The AI did it" is not a defence. It never will be. So before AI touches another client file, it's worth knowing exactly where the buck stops when it's wrong

Here's a question worth sitting with. If AI drafts a set of accounts, a number's wrong, it sails through unnoticed, and the client acts on it — whose problem is that?

The answer clients and courts will give is simple: yours. Your name is on the accounts. It always has been. AI doesn't change who's responsible any more than a calculator or a junior does. "The software got it wrong" has never protected anyone, and it won't start now.

What has changed is that most firms have quietly let AI into serious work — first-draft bookkeeping, file reviews, tax workings — without touching the three things that actually decide who carries the can when it goes wrong. Their engagement letter, their PI cover and their review trail all still describe a firm that doesn't use AI.

Take them in turn. The engagement letter should say AI may be used and that a human reviews the output — so the client knows, and your position is clear. Your PI insurer should understand how you're using it; the last thing you want is to find a claim turns on a disclosure you never made. And the review trail is the one that genuinely saves you: a documented human check, with an owner, on any AI-assisted work. That trail is your evidence that a qualified person stood behind the number — which is the whole of your defence.

I don't raise this to scare anyone off. We use AI on real work every day. The point is the opposite: once these three are sorted, you can use it with confidence, because you know exactly where you stand if something slips. It's the firms using AI with none of this in place that are exposed — and most of them have no idea.

The mental model that keeps us straight: treat AI exactly like a junior. A junior drafts, a senior reviews, the partner signs — and "the junior did it" was never going to wash. Same chain, same accountability. Document the human check the way you'd document reviewing a junior's file — the Reconciled Review Method — and you're on solid ground.

The signature is the product. Not the draft, not the tool — the signature. Protect that, and you've protected the firm.

Spend an hour this week checking the three things that carry the risk:

None of this takes long. All of it is far cheaper to sort now than to explain later.

Clients: yes — through the engagement letter. It's transparent and it protects you. Your insurer: talk to them or your broker, and don't assume your cover is unaffected. Insurers are actively updating how they treat AI-assisted work, and you want any answer in writing rather than discovered at claim time. It's a short conversation now against a very bad surprise later.

A starting point to take to your own reviewer — never straight to a client:

It gets you a solid draft in seconds. The wording that actually goes in front of a client gets signed off by your professional body's template or your solicitor first.

Pull your engagement letter, check your review trail is actually recorded, and put a note in your diary to ask your PI broker about AI. Then hit reply and let us know how you got on — and whether any of the three gave you a nasty surprise.

That's Safe Hands done — the guardrails are up. Next week we start the run we're most excited about: Let Cowork Do the Books, opening with the big one — could Claude actually replace your bookkeeping software?