Somewhere right now, a freelancer is asking ChatGPT what they can deduct. A small landlord is having Claude compute their rental income tax. A sole trader is copy-pasting their bank statements into an AI tool and asking it to produce a tax summary they can file. They are not doing this because they are reckless. They are doing it because the AI sounds authoritative, the output looks professional, and hiring an accountant feels like an unnecessary expense for what seems like a straightforward situation.
None of these people are calling us afterward. They are filing.
This is the part of the AI-in-tax conversation that the profession has largely missed. We spend a lot of time debating whether AI will replace accountants, whether firms should adopt AI tools, whether the technology is ready for professional use. These are reasonable questions. But they all assume the accountant is still in the room. Increasingly, they are not.
What AI actually does when it computes tax
I spent the past year reviewing AI-generated tax workings across multiple jurisdictions. The errors I found were not obvious. They did not look like mistakes. They looked like considered professional work, with clear reasoning attached.
A general-purpose AI model has no reliable mechanism for knowing that a VAT rate changed on a specific date, that a social security threshold was revised for the current tax year, or that a deduction which existed two years ago was quietly removed from legislation. The model was trained on a large dataset that includes tax information, but that information has a cutoff date and no update mechanism. The model does not know what it does not know. It will apply a rate it learned during training with the same confidence whether that rate is current or three years out of date.
What makes this particularly difficult is the presentation. The output is well-formatted, the terminology is correct, and the reasoning sounds coherent. A non-accountant reading it has no basis to question it. They asked a question, they got a detailed answer, the numbers add up. They file.
An accountant would catch most of these errors in minutes. But the accountant was never asked.
The scale of people this affects
The population doing this is not small. In the US alone, tens of millions of self-employed individuals and small business owners file their own returns. Many already used TurboTax or similar tools, which at least maintained current tax rules through a professional update process. AI tools have no such constraint. Anyone can ask any AI anything, get a confident answer, and act on it. There is no guardrail, no audit trail, no one checking the output before it goes to the IRS.
The same pattern plays out across every developed jurisdiction. Freelancers in the UK computing their self-assessment. Sole traders in Australia working out their GST. Self-employed workers across Europe trying to figure out what they owe. AI is accessible, free or cheap, available in any language, and answers instantly. For someone who has never found the tax system easy to navigate, it feels like a solution.
It is a solution. It is just sometimes wrong in ways they cannot detect.
When we do eventually see it
The accountant usually enters the picture at one of three points. An inquiry arrives from the tax authority and the person finally hires someone. They realise they may have underpaid and come looking for advice. Or they did hire an accountant — but only for a one-off review of something they had already prepared, and the accountant treated it as professional work product rather than unverified input from a source that confidently hallucinates.
That last scenario deserves the closest attention. When someone presents polished, well-reasoned workings, the natural response is to review them as you would review work from a colleague. The shift required — to treat fluent, confident output as potentially wrong at the level of specific figures — is not automatic. It needs to be deliberate.
Any work that may have been AI-assisted should be treated as unverified at the figure level until the legislative basis has been confirmed. Not because AI is always wrong, but because when it is wrong, it does not signal the error. It explains the error, calmly and in detail, and moves on.
What the profession can actually do
Telling people not to use AI for tax is not a strategy. They already are, and the tools are only getting more capable. The question is what shape the profession’s response takes.
Part of the answer is practical: engagement letters should reflect the different scope involved in reviewing AI-prepared material versus preparing from source documents. The review process should explicitly check current legislation rather than assume the inputs are correct.
The deeper answer is about who controls the knowledge AI uses. Right now, general-purpose models compute tax by drawing on whatever they absorbed during training. Structured, jurisdiction-specific rule sets — maintained by licensed practitioners and updated when legislation changes — give AI computation a reliable foundation instead of an improvised one. Projects like Open Accountants are building this kind of open-source tax knowledge infrastructure. Whether it becomes standard will depend on whether qualified practitioners engage with it, or leave the field to people with no professional accountability for the output.
The question worth sitting with
Tax compliance has always had people who did it themselves, made mistakes, and either got away with it or eventually faced the consequences. What has changed is the confidence with which those mistakes are now made. A person who sat down with a paper form and got something wrong usually had some sense that they might be getting it wrong. A person who received a detailed, professional-sounding AI answer does not have that uncertainty. They think they got it right.
The errors are not louder than before. They are quieter. And they are being made at scale, by people who have no reason to think they need us.
Michael Cutajar is an ACCA-qualified accountant and CPA based in Malta. He is the founder of Open Accountants, an open-source tax knowledge library, and Accora, an accounting infrastructure company.
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
