An AI assistant will never tell you to be careful. It gives you a number, formats it nicely, and moves on. The judgement about whether that number is safe to file is left entirely to you — the person least equipped to make it.

So here's the checklist I use as an accountant, translated for anyone using AI to help with tax. Run an AI answer through these five questions before you act on it. If it fails any of them, don't file — verify first.

1. Does it cite a source you can check?

A trustworthy tax answer points to where the rule comes from — a specific piece of legislation, a tax-authority page, a dated rate table. If the AI just asserts a number with no source, treat it as a guess, because that's what it is.

Green flag: the answer names the rule and the year. Red flag: confident numbers, zero references.

2. Is it using this year's figures?

Tax thresholds, bands, and reliefs change every year. AI models are trained on a snapshot of the past, so their default is often to quote last year's numbers — or an average of several years — without realising it.

Ask directly: "What tax year are these figures for, and are they the current ones?" If it can't tell you the year, or names the wrong one, stop.

3. Is it answering for your actual country and situation?

Tax logic that's correct in one country is wrong in the next. The same is true within a country — employed vs self-employed, resident vs non-resident, registered for VAT or not. AI frequently blends rules from different jurisdictions or assumes the most common case.

Make it state its assumptions: "What country and what taxpayer type are you assuming?" If the assumptions don't match you, the answer doesn't either.

4. Did it show its working — and does the working hold up?

A final number with no steps is impossible to trust or correct. A good answer walks through the computation so you (or an accountant) can follow the logic and spot where it breaks. Watch especially for tapers, bands, and reliefs that interact — that's where confident answers quietly go wrong.

5. Has a qualified human stood behind the rules it used?

This is the one that actually settles it. Everything above tells you whether an answer looks careful. Only one thing tells you whether it is: did a real, credentialed accountant review the underlying rules and sign off on them?

That's the difference between an AI recalling tax from memory and an AI applying tax that a professional has verified. It's also the only check a non-accountant can't realistically perform alone — which is exactly why it matters most.

The shortcut

You can run all five of these manually, every time. Or you can remove the guesswork from steps 1–4 entirely by pointing your AI agent at rules that are already sourced, current, jurisdiction-specific, and verified by a practising accountant.

That's what the open library of verified tax skills is for: instead of hoping your AI happens to be right, you give it a source that is right. And when a return is high-stakes enough that you want a human signature on it before you file, an accountant can review the work directly.

AI won't tell you when to be careful. Now you have the checklist that does.