Yet again there’s a scare story about what AI is doing. Personally, I've been worrying more about what's happening on BBC Radio 4's The Archers.

A few weeks ago it ran a storyline in which an elderly parishioner had Alan, the vicar, do her online banking. Nobody in the script remarked on it. The joke was that the vicar was overworked and couldn’t say no. Then the parishioner died – and I found myself wondering, with no idea what the writers have in store, whether the poor man might end up implicated in some sort of online fraud.

Around eleven million people in the UK do digital tasks on someone else’s behalf – a parent, a partner, a neighbour. Almost all of them do it the only way available: by logging in as that person, with their password.

It breaches the terms and conditions. It can void the fraud protection. And it is so ordinary that a drama which researches its plots put it on air as an everyday errand.

There is no lawful way to help someone you love manage their money online. Not because banks are obstructive, but because nobody has legislated for it.

Health worked this out years ago. NHS proxy access exists: identity verified, consent recorded, a national application service behind it. It’s imperfect – same GP practice only, and it stops at the door of secondary care. But it exists. So this isn’t a hard problem. It’s an unaddressed one.

Here’s what’s changed while we weren’t addressing it.

I spent four years working on consumer digital and AI at a large US healthcare business, including an AI assistant that could book appointments and request prescriptions on a member’s behalf. Useful, popular, and built because it made things easier – not because anyone had passed a law about it.

Agents that act for people are arriving here now. Banks will build them because they are useful. And when they do, delegated access will have been solved commercially before it was solved legally.

That should worry us more than it does. The thing that makes NHS proxy access safe isn’t the technology. It’s the scaffolding around it: verified identity, recorded consent, defined liability, an audit trail. Agent access without that framework isn’t a fix for the eleven million. It’s the same workaround, faster and at scale, with nobody accountable when it goes wrong.

I’m certainly not arguing against agents. I think they could be the largest accessibility advance in a generation, particularly for people who can’t type or navigate a screen. But I am arguing about sequence. We have a decade of evidence that when a technology arrives before its framework, the people who most need protecting are the ones who find out first.

The question isn’t whether machines will act on our behalf. They already do. It’s whether we decide the rules before or after that becomes normal.