What's the one task you keep hoping AI will sort — but it never quite gets right? Odds are it's not the tool, it's the ask. Hit reply and tell us the job you want AI to nail — send us the prompt you're using and we'll show you how we'd rework it.

Every few weeks, someone in the profession tells me AI doesn't work for accountants. They tried it once, it handed them a confident, completely wrong answer on a benefit-in-kind question, and they filed the whole thing under hype. I get the reaction. I still think it's wrong.

Because I've watched the opposite happen at my own desk. A client rang mid-afternoon last week asking whether to take the next slice of profit as salary or dividend. Eighteen months ago, that's a "leave it with me." Last week the comparison — with workings — was on my screen before we'd finished the call. Same tool the sceptics gave up on. Completely different result.

The difference wasn't a secret model or a clever hack. It was the ask. The people getting burned type "what's the benefit in kind on this car" and trust whatever comes back. The people getting value tell it exactly who it is, what the situation is, and what "done" looks like — then check the workings before anything reaches a client.

That isn't prompt engineering. It's briefing. It's the same thing you do when you hand a job to a good junior: the vaguer the instruction, the worse the work comes back. Nobody blames the junior for that. For some reason, we blame the software.

So here's the line I'll happily defend: AI isn't going to replace accountants. But the ones who can brief it properly are quietly pulling away from the ones who tried it once and wrote it off — and the gap widens every month.

The good news? Briefing it properly isn't a talent. It's a formula — we call it the Reconciled Prompt Formula — and it takes about a minute to learn. Three ingredients: Context, Structure, Output. Miss one and you get generic mush. Include all three and you get something you'd actually put your name near. That's this whole issue: the formula, a before-and-after, and a template you can paste straight in.

Best way to feel the difference is to watch the same task done two ways. Take a real one: emailing a client their year-end tax position.

The lazy version

No context, no format, no idea who's reading it. You'll get a shapeless paragraph you have to rewrite anyway.

The same task, done properly

Same tool. Thirty seconds more typing. A draft you'd barely touch. The difference is the three ingredients:

Your turn: pick one prompt you use that gives you mediocre results. Rewrite it with all three. Run both side by side. Keep the winner — and drop it somewhere your team can reuse it.

Good question — and yes, more than ever, just differently. The models have got better at guessing what you meant, so you no longer need fiddly "prompt engineering" tricks. But guessing is the problem. On a friendly email, a good guess is fine. On a tax figure, a benefit-in-kind calculation, or anything you'll put in front of a client, you don't want it guessing at all.

Context, Structure and Output aren't tricks — they're just clear instructions. That never goes out of date. Think of it less as learning to code and more as learning to delegate well: the clearer the brief, the less you have to fix. As the work you hand over gets bigger, the brief matters more, not less.

Keep this somewhere handy. Fill the brackets and you've hit all three ingredients without thinking about it:

That last line does two jobs: it makes the answer checkable, and it turns the AI from an oracle you have to trust into a junior whose working you can review. Use it once on a real task today. You won't go back to one-liners.

Take one prompt that gives you mediocre results, rewrite it with Context + Structure + Output, and run the two side by side. Then hit reply and let us know how you got on — we read every one, and the best rewrites get featured in a future issue.

You've fixed one prompt. Next week we turn that into your firm's prompt library — write your handful of most common jobs once, and everything client-facing starts at 80% done.