer be the bottleneck you know when it grows to 40 to 50 to 60 so you then have to turn your problem solving into more systems thinking that’s one of the biggest things i’ve learned as a leader is how to adapt to my personal operating system with the role but what got you there To one role, it’s not the one that gets you to the next one, no matter how strong you are on the previous role. And did the onset of agentic systems change your philosophy or your approach in hiring? There are some things that never change.

I think giving bright people the autonomy and fearlessness to tackle big problems. The technology has changed and maybe the fact it’s more directly customer-facing may have done. But the best teams I’ve always led and kind of worked with, they’ve always had that culture of, you know, moving from the data science space, you might run 10 experiments and nine times out of 10, it may fail, may not be good enough. The 10th time might be an absolute game changer. So I think things that hold are people who can take on ambiguous problems, be prepared to fail, but learn from it and adapt until they hit the winning recipe. So I did. 

The quality of people and the ambition in there is always a great hallmark for hiring and skills. As we’ve moved into, technology’s moved on, it’s perhaps where the way you apply those skills has changed. I remember early on, working in more traditional machine learning and data science, there was so much of an art towards life. model selection, trying different models, different approaches, compare and contrast, erotive weaknesses, confusion matrix at the end of it.

Now, as we move into an AI world where many of those models are rented from Google or Anthropic, those skills are probably more valuable at the evaluation end. Where are we optimizing the model? Which kind of parameters? Where are the failure cases? What’s the cost of that failure? So for me, it’s that the people who’ve moved fastest with the technology have learned where to apply their skills for most value, rather than feeling that their careers are changing or moving away from them. So I hear that you’re putting mindset and attitudes first when hiring. 

But how about the technical skills? What are the typical profiles that you’re looking for? 

A key part of the role is to be there as an advisor on risk and reward. So I think what we, two other things you kind of look for are people who can convey that really succinctly and accurate and efficiently, you know, the risk reward of adopting an approach. And then secondly, they know what it takes to put that system into production where you learn the most. So I see so many things where people are, you know, looking at their career and looking at their CV and, you know, how can I… know, do more or is it the variety of techniques or is it the most number of demos? And I think people who’ve actually put a handful of things in production have kind of realized, you know, what you’re effectively doing is showing we’re putting this model forward or this approach to automate a process, but to risk or trade off. It’s not going to be 100 % accurate, but the benefit of the upside far outweighs the downside that we can live with and we can then systemize and scale it in a way. that kind of monetizes it to the right level. So deep all those skills, an eye for production, but also convey, honestly, the risk of all trade-offs are the people we will find themselves the highest in demand. Yeah, you’re talking about putting systems in production and so on. Maybe you can give us an overview quickly of the current challenges in the energy sector, how regulation basically affects those, and maybe what the typical use cases are. for people not familiar with the sector? Yes. So for energy suppliers in the UK, we operate in a market where it’s a standardized commodity.

The electricity and gas that comes through is the same quality wherever. But in terms of the quality of the service that you provide, your ability to help bring information to the customer to help them save money and use it economically, or invest in… Good choices that reduce their carbon footprint and their use of energy is where the value is created. And it may look a very simple commodity on the outside, but behind it is a very complex series of market forces. You’ve got a number of competitors in the upstream generation between traditional fossil fuels with bombs and solar panels and nuclear power stations all have to be optimised. You’ve got many different… suppliers kind of aggregating and providing that service to customers that require contact centers, back office operations. And then you’ve got a whole host of data aggregators, the smart metering, the fact that every half an hour you can get a different data point on the usage of that customer should be very valuable in helping them optimize their home. And then third parties who can bring in solar panels, electric vehicle facilities. There’s a whole ecosystem there that has to balance. security of supply of energy, reliable supply to homes, and cost and affordability because it’s an essential for people, not a luxury good. And the third part of the trilemma is doing that at a weight which is reducing the carbon footprint of the energy sector. So it’s trying to bring a whole complex array of things to customers who ultimately want a reliable service they can afford.

They don’t feel as detrimental to the planet. You mentioned already data aggregators. So how does this connect to data science or to agentic systems? How do you bridge the gap between the use cases and then the data and making great value from that? 

Yeah, what it gives right to is a rich variety of problems where as engineers, as data scientists, can we abstract away complexity from the customer? So they feel they’re getting a reliable, powerful service. So there’s a number of things we’ve built.

I can give examples, the genetic systems that we’ve kind of looked at. If you take something like the contact center, customers call in and there could be a whole host of reasons about their meat are not working, about which price they’re on. It could be a conversation that happened six months ago. You know, they moved into their house and they were looking to set up. Now they’re ready to set up a whole host of payments. So what you have in that human advisor is a whole range of data points that could have come at different times in different formats that need to be presented back to the customer really succinctly into what the customer can do next or how we can help them. And that’s where you see the variety of outcomes before AI or before a data science team gets in.

They’re high volume. They can be high frequency. But there’s also a high variety of problems and things that get talked about on the call. And that naturally lends itself to it being valuable for data scientists to review the problem. Can you make it more specific, a little bit tangible for me so I can better understand what kind of problems would the client face?

So I’m assuming, you know. You have different tariffs maybe or different types of energy that you want to discuss with your provider or maybe there’s a problem or something occurring that you want to have solved. Yes. One of the most frequent requests that we kind of get are customers who want to understand the best way for them to pay for their energy. And the complexity comes in because you’ve got energy in kilowatt hours. that varies by day and by season and the weather. They get smoothed out in terms of fixed price unit rates and variable tariff rates that then get smoothed out into average direct debit payments per month. So all of the maths in that is like really complex. Even myself going down, you know, mathematics degree, had to sit down and work out that it’s okay to be over and under in terms of your payments because of seasonality of energy use.

So one of the best use cases we’ve built an agentic system is actually helping people understand their bill and their payments. So if they have a bill and they kind of say, oh, it’s an increase in their bill, we’ve got an agentic system, we’ll kind of look to break it down in terms of explaining what’s changed in their usage, has the unit price of what they’re paying changed, and has there been an up-to-date meter reading that would make it more accurate.

And therefore, it’s changed. And therefore, you’re able to tell a much more succinct story to customers to say, yes, your bill has changed. These are the three factors as to why it’s changed.

Most importantly, and this is the benefits of an energetic system, here’s what you can do next. We’ve noticed that if you change your monthly payment to £30 a month, then it will be much more affordable, little and often. You offer different tariff types. You can tell them, actually, we’ve noticed that you tend to use more energy at the weekend or you’ve got some flexibility. You can use some of that more in the evenings. Then you’re able to offer the customer perhaps a better pricing structure that fits with how energy is bought and sold. So that’s the importance, I think, where it becomes more of an agentic problem is many variables are moving that you need to synthesize and something succinct.

And then you’ve got multiple options on price and payment method. How do you convey that in simple, clear terms to an advisor who may be talking to a customer who may be struggling to pay their bills? And that’s where you want to get them onto the best option first. So if I look at it a little bit from the analytical, technical perspective, the way I’m imagining it is that you have classical data science tools like predictive. models maybe that take into account seasonality effects and so on which we have have existed for a long time basically right so there’s no need touching that with agentic or whatever they exist they’re good and then you have basically a new agentic approach agentic model that basically is there to create trust with clients with customers to explain you said yeah and to also maybe automate a little bit this uh interaction of the you know an advisory role trying to help clients find the best kind of method for them to to approach their payment schedule and and so on is that the fair way of uh imagining what you have built it is yes it’s effectively i talk about like bringing in teams of people with a creative mindset is What’s the best portfolio of tools and techniques that resolve the problem? So you can imagine if it’s just like verifying someone’s identity, you don’t want a probabilistic system in there. You want, this is the customer, their account number and proof of identity. If you’re talking about their predicted usage, that lends itself very much to kind of machine learning.

How much do I expect to spend over the next few months? You want good, reliable calculators, not large language models. And then if you want to then be able to recommend for that customer as an individual, what’s the best way to explain a series of actions they could take, then that lends itself to more of a generative AI type of opportunity that can explain and call the right service to them. And I think that’s where the companies will kind of get the most value. I’ve seen these cases where people just try to force everything into an agentic system. push as much context into a large language model as possible and make it as accurate as possible. It’s always got that degree of randomness and it’s got a very different cost and training structure to other tools as well.

And that’s where I say that although the language has changed around artificial intelligence, the mindset of people, the right tool for the right job at the right cost is what drives the best answer for customers. And how do customers engage with your system? We’ve built a variety of kind of tools that can maybe talk about. So one of those is helping advise us live on a call where a customer calls in and they start to mention that they don’t understand their bill or they want to find the best way to pay. And so we’ll have an AI agent that will kick in automatically about hearing that from the customer. We’ll start to bring back and compute all the numerical information about their account. and then display that information to an advisor. And the benefit of that is we see calls that are higher satisfaction levels. We also see far fewer repeats.

So we provide the right information at the right time, and it means we’re able to handle the customer’s call once rather than many times. More of the use cases that we’ve moved more into, though, are more in the preventing customer problems. So one of the AI agents we’ve been looking at is Are we producing customer bills of the right quality that we can send before they go out?

And this is one of the best things to do is remove the friction point before it occurs. So if there is a customer provides a meter reading and they accidentally put an extra zero at the end, so rather than then sending a bill that’s much higher and that’s incorrect, these two calls, complaints, could you trap the anomaly, provide a better estimate? send a bill that’s more likely in line with their usage, and then gently go back to the customer and tell them, are they sure about the meat reading? Does it look accurate? And therefore you prevent the urgency and the pain the customer’s feeling. You also provide bills just being stopped, which actually block revenue for the company. Because if we don’t trust the bill to send out, normally it stops. Have an agentic system that can… You know, make sure it’s of the right quality. You unlock billing revenue. You stop a necessary demand failure contact in the contact center. And that you just reduce the number of handoffs in back office teams that have to work out what went wrong. So it’s got a triple benefit from one very defined agentic system.

And it’s probably one of our most valuable use cases, preventing pain before the customer sees it. It’s also one of the hardest to prove the benefits case. Because actually what you’re trying to prove then is like the absence of pain. So if this system is successful, we should see fewer calls and fewer complaints. It then becomes harder to prove the counterfactual to say, well, how do we know a complaint would have occurred? So then that’s when you have to go back to the discipline of data science around holdouts and samples and proper A-B testing to prove a genuine uplift. So moving much more into the preventative, predictive maintenance phases unlocks a lot of value from the genetic systems.

I would like to take a brief moment to tell you about our quarterly industry magazine called The Data Scientist and how you can get a complimentary subscription. My co-host on the podcast, Philip Diesinger is a regular contributor and the magazine is packed full of features from some of the industry’s leading data and AI practitioners. We have articles spanning deep technical topics from across the data science and machine learning spectrum. Plus, there’s careers advice and industry case studies from many of the world’s leading ompanies. So go to datasciencetalent.co.uk forward slash media to get your complimentary magazine subscription. 

And now we head back to the conversation. So you mentioned already multiple authentic systems. How many of those have you built and how do they coordinate with each other? There are probably about 10 different agentic systems that we’ve built. We’ve run over 70 machine learning models. The reason I’m going to hesitate on the question is what we’ve found is that it’s not about the number of agents that you build or the number of systems you build, but just how effective they are. At one point, when I was new in the role, they were trying to come up with key performance indicators and a scorecard. And it’s like, is it the number of agents? Build the most number of agents you can. What we found is that when they were built in isolation, those usually end up duplicating cost and complexity for very small kind of localized solutions. So we’ve got a number of solutions, but we’ve looked to deliver it in two ways. One is through an agentic platform. So every time somebody builds an application with an AI, like an AI… a utility, an evaluation package. Can you then publish it to a central team?

That means any other team can reuse it and save the cost of redod. And then secondly, by having that agent there in a catalog, it doesn’t have to be reinvented twice, but you just invoke that agent in another or service in another setting. So identity and verification, you should kind of build once. If you’ve got the opportunity to check and explain a customer’s bill, why it’s gone up. That’s useful both to the customer on the app. It’s useful to them in a call with an advisor, to explain it simply. It’s also equally useful to a colleague who’s just reviewing the quality of bills being produced in back office. It’s like build one agent, use many

times. So we don’t tend to guide ourselves by the number of agents.

What’s the number of capabilities per colleague that we’re introducing? As we can add more tools into an existing agent, that’s far more efficient than having to build. another system on top of it you mentioned earlier that the advisors also have basically a virtual assistant in the form of an agentic system that helps them in the calls with the customers how did you introduce that to the advisors and how was what was their reaction yes this is one of the biggest things to do is to understand the the culture shift that’s needed because In any kind of value from projects, I say 20% of the value comes from the technology and 80% of it comes through adoption. You can build the smartest agents in the world, the best kind of ecosystem, but if you contact center advisors, don’t trust it or won’t use it, it’s not going to realize a single penny of value. So one of the clear things for us from the outset was we did two things. Made sure we had joined up operating teams. and they were involved in design. And then secondly, in the design of the AI agents that we’re building, make sure they’re kind of scoped to the right level that kind of works in the workflow. So an example of this would be out. We built an AI assistant that detects if there’s a customer has any sign of vulnerability. And what that means is from a regulatory point of view, if there’s any risk on their being able to afford their bill. or a medical condition, or young children at home, or they’re a carer, then they have to be a priority service to make sure they don’t get disconnected, that we help them out financially and they’re made fully aware of all the support because it’s an essential

commodity. Those two things is actually bringing in contact centre advisors and helping them design what was needed when customers call in. What do they need? Why do they need it?

What more could you do for them if you were given access to technology? And that helps break down the kind of barriers of making it useful and bringing people together for the design. The other part I mentioned was the scoping of the AI themselves. So if you think of agent maturity as a series of levels from something that provides information to an advisor to level zero to something that… Then can kind of recommend actions to then has human in the loop approving the actions and human on the loop observing it to something fully autonomous. If you’re designing something like a customer vulnerability service, you deliberately want that to be approved by a human.

And that meant we were designing an agent that was in the workflow of the advisor to say, I’m speaking to a customer who wants to be heard. They may have very tough personal circumstances at home. If it’s someone who’s lost their job or has a medical condition and they’re worried about paying for their bill, they need to be listened to, it’s not about the fastest AI-generated answer that you can give them. And then it’s helpful for that advisor to then synthesize, have the information presented in front of them as to the ways they can genuinely help, but then reflect on that and make sure they’re conveying it in the best, most empathetic manner at the right time.

So designing an AI agent that was deliberately for… consumption by a human to then consume and then replace the customer was very different than if we tried to scope it as autonomously as possible, resolve this problem by customer circumstances, get the fastest answer out there. You end up just only subtracting value from the use case because customers just want that level of empathy.

So what we actually found, even though it was designed with, you know, presented to a human for verification, to then articulate. We reduced repeat contact by a third. We saw 30% increase in customer satisfaction. And the biggest source of feedback was that the customer felt listened to. And this is when people are in tough circumstances and they’re making the time to tell their story. One of the reasons we had lots of repeat contact was if they felt the advisor wasn’t listening to them. Whereas when they’ve got a AI assistant on the phone that can bring help that’s relevant but also maybe replay at the end of the call okay i’ve heard you these are three things that matter to you you know neatly summarized by an ai but it’s the human talking it over gives a lot of reassurance that they’ve been listened to concerns recorded so that’s been the best way to build trust in an environment where naturally people are worried about ai automating a very wide job or i don’t understand this concept and therefore i don’t trust it or don’t use it Or at the other extreme, people just use it in blind faith.

They’re not aware these are problemistic, you know, need some form of verification and repeat it without reviewing first. So yeah, it’s been a very interesting study of getting human behavior to adopt technology to make it work. Yeah, I think building trust in conversational systems is like a key factor. It’s very easy to overlook or to skip because it’s hard to measure and hard to make tangible. But that’s usually what drives success, especially if you have something that’s client -facing. That’s an absolute must. So I want to get back a little bit to the advisors, basically, leveraging the agentic system. Did you measure any KPIs to see if the performance went up of the advisors? So you mentioned already the clients were more satisfied, basically. But do you have any business KPIs that you measured? number of cases per hour or something like that. Yeah, it’s so important to get the evaluation right, like for trust in the product, but also where to optimize it. So there are natural high-level KPIs in the business around the financial performance of what we’re delivering, that we should see a

better commercial, we should be a better business. Customer satisfaction should be increased. 

So we shouldn’t be making more money at the detriment of customer experience. And we should be improving the quality of data as we’re going along and get to know our customers better. Challenge with those is they’re all kind of quite lagging metrics. So it’s really important to think of what are the leading metrics when you’re kind of building and designing these systems. So how many customers are in scope? First of all, you know, the start, I think a bit like a sales funnel, how many customers are in scope? How many of those interactions have been passed through an AI tool? And then how many were a successful output generated? And then of those outputs that were generated, how many were a recommendation that was then used and acted? And that’s where you can use things like the number of AI requests per call or measure the number of contacts and frequency and repeats of contact. You can then go back to that kind of funnel view where you’ve got all those leading metrics and you say, well, there’s a repeat contact. Was AI used in the first place? Yes.

If not, it’s an adoption problem. We haven’t got through to as many advisors as we wanted to cover it. If it was, the AI was used. then it’s an optimization problem. What can we learn from the first time? It wasn’t a success in getting the result. It led to a repeat contact and a follow-up. And therefore, what did we learn that could make the product better again? And those are quite two different things. And that’s why, as you described, measurement of the system is really important. And when you’re designing these systems, in your practical experience, what separates good design from bad design? Yes. I think a number of key principles can work with AI. I think first of all is what’s the simplest, leanest system to deliver the result? If it can be done deterministically, is it a calculator or is it something that requires more of a decisioning layer over it? So if it doesn’t require complexity, don’t build it in. Use the most easiest deterministic service if that’s possible.

The other one is like scoping things at the right level to get the business case right. Like in the vulnerability example, if you’d have scoped, say, okay, look at all of these contact center,

all of these real customer problems on affordability, and we can make them all go away with Agentec AI, which created a huge business case, but then you’ve completely only delivered like 10% of it.

So if you kind of scope the AI to be right size to the business problem. No, we want this to be presented to a human advisor to review it and make sure it’s correct and present it empathetically. You’re then scoping both the agents and the business use correctly. And then the other one is having the key metrics in place are kind of a must. Because as I said, the challenge in data science has kind of moved from model selection. If you look at all the LLM

frontier models are now within 5% to 10% of each other, you’re not really buying any additional intelligence. but in how quickly you can help those models learn. So the key design up front is rather than putting a system live and then recording experiments and several months later, how can you build evaluation at the start and design at the start of the project to make sure that you’re getting that feedback loop. And I can give an example of the importance of that. It’s almost like one of our failures in this space. We saw a lot of customers who were very digitally active, digitally enabled, and then still calling into the contact center. So our thought was, oh, they’re struggling to get information. AI can present that information better, faster, richer than ever before. And this will mean the customer’s fully served on the app. What actually happened was the reverse. We put an AI assistant on the app. We found it was actually driving more contact into the contact center. And so because we have an evaluation, we could measure it. It wasn’t a great result that we wanted. It’s going the opposite way of the business case. But what we were finding was that we were giving customers better access to information. They could be clear about their account and clear what they’re asking for. But then there was a missing service of when they wanted to speak to someone to get that service done. So what it kind of taught us was that actually putting in one of the next best follow-up actions after information, that means that customers feel really satisfied when using an app. We wouldn’t have caught that if we hadn’t put measurements at the front of it. And in this whole design process, where do the regulations come in? Like what are limitations that those impose on design or the way you would like to maybe build a system, but you have to find ways to deal with the regulated environment? Privilege to work in both energy and financial services, heavily kind of regulated. My strong advice, bring all the regulation to the front of the process. Embrace it at the start and find ways to codify it. Because if you think of the opposite approach that maybe I’ve seen in other use cases when it comes to machine learning models, it’s quite often they’ve built and designed optimization models that have helped customers with the process. And then at the end of the process, it’s then been restricted to which customers you can apply it to because… There’s already a complaint on the account and therefore we don’t want to have a sales conversation or its customer has a vulnerability marker flag and therefore it might be risky. And the business case just shrinks and shrinks and people just then find ways to constrain it out of natural caution as a business for your reputation. Whereas if you look the other way around, what’s possible with context engineering now is if you take all of the regulation complexity there, if you start to overlay that with your the goals for the customer, the information about the customer and the outcomes you’re looking to drive. You’re actually

building like a regulatory framework around the model by design. And what we found is we realize more value because you’re taking into account all of the regulation and building it up front. You’re still able to provide a recommendation. You’re also building more trust with the board and with risk and regulation teams. So rather than having to prove at the end your model is less harmful or you’re restricting it, you’re working with them and saying, what’s your experience of the regulation? How would you interpret it? And how can you instill it into agentic systems that we’re building? And therefore, it just leads to a faster time to value because you bring in that knowledge and that acceptance up front. So as I say, bring governance to the front, treat it like context engineering rather than a bottleneck at the end of the process. And it helps both with the time to value and the opportunity space that you can put these agents at. You mentioned building trust with the board and in your experience, like what are things that earn credibility with senior stakeholders? Yes, there are a number of key things as I’ve kind of built up trust with the board.

I think being honest about trade-offs, when you’re accountable to the board, you’re effectively asked for a book you’ve been given a portfolio to invest in. And you need to build up trust that you’re making the wise. choices with those so actually some of the best decisions that could have made have been where you’ve actually turned off or reduced a system or service that was working that was valuable because we know that we could take the capacity of the team and to build something better so an example would be we had a very valuable email auto reply service customers would email us we could triage it write very good course emails and turn those out on a Norway way very quickly. It’s very high maintenance because you can imagine the sheer variety and complexity of language that people use. But as we move into more of the use case of, well, actually, why don’t we use AI agents to prevent problems in the first place so they don’t have to write in and complain to us? That meant that we had to stop and pause or limit one project to be able to kind of self-fund another one. So I’ve built a lot of trust at the board that you’re prepared to make trade-offs and options rather than always asking for more. The other thing that’s been very valuable in building trust is actually having the independent verification of the value.

We work quite hard when we set out a three-year business plan for AI and some of the partnerships we’ve used with companies like Google. is to bring finance into the process at the start and say, if you look to this from a business case, how would you value it? How do you think it’s credible?

If we improve cash flow by making sure bills were of the right quality, how would you genuinely value that on there? And I think it just gives that credibility. So it’s almost like give up the proprietary of owning all the numbers and owning all of the data and be prepared to have somebody. build a system that can verify tools that you’ve built. And then it’s an additional voice in the boardroom to say, yes, we’ve delivered this program of AI tools. And the CFO is happy to say, yes, we can clearly measure the value and we’ve shaken hands for that’s true.

I think that kind of builds the credibility. The final thing for me is probably more career advice for someone who’s grown up, you know, with a master’s degree, working in data science and then leading teams. And now kind of being accountable at board level for delivering an AI program is just being mindful of the choices that you kind of, how you present yourself, present options. So I mentioned at the start is how maturing teams is people, systems of value. You find really good quality teams you believe in.

You build scalable operating practices. That means those teams can learn and scale and do AI responsibly. And then you allocate. those teams to the highest value project as it grows. I found when talking to build trust with the boards, it’s talking about that in reverse, the value systems and then the people, because that’s cutting to the chase. So we can deliver 50 million pounds worth of improvement over the next two years. Here are the systems and trade-offs, an operating model we need in place. That means we need people working in operations to be working side by side in blended teams with technology practitioners to build trust and contextual knowledge into them. And here’s the right skills mix that we need to do it. And what I see from a number of technical practitioners, they step up in leadership roles. They still retain that excitement about talking about the tools, how they built it, the model selections that are found to build trust with boards. You just reverse the order. Talk about the value first. how you scale that kind of value in the trade-offs. And then that opens the door for you to talk about the excitement of the technology and where it may go next. So Andrew, there’s so many potential use cases for data science and AI. How do you decide where to put your team’s time and what use cases you go to the board to get investment for? What’s your process for that? It’s a really great way to get some energy around the business is to treat anyone of the business as having the opportunity for an idea. So quite often we’ll go to where we’re doing our planning, looking, say, over the next 12 months. We’ll often talk to many of the business leaders about their, what are your pain points?

What are your opportunities that you kind of see? How can we help you and your teams learn about AI? And I’m often providing them with more kind of low-code tooling, like Gemini Enterprise, clawed on their desktop, just so they start to understand the potential of these tools as well. So I think going out and talking about people in an exciting way, give them first-hand experience of these tools, means that they’re interested in AI and they can learn and embody that for themselves as well. In terms of then kind of picking use cases, it’s almost like a… approaches. One, you need to understand constraints from the board. Is it value going after, customer experience? What are you optimizing for? And what are the constraints? Is it over three years? Is it two years? Is it over the next six months? You have a critical capital issue. You want AI to address it. And then finally, when you’re actually soliciting those kind of ideas from teams, it’s always good useful to have a number of… Just really helpful techniques to solicit and privatize those ideas. So one of those can be things like a lean value tree method. So sometimes you speak to the head of billing and payments and they’ll just be a really big problem. We have too many people who we could have spotted going to debt and we’re not doing enough for them at an earlier stage. It’s a really big problem. It’s not like one. You then go through and break it down to different pain points. different bets and then you start to then overlay where you think different tools where the interventions could be sometimes you get the other where you get very people who think only think about processes in today’s terms like okay we a complaint comes in we log it you know how can i do that faster how can it do more of the login for me and record a complaint faster so you save a little bit of productivity and actually helping the customer And then that’s where like more strategic framing about playing to win, saying, okay, if your role in this might feel small in a larger company, processing complaint, but actually you are the most informed person on the volume variety and what we could have done better. So how can you help us to say what could be a game changer? How could we have prevented it? So it’s a lot of engagement around the business. And I think a number of techniques I mentioned about either taking big ideas and breaking them down to things that you can picture as solutions or agentic products it’s all the reverse of taking small pain points and then helping people imagine what they could do further and further upstream we then run like a quarterly planning process where you kind of review all of the portfolio options that you have i go through all of those candidate projects and we We prioritize them against value, feasibility, and anything that needs regulation compliance. And it just gives you a good portfolio. You can demonstrate to the board that we’ve got enough mix and variety of projects to go after. There’s enough value on there for it to be interesting. But there’s also a mix of feasibilities in

there from well-known interventions that could predict when a customer’s likely to churn from a pregnancy model. to something that could be more of a bigger bet on voice-to-voice AI where you may not know the customer experience. So if we look at the model economics, where do you see that going? And what’s your strategy for dealing with potential, I guess, model cost, token cost increases?

So I believe that FinOps will become like a optimization problem and a science in its own right. in the rush to kind of scale to AI and to see the value from it. Although many other companies have kind of adopted a range of models, put them to many use cases. And historically, the approach has been to kind of find a vendor that you trust, partner with one vendor, whether that’s Anthropic or with Google, OpenAI, build a deep partnership and use their infrastructure and you’ll be in a strong position. I think what we’re seeing now and over the next few years is actually optionality is valuable for two reasons. One is that there are many competing companies out there producing models with different capabilities. So as we’ve seen the likes of Anthropic become very proficient at helping software engineers develop software that makes sense to have with Gemini models. that are very good at doing image recognition. You know, it’s kind of helped with kind of meter reading as well. So it’s becoming more of a portfolio science in its own right. And the second one is building the flexibility for cost optimization. So in particular, if you’ve got, there’s always a tendency to want to adopt the latest and brightest model, but these models have been developed, become more accurate because they’re doing more reasoning and therefore they’re… you know, cheaper to run, but they’re burning more tokens per useful output because they’re doing more reasoning. And therefore the overall cost of tasks, you know, putting it to that high complexity model is becoming higher. And what we found from our own experiments that we kind of done with the team is that if you take some of the models that could be more self-hosted, if they’re high volume tasks and the decisioning is kind of less complicated. then you can pay like a fifth of the cost and just get only within a few percentage points of the similar accuracy. So I think businesses looking forward to invest in having the optionality to kind of choose providers and get best of breed, but also the humility of AI teams to then think what’s the right size model for the job. And that’s where you bring cost effectiveness into the equation is important too. So it’s like the earlier conversation we had about building evaluation up front, evaluating the product, the leading and lagging metrics, but what was the cost per outcome, cost per successful outcome that was generated. And if you can get an equivalent with a lower latency or a cheaper cost from a different provider, that will probably be the competitive advantage to Forge.

That’s where I feel the strategy is moving in terms of the unity economics of these models. You mentioned Cloud Code. What are you guys planning in terms of the future with technology like Cloud Code? I see huge potential in Cloud Code. I see this as a disruptor of the ability for software engineers to produce software that generates more systems. And it’s kind of very different to then build versus buy decisions. but it also changes the software development lifecycle. So we’ve got to a point where we’ve rolled out Claude Code to a number of our engineers. We’ve got 95% adoption. We’ve got a community of practice. But actually what we’re seeing in investing more and more is actually harnesses over Claude Code where you can give people a safe environment to develop and actually reusable tools that mean they can develop faster. So an example would be if you’re building a new organic AI system, I think rather than going through end-to-end the typical phases over several weeks of the context and the knowledge, designing the agents, building the evaluation, if you can actually build some of that recipe into a cloud code harness, you effectively create a starter kit. And so we recently deployed an agent that would have taken eight weeks previously to down to two weeks because a lot of the learnings about evaluation and how to format the context was already in place. So these tools can really improve the time to value quite significantly. The other side, though, is then you then have to think about how that changes how you run as a business and how you work with those teams. So typically, if it was like an eight weeks to build and design and prepare something, and then you can plan in and schedule, when do I need to speak to the user experience designers? When do I need to speak to the operations team eight weeks down the line to get them involved? If then that happens in a matter of a week, your bottleneck then becomes people availability and then you’re moving too fast and you’ve got lots of downtime for your engineers. So actually that then makes you think of why don’t you bring these skills and people together earlier on in the process because you know the speed at which lines of code can be written can be ever faster. So there’s huge potential with Claude Code to… Bring down time to value, but it does just change where the bottleneck is.

You then have to redesign your end-to-end design team around it. So, and when you talk about harness, you’re talking about some type of deterministic system or engine that’s key to the workflow that perhaps moves Claude into more of an orchestrator or language interface.

Is that what you mean by harness? Yes. If you can… codify the kind of standards and quality that you would ship a project and also the key principles about how you may measure accuracy. Your design and architecture choices as a company, you know, to be modular, API first, a choice of languages used. A lot of that can then be guided through experienced software engineers getting much faster to the result. So yes, it’s almost like a… a deterministic set of guardrails that means that if you let it run autonomously to help you generate code, you can be confident that it’s keeping within the limits. You’re just being very clear what it can and can’t do. Whereas the temptation, what I’ve seen in the early days, first

time people get called code is, build something for me, and they realize there’s just so many unintended effects. It delivers 40,000 lines of code, uses various language and mixes of them.

It delivers huge numbers of changes in the code base and you spend three times as much time reviewing it and picking what was done. So it just brings the software engineers and the AI engineers role much further up front is to say, actually, what would the guidance I give to a

junior intern? If a new junior software engineer joined the team, what guidance would I give them about how we operate, the importance of unit tests, choice of languages and interfaces that we use? and codify it you then have something which is scalable and then thinking about the agents that you currently agentic systems you currently have running there will be certain parts of your regulatory requirements where you have to have some type of explainability or auditability around a decision that’s made is that an area where you’ve moved the agentic stuff into yet? Or have you steered clear of that? Because it’s one thing having a reliable agentic system, it’s another having one that you can then explain that reliable decision. Am I making sense? A hundred percent. I mean, this is something that’s been the benefit of my experience of working in financial services is quite often when the financial services regulator gets involved. They will want to know what decision was made for the customer, what information you had at the time, and what were the factors that influenced that decision. So it’s never that anything goes wrong. Things do, and that’s real life sometimes. What the regulator wants to ensure is that you’ve been fair, considered all the information, and acted under the right level of principles kind of ethically. So having that grounding in financial services where the regulators… so hot and so precise. As many, we’ve been able to design some of that up front with observability. And I think this is one of the best applications of where AI can help itself. using AI tools to cope with a range of observability. If you’re kind of logging every interaction, if you’ve got a really clear scope of each agent. So you can imagine each AI agent we build has a little card attached to it that says, what’s the role of this agent? What services is it allowed to call? Which ones it’s not? What’s the intended purpose? And then you’ve got logging and telemetry that can determine when it was called upon and when it wasn’t. You then create a body of evidence to be able to demonstrate within the system. And this is going to be like so important because you see the risks if you’ve got these tools that a customer facing and it makes a promise to the customer or said something, you know, needed to be checked and verified. What information do you provide it? And so observability is absolutely key as a criteria to go live. We shouldn’t be sending anything out to the customer that we can’t verify. And actually AI can play a role in that. That’s where I feel, yeah, more advanced tools that can look at the variety of outcomes that can interpret the logs and the decisions and actually understand is the decision-making flow that we intended? Did it check the vulnerability flag? Did it check whether that customer had an existing complaint and verify that’s the truth before you graduate it from, you know,

human in the loop to human on loop is really important. Yeah, both in the observability is key and then having a very clear graduation process for your agents from junior intern to more proficient team leader without humanizing them too much.

I think it’s a good way to think of it. Would you trust a junior intern to do this and explain it to them before you let them take broader decisions that you’re accountable for to the regulators?

So you have a large team, basically, you have lots of systems running and so on. They are in production already. So the company, the users, the clients, they rely on them. How do we deal with the system at such scale with the shift in the landscape? So that means shifting data, of course, but also shifting LLMs, changes on the provider side, potential regulations that change and so on. So how do you keep everything running?

Challenger I see, like many companies I speak to and I own as well, is there can be, how do you break through the plateau? There’s a point sometimes where you have a team, you kind of allocate resources, you build a large number of agents, but then there becomes a plateau because you then have to maintain those systems, data sources change, regulation changes, and in particular if you’re renting third-party models, which we found actually one of the… biggest costs we had it was almost taking like one two engineers per year worth of effort was put on model upgrades alone they go from version 3 to version 3.5 the amount of time we added up of those teams to have to then replace the kind of models so best piece of advice i can give on this front is just because you can build it also means that you run and operate it so if you use clawed code to build the software You know, we don’t need to buy in a system. That’s great. You save the procurement costs, but then tag you’re it. You then build, run and operate it. And therefore you have to think about what the system is for building and operating it. So more and more as a team, we’ve devoted more of our research into how can we decompose the model upgrade process? You know, do we keep our test data away from the inputs and the outputs? We’ve invested in golden data sources. That means that once you’ve moved from one model to another, you can then have a repeatable set of tests and find very quickly where the difference is, replay it. It’s also then investing in greater automation around logging and incident management. There’s always things where you’re testing new systems and new code that things can break, things can change, or you’re dealing with a third-party system or CRM. And it’s actually some of the most unglamorous things have the highest value because of the volume or impacts they create. So taking the time to invest in co-defining your incident playbook and building that into a skill as Claude Code has been hugely valuable because it just reduces the effect of the incident happening again, but every time an incident happens or it’s a different member of the team, they then have a system that gives them a really fast start. So for me, that’s really important to break through the plateau, is that no matter how many of the team resources you give, work will fill the space, complexity will fill the space. How can you use AI to drive its own efficiencies? And my advice would be automate the non-glamorous stuff and be as proud of that as you would some of the intelligent model and custom-facing impacts we’ve spoken about. We haven’t touched on the open source side of things. Just wondering, Andrew, very quickly, how much that is part of your model portfolio? We’ve deliberately partnered publicly with Google on a lot of the model development, firstly because they’re depth of expertise, but also trust in the data itself. We know that our customers want to make sure our data, their data, is not shared and reused or exposed anywhere else. So we’re also always very… wary of model choice and selection is not just the performance of it but where does the data residency live and generally we are looking more and more at self -hosting models because the economics is trending towards that and we’ve run a number of tests where we can actually demonstrate so you don’t have to use frontier models to get comparable performance you can get lower costs but again it just moves the tail into then running and operating is you may be able to kind of build and utilize different models from different providers and run at a lower cost. But it introduces complexity when you need the auditability for the regulator, or if there is an incident and something hasn’t happened as you’d like, complexity goes up as to which system it was and why. So it’s just thinking about the scenario planning ahead when you’re evaluating these tools. Economics is one, it can bring the cost down. total cost of ownership can go down but then if you’re looking at the complexity it may bring in when you need to change systems diagnose results surface observability to the regulator is the extra complexity worth it as well so control diversity is the best bet you know enough diversity that you can optimize your economics and be as optimal as possible but not so much diversity that you end up with an creating a series of optimization complexity problems that cause you to use even more people in it. And that’s something where you just need the judgment and expertise of the people building these systems to help you with. There’s not one single answer, one single number that’s right for every company. I think these are where the best judgment calls are made in the strategy and planning. Andrew, what are you most excited about in terms of the future development of AI? in your space i’m really excited about when you look at the achievements anthropic have made in building software where you can continuously add more skills without breaking the system we’re probably in an era where you kind of genuinely building self-learning systems and i think that’s what really excites me about the design so rather than taking like a single problem Model, solution, new product. Another problem, product, model, solution, deployment. It could be as simple as just adding new levels of skills to an existing system. It would mean you can develop faster, but also just personalize much more deeply for the customer. So if I was to kind of look in the energy space, many customers want to do the right thing. They want to be able to afford their bills. They want to be reassured they’re on the right price. They want their energy company to make sure that there’s a good level of service, but they also want to make sure that they feel good that their energy supply is doing the right thing by them in the environment and their usage of energy, and they’re not wasting it to lower their carbon footprint. And so having the point where AI can help individual customers make better decisions or remove the friction points before them, because it can self-learn to say, you as an individual, have to use energy at these times because of young children at home but you’ve also got an electric vehicle and the opportunity to kind of have a battery storage as well you could take that complexity and put in you know the computing power of a mobile phone and just show the customer hey we’ve saved you some money today and there’s opportunities and choices for you to do that and do with ever more green green energy and i think that’s what’s always been missing particularly like in the energy sector is it’s a contribution of many small actions that have to ladder up to big global problems, security of energy or decarbonization. But we could really use AI to drive better solutions, more personal solutions for customers. So that’s very much at the front of my mind, just bringing in that deeper personalization and ever abstracting away the complexity from the customer that they generally feel weird. proactively serving them and proactively giving them better choices in the palm of their hand in something as commoditized as the electricity that comes through to your home. I think that would be a wonderful way to celebrate technology for good. Good. And a very fitting way to conclude. So, Andrew, thank you so much for joining us today. It was an absolute pleasure talking to you. Oh, you’re very welcome. It’s been such an exciting array of topics and how we could have fit this all into one cast. Good. And thank you also to my co -host, Philip Diesinger. And thank you, of course, to you for listening.

Do check out our other episodes at datascienceconversations.com.

And we’ll see you soon on the next show.