A friend of Anne Boden’s asked Meta’s new AI assistant, Muse, to get her some vanilla ice cream. It found some on Deliveroo and sent her a link so she could pay for it.
Boden, who founded Starling Bank, tells the story in the FT this week, and she isn’t impressed. “That’s not an autonomous agentic economy,” she writes. “It’s a marginally faster human economy with extra steps.”
It’s a small story, but her whole argument sits inside it. The assistant could do everything except the one step that mattered, and it stopped there because the payments system is built to make sure a person is present when money moves. As she puts it, “The central point of the solution we’ve all adopted has been to insist on a human presence.” For years that has been exactly the right answer. Boden thinks it’s about to become a problem.
Her piece is about banking. I read it as a mortgage professional and kept seeing our industry in it.
When she pitched a mobile-first bank 12 years ago, she says, the industry ignored her, then laughed, then grew sceptical, then started to worry. Her prediction now is that the next phase “won’t even involve an app”. Instead, “a very able AI agent is almost certain to step into the gap between banks and their customers”. These are agents that research products, compare prices and complete tasks for us, within permissions we set. She points to Muse as evidence that people want them. It launched on 8 September and, she writes, passed 6.6 million downloads and topped the US app charts within two weeks.
For brokers there’s an obvious catch. In mortgages, the gap between lenders and their customers already has someone standing in it. Most UK mortgages are arranged through intermediaries. So when Boden says an agent will step into that gap, the question follows naturally. Does it stand alongside the broker, or instead of them?
Think about what an agent could plausibly do for a homeowner whose fixed rate ends next spring. It could read their bank statements through open banking, pull current rates from lender and comparison sites, and come back with a shortlist. All of that is within reach. Completing the application is another matter, and that’s the ice cream problem again. Mortgage processes lean on human presence even more heavily than card payments do. Identity checks, fraud screening, declarations the applicant signs personally, an offer that has to be accepted and a deed someone has to execute. Each of those exists for a good reason, and each one assumes a person at the other end.
Then there’s the scale point, which Boden makes about payments and which carries over uncomfortably well. “Human beings sleep, but agents do not.” Our fraud monitoring systems, she says, “have been carefully calibrated around human behaviour” and will treat agent activity as suspicious. Picture an agent running decisions in principle with eight lenders at three in the morning to find the best fit for its owner. To a lender’s fraud team today, that looks a lot more like an attack than a diligent customer.
Boden doesn’t pretend the risk is imaginary. She quotes Britain’s AI minister, Kanishka Narayan, who warned this week that cyber attacks on the UK by rogue AI agents were “likely, if not inevitable”, and she expects banks to be an obvious target. They’ll have to keep the bad agents out while letting in the ones customers have authorised. Mortgage lenders, with large sums moving on completion day, face the same balancing act with less room for error.
So where does that leave the broker? I think her argument cuts in our favour, at least for a while.
Her main complaint is that the sector is “looking the wrong way”. When banks talk about AI, she says, “they are mostly talking about efficiency and trimming costs.” Plenty of broker firms are doing much the same, and with good reason. Meeting notes, file checks and research summaries save real hours. But nearly all of it is back-office work. Very little of it asks what happens when a client turns up with an agent instead of a phone call.
And a mortgage isn’t a tub of ice cream. Boden says some account holders “would be quite happy for a non-human to act on their behalf for certain purchases”. I suspect a 25-year debt secured on someone’s home won’t make that list for a long time. The more likely pattern is an agent doing the legwork and then handing over: a shortlist, a full picture of the client’s finances, a set of permissions and a request for advice. In that version of events, the broker is where the agent ends up. That only works if brokers are ready to receive it.
The routine end of the market is the part to watch. A like-for-like product transfer at the end of a fix, same lender, nothing changed, is exactly the sort of rules-based, low-drama decision a customer might be happy to let an agent handle alone. That’s a steady part of many brokers’ books, and it’s the part most exposed.
Boden finishes by calling for an initiative that brings banks, regulators and AI companies such as OpenAI and Anthropic to the same table. For mortgages I’d want one more chair pulled up. Lenders will decide what their front door looks like and the FCA will decide what an agent may do on someone’s behalf. Brokers are the people who know what a good mortgage conversation actually involves, from the nervous first-time buyer to the self-employed client whose income doesn’t fit the boxes. If they aren’t in the room when the rules for agents get written, they’ll inherit a process designed by people who have never had that conversation.
Twelve years ago the banks laughed at Boden, and then they copied her. When your client’s agent goes looking for a mortgage, who will it ask first?
Sources:
Financial Times, “The battle over the AI-led future of banking has begun”, Anne Boden, 14 October 2026
GOV.UK, “Building a Safer, More Secure Britain in the age of AI”, speech by Kanishka Narayan, 12 October 2026