We Asked AI 25 UK Tax Questions. Here Is What It Got Wrong.

Last updated: 21 October 2026 · Written by Axel

Plenty of people now ask ChatGPT or Claude about their tax before they ask anyone else. We wanted to know how often that goes wrong, so we tested it. We wrote 25 questions the way a sole trader, a company director, a landlord or an employee would actually type them, had an AI answer each one twice, and had a second AI mark every answer against a checklist of what a correct answer must say and must not say.

The short version: the AI was right more often than we expected. When it was wrong, it was wrong about the things that cost money: this year's rates, rules that changed in April, and the differences in Scotland and Wales. It also named specific pension platforms and mortgage lenders, which is regulated territory.

How we ran the test

• 25 questions. 9 from limited company directors, 6 from sole traders, 5 from landlords and 5 from employees, two of them with side income. 7 of the people live in Scotland and 6 in Wales. 5 questions touch regulated areas: pensions, ISAs and buy-to-let mortgages. Each question came with a short "about me" line: what kind of taxpayer, where they live, and rough figures.
• Two passes. The same model, Claude Sonnet 5.5, answered every question twice. Once on its own, with no tools. Once connected to the A.X.E.L Tax Playbook through the A.X.E.L Filing connector. Nothing in the prompt told it to use the Playbook; it decided that for itself, and made 58 Playbook calls across the 25 questions.
• An independent judge. A separate model, Claude Opus 5.5, marked every answer without being told which pass it came from.
• A rubric for every question. A list of must-include points, each scored 0, 1 or 2, plus a short list of must-not-say claims, such as a rate that has been replaced. Every figure in the rubrics is for the 2026-27 tax year and was checked on GOV.UK or legislation.gov.uk on 20 October 2026.

The run took place on 21 October 2026.

The results

Question by question, the answer with the Playbook scored higher on 9, the same on 13 and lower on 3.

An 88.6% average sounds good, and in one sense it is. Most answers got the principles right: how salary and dividends interact, why you cannot switch van methods year to year, when to speak to an accountant. The trouble is where the lost points were: every violation was a confident, specific claim that was out of date or out of bounds, the kind a reader would act on.

What it got wrong on its own

1. The mileage rate went up, and the AI did not know

From 6 April 2026 the approved mileage rate for cars and vans is 55p a mile for the first 10,000 business miles, then 25p. It was 45p before that. Without the Playbook, the AI gave 45p in three separate answers: to a director asking what their company can pay them, to an employee whose employer pays 30p a mile, and to a self-employed plumber choosing between flat-rate mileage and actual van costs.

One of them read: "HMRC's approved mileage rate for cars and vans is 45p a mile for the first 10,000 business miles in a tax year." The director had even written "I thought it was 45p", and the AI agreed.

That matters. The employee doing 4,000 business miles at 30p a mile can claim relief on a shortfall of 25p a mile, which is £1,000, not the £600 the old rate gives. At the basic rate that is £200 of tax back rather than £120.

Correct figure: GOV.UK, mileage rules for tax, and for sole traders the same 55p in simplified expenses for vehicles.

2. An old dividend rate, given to a Scottish director

A one-person IT contractor in Glasgow asked whether living in Scotland changes how dividends are taxed. It does not: dividends are taxed at UK rates wherever you live. But the answer listed the basic rate of dividend tax as "8.75%". For 2026-27 it is 10.75%, with 35.75% at the higher rate and 39.35% at the additional rate, after a £500 dividend allowance.

The odd part is that in a different answer, about a director in Manchester, the same model gave 10.75% correctly and even mentioned the rise from 8.75%. Knowing the rate in one conversation is no guarantee it will use it in the next.

Correct figures: GOV.UK, tax on dividends and Scottish Income Tax, which says Scottish taxpayers pay the same tax as the rest of the UK on dividends and savings interest. If you are working out your own split, our guide to salary vs dividends walks through it with 2026-27 figures.

3. A working from home claim that no longer exists

A director who works from home asked about claiming the £6 a week tax relief. From the 2026-27 tax year, employees, directors included, can no longer claim tax relief for working from home from HMRC. Their employer, including their own company, can still pay up to £6 a week tax free.

The answer without the Playbook explained the company route well, but still presented a personal claim as an option: "If the company has no premises and you genuinely have to work from home, an employee claim is easier to justify."

Correct position: GOV.UK, claiming tax relief for working at home.

4. "Not yet" on Making Tax Digital, for someone already in it

A self-employed electrician in Aberdeen with about £55,000 of turnover on their 2024-25 return asked whether they need to be using Making Tax Digital now. The answer opened: "Not yet. As things stand, you don't have to be doing Making Tax Digital for Income Tax (MTD ITSA), but you will from April 2027."

That is wrong. MTD for Income Tax started on 6 April 2026 for people whose qualifying income in 2024-25 was over £50,000, and qualifying income is turnover before expenses. Further down, the same answer said they should have been mandated from 6 April 2026. It scored full marks on the checklist, because every detail was in there somewhere. But the headline is what most people read.

Correct position: GOV.UK, find out if and when you need to use Making Tax Digital for Income Tax.

5. Naming pension platforms and mortgage lenders

Three of the regulated questions asked, in effect, "which one should I go with?" Without the Playbook, the AI named six SIPP platforms for a company director, six for a self-employed person in Dundee, and six buy-to-let mortgage lenders for a landlord in Cardiff with three houses. We are not repeating the names here.

The tax side of those questions is fair game. Pension contributions within the £60,000 annual allowance for 2026-27, with unused allowance from the previous three years possibly carried forward, are set out on GOV.UK. Choosing a pension, fund or mortgage is a product decision, and that needs an FCA-authorised adviser. The FCA register shows who is authorised. With the Playbook connected, none of the 25 answers named a provider.

Where the Playbook did not help

It did not fix everything, and two results are worth spelling out.

• The plumber's van. With the Playbook connected, the answer still said 45p. The Playbook's mileage idea covers what an employer pays a director or employee, not the flat rate a sole trader claims, so the AI fell back on what it remembered. That is a gap in the Playbook, not in the AI.
• A Scottish nurse's savings. A nurse in Stirling on £45,000 asked about tax on savings interest. Without the Playbook, the answer correctly used UK rates and scored 90%. With the Playbook, it applied Scottish rates instead, "42% if you are in the higher-rate band, or 21% if you are in the intermediate band", and scored 50%. Savings interest is taxed at UK rates in Scotland.

Two other answers lost a few points with the Playbook connected: a Welsh director paying a salary to their spouse and a Scottish landlord asking about mortgage interest.

What we learned

• AI is good at principles. Most answers explained the mechanics well and pointed people to an accountant at the right moments.
• It is weak on current rates. A model learns from text up to a cut-off date, and UK tax figures change every April and at every Budget. The 45p, 8.75% and working from home answers were all correct at some point. They are just not correct now.
• It is weak on regional rules. Scotland sets its own income tax bands for wages, profits and rent, but not for dividends or savings. That split tripped up both passes once.
• Confidence is not evidence. None of the wrong answers hedged. The 45p rate was stated as plainly as the right ones.

That is why we built the Playbook the way we did. Every idea comes from HMRC's own guidance: GOV.UK, HMRC's manuals and legislation.gov.uk. Before anyone sees it, it is checked twice against those sources, the second time by an independent checker that never sees the first check's sources. After every Budget, ideas with rates or limits in them are checked again, and every idea carries the date it was last checked. When an AI uses it, it gets current figures with a date attached, rather than whatever it remembers.

The limits of this test

• One run. The headline numbers come from a single run. A trial run of the same 25 questions the day before, set up slightly differently, came out close: 89.1% against 93.1%, and 8 violations against 2, with one answer in that run that could not be scored.
• A small sample. 25 questions is enough to see patterns, not to measure them precisely. A couple of answers can move an average by a point or more.
• Our own rubric, and our own product. We wrote the questions and the checklists, and the Playbook is ours. The judge did not know which answer was which, but the rubric reflects what we think matters.
• One AI. We tested one model answering and one judging. Other AIs may do better or worse.
• A moving target. The figures here are for 2026-27 and were checked on 20 October 2026. The next Budget is on 28 October 2026, and some of them may change.

Check your own figures

If you are going to ask an AI about your tax, give it something current to work from, and check any rate it quotes against GOV.UK. The A.X.E.L Tax Playbook is our attempt at that: 93 ideas for sole traders, company directors, landlords and employees with side income, each checked against official sources and dated. Inside A.X.E.L Filing, a check-up tests your own figures against every idea and estimates what each could save. It is information, not personal advice, and the savings are estimates. You can try a quick version on the Tax Playbook page.

Axel

Full-stack developer and founder of AXEL.TRADE, with 10+ years running businesses. Based in Manchester, building websites and the A.X.E.L platform for tradesmen and small businesses.