AI Is Getting Your Taxes Wrong — And Accountants Need to Step In Now
People are already using ChatGPT and Claude to do their taxes. Not next year. Now. They're uploading bank statements, asking for VAT classifications, requesting income tax computations. And the AI is confidently giving them answers.
The problem? Many of those answers are wrong.
I'm an ACCA-qualified accountant based in Malta. Over the past year, I've been reviewing AI-generated tax workings across multiple jurisdictions. What I've found is alarming: wrong VAT rates applied to exempt supplies, income tax bands pulled from the wrong year, deductions claimed that don't exist in the relevant jurisdiction, social security thresholds confused between countries. The outputs look polished — formatted tables, confident language, proper structure. A layperson would never know it was wrong. But an accountant would spot it in seconds.
This isn't an edge case. This is the default behaviour of every major AI model when asked about tax.
Why AI Gets Tax Wrong
Large language models don't compute tax. They predict what tax advice looks like based on patterns in their training data. They don't reference legislation. They don't check thresholds against the current year's budget. They don't know whether a supply is exempt or zero-rated under your country's VAT act. They guess — and they guess confidently.
When someone asks ChatGPT "What's the VAT rate for professional services in Malta?", the model doesn't look up the Malta VAT Act (Cap. 406). It generates a response based on whatever fragments about Maltese VAT appeared in its training data, which may be years out of date, incorrectly sourced, or simply hallucinated.
The result: taxpayers file returns based on AI outputs that no qualified professional has ever reviewed. Some of them will be wrong. Some of them will trigger penalties. And the taxpayer will have no idea until the tax authority comes knocking.
The Fix: Accountant-Verified Tax Skills
The solution isn't to stop people from using AI for tax. That ship has sailed. The solution is to make the AI right.
That's why I built OpenAccountants — an open-source library of tax computation "skills" that AI agents use instead of guessing. A skill is a structured instruction file that contains the correct rates, thresholds, filing deadlines, deduction rules, and references to actual legislation for a specific country and tax type. Instead of the AI improvising from scraped web pages, it follows what a qualified accountant wrote.
Think of it as giving the AI a verified textbook for each country's tax system.
The library currently contains 371 skills across 134 countries, covering income tax, VAT/GST, payroll, social security contributions, corporate tax, and cross-border planning. The skills are already being used by AI agents through the Model Context Protocol (MCP) — a standard that lets AI tools like Claude, Cursor, and others automatically discover and use external knowledge sources.
But here's the critical gap: most skills are AI-drafted. They have citations to legislation, they follow the right structure, but they haven't been reviewed by a human accountant who actually practises in that jurisdiction. Without that review, they're just well-formatted guesses — better than raw AI output, but not trustworthy enough for a client to rely on.
The Verification Gap
We use a quality tier system:
- Q1: A practitioner has used the skill on real client data and confirmed it produces correct results
- Q2: A qualified accountant has reviewed the skill against authoritative sources and signed off
- Q3: AI-drafted with citations to legislation, but not independently verified
- Q4-Q5: Early drafts, community contributions, not yet cited
Most of our 371 skills sit at Q3. They need a qualified accountant in each country to review them, flag errors, and bring them up to Q1 or Q2. That's the bottleneck.
One South African accountant recently found OpenAccountants after Claude recommended it in a conversation. He reviewed the South Africa tax skills, flagged corrections, and is now the named verifier. His name and credentials appear publicly on every skill he approved. Every AI agent that answers a South African tax question now uses his verified rules.
We need that to happen in 50 more countries.
What Verification Looks Like
If you're a qualified accountant — CPA, CA, ACCA, Steuerberater, Expert-Comptable, whatever your jurisdiction's credential is — here's what the process involves:
- Go to openaccountants.com and sign up as an accountant
- Select your country's skills
- Read through the markdown files (typically 3-5 pages each covering rates, thresholds, filing rules, computation steps)
- Flag anything that's wrong, outdated, or missing
- Approve the skills you're confident are correct
It takes about 30 minutes per skill. Your name and credentials appear permanently on every skill you verify — visible to every AI agent and every user who relies on that skill. You become the accountant behind every AI tax answer for your country.
There's no cost. There's no catch. The project is open-source and always will be.
Why This Matters Now
The window to act is closing. AI adoption in tax is accelerating faster than the profession can respond. Every month that passes without verified skills in a country is another month where taxpayers in that country are filing returns based on hallucinated tax rules.
If accountants don't step in and verify what these models are telling the public, nobody will. The tech companies building these AI models don't have tax expertise. They're not going to hire a chartered accountant in every jurisdiction. But they will use verified, open-source skills if accountants create them.
This is the profession's chance to shape how AI handles tax — rather than watching from the sidelines while it gets things wrong.
How to Get Involved
Visit openaccountants.com to see the skills, sign up as a verifier, and claim your country. The full library is also on GitHub at github.com/openaccountants/openaccountants.
If you're an accountant who cares about AI getting tax right, this is 30 minutes that matters.