Let me start with a small confession. This weekend, while thinking about this post, I asked an AI assistant to find me the best easy-access savings account, and then move my money into it.
I half expected it to fall over. It didn’t.
It went about the job like a very diligent, very unsentimental customer. It skipped past the out-of-date “best buy” articles that still clutter the search results, found a comparison table with rates checked that same day, and came back with an answer in a couple of minutes. Two providers at the top paying 5.00%, followed by a long string of challenger banks and fintechs.
Here’s what struck me:
- Building societies barely featured. Of the twelve accounts in the main table, just one was a building society – in eleventh place. If the agent had simply taken the top three, no mutual would have been in the running.
- The best deals were for strangers. Half of the top twelve were for new customers only, and the top rate was a six-month bonus. Taken to its logical conclusion, the agent would turn me into a permanent new customer, moving my money every few months.
- It read the small print. It flagged that some “easy access” accounts limit withdrawals, and that one needed £25,000 just to open. It didn’t compare headlines; it compared terms.
- The comparison site was already becoming the agent. The site it used has its own AI savings assistant, rate alerts, an API for developers – and affiliate links. The gatekeeper was already sitting there.
What it didn’t do was ask about service, branches, who owns the provider, or how long I’d been with my current one. Not once. The only thing beyond the rate that came up was FSCS protection.
And then it stopped. Opening an account needs your identity, address history and National Insurance number. The agent could find the deal, but it couldn’t (yet) be me. That’s the only thing standing between today’s search and tomorrow’s switch – and, as we’ll see, the regulator is already working on it.
In my last post, I argued that the more digital financial services becomes, the more valuable the human becomes. I stand by every word. But this little experiment made me realise there’s a second half to the story I hadn’t fully thought through. What happens when the customer isn’t human either?
The Regulator Has Already Called It
This isn’t just a me with too much time on a weekend. In July 2026, the FCA published the Mills Review – the first study of its kind by any financial regulator, anywhere looking at how AI will reshape retail financial services by 2030.
Its central finding is that financial services are moving from human-led and episodic to AI-enabled, continuous and delegated. In plain English: customers will increasingly hand their financial decisions to software that acts on their behalf, all the time, rather than making a decision themselves once every few years (usually when a letter arrives telling them their fixed rate is ending).
Two things jumped out at me:
- Consumers are already halfway there. Around a quarter of the consumers surveyed said they trust general-purpose AI tools for financial advice. That’s before most people have even seen a proper agent in action.
- Control of the interface is the new market power. The review flags that whoever controls the AI-mediated customer interface is likely to hold significant market power. Read that again. Not the bank. Not the building society. Whoever owns the agent.
The FCA isn’t writing new AI rules – Consumer Duty remains the anchor, but it does want a trusted framework for how agents are identified, authorised and held accountable. The direction of travel is clear, and it has the regulator’s backing.
The Inertia Economy
Here’s the uncomfortable bit.
A meaningful slice of retail banking profit relies on customers not moving. Savings balances sitting on yesterday’s rate. “Loyalty” that is really just inertia. Switching that feels like too much hassle for an extra half a per cent.
We don’t tend to say it out loud, but the whole industry benefits from this – and yes, mutuals included. Every building society has a back book.
Consumer Duty has already started to squeeze the gap between what new and existing savers earn. But regulation moves at the pace of supervision. An AI agent that checks rates every week, never gets bored and never forgets would finish the job in months, not years.
I’m not saying that’s a bad thing. As a mutual member myself, I’d much rather my society earned my loyalty than relied on my laziness. But it does mean many providers’ business models are about to be tested in a way they never have been.
Three Ways This Could Play Out
Nobody knows exactly how this unfolds (and anyone who tells you otherwise is probably selling something). But I can see three broad scenarios-– and in reality, we’ll probably get a mix of all three.
- The rate table wins. Agents reduce every product to a number. The best-buy table becomes the market, and the cheapest provider takes the money – until someone undercuts them. Brand, service and history count for nothing. Great for consumers in the short term; brutal for anyone without a low cost base.
- Trust becomes the filter. Agents and the people instructing them, don’t just optimise for rate. They weigh security, service quality, complaint records and values. In this world, the things mutuals have always been good at stop being “soft” qualities and become hard data an agent can compare. How do organisations articulate and make it externally visible to customers (and agents) will become become key.
- The gatekeeper takes a toll. Whoever owns the agent owns the customer, and providers pay for access – much as retailers pay to appear on a marketplace. I’ve seen this earlier in my career in payments: Big Tech won the payments interface but never really wanted the balance sheet. The firms building agents may not want to become banks either – but they will very happily rent access to the people who use them.
Which scenario wins depends partly on regulation, partly on technology and partly on what we do now. Which brings me to marketing (the most exciting, bit).
Marketing to Machines
I’ve spent most of my career trying to persuade people. The idea of persuading software is a strange one – but it’s coming.
If an agent is choosing on a customer’s behalf, what does it actually see? Not your advert. Not your community sponsorships. Not the friendly team in your branch. It sees structured product data (rates, terms and eligibility, in a form a machine can read), how transparently you publish fees and rate changes, whether it can actually open an account through an API or hits a PDF form and gives up, and whatever evidence exists on service and complaints.
Brand doesn’t disappear in this world, but it has to become legible to machines. That’s a whole new discipline for marketing teams, and almost nobody in our sector is talking about it yet. Being invisible to agents will be the 2030 equivalent of not having a website in 2005.
And here’s the twist: this could be a leveller. An agent doesn’t care about advertising budgets. A smaller provider with fair products and clean data could be found as easily as the biggest brand on the high street. More on that shortly.
What an Agent Can’t Do
Agents are brilliant at routine, repeatable decisions with clear criteria. Which savings account pays the most? Which mortgage has the lowest total cost? When should I move my money? Give those to the machine – it will do them better than most of us.
But they’re poor at the moments that actually matter in someone’s financial life: sitting with a family after a bereavement; helping someone who’s fallen behind and is frightened to open their post; making a judgement call on a complex mortgage that doesn’t fit the box; reassuring someone who’s just been scammed and no longer trusts anything on a screen.
And the demand for people isn’t going away. Nationwide saw an 11% increase in customers using its branches in the year to September 2025, with a third of its current accounts opened in branch. Meanwhile, Which? counts more than 6,500 bank branches closed since 2015. As the rest of the industry retreats, the human layer gets scarcer – and whoever still offers it stands out.
So agents don’t kill the human touch argument. They sharpen it. If the machine handles the routine, the only thing left to compete on – beyond rate – is how you show up when it really matters.
Funding the Human: The Automation Dividend
In Human Touch, I said every pound spent automating the routine should be explicitly reinvested in the quality of the human interaction. Fair challenge: what does that actually look like in practice? Let me try to put some flesh on it.
AI is going to take real cost out of banking, document processing, mortgage administration, call handling, complaint triage, compliance checks. The question is where those savings go. The default answer is straight to the cost-to-income ratio (or manes in mutual speak). I’d argue for treating them as an automation dividend instead, and deliberately reinvesting a share of it in human service.
In practice, I think that human layer has five parts:
- Branches built around life moments, not transactions. Fewer counters, more private rooms. Mortgages, bereavement, financial difficulty, later-life planning, fraud recovery. And where you’re the last branch in town, community infrastructure too.
- Specialists, not generalists. Underwriters for complex cases, later-life lending specialists, financial difficulty and bereavement teams.
- Humans as the trust check. As voice cloning makes the phone less reliable, a known person and a physical place become part of fraud defence. “Pop into the branch and we’ll confirm it’s really us” may turn out to be one of the most valuable sentences in banking.
- A guaranteed route from agent to person. If a member’s AI agent is managing their money, there should always be a clear handover to a human for big or emotional decisions. Design it in from the start, rather than leaving people stuck in a loop between two machines.
- Colleagues freed up and better equipped. Automation takes away the admin; AI co-pilots give colleagues the full picture in seconds. What’s left is the part of the job people actually joined for.
How do you fund it? Ring-fence a share of the AI savings and say so publicly – the exact percentage matters less than the commitment. Turn it into a promise, the way Nationwide turned branch retention into a brand asset with its Branch Promise, now extended to 2030. Measure the human, not just the cost – pair cost-to-income with Consumer Duty outcomes, complaints and support for vulnerable members. And use the mutual advantage: a plc has to explain to shareholders why AI savings aren’t being handed back to them. A mutual can say they are being handed back to the owners – as service.
It’s the 2008 lesson again: investing through adversity beats austerity (something our politicians should learn). AI is a different kind of disruption, but the principle holds.
What About the Medium-Sized Societies?
It’s easy to write all this from Nationwide’s perspective – hundreds of branches, a technology budget most of the sector can only dream of, and the scale to make network-wide promises. But most building societies look nothing like that. So what does it mean for a medium-sized society, with a modest branch network, a regional heartland and a balance sheet of a few billion pounds or less?
In my view, the case is actually stronger, because societies of this size already compete on people rather than scale. Service, individual underwriting and local relationships are often the whole proposition. But the practicalities are different:
- Agents could level the playing field. A medium-sized society can’t outspend the big brands on advertising but an agent doesn’t care about advertising. Fair rates and clean product data could get it found as easily as anyone. The flip side? If its data isn’t machine-readable, it simply won’t show up.
- Savings inertia is a bigger exposure. Most mid-sized societies are funded largely by retail savings. If agents start sweeping back-book balances to better rates, honest and consistent pricing becomes a strategic defence, not just good conduct.
- Intermediaries may meet the agent first. Most mortgage business comes through brokers, who are early adopters of digital driven sourcing tools. For niche and complex lending, the human layer is the underwriter who picks up the phone and the BDM who knows the broker. The AI sourcing tool can narrow the options; it’s still a person who makes the case for the tricky application.
- Don’t build – buy and collaborate. No mid-sized society needs an in-house AI lab. Proven vendor tools in the back office are where the savings are, and the expensive agent-readiness plumbing (data feeds, APIs, standards) is far better tackled together – through the BSA or partnerships between like minded peer societies – than alone.
The honest tension is cost-to-income/manex. Smaller societies already run higher ratios than the giants, so the human layer has to earn its keep – in retention, conversion, broker satisfaction and Consumer Duty outcomes. The automation dividend framing helps here: it’s reinvested savings, not extra cost.
The ‘So What?’
Four things, if you’re sitting in a society today:
- Price honestly. If your savings book only works because people don’t move, agents will find that out. Fairer back-book pricing turns a weakness into something you can defend.
- Get agent-ready. Clean data, open APIs, readable human data points. Being invisible to agents will be like having no website.
- Fund the human. Treat the savings from AI as a dividend, and reinvest part of it in the moments that matter.
- Help write the rules. The FCA is building the framework for AI agents now. The sector should help shape it, rather than react to it afterwards.
AI will make banking cheaper. The question for mutuals is what we do with the change.
So I’ll leave you with the question I’ve been asking myself since that experiment: if your customer’s AI agent reviewed your products tomorrow, would it recommend you, or move the money?
Feel free to disagree, agree or suggest alternative views – happy to hear them all.
best
MN
