You can already do a meaningful amount of your own tax work with an AI agent. The catch is that the default experience — open a chat, ask a question, trust the number — is exactly the one that gets people penalised.

This guide shows you how to do it properly: get real work out of an AI agent, and make sure the answer is correct before you act on it. It works with whatever assistant you prefer — the principles are the same across all of them.

Step 1: Pick your AI agent

Any capable AI agent can help — ChatGPT, Claude, Gemini, Cursor, and others all work. There's no single "right" tool, and you don't need a tax-specific app. What matters far more than which agent you use is what you let it rely on for the rules. A strong model with bad tax data still gives bad answers.

Step 2: Give it the right rules to follow

This is the step almost everyone skips, and it's the one that decides whether you get a correct answer.

Left alone, an AI agent answers tax questions from memory — an averaged impression of the law from whenever it was trained. That's how you get last year's thresholds and deductions that don't exist where you live.

Instead, connect your agent to a verified tax skill: a structured set of the current rates, thresholds, deductions and computation steps for your country and tax type, written and signed off by a qualified accountant. Your agent reads the rules from a real source instead of guessing. The skills are open and free to use, and the install guide walks through connecting your agent — it takes a few minutes.

Step 3: Give it clean, complete facts

AI is only as good as what you tell it. Before you ask for a number, gather:

Your country (and region/state if relevant) and tax year.
Your taxpayer type — employed, self-employed, landlord, company, VAT-registered or not.
Your income and expenses, organised — totals by category beat a pile of receipts.
Anything unusual — a second job, foreign income, a big one-off sale, a change in circumstances.

Vague inputs produce confident, wrong outputs. Spell out your situation.

Step 4: Ask for the working, not just the answer

Never accept a bare number. Prompt for the full computation:

"Show your full working step by step, state which tax year and country you're assuming, and cite the rule behind each figure."

This does two things: it lets you sanity-check the logic, and it forces the agent to surface its assumptions — which is where most errors hide. If it won't name the year or the rule, that's your signal to stop.

Step 5: Run the five-point safety check

Before you treat any AI tax answer as final, confirm it:

Cites a source you can check.
Uses this year's figures.
Matches your country and situation.
Shows working that holds up.
Relies on rules a qualified human has verified.

If you connected a verified skill in Step 2, the first four are largely handled — the rules are already sourced, current and jurisdiction-specific. (There's a longer walk-through of this checklist here.)

Step 6: Get a human in the loop when it counts

For a simple, low-stakes situation, a well-set-up agent following verified rules will get you a long way. But some returns deserve a professional signature before you file — a complex year, a large bill, anything you'd lose sleep over getting wrong.

That's the point of having an accountant available on top of the AI, not instead of it. You can have a qualified accountant review the work your agent produced — so you get the speed of AI and the assurance of a human who'll stand behind the result.

The short version

Use any agent you like. Give it verified rules instead of letting it guess. Feed it clean facts. Demand the working. Run the safety check. And bring in a human when the stakes are real.

Do that, and AI stops being a liability at tax time and starts being genuinely useful. Set your agent up here.