The government has a clear action plan for agentic commerce

HM Treasury's Financial Services AI Adoption Plan, published 14 July as part of the Mansion House package, sets out 10 recommendations for scaling AI safely across financial services, with agentic payments named as one of five priority areas. Much of its evidence base comes from the FCA's Mills Review, published earlier this month, which treats widespread agentic AI and even longer-horizon shifts like quantum computing as a reality for 2030. As a regulated open banking payments provider already building for a world where AI agents transact on people's behalf, we've read both closely.

Payments are where agentic AI gets tested first

Of everything AI touches in financial services, payments is the sharpest edge. It's the point where consumer trust and money meet directly. The review's own consumer research found that one in five UK adults is already open to AI making decisions for them, with demand strongest for complex or high-stakes choices.

Here's where we think the pattern will diverge. Consumers are likely to delegate recurring payments to AI agents faster than one-off purchases. Similar consumer behaviours already exist; anyone who's set up a Direct Debit has essentially delegated a recurring payment decision, which they don't actively supervise. Agentic commerce for subscriptions and bills is a smaller leap than agentic commerce for one-off spontaneous purchases.

Agentic commerce needs a sound foundation

The Mills Review is refreshingly direct about the gap. It states plainly that the current payments rulebook was designed around a human approving each transaction, and isn't yet aligned to agents executing payments on a consumer's pre-authorised, standing instruction.

That gap matters because trust, once lost, doesn't come back through marketing. The Mills Review itself is candid that the existing regulatory framework remains sound and isn't being rewritten. What's missing is the foundation underneath: the practical plumbing that lets an agent prove the identity of who it's acting for, act within limits the customer has agreed to, and be switched off or overridden the moment the customer wants to step back in. If agentic payments go wrong once before that infrastructure exists, payers revert to old habits and take years to try again. Getting the foundation right for agentic commerce before it becomes mainstream, not after, is the only way this doesn't stall on its first stumble.

That commitment is already becoming action. The Adoption Plan's recommendation on agentic payments puts a timeline on closing the gap: the government's forthcoming payments regulation consultation will build a "trust framework" for agentic payments, covering legal liability when agents transact, Know Your Agent identity protocols, and machine-to-machine authentication standards. That's the practical plumbing the Mills Review said was missing, now with a government commitment attached.

Delegation only works if the consumer keeps the final say

The Mills Review has a helpful way of showing how that trust gets built gradually, rather than all at once.

How AI autonomy builds up, and what that looks like for payments:

1. Operator — AI is just a tool. The customer makes every decision and does the work themselves. Payments today: paying each bill manually.

2. Collaborator — Human and AI act together. AI helps set things up, but the customer still drives it. Payments today: setting up a Direct Debit or standing order, once.

3. Consultant — AI recommends. The customer decides. Agentic payments: AI flags a cheaper provider, or a bill that looks off.

4. Approver — AI prepares an action. The customer signs off. Agentic payments: AI gets a switch or payment ready. The customer approves it.

5. Observer — AI acts within agreed limits. The customer monitors outcomes. Agentic payments: AI manages recurring payments within set limits. The customer checks in.

Source: adapted from the FCA Mills Review, "AI and the future of retail financial services" (July 2026), Figure 2, p.13.

Wherever agentic payments sit on that spectrum, the principle holds: the customer keeps the final say, not just at the start, but at every stage along the process.

That's what we're hearing directly from consumers, too. In our own research, 64% of UK consumers surveyed said they're open to AI managing their recurring payments, provided they keep control over limits or final approval. Requiring approval on every payment was the single biggest trust requirement, picked by 45%, ahead of consumer protection and real-time notifications. That's not a ceiling on where agentic payments can go; it's a starting point. 

The same pattern has played out with every new payment method: people want tight control at first, then relax it as trust builds. Contactless payments followed the same path, with low limits at launch that gradually increased as people got comfortable and trusted the method more.

We'd expect this requirement to ease as consumers get more comfortable with agentic payments actually working as promised, in the same way people now trust Direct Debit to run quietly in the background without checking every transaction. And consumers are already showing signs of that comfort: according to our research, they're far more comfortable with AI adjusting, pausing or cancelling payments they've already set up than with it opening new mandates or switching providers on their behalf.

In other words, giving consumers control doesn't limit how much they'll delegate; it's what makes them willing to delegate more in the first place. The more customers can see what an agent is doing, adjust it, or cancel it whenever they want, the more comfortable they get handing over the next thing.

That's the trade-off at the centre of the Mills Review: real appetite to delegate, provided the infrastructure keeps customers in charge at every step. Building that infrastructure is only half the job. The other half is who's trusted to run it, and what happens if the wrong people end up controlling it.