Can You Trust Claude or ChatGPT for UK Tax? Mostly No — Here's the Fix

If you're a UK sole trader or freelancer using Claude or ChatGPT for tax questions, you're not alone — and you're not being reckless. These models are genuinely useful for a lot of accounting work. But there's a specific, important category of question where they'll give you a confident, well-formatted, and quietly wrong answer. This post is about that category, and what to do about it.

The Core Problem: AI Has a Stale Rates Problem

Claude's training data has a cutoff. ChatGPT's does too. Neither model is querying HMRC live when you ask a question. They're drawing on whatever tax information was scraped from the web before their cutoff date — and tax thresholds change every year in the Budget.

Here are three specific questions where base Claude will likely give you the wrong number right now:

"What's my personal allowance for 2025/26?" The personal allowance is £12,570. That's been frozen since 2021/22. But a model trained on older data — or one that misread a source — might quote a lower figure, or might hedge by giving you the right number without flagging that it's been frozen (meaning you don't know whether it's current or five years old).

"What's the higher rate Income Tax threshold?" The higher rate kicks in at £50,270 in England, Wales, and Northern Ireland for 2025/26. But this is also a frozen figure, and the Scottish higher rate threshold is different (£43,662 in 2025/26). If you don't specify where you live and the model doesn't ask, you could get the wrong number entirely.

"Do I need to register for VAT?" The VAT registration threshold is £90,000 (from April 2024). Before April 2024 it was £85,000. If the model was trained on pre-April 2024 data, it may still tell you £85,000. A few thousand pounds of difference — and crossing the threshold without knowing it can mean back-charging output VAT you've already spent.

Class 4 NIC rates, the dividend allowance (cut to £500 from 2024/25), the CGT annual exempt amount (cut to £3,000 from 2024/25) — all of these have changed recently, and all of them are the kind of number where a single wrong digit changes someone's tax bill.

What Claude Does Well Without Help

To be clear, I'm not saying avoid AI for tax. I use it myself. These models are excellent at things that are fundamentally about reasoning rather than current data:

Structuring your thinking. "I'm a UK sole trader who also has rental income and a PAYE job — what do I need to declare on my Self Assessment?" Claude will give you a good map of the territory: SA100, SA103F for the self-employment income, the property pages, the employment pages. That structural knowledge doesn't go stale.

Working out if something is in principle allowable. "Is my home office setup allowable as a business expense?" Claude understands the principles — wholly and exclusively for business, simplified expenses vs actual costs, the HMRC guidance framework. The principles don't change much year to year.

Drafting correspondence to HMRC. If you need to write a letter explaining late filing circumstances or querying a tax code, Claude is genuinely excellent at this. It understands the register, the right tone, what HMRC looks for in reasonable excuse submissions.

Explaining concepts. What's the difference between a payment on account and a balancing payment? How does MTD for Income Tax work? What does "cessation of trade" mean for your SA103F? These explanations are largely correct and very useful.

What Claude Gets Wrong Without Help

The dangerous category is anything that requires a current number:

• Income Tax bands and rates — the 20%, 40%, 45% bands, their exact thresholds, the Scottish divergence
• NIC thresholds — Class 2 NICs were abolished from April 2024 (small profits threshold no longer triggers them); Class 4 rates and the Lower and Upper Profits Limits change
• VAT registration threshold — as above
• MTD obligations — when your turnover figure triggers MTD for Income Tax (phase-in is by income threshold, and the schedule has shifted)
• US law bleed-through — this one catches people out. Claude has seen vastly more US tax content than UK content. Ask about deductible business expenses and you may get advice that's correct under the IRS's "ordinary and necessary" test but wrong under the HMRC "wholly and exclusively" test. The standard is subtly different, and the difference matters in edge cases.
• Which SA schedule to use — SA103F (full) vs SA103S (short) depends on your turnover. These rules are simple but the model may confuse them or give you outdated thresholds.

The Fix: Give Claude Verified UK Tax Data

The reason Claude gets the numbers wrong isn't that it's stupid — it's that it's working from stale or incomplete training data. The fix is to give it current, verified data at query time.

OpenAccountants is an open-source library of accountant-verified tax skills. Each skill is a structured file covering the correct rates, thresholds, filing rules, and HMRC source references for a specific tax type. The UK skills cover Income Tax, VAT, PAYE, and Self Assessment workflows — reviewed by ACCA/CPA-qualified accountants.

You connect them to Claude via the MCP (Model Context Protocol), which lets Claude read the skill file before answering. Instead of guessing the higher rate threshold from training data, it reads it from a verified source.

Install it in Claude Desktop by adding this to your claude_desktop_config.json:

{
  "mcpServers": {
    "openaccountants": {
      "command": "npx",
      "args": ["-y", "@openaccountants/mcp"]
    }
  }
}

Or visit openaccountants.com/connect for the full setup guide.

Before and After: A Real Example

Without OpenAccountants — "What's the Income Tax on a £75,000 salary in England?"

You'll get a broadly reasonable answer: something about personal allowance, basic rate, higher rate. But you might not get the precise 2025/26 figures — £12,570 personal allowance, 20% on £12,571–£50,270, 40% on £50,271–£75,000. And you almost certainly won't get the Class 1 NIC calculation layered in, or a note that the figures are from a specific tax year.

With OpenAccountants connected — same question

Claude reads the verified UK Income Tax skill and gives you: the exact 2025/26 bands sourced from HMRC, the personal allowance, the basic rate and higher rate components, the Class 1 NIC (12% on earnings between the Primary Threshold and Upper Earnings Limit, 2% above), and a net take-home figure. It will tell you which tax year it's using, and the skill it read to get there.

The difference isn't that Claude became smarter. It's that it stopped guessing and started reading verified numbers.

When You Still Need a Human Accountant

Even with verified skills, there are situations where you need a qualified person:

• CGT on property — the rates, reliefs (Private Residence Relief, lettings relief), and interaction with your Income Tax band require proper analysis
• IR35 status — off-payroll working rules are judgment calls that depend on the specific contract and working arrangements, not just rules
• SDLT — Stamp Duty Land Tax on property transactions, especially anything involving mixed use, higher rates for additional dwellings, or reliefs
• Anything involving a company — Corporation Tax, director's loans, salary vs dividend optimisation for a Ltd company
• First-time Self Assessment — if this is your first return with a UTR, a one-off session with an accountant is worth it

For these situations, the review service at OpenAccountants lets you share what Claude gave you with a qualified accountant who'll check it and flag anything that needs changing. The AI does the first pass; the accountant checks the work.

Getting Started

Install the OpenAccountants connector free at openaccountants.com/connect. It works with Claude Desktop, Cursor, and any other tool that supports MCP.

If you want a human accountant to review what Claude produced — whether that's a Self Assessment computation, a VAT calculation, or an expenses claim — the review service is built into the same platform.

The goal isn't to replace accountants with AI. It's to make the AI actually right, so that when you do need a human to review it, they're checking real work — not correcting hallucinated thresholds from three Budget cycles ago.