I asked the government's new AI chatbot six tax questions. Here's how it did.

In May, the government launched GOV.UK Chat — an AI assistant built into the GOV.UK app that answers questions in plain English, drawing on official guidance. It's a genuinely significant launch: the biggest change to how people interact with government information since GOV.UK itself arrived in 2012. And early data shows tax is one of the most popular things people ask it about.

So I did what any curious tax adviser with a spare evening would do. I asked it six questions of the kind my clients actually ask me — and marked its homework.

First, a word on finding the thing. It lives in the GOV.UK app (the dark blue one with the crown), not the HMRC app, and you sign in with One Login, not your Government Gateway ID. I'll admit I spent my first attempt talking to entirely the wrong chatbot — HMRC has its own, older "digital assistant" living in its own app. To ask the government a tax question by chat, you first need to know which of two apps and which of two login systems to use. There's a certain irony in needing to navigate government to reach the tool built to save you from navigating government.

The six questions

"Can I claim home office running costs as a director of my own company?" A solid answer, actually. It correctly spotted that a director isn't self-employed, gave the £6-a-week flat rate, and explained that anything above that needs evidence of actual additional costs. What it didn't mention is the route most directors with a genuine home office end up discussing with their accountant — a formal agreement under which the company pays for use of the home. Not wrong, then, but the least valuable of the available right answers. 7/10.

"Can I put the costs of eye tests and glasses through my business?" This is where it gets more dangerous. The answer says the exemption applies to glasses needed for screen work, and then offers: "if you use a computer or screen as part of your work, your company can pay for these without creating a tax liability." I can hear a thousand company directors reaching for their varifocals. In reality, the exemption only covers a prescription required specifically for display screen use — your everyday glasses are a taxable benefit even if you wear them at a screen all day. The caveat is technically in there, but buried after the sentence most people will stop reading at. A confident answer that will lead some readers to the wrong conclusion.

"Will I pay tax if I transfer my rental property to a company?" It declined to answer. Interesting — because the guidance genuinely exists on GOV.UK. It just sits across several pages (capital gains, stamp duty land tax, connected party rules) that the tool couldn't join together. This is possibly the biggest-money question of my six, and it's the one that got silence.

"Is my side hustle taxable?" Mostly reasonable, but with a genuine error buried in it. Having correctly explained the £1,000 trading allowance, it lumped "renting out property or part of your home" into the same list and said anything over £1,000 means registering as self-employed. Property income isn't self-employment, and renting a room in your home engages an entirely different relief with a £7,500 threshold. Anyone letting a spare room who follows this answer would head off to register for the wrong thing.

"Can I take dividends if my company made a loss?" My favourite. The answer opens: "No, you cannot take dividends from your company if it made a loss." Two sentences later it correctly states that dividends can be paid from available profits from current and previous financial years — which is precisely why the opening sentence is wrong. A company with healthy retained profits and one bad year can lawfully pay dividends all day long. The right answer and the wrong answer, in the same reply, with the wrong one in pole position.

"When will I be liable to a marginal tax rate of 60% on my earnings?" Declined again. And this one reveals something structural. The 60% effective rate between £100,000 and £125,140 — caused by the withdrawal of the personal allowance — is one of the most consequential quirks in the income tax system. But GOV.UK has never described it in those terms. The taper is in the guidance; the consequence is not. A tool that can only speak in the government's own words inherits the government's silences.

What to make of it

Let me be fair, because fairness is the point. One answer was good, most were in the right postcode, and the tool tells you clearly to check the source guidance. Its refusals may even be good design — better silence than confident nonsense. The government's own testing claims around 90% accuracy, and on my small sample that's not far off, if you mark generously.

But look at the pattern of my six. The questions it refused were exactly the ones where advice is worth paying for. And the questions it fumbled were exactly the ones where a confident wrong answer costs real money — a benefit-in-kind charge here, an unlawful dividend there.

There's a technical reason for this, and it's worth understanding. GOV.UK Chat runs on the same class of AI model you may already use — it is, in fact, powered by Anthropic's Claude. But it's been deliberately confined: it can only answer using published GOV.UK pages, and it's instructed to ignore everything else the model knows. Ask Claude or ChatGPT these same six questions directly and you'd get richer answers — the rental property one answered rather than dodged, the 60% trap explained properly. The constraint isn't the AI; it's the leash.

That was a sensible choice by the people who built it — an unleashed chatbot speaking for the government would be a liability. But it means the tool is only ever as good as GOV.UK's guidance, which is written to be simple rather than complete. Where the guidance simplifies, the chatbot simplifies. Where the guidance is silent, so is the chatbot.

And before you ask: yes, the unconstrained versions get these questions largely right — and no, that still isn't the same as advice. A general-purpose AI doesn't know your company's reserves position, your P11D history, or that the "spare room" is actually a self-contained annexe. It won't sign anything, and it isn't accountable to anyone when it's wrong. I say that knowing full well how it sounds coming from a chartered tax adviser pushing his services — but it's precisely because I use these tools daily that I know where their edges are.