# Why Full Autonomy Beats Co-Pilots for AI in Banking with Dimitri Masin, CEO of Gradient Labs

- **Show:** Fintech One-On-One
- **Published:** 2 Jul 2026
- **Source:** https://www.buzzsprout.com/1966250/episodes/19436862-why-full-autonomy-beats-co-pilots-for-ai-in-banking-with-dimitri-masin-ceo-of-gradient-labs.mp3
- **Transcript:** https://podcasts.opennash.com/shows/fintech-one-on-one/why-full-autonomy-beats-co-pilots-for-ai-in-banking-with-dimitri-masin-ceo-of-gr-fa0bca/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Dimitri Masin argues that the biggest remaining opportunity in banking is the manual customer operations behind the app, such as support, fraud checks and onboarding. He says fully autonomous AI agents beat co-pilots, which are AI assistants that suggest answers to human staff. His company measures them against the human team and aims to be clearly better, not perfect.

The central tension is between what the technology can already do and how slowly banks adopt it. Masin says the key AI capability has existed since early 2023, yet even easy customer support work takes about a year to move from 50 to 80 percent automation. He puts large-scale change at five to ten years, held back by how organizations decide and adapt, not by technology.

## Key insights

1. **Co-pilots can make quality and compliance worse** Masin says that when an assistant is right about 90 percent of the time, staff start accepting its suggestions blindly, so the 10 percent of errors slip through. He says his team saw this firsthand, and it led them to choose full autonomy.
2. **Judge AI against humans, not perfection** Masin says no system is correct 100 percent of the time, including humans. His company uses the current human team as the benchmark and aims to beat it, which he says gives risk officers and regulators something concrete to approve.
3. **Easy tickets save less time than expected** Masin says automating about 50 percent of support queries yields only about 20 percent of time savings, because those are the quickest tickets. The savings come from automating the harder remaining half, which he says takes roughly a year to move from 50 to 80 percent.
4. **Adoption speed, not technology, sets the timeline** He says the core model capability was available in early 2023, and newer frontier models add little for operations work. He puts large-scale transformation at five to ten years because of how banks decide and diffuse new technology.
5. **Banks learn that building in-house is hard** Masin says large banks first said they would build AI agents themselves, then found after 12 to 18 months that keeping pace with specialist vendors is very hard. Some are now building infrastructure to plug in outside providers and swap them out.

## Questions this episode answers

### Why does Gradient Labs prefer fully autonomous AI agents over co-pilots?

Dimitri Masin says co-pilots, which suggest answers for humans to approve, degrade quality because people start accepting suggestions blindly. His company says it has spent three years showing that full autonomy with the right controls is safer and higher quality.

### How can a bank trust an AI agent if it is not always right?

Masin says nothing is correct 100 percent of the time, humans included. His company compares the AI against the current human team and aims to beat that benchmark, which he says should satisfy risk officers and regulators.

### Do banks lay off staff when they automate customer operations with AI?

Masin says none of his customers have planned or made layoffs in this area. He says they move people to more complex and empathetic work, such as financial difficulty and vulnerable customers.

### How much money or time can AI save in banking operations?

Masin says the main motivation is usually better customer experience, with about a 15 percent uplift over the human team. For handling time, he says a thorough job can save up to 70 percent and a minimal one about 20 percent.

### How long will it take for AI to transform banking?

Masin expects a five to ten year horizon for large-scale change. He says the delay comes from how slowly organizations adopt technology, since even the easiest use case, customer support, takes about a year to go from 50 to 80 percent automation.

## Facts

- **Dimitri Masin** [0:02:44]: "I've joined them quite early on, probably as one of the first 30 people, and was kind of lucky to bet on the right horse, to be honest."
- **Dimitri Masin** [0:02:44]: "And I spent there seven or eight years literally building the bank from 30 people up to 4 ,000 people, I think by the time I left."
- **Dimitri Masin** [0:05:17]: "at the beginning of 2023, when GPT-4 came around, it became very, very clear to us that that part of the problem"
- **Dimitri Masin** [0:07:40]: "there will be, let's say, 15 % of cases which cannot be straight through process because they need to be to look deeper at."
- **Dimitri Masin** [0:09:19]: "we have about 30 customers within FinTech and financial services space that we're working with."
- **Dimitri Masin** [0:10:22]: "you can save maybe up to 70 % of handling time, operational handling time on a particular process."
- **Dimitri Masin** [0:12:40]: "So a bank would try one or two vendors and compare the performance that they can achieve. That POC phase is typically like four weeks"
- **Dimitri Masin** [0:14:08]: "And we actually tried to do it at Monzo years back, even before GPT-4 came around."
- **Dimitri Masin** [0:14:59]: "what tends to happen is that people start just accepting blindly what it suggests."
- **Dimitri Masin** [0:15:55]: "But now we essentially, for the last three years, we have spent the last three years proving that full autonomy is safer and higher quality than co-pilots."
- **Dimitri Masin** [0:18:11]: "none of the companies we have worked with actually planned or laid off any people in that space, which is maybe like counterintuitive at first."
- **Dimitri Masin** [0:19:30]: "it was already more than a year ago in January, we realized we have actually had the first full-scale AI agent deployment was one of the largest regulated banks in the UK."
- **Dimitri Masin** [0:20:30]: "I think it was like 15 % difference or better than the human team, the human average team in that particular bank previously."
- **Dimitri Masin** [0:22:44]: "the typical conversation that I had with fintechs or banks probably one and a half years ago, one year ago, was like, especially tier one, tier two banks, we're definitely going to build it ourselves."
- **Dimitri Masin** [0:25:15]: "But the first, those 50 % of the queries that you're automating actually yield roughly 20 % of the time savings that you achieve through that."
- **Dimitri Masin** [0:26:10]: "And it still takes a year to go from 50 % to 80 % where you really kind of see step changes."
- **Dimitri Masin** [0:27:31]: "In the US, I found that companies are a lot more performance-oriented. So essentially they do those bake-offs, essentially."
- **Dimitri Masin** [0:28:27]: "We actually have won every single bake-off that we run against any other company."
- **Peter Renton** [0:20:59]: "I was at New York FinTech Week in April, and I was listening to a panel with a bunch of the largest banks in the US."

## Transcript

**[0:00:10] Dimitri Masin:** What is the most important thing about AI? What is the most important thing about AI? Nothing is correct 100 % of the time. Right? So I think that's the most important bit here to kind of that companies that want to adopt AI kind of need to accept. But that's fine, right? Nothing is kind of ever 100 % of the time correct. And those are non-deterministic systems and humans make a lot of mistakes. So really the real question is, can you create a significant uplift on the quality that the cumin can produce?

**[0:00:35] Dimitri Masin:** Because then you can make an argument, look, I'm automating this process, but it's much faster. The customer gets a loan or mortgage decision instantly, or like close to real time. It's also a lot more compliant because you can objectively show that your process complies a lot more often with a set out standard and policy. And really the way treating it, we typically start on the project and we say, okay, the human team is a benchmark. What we want to do, we want to do it significantly better than the current benchmark.

**[0:01:04] Peter Renton:** This is the FinTech One-on-One Podcast, the show for FinTech enthusiasts looking to better understand the leaders shaping FinTech and banking today. My name is Peter Renton, and since 2013, I've been conducting in-depth interviews with FinTech founders and banking executives. My guest today is Dimitri Masin, co-founder and CEO of Gradient Labs, a company building autonomous AI agents for customer operations in financial services. Dimitri was one of the first 30 employees at Monzo, where he led AI and data science from day one and watched the bank grow from 30 to 4 ,000 people.

**[0:01:44] Peter Renton:** In our conversation, we talk about why the real opportunity in banking is automating the manual operational work behind the scenes, not just the app experience. We get into why co-pilots can actually degrade quality and compliance and why Dimitri believes full autonomy is the safer path. He shares a story behind what may be the largest known AI agent deployment in banking, serving millions of customers. We also cover why banks aren't seeing productivity gains. His five to 10 year view on the transformation and how the U.S.

**[0:02:19] Peter Renton:** Bake Off culture plays to his advantage. Now let's get on with the show. Welcome to the podcast, Dimitri.

**[0:02:30] Dimitri Masin:** Thank you, Peter. Nice to be here.

**[0:02:32] Peter Renton:** Great to have you. So let's get started by giving listeners a little bit of background about yourself. I know that you've worked at Google and you were early at Monzo. Tell us a little bit about what you've done in your career to date.

**[0:02:44] Dimitri Masin:** Yeah, amazing. Let's start with Monzo. that's probably the most exciting part. I've joined them quite early on, probably as one of the first 30 people, and was kind of lucky to bet on the right horse, to be honest. I mean, nobody can predict which companies will succeed in the end, but Monzo was a very, very successful bet. And I joined them really to lead all things AI and data science from day one. And I spent there seven or eight years literally building the bank from 30 people up to 4 ,000 people, I think by the time I left.

**[0:03:13] Dimitri Masin:** And so really kind of had an amazing experience throughout those seven years, kind of looking into every single area of the bank very deeply and kind of understanding how banks work from the ground up and kind of building one yourself. It was an amazing experience.

**[0:03:27] Peter Renton:** That's great. Yeah. And Monzo, obviously one of the most successful UK digital banks. They've tried to make it here in the US and haven't really been successful. But so then what was it that you saw, like you were still at Monzo, I believe, when you had the idea for Gradient Labs? What was the thing that needed addressing?

**[0:03:45] Dimitri Masin:** That's a great question. I think retrospectively, if I look at this, there's obviously this big transformation that fintechs were trying to achieve in financial services, right? So they kind of introduced the kind of at your fingertips, amazing, flawless experiences where you can really bank just by a few taps on your mobile phone, whereas previously you would need to go into a branch or like call somebody up, which is even worse probably. And really fintechs, like the whole purpose behind fintechs like Monzo were to reinvent that experience to make it really amazing.

**[0:04:16] Dimitri Masin:** And I think what most of those companies though, like Monzo and many others have realized is that while you could reinvent the happy past and the experience and kind of create very slick apps and a very good customer support, for example, really a big part of their financial services organization could not be solved very easily. And that's kind of the manual repetitive operations, let's call it customer operations behind the scenes that kind of is needed to run any financial services organization.

**[0:04:46] Dimitri Masin:** That part kind of remained the same, right? So in every larger retail bank, you would have probably more than half of the stuff really responsible for keeping the lights on of accounts behind the scenes. And even those fintechs were tech first, that part, the customer operations, and by the way, when I say customer operations, I mean things like customer support, AML, fraud investigations, KYC, onboarding, screening. So really all the processes that are needed to make a bank work, they were pretty much not reinvented.

**[0:05:17] Dimitri Masin:** So they were like fairly similar to how they were before. And at the beginning of 2023, when GPT-4 came around, it became very, very clear to us that that part of the problem, i.e. the second half of the transformation of financial services to like a very slick modern experiences is now possible as well. So the first leg was putting your experience into the app, make everything very nice and easy. And the second part of the transformation is almost like now creating a significantly more scalable, more efficient organization by doing all of the customer operational work in a much more efficient in a much more delightful way for the customers.

**[0:05:59] Peter Renton:** So you started the company, I believe, with two other co-founders, all from Monzo. What did you kind of bring from your time at Monzo that really helped you get this off the ground?

**[0:06:11] Dimitri Masin:** Both of my co-founders, Neland and I, who are both amazing, they're part of my rider team as well at Monzo. So I knew them amazingly well. We've worked together for many, many years. And so we really kind of decided together to build a new company. And really what triggered it is that we were very, very deep in automating customer operations processes, right? And like probably in every other bank, we kind of had our successes doing like achieving 10 % efficiency gains here and there.

**[0:06:38] Dimitri Masin:** But then we acquired essentially very deep expertise of how everything works under the hood. And then, as I mentioned, when GPT-4 came around at the beginning of 2023, industry, we essentially realized that this whole area will be completely reinvented over the next five to 10 years. And so we want to drive that change in the industry and enhance kind of started gradient labs, which is focusing on kind of, if you wish, in a nutshell, putting banks on autopilot, i.e.

**[0:07:07] Dimitri Masin:** creating almost like a custom operations platform where you can write down your processes and procedures, and they will be executed by AI agents in a completely autonomous way. But as safe as humans or ideally much safer than human teams can, but also with much better customer experience at the end.

**[0:07:24] Peter Renton:** Okay. So then take us through that. If we could go a little bit deeper there, like when you're in conversations with banks, are you selling a suite of tools that are sort of AI agents can go out and do those things? Or is it, what are you actually providing to the market?

**[0:07:40] Dimitri Masin:** I think the closest analogy probably would be if you think about the human operator's role within the custom operation system, right? So like if you want to, let's take a KYC process, there will typically be some KYC system already in place that does straight through processes for the majority of the accounts. But then there will be, let's say, 15 % of cases which cannot be straight through process because they need to be to look deeper at.

**[0:08:06] Dimitri Masin:** Maybe there's like a PEP or sanction suspicious and you really need to investigate it in account. And typically you would have a human team that kind of looks at this exception case and does all the manual works, goes to different websites, maybe applies some decisioning metrics on the information that they're gathering. And the way to think about us is essentially we're trying to mimic how humans work within the operations teams today. So we're assuming that all the existing systems, source of truth or system of records rather, are staying as they are today.

**[0:08:39] Dimitri Masin:** But then the way if you think about how human operator works, They go from system to system, collect data here and here to complete an end-to-end workflow with a particular purpose like KRC exception handling. And that's exactly how we try to mimic it. So we essentially sit on top of the existing system. And you can think about us like an orchestration layer of end-to-end workflows across those different systems.

**[0:09:02] Peter Renton:** I know Monzo is one of your clients. I read about that. what are they actually doing and do you have any kind of metrics or something that you can share about how much they're improving their efficiency? I can't talk about Monzo, but I'm happy to talk

**[0:09:19] Dimitri Masin:** about many of the other customers. So the primary motivation of working with us by any of those customers, you would think maybe it's cost-cutting, but it's actually not. I think the primary motivation is typically to improve customer experience. And what we were able to show in every single deployment so far, and that's literally, we have about 30 customers within FinTech and financial services space that we're working with. In every single deployment, you can achieve a significantly better customer experience in support, for example, in customer service and customer support than the human team can provide.

**[0:09:53] Dimitri Masin:** So typically you see an uplift of around 15 % on top of whatever the human team can achieve in that organization through automation of AI agents, past three agents. So really, for most customers that we're working with, they primarily care about improving customer experience, but also making them more scalable because most of them are like fast growing FinTechs. They want to be able to say like, look, we want to keep headcount steady, but we want to be able to support twice as many customers as through that headcount.

**[0:10:22] Dimitri Masin:** And so really kind of being able to scale easier with the same headcount and improved customer experience is what really drives the motivation. On the cost side, frankly, we can go deeper into that. The picture is very nuanced. So I try to kind of advocate usually prospects and what they can expect on the cost side. But I would say there, if you do the job really, really well, you can save maybe up to 70 % of handling time, operational handling time on a particular process.

**[0:10:52] Dimitri Masin:** And if you just scratch the service and do like the bare minimum, maybe you save like 20%. So it's not quite a step

**[0:10:57] Peter Renton:** change every time. So then when you're going into a bank or a fintech for the first time, and they said, yes, we want to implement, what's that look like? What's sort of the implementation? Tell us a little bit about how long that takes and what's it look like?

**[0:11:12] Dimitri Masin:** So by now we have quite a wide suite of products, right? Because our vision is really to put operations and autopilot. So an operation is like essentially like a long tail of all the kinds of processes within a bank. But every company has usually at any given point, some sort of focus area or burning point paints, the pain points that they're focusing on. And so really typically the process starts as we kind of start a conversation and try to uncover what are kind of two, three most important bits that really are expensive or somehow, I don't know, not working well from a customer experience perspective.

**[0:11:46] Dimitri Masin:** So an example can be like, we're talking to quite a few mortgage lenders right now. And one of the key problems is like, it just takes so much time to review a mortgage application, right? Because you need to submit like 20 different documents and whatnot. And really like for them, it's about like, how do we shorten that time and maybe do it almost like close to real time. Somebody uploads the documents, gives us all the information, we can tell them, okay, X, Y, Z is missing, or if nothing is missing, here's an offer type of thing.

**[0:12:13] Dimitri Masin:** And so really as a first step, we try to discover what is the most burning thing, because our range of services by now so wide. And then once we kind of narrow down on two, three, like very specific high value use cases, we essentially say, okay, let's go and try to run a POC. And typically there's also like multiple vendors involved or like often enough, there are multiple vendors involved where you kind of almost like bake off for a particular use case, right?

**[0:12:40] Dimitri Masin:** So a bank would try one or two vendors and compare the performance that they can achieve. That POC phase is typically like four weeks but obviously before the POC stage you kind of all want to have all the legal work and security digital work done and so on so that's kind of probably the longest time but really the implementation is very very short the actual like implementing the product to show that it works is like four weeks initially at least for the kind of initial rollout and then it depends how deep you want to go because typically our motion is we implement the first use case, let's say within the four weeks.

**[0:13:16] Dimitri Masin:** And then we essentially say, okay, let's implement the second and the third and the fourth, because everything is built on the same platform.

**[0:13:23] Peter Renton:** I'm sure you've seen this in the press this year, where there's a lot of talk about banks, they're implementing AI and they're not seeing the productivity gains. What do you think is the reason that we're seeing so much talk about AI and the lack of productivity gains?

**[0:13:39] Dimitri Masin:** Two, three years ago, the answer was clear to that as well. But nobody was asking that, Maybe because people were not spending so much money on AI. The answer is still the same, and we kind of knew it already back then. So essentially, the cost efficiencies that you can create, i.e. productivity, is kind of very much related. It has a fairly nuanced view. So A, co-pilots. Most of the banks implement co-pilots. Co-pilots are essentially meant to make current workers faster.

**[0:14:08] Dimitri Masin:** And we actually tried to do it at Monzo years back, even before GPT-4 came around. where essentially we're all just adjusting responses to people and the frontline to kind of reply back to the customers. An interesting learning back then was already that if you do that, two things can happen. Either kind of people use those replies and not necessarily open a new chat and kind of work faster, but they maybe just like do something else.

**[0:14:34] Dimitri Masin:** So unless you adjust their targets as well and everything else around them, because it's a complex human system, you might not actually even able to see that they're doing more responses or chats per hour, let's say, or per a time frame. So that was one thing. So you kind of need to adjust the whole system. And that's not trivial. It's not just giving them a tool. But a second thing, which is maybe even more interesting, is that it has an interest.

**[0:14:59] Dimitri Masin:** Copilots have an interesting effect. If they're like 90 % of the time correct, what tends to happen is that people start just accepting blindly what it suggests. And so it might be fine. in 90 % of the cases, but in 10 % of the cases, it will be wrong. And so what we started seeing and what we experienced with our customers is that if you do that, you actually degrade the quality of the service and the compliance that you will achieve, which is counterintuitive, right?

**[0:15:25] Dimitri Masin:** So you wouldn't guess it, but we have seen it firsthand. And so that's why the core decision for us when we were building Gradient Labs was like, okay, we actually have seen firsthand that co-pilot degrade the quality of the experience and degrade the compliance rate because people just at some point start accepting. What the real solution is in our minds is like complete autonomous kind of autopilot end-to-end automation, which doesn't require a human in the loop.

**[0:15:55] Dimitri Masin:** But if you want to do that, you obviously need to convince many, many different parties that it's safe to do. You need to have the right controls in place. You need to have a way to kind of roll it out very safely. So you need to think a lot about the controls that you need to have in place. in order to have a fully autonomous system. But now we essentially, for the last three years, we have spent the last three years proving that full autonomy is safer and higher quality than co-pilots.

**[0:16:22] Peter Renton:** So then how are you ensuring that your AI agents are correct 100 % of the time?

**[0:16:28] Dimitri Masin:** The answer here is nothing is correct 100 % of the time, right? So I think that's the most important bit here to kind of, that companies that want to adopt AI kind of need to accept. But that's fine, right? nothing is kind of ever 100 % of the time correct. And those are non-deterministic systems. And humans make a lot of mistakes. So really the real question is, can you create a significant uplift on the quality that a human can produce?

**[0:16:55] Dimitri Masin:** Because then you can make an argument, look, I'm automating this process, but it's much faster. The customer gets a loan or more to get a decision instantly or like close to real time. It's also a lot more compliant because you can objectively show that your process complies a lot more often with a set-out standard and policy. And really, the way you're treating it, we typically start on the project and we say, okay, the human team is a benchmark.

**[0:17:19] Dimitri Masin:** What we want to do, we want to do it significantly better than the current benchmark. And this way, hopefully, your chief risk officer and your governance team will be happy, the regulator will be more happy, and the end customers will typically be happy because something will be also in it for them because it's going to be much faster, much clearer, and so on. So really, the answer here is to benchmark it to the current human team, because that's kind of the work that you're replacing rather than expecting a completely 100 % true and accurate solution.

**[0:17:49] Peter Renton:** when you're then implementing this end-to-end autonomous solution what do the people that used to do that manually do and what have you found at the implementations that you've done how are they utilizing the employees that used to run that run these tasks i think the first

**[0:18:11] Dimitri Masin:** assumption would be oh like there must be like layers happening because of that actually the counterintuitive truth is that none of the companies we have worked with actually planned or laid off any people in that space, which is maybe like counterintuitive at first. But in most cases, companies redeploying those people to create better customer experiences for their customers, right? So sometimes it's services that they didn't provide before. So for example, you might have an account manager from a certain size of the account and that threshold kind of goes down.

**[0:18:43] Dimitri Masin:** or you spend more time actually investigating more complex issues or dealing with cases which require a lot more human empathy, like financial difficulties and vulnerability. So in reality, for every single customer of ours, typically the teams got relocated on slightly more complex, more difficult, more customer-centric tasks in every possible case.

**[0:19:06] Peter Renton:** So I want to dive into something I read about, which I was really interested in to get your take here. It was described in the article I read as the largest known AI agent deployment in banking, handling nearly half a million users with quality that beat the bank's own human agents. So what can you tell us about this largest known AI agent deployment?

**[0:19:30] Dimitri Masin:** Well, I think what's worth noting that, I can't obviously tell the name, otherwise we would have disclosed it in the article. But essentially at some point at the beginning of 2025, it was already more than a year ago in January, we realized we have actually had the first full-scale AI agent deployment was one of the largest regulated banks in the UK. And at some point we kind of just glassed over it. We didn't even clock it, but at some point it's like, hey, there's literally no other large regulated bank who would have deployed it at full scale to their customers.

**[0:20:03] Dimitri Masin:** And we started kind of writing and talking about it because it was the first, at least the ones that I would know about. Most large regulated banks with tech teams would have kind of played with AI agency and they were like half niche cases automated. But in this particular case, it was really deployed to like 12 or 15 million customers. And throughout the process, Obviously, we needed to jump through a lot of hoops to kind of prove that it's going to be safe and working well.

**[0:20:30] Dimitri Masin:** But in the end, we realized that the customer experience through doing that has significantly improved for this particular bank. So I think it was like 15 % difference or better than the human team, the human average team in that particular bank previously. So it was really kind of overall a full-on success story. And we started really beating the drum about it because we realized, oh my God, this is literally like the first bank. And everyone is trying to do that.

**[0:20:56] Dimitri Masin:** And we actually had one at full scale.

**[0:20:59] Peter Renton:** I was at New York FinTech Week in April, and I was listening to a panel with a bunch of the largest banks in the US. And it was interesting because they said that they're all doing pilots right now. And these are obviously very big, complicated institutions that do things very carefully. But what was interesting to me, and I'd love to get your take on this, is that they said, right now we're doing pilots. And then someone asked the question, okay, in two years' time, what percentage of the bank's processes will be done by AI agents?

**[0:21:38] Peter Renton:** And they said 50%. And right now it's zero. So that seemed to me I like to be a little bit optimistic. What's your take?

**[0:21:47] Dimitri Masin:** I'm pretty sure it's a bit too optimistic. I mean, we're at it now for three years and we kind of, it's worth saying that the core technology that needed to unlock that use case was there at the beginning of 2023. I mean, the GPT-4 was powerful enough to do that use case well, right? So it's like whenever the frontier models improve now, it doesn't necessarily drive too much of like gain or improvement in the products like ours because like the operation space is like, you don't need a PhD LLM model to do that, right?

**[0:22:18] Dimitri Masin:** And so really what's happening is that even though the product is there and some companies, especially fintechs at the beginning, have proved that it can be very safe or like safer than human, more compliant than human and create a much better experience. And yet you wouldn't say that every bank has somehow adopted it, even though those products exist. And I think it's more about the kind of technological diffusion and how decisions and things happen in those companies.

**[0:22:44] Dimitri Masin:** To give you actually a very, very specific example, that happens very frequently and more recently even started to happen more frequently. So the typical decision, the typical conversation that I had with fintechs or banks probably one and a half years ago, one year ago, was like, especially tier one, tier two banks, we're definitely going to build it ourselves. We have thousands of engineers, why would we ever partner with somebody else? We're just going to build it.

**[0:23:08] Dimitri Masin:** We can build anything. That was typically the answer. And then I think typically companies have spent maybe a year or 18 months on running pilots and have in many cases learned that while they can build something internally, the technology is evolving so fast that it's almost impossible to keep up with specialized providers and vendors who are essentially just doing that one thing. And so essentially, even companies who build themselves started realizing, okay, we're nowhere close to the frontier capability to the safety that you can achieve through some of the providers that are out there.

**[0:23:43] Dimitri Masin:** And so really, the situation started to flip, probably at the beginning of this year, where I've started seeing the first cases where like, okay, we've done it for 12 or 18 months. we are now realizing that we either invest hundreds of people to do it properly, to keep up with the technological change, which is really kind of happening every two months, or we are building infrastructure in such a way that enables us to plug in the best providers from agent providers from the outside in a safe and easy way into our infrastructure.

**[0:24:14] Dimitri Masin:** This way, we can kind of bat on the best provider. And if they're not the best anymore, we can switch them out for something else. But we really, as a bank, we're going to build the core infrastructure to be able to plug those providers in. So that's kind of the mindset shift that happened. My answer is, it wasn't like a straight pass or just build, deploy, and it works. It's like a bit of a zigzag where you try one thing, you realize after, oh, I didn't quite yield results, now I'm

**[0:24:40] Peter Renton:** trying something else. Right. So two years may be optimistic, but I'm guessing you think like five years is probably pessimistic, right?

**[0:24:47] Dimitri Masin:** Well, for me, like when I talk about it, I talk about the five to 10 years time horizon where they're like larger scale transformation will happen. Even on the support side, right? So like to give you an anecdote, support side, like customer support, customer servicing is the easiest use case that we started with three years ago, right? So that's kind of what you start with typically. Because like for most people, they've usually said, It's just like a chat GPT style application.

**[0:25:15] Dimitri Masin:** So what could go overall? That's kind of the easiest application. And there it has a very interesting dynamic, right? So if you look at all the customer support, even with kind of companies that they're working with, you can automate fairly quickly 50 % of all the inbound queries that customers are contacting you with, let's say on the calls or by chat. But the first, those 50 % of the queries that you're automating actually yield roughly 20 % of the time savings that you achieve through that.

**[0:25:44] Dimitri Masin:** So what it means is that just like those 50 % of the queries that you're automating are just the easiest ones anyways, that would have taken the human very, very quickly to resolve as well. And so really you automate half of the tickets, but only 20 % of the time. And so really the crux of the problem is essentially like, how do you automate the remaining 50 %? And we're doing it with a few very large customers today.

**[0:26:10] Dimitri Masin:** And we are on the journey from 50 to 80 or 50 to 90 already like for close to a year. And they are like fast running tech forward companies. And it still takes a year to go from 50 % to 80 % where you really kind of see step changes. So based on that, and that's the easiest possible case, right? So and now if you kind of take a step back and say, okay, customer support is the easiest possible case but like there are all kinds of like heavier compliance processes like i don't know money laundering uh reviews or like account freezes and things like that like they're a lot harder to automate and if you like actually go after

**[0:26:46] Dimitri Masin:** the the long tail of all those things i can't see it to take two years like it just uh even for the easiest stuff it takes already a year and you're like maybe automating 5% of all the operations in a bank so really five to ten years is for me the horizon. Not because of technology, it's not that technology is there already. It's about the kind of technological diffusion in how organizations move and adapt

**[0:27:10] Peter Renton:** is where the bottleneck is, I think. Okay, so you're based in London, but you have been here in the US, I believe for almost a year, and you already have some fairly big name fintech customers, Current and Stash, both seem to be publicly named customers of yours. So what's the plan here in the US?

**[0:27:31] Dimitri Masin:** So US is a pretty big market. It's a very interesting one because it has fairly large fintechs, right? And that's kind of one of the ICP segments. It is also interesting from a perspective that US fintechs buying slightly differently. So there's like this difference between US and Europe. In the US, I found that companies are a lot more performance-oriented. So essentially they do those bake-offs, essentially. Like when we are talking to big, large kind of fintechs in the US, they say, oh, we're going to try two, three, maybe even four providers sometimes.

**[0:28:03] Dimitri Masin:** And so they are thinking about it's like, okay, what is the infrastructure that needs to be in place for us to test as many providers as quickly as possible? Because they have learned that it's very, very difficult to evaluate providers as a tabletop exercise, because kind of 90 % of the agentic work is kind of hidden below the surface, right? So you don't know how it's going to do in the wild before you actually try it out.

**[0:28:27] Dimitri Masin:** And so they made it all about like bake-offs. And that actually suits us particularly well because so far, because we are such a specialized provider, we serve only fintechs and only financial services. We actually have won every single bake-off that we run against any other company. So that suits us very well. So that's why we are going after the US market, because we actually want to have bake-offs. because whenever we have a bake-off, we know that we can win.

**[0:28:53] Dimitri Masin:** Where it's harder for us is because we are a smaller brand, because we are like niche and vertically focused. We obviously don't have like a large tech brand. So it's harder to get into the door. But once we're in, we always win on the technical merits. So from that perspective, U.S. market is a very interesting one for us because in Europe, you typically pick one provider and kind of go with them and hope for the best.

**[0:29:13] Dimitri Masin:** Whereas U.S., not always, but many times, but in the U.S., like the bake-off is a standard mode.

**[0:29:19] Peter Renton:** Okay, so then last question. What's your vision for Gradient Labs? I mean, the opportunity seems monstrously large, but what's your vision?

**[0:29:29] Dimitri Masin:** Oh, no, 100%. In a nutshell, if you think about putting a bank on autopilot, essentially making it significantly easier to operate and kind of reinventing the whole industry and how it works today, that would be quite something to achieve, right? Because that's what drives me every day. So I kind of think about like transitioning from traditional banking to fintechs. That was kind of the first lack of the journey. And the second lack of the journey is like transforming how companies or banks operate from the ground up and making it infinitely scalable.

**[0:29:59] Dimitri Masin:** That's kind of where the time is going, where the trend is going. And that's what we want to be driving, right? So we want to invent that future or drive that future built towards the future. That's what we are building.

**[0:30:09] Peter Renton:** It really is interesting. I mean, you are at the absolute cutting edge of technological change here in financial services and really interesting learning more about your company, Dimitri. And thanks so much for coming on the show and best of luck.

**[0:30:23] Dimitri Masin:** Amazing. Thanks, Al Peter, for having me.

**[0:30:30] Peter Renton:** Every day we hear news about one large company or another doing a round of layoffs in response to AI automation. That is why I was so interested in Dimitri's answer when I asked what happens to the people who used to do this work manually at their fintech and bank clients. He said not one of his customers has actually cut staff in that space. Instead, they redeploy people to the harder, more human-centric work, the complex investigations, the cases involving financial difficulty and vulnerability.

**[0:31:04] Peter Renton:** It's a useful counterpoint to the doom narrative around AI and jobs, at least for now. Whether that holds as automation continues to improve remains to be seen, but it is quite possible this doom narrative is overblown. Anyway, that's it for today's show. If you enjoy these episodes, please go ahead and subscribe, tell a friend, or leave a review. And thanks so much for listening.
