Your Board Report Says Customers Got Good Outcomes. Ewen Fleming Asks How You Know

The regulator wants evidence, not dashboards. Here is what reading every customer conversation, instead of one percent, actually changes.

UK financial services firms reported 1.74 million complaints in the second half of 2025¹ and paid out £236 million in redress. Firms upheld more than half of the cases they closed. Every one of those started as a conversation that went wrong somewhere, and in most firms nobody was listening to it at the time.

Ewen Fleming has spent more than forty years in that industry, first inside the banks, then advising them, and now selling the technology that reads what nobody previously had the capacity to read.

The timing matters. In July 2026 the Financial Conduct Authority published a review of how firms monitor customer outcomes², drawing on board reports, information requests and a survey of 56 firms. Its opening point was that collecting data, listing metrics and reporting management information do not, by themselves, show that customers are getting good outcomes. Firms have to explain what the information tells them, what they did about it, and whether the action worked.

What follows is the argument Fleming made before he joined the company he now helps run, what the work involves once a firm says yes, and why he thinks the regulation is the least interesting part of the case.

What a One Percent Sample Cannot Tell You

Most firms quality-assure customer service the same way. A team pulls a random selection of calls and chats, scores them against a checklist, and reports the result upward. By Fleming’s estimate, that covers under two percent of customer interactions in a typical firm. The rest are never read by anyone.

That was defensible when reading them all was impossible.

It is harder to defend now, and the regulator has noticed. The FCA review found that firms could often describe their monitoring frameworks in convincing detail but could not produce the trail from an issue to a decision to a result. Some relied on lagging indicators. Some used operational measures, conversion rates or review completion, as a stand-in for whether the customer was actually better off. The gap the regulator keeps pointing at is not a shortage of data. It is the distance between having data and being able to say what it proves.

Fleming saw that gap from the consulting side, working with a bank’s own numbers, before he saw the technology that closes it.

At Johnston Carmichael, I was working on the new UK Consumer Duty regulation and saw an opportunity to use compliance as a catalyst for delivering better outcomes for customers, while simultaneously reducing the operational cost to serve by reducing failure demand and eliminating potential complaints.

That is the whole argument in one sentence, and the order of it is deliberate. Compliance is the catalyst, not the product. The product is fewer things going wrong.

Reading All of It, Not a Sample of It

The alternative he recommended, before he had any commercial interest in it, was to review 100% of customer interactions using AI rather than sampling them. He described it as a game-changing opportunity, and the reason was scale: a consulting engagement improves one process inside one client, while a platform improves the same process everywhere it exists.

He had been circling this problem for a long time. He was responsible for premium banking across both NatWest and RBS, when the group was still called RBS, and led the productivity work inside it. That job is about finding where effort disappears and getting it back.

The difference between then and now is that the effort is finally visible.

What the Work Involves Once It Starts

The first step is not technical. It is understanding the journeys customers actually take, as opposed to “the theoretical journeys that follow a product sales and fulfilment process”. Those two maps are rarely the same document, and the difference between them is where complaints live.

Then the interactions become viewable at two levels at once, holistic and granular, so friction and frustration surface early rather than in a quarterly report. Then the firm takes measured, specific steps at the root cause, which he describes as “eliminating guesswork or gut feel from customer service”. Then, and only then, the firm can evidence 100% of the outcomes it delivered, to what he calls “the evidential standard required by Consumer Duty”.

The changes this produces are smaller than the technology sounds. Ask him for evidence and he offers two: a better process for customers reporting a recent bereavement, and new agent headsets to improve call clarity for customers with impaired hearing.

Nobody writes a press release about headsets. The customer who could finally hear the answer does not care what model found the problem.

Why the Model Should Be Small

The wider market is not short of AI. In the most recent Bank of England and FCA survey of AI in UK financial services³, 75% of firms said they were already using it, with another 10% planning to within three years. The same survey found 46% of firms had only a partial understanding of the AI they were using, against 34% claiming a complete one.

That is the room Fleming is selling into: high adoption, thin comprehension, and a regulator asking firms to evidence outcomes they cannot yet explain.

His answer is a smaller model doing one thing exactly. Where the general commentary fixes on large language models because they are the ones everyone can open in a browser, he argues for small language models on specificity, precision and speed against a complex, narrow, high-volume problem. A model that reads a million customer conversations in a regulated firm does not need to write poetry. It needs to be right about this, consistently, and cheap enough to run on all of it.

Ninety-eight percent coverage of nothing is still nothing.

Forty Years of Seeing What Short-Termism Costs

The credentials sit behind the argument rather than in front of it. Six years as a partner at Grant Thornton leading the retail and SME banking consulting team. Almost six more as Johnston Carmichael’s financial services sector leader, advising on strategy, performance improvement and transformational change. Fourteen recommendations on his profile, the earliest from people who reported to him. One of them wrote: “He leads by example and is very clear of the goals and expectations at the outset. He asks simple, insightful questions.”

He is also unusually specific about what the years taught him. His career, he says, “has spanned many downturns, including the nadir in the aftermath of the global financial crisis”, when trust in the sector collapsed and regulators fined firms heavily for mis-selling and imprudent balance sheet management.

In my 40+ years in financial services, both in industry and as a consultant to the sector, I have seen the economic imperative for a strong financial services sector. But I have also felt that the focus has often been too short-term, missing opportunities to find the optimal balance that delivers for customers, colleagues, investors and regulators without ever delivering for one stakeholder at the expense of another.

This is what a technologist arriving from outside would struggle to supply. He joined Malted AI as head of financial services and became chief commercial officer six months later, and the thing he brought was not the model. It was the ability to say, internally and to a buyer, where the return comes from and how long it takes to arrive.

From my experience, I also knew that the use case was largely universal across banking, insurance and wealth management firms serving both retail and business customers. In many respects, the regulatory angle is simply a reason to prioritise the opportunity; the underlying business benefits exist even without the regulatory compliance imperative.

Read that twice if you are the person who has to approve the budget. He is telling you the regulation is a reason to do it this year, not the reason to do it.

Where Ewen Is Now

He works from Edinburgh, where Malted AI is based, and spends his time with banks, building societies and wealth managers who want to see every customer interaction rather than a sample of them. The order of the plan is deliberate: product market fit in the UK first, then expansion into the US, which is set for early 2027. He also sits on the board of trustees of Cancer Card, a Scottish charity that runs a patient-led directory of support services, having previously served as its interim chair.

Ask what he wants the next twelve months to do and the answer is partly about the market, not the company: to “raise the profile of Small Language Models (SLMs) amid the generic AI commentary that tends to focus primarily on easily accessible Large Language Models (LLMs), highlighting SLMs specificity, precision and speed in addressing complex challenges”.

Ewen Fleming is Chief Commercial Officer of Malted AI at malted.ai.