Last month we hosted a dinner for CFOs and heads of finance from across UK fintech and payments, with André Balleyguier, who leads Applied AI across EMEA at Anthropic, as our guest speaker.
The room had already moved past the question of whether AI can do the work. The questions were about how to use it well. How do we see what it did? Is our data safe? Who is accountable when it gets something wrong? Should we buy a solution or build our own? How do we bring the whole team along?
Those are the questions of people getting ready to act, and they tell us where finance currently is with AI. The capability is there. What finance leaders want now is a way to trust it. Our view is that the firms that pull ahead won't be the ones with the best model, but the ones with the strongest record underneath it.
Capability is no longer the constraint
AI is improving faster than most finance teams plan for. METR, an independent research lab, measures how long a task AI can complete on its own, counted in the time it would take a skilled person. In mid-2024 that was about 11 minutes of work. By spring 2026 it was more than 16 hours, and it has roughly doubled every four months.
Adoption hasn't kept pace. According to Ardent Partners, 58% of accounts payable teams are already using or piloting AI, yet only 1% run it at the core of their process. EY found just 8% of UK CFOs say AI is fully integrated across finance. And according to SSON Research, 86% of finance teams still reconcile in spreadsheets, even when they own a close platform.
The gap between trying AI and depending on it isn't technical. Finance teams already have the tools. What they don't have yet is the confidence to put those tools at the centre of a regulated process.
Where is your finance team with AI today?
The ‘black box effect’ and data privacy
Two questions that came up at the dinner were about visibility and data. Will AI become a black box we can't explain to a regulator? And can we be sure our data won't be used to train models?
André's answer to the first was to treat AI like a junior colleague. No one would accept a decision from a new team member without asking why, or without a trail to back it up. AI can leave one, showing the source it used, the evidence it found and the decision it reached. Often that's a fuller record than a person working in a spreadsheet would leave.
On data, enterprise agreements already rule out training on customer data, and that's written into contracts and audited. In André's view the bigger risk is shadow AI: staff pasting company data into personal accounts because no approved tool exists. His advice was to give an AI agent less access than the person using it and widen it only as it earns trust.
The research backs up the room's instinct. In a 2026 Grant Thornton survey, 78% of senior leaders said they weren't fully confident they could pass an independent AI governance audit within 90 days. Almost half, 46%, said AI underperforms because their controls and compliance aren't working.
That's the point we'd underline. Trust doesn't come from the model. It comes from the system around it: the data it reads, the controls it runs under and the evidence it leaves behind. For payment and e-money firms, the FCA's CASS 15 safeguarding rules make this concrete. Since May 2026 they require daily reconciliations, a monthly safeguarding return within 15 business days, and an annual audit for most firms. In that environment, "the AI said so" will never count as evidence.
Can AI run reconciliations and finance operations?
André's rule of thumb for where AI helps first was simple. The easier a task is to check, and the better it's documented, the sooner AI can take it on. Reconciliation is near the front of the queue for exactly that reason. The numbers either match or they don't, so AI's work can always be checked.
But that doesn't mean AI should do the matching. Matching needs the same answer every time, and rules engines have been tuned and regulated for decades to give it. Where AI earns its place is in the exceptions those engines leave behind. If 14 lines out of 1,300 don't match, AI can pull together the last month's reports, chargebacks and refunds. It can work out why each line broke and draft the explanation for the audit trail. That's where most of a reconciliation team's time goes.
Rules engines should stay in charge of the matching. AI's job is to make sense of what they leave behind.
Would you let AI do the matching in your reconciliations?
Buy vs build
One guest asked the question many in the room were thinking. Why buy from a provider at all when you could build everything with AI?
André's answer was candid. Anthropic provides the intelligence layer, he said, and it isn't going to pretend it can do matching the way deterministic rules can. A regulated process needs more than intelligence. It needs traceability and clear accountability at every step, showing whether a person or the AI did the work, who approved it and when.
That's the distinction finance leaders should hold on to. The model and the regulated layer around it are separate things, and they come from different places. AI added as a separate chatbot, where data gets copied in and answers get copied back, saves minutes. AI built into the system that already holds your data, controls and audit trail saves days. The Hackett Group found that when AI is built into the order-to-cash process, average overdue days fall by 85%.
Accountability is ALWAYS on the finance team
"When did you last fly on a plane without a pilot? You never do. It should be the same for finance teams."
André Balleyguier
Head of Applied AI, Anthropic
Nobody boards a plane without a pilot, even though autopilot system flies around 95% of the journey. The pilot handles take-off, landing and anything unusual, and stays accountable throughout. André's view was that finance should work the same way. AI flags the issues and drafts the evidence. People make the judgement calls and sign off.
That also changes how teams learn. Many of today's finance leaders learned the job by doing reconciliations by hand. The next generation will learn by reviewing; checking the AI's work, explaining each break in their own words and defending it. Someone senior still owns the decision and must be able to explain it.
Adoption is a people question too. One guest described a team split in two, with some people diving straight in while others have the tool but don't know where to start. The answer that came up was practical. Give everyone access to an approved tool and keep sharing real examples of time saved. People who use AI report it helps. In Anthropic's 2026 Economic Index survey, 68% of respondents said they were learning more with AI, and 57% said it had made their skills more valuable.
Where to begin?
These are the first steps suggested by André suggested on the night.
Give the whole team an approved AI tool. André's first step is building AI literacy across the team. That means giving everyone access to an approved assistant so they can experiment and work out where it helps. It also tackles shadow AI, the bigger risk of staff putting company data into personal accounts.
Write down what lives in people's heads. A lot of finance knowledge sits with one person, and if they leave, it goes with them. André's advice was to document processes and exceptions as early as possible. That way the work stays traceable, and AI follows rules the team has set.
Start where you can check AI's work. He warned against putting AI into a full end-to-end process straight away. Begin where you can trace what it's doing, with someone reviewing its work early on. Then track where it gets things wrong.
Give AI less access than the person using it. Start small and widen access as the controls prove themselves. Data the AI doesn't need, especially personal data, shouldn't go in at all.
Share real wins. To bring hesitant colleagues along, André recommended training for everyone and visible support from leadership. Most of all, people need to keep hearing concrete examples, such as a colleague doing something in ten minutes that used to take two hours.
Look for AI inside the tools you already use. For regulated processes, the bigger gains come from AI built into your existing systems, with their data, controls and audit trail. Check that the tools you rely on are building AI in with the same level of security and safety.
The bottom line
The capability is already here. What finance teams need now is a way to trust it, and that won't come from a better model. It comes from clean data, clear controls and evidence for every step. The firms that get that foundation right first will be the ones that move fastest with AI.
Thank you to André Balleyguier and to every finance leader who joined us and brought such good questions.
