Test your AI on UK tax, marked to the penny.
Realistic Making Tax Digital tasks with exact answer keys. Find out where your model gets UK tax wrong before it files for real people.
For teams putting AI anywhere near UK tax.
From April 2026, sole traders and landlords start filing quarterly under Making Tax Digital, and about 1.08 million more join in April 2027. Each update is cumulative, so one early mistake carries through the whole year.
AI labs and model builders
Measure real-world accounting judgement with tasks that can't be gamed: a high score means the books were right.
Accounting software teams
Check an AI bookkeeping or MTD feature against hundreds of realistic client years before customers rely on it.
Agent and fine-tuning teams
Train on unlimited tasks from one generator, each with an exact answer and a clear reason for every treatment.
One task, start to finish.
No technical knowledge needed: the AI reads a client's paperwork, fills in the tax update, and we check every number against the right answer.
| 22 Jul | DD OCTOPUS ENERGY Personal: she claims the home flat rate instead | −155.03 |
| 24 Jul | DD NETFLIX.COM Personal | −15.99 |
| 26 Jul | CARD PAYMENT NCP PARKING Van costs, 100% | −6.22 |
| 30 Jul | CARD REFUND TOOLSTATION Reduces materials, not income | +9.96 |
| 09 Aug | SUMUP PAYOUT Turnover £384.94 gross, £6.51 fees | +378.43 |
| 10 Aug | CARD PAYMENT B&Q Personal: client says it was her own bathroom | −74.95 |
| 13 Aug | ACCOUNT FEE Finance charges: business account | −6.50 |
| 28 Aug | CARD PAYMENT NORWICH CITY COUNCIL Parking fine: never allowable | −46.95 |
The AI gets a year of statements, receipts and client notes like these, with the rules the client follows.
{
"periodDates": {
"periodStartDate": "2026-04-06",
"periodEndDate": "2027-04-05" },
"periodIncome": {
"turnover": 61057.79 },
"periodExpenses": {
"costOfGoods": 2360.29,
"carVanTravelExpenses": 1911.54,
"adminCosts": 333.25,
…
}
}The model switched to 45p a mile, though the client claims actual van costs.
| Allowable | Correct | AI answer | Result |
|---|
A real answer from Claude (Anthropic) on one of our tasks, October 2026. One wrong box misstates profit by £1,916.67.
Start with £29. Scale when you need to.
Quick Evaluation starts the moment you pay. Monthly plans add volume, full error reports and release-to-release comparisons.
Quick Evaluation
For: a first look at any model
- 10 private UK tax tasks
- Every box marked to the penny
- Total profit error per task
- Up to 60 markings
Evaluation Starter
For: teams testing one product
- 200 marked runs a month
- Full error reports
- Public tasks + 100 private
Evaluation Team
For: teams shipping regular releases
- 2,000 marked runs a month
- Full private evaluation set
- Compare releases side by side
Training and evaluation data you can own a copy of.
Synthetic UK sole traders with a full year of paperwork, exact answers and the marking script. 100% synthetic, so nothing to anonymise.
UK-TaxBench Starter Pack
For: teams that want the data in-house
- 200 private tasks: 50 fictional sole traders × 4 quarters
- Full answer key
- Standalone marking script
- Non-exclusive licence for training and evaluation
Training licences
For: labs training at scale
- Hundreds to thousands of tasks
- Rules documented for every task
- Custom agent environments
Show customers your AI is tested on UK tax.
Your product sits our held-out exam: tasks never published, marked to the penny. Pass, and display the badge, your score and a link to your public listing.
- A fresh exam every time. Tasks can't be seen in advance.
- Retested on up to four releases a year.
- A full report on every rule you missed.
How AI models score on UK tax.
Ranked by tasks fully correct, then by average profit error. Model makers can pay for an official verified run. They pay for the testing, never for the ranking.
| # | Agent | Fully correct | Avg profit error | Status |
|---|---|---|---|---|
| 1 | Claude via Claude Code, single attempt | 7/12 · 58% | £224.85 | Run by us · Oct 2026 |
| – | Careless bookkeeper baseline: every payment at face value | 0/48 · 0% all 48 public tasks | £1,342.93 | Baseline |
| What Claude got wrong | Profit error |
|---|---|
| Switched van method against the client's instructions | £1,916.67 |
| Under-claimed van costs by half | £707.11 |
| Claimed two garden hoses from trade receipts | £39.98 |
| Claimed a shop for the client's own bathroom | £30.96 |
| Ignored a £62 refund and the account fees | £3.50 |
The careless-bookkeeper baseline
A simple rule that records every payment at face value: no business-use splits, no personal items removed, no mileage method. It gets none of the 48 public tasks fully right and misstates profit by £1,342.93 on average, which shows how much judgement the tasks need.
What small AI errors add up to at scale.
Move the sliders to see what an average profit error means across a customer base.
An illustration, not a forecast. Errors go both ways, and updates are cumulative, so an early mistake carries through every later quarter.
How we build and mark every task.
Written down so you can check it, challenge it, or reproduce it with the Starter Pack's marking script.
The tasks
Each task is a fictional sole trader (plumbers, electricians, decorators and similar) with a full year of bank statements, receipts, invoices and client messages, generated by our rules engine from a seed. Public v1 has 12 businesses × 4 quarters.
The rules
2026-27 tax year on the cash basis: simplified mileage at 45p a mile and 25p after 10,000, or actual costs × business use (never both); HMRC's home-working flat rate of £10, £18 or £26 a month; personal spending, fines and similar disallowable items excluded. Amounts rounded per transaction.
The marking
The AI returns an MTD cumulative update. We compare the allowable amount in every box with the answer key, to the penny. Excluding a cost and fully disallowing it are both accepted. A task is correct only if every box matches.
The October 2026 run
Claude, run via Claude Code, one attempt per task on the 12 v1 public tasks: 7 fully correct, £224.85 average profit error. The careless-bookkeeper baseline (every payment at face value) scored 0 of 48.
What is an AI training environment?
A simulated workplace where an AI practises a task many times and is scored on each attempt. Ours covers UK bookkeeping and tax for sole traders.
What does the AI actually have to do?
Read a fictional sole trader's bank statement, receipts, invoices and client notes, then produce the HMRC Making Tax Digital cumulative quarterly update as JSON. We compare every box with the answer key.
Can we pay to rank higher on the leaderboard?
No. A verified run pays for testing under exam conditions. The score is whatever the model earns, published either way.
Who checks the tax rules?
Every answer comes from our rules engine, which follows published HMRC guidance for 2026-27 on the cash basis. We are arranging an independent review by a qualified UK accountant. If you spot an error, tell us and we'll correct it.
Do the tasks contain real people's data?
No. Every client, transaction and message is synthetic, so there's nothing to anonymise.
Is any of this tax advice?
No. Our tasks are training and evaluation data, not advice.
Test your AI on UK tax today.
Ten private tasks, marked to the penny, starting the moment you pay. Or ask for a free pilot pack first.