New run this week: Safe Hands — how to use AI without the blow-up. It starts with the one decision most firms skip: deciding what AI is actually allowed to touch, before you let it near anything.

The most useful thing we ever did with AI wasn't a clever prompt. It was drawing a line.

Early on, the debate in our office was the wrong one: "is AI good enough to do this?" Good enough compared to what? The honest answer is that on some jobs it's already sharper than a tired human at 5pm, and on others it should never be within a mile of the final call. Arguing about the tool in the abstract gets you nowhere.

So we stopped asking "can it" and started asking "should it lead." We wrote every recurring task in the firm on a list and gave each one a colour — what we now call the Reconciled Traffic-Light System.

Green — AI takes the first pass: drafting a client email, summarising a meeting, a first cut at a bank feed, research we then check against source. Low stakes, high volume, easy to verify. Hand it over.

Amber — AI helps, a human signs: a tax computation it's drafted, a file it's had a first look at. Genuinely useful, but a qualified person owns the output and checks it properly before it moves.

Red — AI does not lead, full stop: the final sign-off on a set of accounts, the advice call where a client's genuinely worried, anything where judgement rather than information is the product. That's what clients pay us for. It stays with a human.

What surprised me was how freeing the red list is. Once everyone knows what AI must never lead on, they stop being nervous about the green stuff. The "where does this end" fear disappears, because you've already answered it. The team ends up using AI more, not less, because the guardrails are clear.

The principle underneath it: AI handles volume, you handle judgement. The firms getting this wrong treat it as all-or-nothing — banning it out of fear, or letting it run unchecked out of excitement. The ones getting it right have simply decided, in writing, which jobs sit where.

You don't need a policy document for this yet. You need three coloured columns and an honest hour.

List the tasks your team does most weeks, then give each one a colour. Here's the shape of it:

Hand the green ones to AI this month and leave the red ones exactly where they are. The amber list is next week's subject.

Two questions decide the colour. How costly is it if this goes wrong? And how easily could you spot that it had? Low stakes and easy to check is green — hand it over. High stakes or hard to catch an error is red — keep it human. Everything in between is amber: useful, but a named person owns and verifies it. Notice it's about the task, not the tool. The same model can be green on one job and red on another.

Paste in your task list and let it draft the sort — then argue with it. The disagreements are the useful part:

Treat the output as a first draft of your map, not the map itself. Where you disagree with its colour, that's your firm's judgement — write down why.

Colour-code your team's recurring tasks and hand the green ones to AI. Then hit reply and let us know how you got on — especially what landed in your red column. The lists people draw are fascinating.

You've got an amber list — jobs AI drafts and a human checks. Next week, the habit that makes that check reliable: how to catch the confident, fluent, completely wrong answers AI will never flag itself.