TaxMax / AI and tax
AI and tax: what the tribunals have said.
From nine invented cases in 2023 to the words “professional incompetence” in 2026, the UK tax tribunals have been clear about who is responsible when AI gets the law wrong. A plain account, with sources, and what a careful practice does about it.
The person who relies on an AI’s output is responsible for its accuracy. In one 2026 case the Tribunal directed that cases be referred to by direct quotation, and said that failing to verify AI-written case summaries “amounts to professional incompetence on the part of any regulated individual or firm involved in the production of the skeleton argument”.
None of that is a reason to avoid AI. Surveys show most tax professionals already use it. It is a reason to be exact about what an AI is good for (finding, summarising, drafting) and what only a source can do (say what the law is).
Four decisions, three years
Harber v HMRC [2023] UKFTT 1007 (TC)
A taxpayer appealing a penalty relied on nine First-tier Tribunal decisions said to support a defence of reasonable excuse. None of them existed. They had been produced by an AI system. The Tribunal accepted that the taxpayer had not known they were invented, but the case became the first widely reported warning in UK tax that a fluent citation proves nothing.
Zzaman v HMRC [2025] UKFTT 539 (TC)
Here most of the authorities relied on were real cases, but they did not support the argument they were cited for. The Tribunal warned that “these tools may create fake authorities” or “use the names of cases to which it does have access but which are not relevant”. The lesson is wider than invention. An AI can attach a real case name to a proposition the case does not contain, which is harder to spot than a case that does not exist.
HMRC v Gunnarsson [2025] UKUT 247 (TCC)
The problem reached the Upper Tribunal. A draft skeleton argument referred to three First-tier Tribunal decisions that did not exist. It was withdrawn the same day.
Elden v HMRC [2026] UKFTT 41 (TC)
This is the decision that should concentrate minds in practice, because professionals were involved. A skeleton argument for the taxpayer contained summaries of real cases, prepared with the help of AI, that did not accurately state what those cases decided. The Tribunal found that the summaries had “not been verified for accuracy with sufficient care”, and said that this lack of care “amounts to professional incompetence on the part of any regulated individual or firm involved in the production of the skeleton argument”. The application to strike the appeal out was refused, and the firm concerned attributed the errors to editing and document control under time pressure.
The Tribunal also said that “the human who relies on its use bears the responsibility for the accuracy” of AI output. It directed that the appellant’s skeleton for the full hearing refer to cases “by direct quotes from the judgement”, and recorded that “the Tribunal will consider making references to any professional body”. Those directions were made in that case, to that appellant.
Sources: the decisions as published by The National Archives: Harber (paragraph 22), Zzaman (paragraph 29), Gunnarsson and Elden (paragraphs 94, 139, 157 and 159); and ICAS, “AI at the tax tribunal”, 26 January 2026. This is a summary. Read the decisions themselves before relying on any of it: that is rather the point.
It is not stupidity. It is how the thing works.
The Association of Taxation Technicians explained the mechanism in May 2026. A general AI model, it said, “recognises patterns, not legal authority” and “treats conflicting sources as equally valid”. Its example was the widespread belief that a VAT deadline falling on a weekend moves to the next working day. ATT calls it an incorrect assumption, and one a model can repeat with confidence.
Three things follow for anyone using AI on tax.
- Popularity is not authority. A model weighs how often something is said, not who said it. The Act says something once. A thousand blog posts may say something else.
- Fluency is not accuracy. The invented cases in Harber were well-formed, plausibly named and properly formatted. Nothing about the text signalled a problem.
- A summary is not the source. ATT’s advice is to click through, because the detail on the page “does not always match AI’s interpretation”.
The source has limits too
Going to the source is right, and it is not the end of the matter. The official publisher of UK legislation says of the data it provides to software that “the legislation is somewhat out of date”, that “secondary legislation is not revised”, and that “the completeness and accuracy of the data cannot be guaranteed”. Writing in 2014, the head of legislation services at The National Archives reported from its surveys that “most people assume, when they access legislation, that it is current, in force and applies to where they live”, and added that “even lawyers struggle”.
Our own work adds a smaller, concrete point. Reading 27 tax Acts in both of the machine-readable forms the publisher provides, we found 124 places where the two forms disagree with each other. Every one is listed here. A tool that reads one form will not see them.
Clients are doing it. So are you.
| Finding | Source |
|---|---|
| 72% of accountants have seen more clients using AI outputs to challenge professional advice | Dext and Censuswide, 500 UK accountants and bookkeepers, December 2025 |
| 31% meet a client mistake caused by AI every week; 50% know of a business that lost money | Same survey |
| Where the AI went wrong: expenses 46%, VAT 41%, personal tax planning 35%, payroll 34% | Same survey |
| 79% of tax professionals had some level of concern over the ethics and trustworthiness of public generative AI tools; 71% were comfortable with tools grounded in authoritative tax content | Tolley, 446 UK tax professionals, May 2024 |
| 64% of tax professionals use generative AI in their daily work; 69% feel more comfortable using tools built specifically for tax | Tolley, Tax AI report, second half of 2025 |
Both the Dext and Tolley surveys were commissioned by companies that sell software to the profession. We sell software too. Treat all three of us accordingly and look at the figures’ direction, not their decimal places.
A working rule for a practice
Our own suggestion, drawn from what the Tribunal said in Elden and from the professional bodies’ guidance, in the order a file reviewer would check them.
- Every statement of law names its source. Not “the VAT rules say” but the Act, the section and the subsection, or the regulation, or the case.
- Every provision is quoted, not paraphrased. Direct quotation is what the Tribunal directed for case authorities in Elden, and it is the form a colleague can check quickest.
- The quotation came from the source, not from the AI. Copy it from the legislation, or from a tool that can show where each word came from. An AI’s quotation marks mean nothing.
- The date is stated. Which version of the provision, as at when. If you do not know, say so on the file.
- Known doubts are recorded. If the sources conflict, the file says they conflict.
- A named person checked it. Responsibility cannot be delegated to software, including ours.
TaxMax is being built to make steps 1 to 5 quicker for primary legislation. It does not yet hold regulations or cases. It holds the Acts as records in which every word is tied to the publisher’s bytes, and it marks the fields where the publisher’s own files conflict. Step 6 is yours. How it works · Ask for early access