.png)
How Lee Provoost Is Rebuilding Flagstone for an AI-Native World
May 21, 2026
There is a specific kind of conversation that is taking place right now amongst CTOs and business leaders, somewhere between strategy and survival. It is not the conversation about productivity gains or efficiency ratios, but the one that starts with an honest assessment of whether the business you are running today is the same business that will exist in eighteen months. Lee Provoost, CTO at Flagstone, has been having that conversation since Christmas 2025.
Flagstone is one of the UK's leading cash deposit platforms, operating at the intersection of financial services, regulatory complexity, and high-stakes customer trust. It is not the kind of business where you lean into technological change recklessly. And yet, speaking to Provoost in May 2026, what comes through most clearly is not caution but conviction: a settled, clear-eyed belief that the transformation now underway is not optional, that the risk of moving too slowly is greater than the risk of moving at all, and that the businesses which hesitate for another six to nine months are making a choice they will not easily undo.
Understanding how he arrived at that position requires going back a year.
Provoost is candid about where Flagstone stood during 2025. The business made sensible, considered decisions: GitHub Copilot licences for the engineering team, a managed trial of Microsoft 365 Copilot for the broader organisation, the kind of structured experimentation that a regulated financial services platform running under data residency constraints is almost obliged to run. The rationale for Microsoft's tooling was straightforward and legitimate. None of the frontier model providers, including Anthropic with Claude and OpenAI with ChatGPT, offered genuine UK or European data residency at the time. "Even though ChatGPT and OpenAI claim that their enterprise offering allows you to host your data in the EU or UK," he explains, "there is actually quite an important detail there. The actual inference, the processing of your data, is still in the US." For a business like Flagstone, where the handling of customer financial data is not merely a preference but a regulatory obligation, that distinction mattered.
The honest assessment of the AI tooling landscape at that point was one of horses for courses. Some tools performed well in day-to-day organisational tasks, meeting transcription, email drafting, calendar management, while others proved more valuable in the hands of the engineering team. What mattered more than any individual product, though, was the broader trajectory: frontier models were advancing at a pace that would rapidly reframe every tooling decision made in 2025. That context is everything.
Christmas 2025 brought what Provoost describes as a step-change moment, one that he, like many technology leaders, experienced firsthand during the break. The catalyst was Anthropic.
The technology industry has a habit of marking moments in retrospect, identifying turning points only after the direction has become clear. The emergence of Claude Opus and Sonnet 4.5 in late 2025 is one of those moments, but what is striking about Provoost's account is that he experienced it in real time. He went on holiday. He tinkered with the tools, as many in the industry did. And he came back not with a refined view on the experiment but with a fundamentally altered assessment of the trajectory.
"When 4.5 was released," he says. "That was the very first time when everyone realised: holy sh*t, our whole industry and our future is going to change. Because before that point, it was very much trying to hack together, trying to deal with your frustrations. Christmas 2025 was a pivotal moment for me."
The response was not a working group or a discovery phase. It was a direct conversation with the executive team in the first week of January, and a simple, sobering message: everything planned for 2026 had to change. The budget set in September, months before any of this had crystallised, was nowhere near what the opportunity now demanded. "Everything we had planned for 2026 has to radically change," he told them. And the opportunity, as Provoost framed it to his colleagues, was inseparable from the threat.
The business case conversation for AI adoption in large organisations has become something of a running joke among honest technology leaders, and Provoost does not shy away from saying so. The spreadsheet models that circulate, the ones showing millions saved by multiplying employee headcount by theoretical hours reclaimed, are not serious analyses. If someone saves three hours a week, they might in fact just take a longer lunch, send a few more emails. The efficiency does not translate automatically into measurable value, and everyone in the room knows it.
The case Provoost made was different in kind. It was not about productivity ratios. It was about whether, in eighteen months, Flagstone would still be commercially viable if its competitors had embraced these capabilities and it had not. That framing reorients the entire calculation. You are no longer asking whether the investment pays back. You are asking whether the inaction is survivable.
"This is really around existential threat," he says. "And that puts you in a very different mindset of justification of investment and the risk-reward type of calculation you do."
The board and executive team, led by a CEO Provoost describes as deeply attuned to technology trends, required little persuasion. There was already alignment at the top. What remained was the practical question of where the money would come from, given a budget cycle that had closed months before the landscape shifted. The answer was reallocation: redirecting spend from other areas of the technology estate, and choosing not to backfill several positions that had become vacant through natural attrition. By the time Q1 had run its course, that decision looked well-founded. Enough material improvement in engineering productivity had emerged to justify the conviction, and enough headroom had been created to fund the ongoing investment.
One of the most striking passages in the conversation with Provoost is his attempt to place the current moment in historical context. The internet, e-commerce, social media, smartphones: each of those transformations played out over a decade or more, giving industries and institutions time to adapt, to develop norms, to make mistakes and recover from them. The current change is not operating on that timescale.
"The impact of AI on coding, the first major aha moment, was Christmas last year. Now we are in May. Four months ago, people woke up to a new world." In those four months, major companies lost significant portions of their valuations. Businesses that had barely existed twelve months ago were being ascribed extraordinary worth. And the pace of change is not slowing.
"What I've said to everyone in the business," Provoost notes, "is that Flagstone is going to be a different business by December this year."
That is not hyperbole deployed for internal motivation. It reflects a genuine belief that the rate of adoption and the rate of capability improvement are combining in ways that compress the timelines most organisations are accustomed to working within. The question is not whether to adapt, but how quickly.
Flagstone's approach to rolling out AI in the engineering organisation was deliberately unstructured in its early phase, and Provoost is clear that this was intentional rather than accidental. A conventional change programme, with defined workstreams, governance gates, and structured rollout plans, would have been too slow and too rigid for a moment where the tools themselves were changing faster than any plan could accommodate.
Instead, the organisation leaned into a period of genuine divergence. Engineers were encouraged to develop their own workflows, their own MCP configurations, their own ways of harnessing the tools that were now available. Champions emerged organically and learning spread through informal channels. The goal was not uniformity but acceleration, getting the organisation to a place where the genuine impact could be felt and assessed, rather than theorised.
That phase is now closing. Coming out of Q1, Provoost published an internal thought piece setting out a new direction for the rest of 2026, and the language is instructive: standardisation of ways of working, accelerated governance and control, pre-built container environments with AI coding agents embedded and restricted appropriately. The diamond of innovation, as he describes it, first widens and then narrows. Flagstone has widened, and is now narrowing with intent.
The areas of greatest impact so far have been those where experienced engineers apply AI as a force multiplier on work that previously consumed disproportionate time. Complex code base refactoring, architectural analysis, understanding the domain topology of legacy systems: tasks that previously required two or three weeks of painstaking coordination across teams are now achievable in days by staff engineers who bring the contextual understanding that makes the AI output actually usable. The point Provoost makes here is important and often missed in discussions of AI in engineering: the value of these tools is not democratisation of expertise. It is amplification of it. A staff engineer with fifteen years of experience can now do in three days what previously took three weeks. A junior engineer given the same tools does not produce the same result, because the tools are not a substitute for judgment, only an accelerant of it.
Provoost has thought carefully about how to communicate the transformation of the engineering role in terms that do not threaten but instead motivate. The metaphor he reaches for is the shift from conveyor belt to factory architect. For the past two decades, engineers were, in meaningful ways, the conveyor belt: the people who sat at the mechanism of production, writing the code line by line, holding the process together through their presence in it. In the emerging model, a significant portion of that production moves into what he calls a dark factory, automated and running without constant human intervention. The engineers who previously operated the belt now design, build and govern the factories and the robots.
"The complexity and the interestingness of the job will significantly increase," he says, "but also change."
Flagstone currently employs around 200 people across its R&D function, covering engineering, product, design, and data. Under the old model, growth in output required growth in headcount: another hundred people to deliver more. That assumption no longer holds. Provoost's operating thesis for the next phase is that those 200 people, properly equipped and properly structured, can produce the output of 400. The business does not need to grow the team to grow the capability. It needs to grow what each person in that team can do.
That has significant implications for how roles are defined, how performance is assessed, and how the culture of the engineering organisation is shaped. It also creates a genuine challenge for those who are unwilling or unable to adapt. Provoost is honest about this, describing it not as a threat but as mathematics: if twenty colleagues are doubling or tripling their impact through AI adoption and you are not, the gap is real regardless of how it is characterised.
The constraints that shaped Flagstone's approach to AI adoption are not merely theoretical. The business operates under genuine regulatory obligations around data handling, and the implications of getting this wrong are not abstract. The starting point for Provoost was therefore to contain the initial rollout within the engineering organisation, where the data in question is source code rather than sensitive financial or customer information. That is still a question of intellectual property and a calculated business decision, but it is a fundamentally different category of risk from exposing customer financial data to models running inference in the United States.
When it came to the source code question, the leadership team at Flagstone made a considered call. Processing proprietary code through US-based inference infrastructure carries inherent risk for any business, but Provoost is clear-eyed about where Flagstone's actual competitive advantage lies. The codebase is one part of the picture, but it is not the moat. The moat is the commercial relationships, the regulatory standing, and the operational track record built over years, none of which a competitor could acquire simply by accessing the underlying code. Framed that way, the decision to proceed became a proportionate one, and it gave the engineering team the runway to move quickly without needing the heavier governance infrastructure that any use case touching customer or financial data would inevitably require.
The boundaries of that framework are now being tested. Teams across the business are beginning to ask questions about analysing operational data, about using AI in contexts that go well beyond the engineering environment. The answer at the moment is a clear boundary, but it is one Provoost is actively working to understand and, in time, to move. The engineering rollout was the beginning of that journey, not the destination.
The last topic Provoost raises, and the one he describes with equal measures of excitement and apprehension, is the token apocalypse: the escalating cost of running AI at the scale a genuinely committed organisation now requires. His forecast for 2027 is that Flagstone will spend over a million in AI tokens and licences. For a business of Flagstone's size, that is a material budget line requiring proper justification rather than the reallocation of existing technology spend that has funded the current phase.
The response he is actively investigating is one that would have seemed improbable even recently: the possibility of setting up on-premises hardware, physical racks in a data centre, running a private AI Linux cluster that could potentially remove several hundred thousand from the annual token bill. In a world that has moved almost entirely to cloud infrastructure, the prospect of touching hardware again is something he describes with genuine enthusiasm. It is also, he notes, a signal of how much has changed, and how quickly.
"I've always loved playing with hardware," he says. "In a cloud world, you don't tend to get to touch hardware anymore. That is exciting for me."
When asked what he would say to other CTOs who have not yet moved as far along this journey, Provoost is measured and generous. He understands the hesitation of leaders in regulated businesses, where the data residency issue is real. For some organisations it genuinely is a deal-breaker, and he does not dismiss that. His advice is to find a way to start within the engineering organisation, where the data constraints are less acute, and then to accept a degree of risk that may feel uncomfortable.
"Ask for forgiveness rather than wait till it's all solved," he says. "You can either wait another six to nine months before you think it's all going to be resolved, or you can lean into it, embrace it, accept that a few mistakes will be made. But in six to nine months, you're a different business rather than only just starting your journey."
The closing thought, and perhaps the most useful piece of advice for any technology leader navigating this moment, is about permission. Permission to not have it all figured out. Permission to feel behind, because the honest truth is that almost everyone is behind. There is no company, with perhaps a small number of exceptions, that is genuinely on top of what is happening. The leaders who are doing well are not those who have built the perfect framework. They are the ones who have accepted that the framework is still being built, that mistakes are part of the process, and that the cost of waiting for certainty is paid in the kind of ground that, once lost, is very difficult to recover.
"All CTOs need to be kind to themselves and accept that this is a wild west, where very few people have really figured it out," Provoost says. "You will always feel behind. And that is normal. We will fix the problems one at a time."
Lee Provoost is CTO at Flagstone, the UK's leading cash deposit platform. This interview was conducted as part of Signal by Gathered & Found's annual AI research initiative report exploring how senior technology leaders are navigating AI transformation. Signal by Gathered & Found (2026 AI Report) | Leadership in Focus
Join our newsletter and stay ahead!