AI can now draft your emails, invoices and offers in seconds — but it still gets things quietly wrong. The newest tools are adding a clever fix: a second AI that reads over the first AI's work and flags the mistakes before you do. Here is what that “AI watches AI” idea means for a Swiss small office, and how you can use it today with the tools you already have.

The mistake nobody catches

Ask an AI to write a payment reminder or a client letter and it will sound confident — even when it is wrong. It might invent a figure, swap two clients, or round a VAT rate badly. Because the wording is smooth, the error hides in plain sight. For a notary, a doctor or a fiduciary, one wrong number in a letter is exactly the kind of small mistake that quietly costs trust.

People miss these slips because they skim. And the same AI that wrote the text will usually miss them too, because it is convinced it is right. That is the problem the newest tools set out to solve.

A second AI as the proofreader

The idea is simple: use one AI to write, and a different one to check. Anthropic recently added a feature called “You Should Know” to its coding assistant Claude Code — a small helper that watches the main assistant's every step and points out what you would otherwise scroll past. It is aimed at programmers, but the principle transfers straight to office work.

In everyday terms, this is the “two pairs of eyes” rule your practice already knows. One person writes the letter, a second reads it before it goes out. Now you can get that same second opinion on anything an AI drafts — without hiring a second person.

How to do it with what you already have

You do not need Claude Code or any new software. The pattern works with two ordinary chats:

  1. Open ChatGPT or Claude and ask it to draft your invoice, email or offer.
  2. Open a second conversation — or a different tool entirely — and paste the draft in.
  3. Say: “You are a careful proofreader. Find any wrong number, wrong name, unclear sentence or claim that is not supported. Do not rewrite anything, just list what is wrong.”

That second pass catches most slips in under a minute. It costs nothing extra and works in German, French and English. Save the instruction once, and it becomes a ten-second habit before anything important goes out.

What it will not fix

An AI reviewer is a safety net, not a guarantee. It can still miss a subtle mistake, and it cannot tell you whether a number is true — only whether it is internally consistent. So the final check is still yours: read the draft once, with the AI's list beside you.

That is the honest, zero-hype version. Two AIs catch far more than one, and neither replaces your own judgement. But for the price of a second paste, it is one of the simplest safeguards a Swiss office can add this month.