Can AI assistants do UK payroll sums?

The short answer

On 9 October 2026 we put 10 UK pay questions for 2026/27 to 5 AI models, each stating the inputs a payroll would use. 36 of the 50 answers were right, allowing for payroll’s rounding. The best two got all 10; the others missed between three and six, mostly by using out-of-date rates or rules (some from 2024/25), or by giving no figure at all. And on the four take-home questions, no model’s figure matched to the penny what payroll actually pays.

Take-home on £30,000

“An employee in England earns a salary of £30,000 a year, paid monthly, tax code 1257L, with no pension contributions and no student loan. What is their take-home pay for the 2026/27 tax year, after income tax and employee National Insurance?”

Take-home on £45,000 in Scotland, with a student loan and a pension

“An employee who lives in Scotland earns £45,000 a year, paid monthly, tax code S1257L. They repay a Plan 2 student loan and pay 5% of their salary into a relief at source pension. What is their take-home pay for the 2026/27 tax year?”

Where they went wrong: Gemini used the 2024/25 Scottish bands (a starter band to £14,876, not £16,537) and the 2024/25 Plan 2 threshold (£27,295, not £29,385). Grok took the gross £2,250 pension from pay, though relief at source takes the net £1,800, and used the 2025/26 Plan 2 threshold. Sonar Pro took the pension off taxable pay, as a net pay scheme would, and its NI was £190 short.

Take-home on £110,000 (the allowance taper)

“An employee in England earns £110,000 a year, paid monthly, with no pension contributions, no student loan and no other income. What is their take-home pay for the 2026/27 tax year, after income tax (including the personal allowance taper) and employee National Insurance?”

Take-home on £40,000 with two student loans

“An employee in England earns £40,000 a year, paid monthly, tax code 1257L, no pension. They repay both a Plan 2 student loan and a Postgraduate loan. What is their take-home pay for the 2026/27 tax year?”

Where they went wrong: Gemini used the 2024/25 Plan 2 threshold (£27,295). Sonar Pro gave a month’s pay where a year was asked, and its year (£29,061.72) was off too. The others are up to £18 from payroll, which rounds each loan deduction down every month: close enough to count as right here.

The salary that takes home £3,000 a month

“An employee in England, paid monthly with tax code 1257L, no pension and no student loan, wants to take home £3,000 a month. What gross annual salary do they need in the 2026/27 tax year?”

What a £40,000 employee costs

“An employer in England pays an employee £40,000 a year (monthly payroll) and pays the auto-enrolment minimum employer pension contribution of 3% of qualifying earnings. The employer cannot claim the Employment Allowance and does not pay the Apprenticeship Levy. What is the total annual cost of employing them in 2026/27: salary plus employer National Insurance plus the employer pension contribution?”

Statutory Sick Pay for 3 days

“An employee with average weekly earnings of £600, who usually works 5 days a week, is off sick for 3 working days in May 2026. How much Statutory Sick Pay are they due for those 3 days?”

Where they went wrong: Gemini applied both rules from before 6 April 2026, that sick pay needs 4 days off in a row and starts after 3 waiting days. Grok knew the waiting days had gone but kept the 4-day rule. Since 6 April 2026 sick pay is due from the first full working day off sick (gov.uk).

Statutory Maternity Pay in total

“An employee with average weekly earnings of £400 takes the full 39 weeks of Statutory Maternity Pay, all paid in the 2026/27 tax year. What is the total Statutory Maternity Pay they receive?”

Where they went wrong: Gemini used last year’s weekly rate (£187.18, not £194.32). Grok said it didn’t know the 2026/27 rate and gave no figure.

Statutory redundancy pay

“An employee aged 45 is made redundant in England after 10 complete years of continuous employment, earning £800 a week. What statutory redundancy pay are they entitled to for a redundancy in the 2026/27 tax year?”

Where they went wrong: Gemini used an old cap on a week’s pay (£700; it is £751 from 6 April 2026). Sonar Pro counted all 10 years at a week and a half, though only the years from age 41 count that way. Grok said it didn’t know the cap and gave no figure.

Self-employed take-home on £50,000 profit

“A self-employed sole trader in England has taxable profits of £50,000 for the 2026/27 tax year, no other income, no pension contributions, and repays a Plan 2 student loan through Self Assessment. What do they take home after income tax, Class 4 National Insurance and the student loan repayment?”

Where they went wrong: Gemini and Grok both used the 2024/25 Plan 2 threshold (£27,295, not £29,385), so they took £188 too much for the student loan.

Why an API

A language model works a pay question out from what it remembers of the rules, and the rules change every April. The API runs payroll the way HMRC specifies it: each pay period through the tax code, NI and student loan routines, with the year’s rates as reviewed code. So the same inputs always give the same answer, and the answer is the one payroll would give.

Building with an AI agent? Give it the API as a tool rather than asking it to do the sums: the MCP server has every endpoint, and an agent can sign itself up for a free key.

How this was run

• Run on 9 October 2026, through OpenRouter’s API: each model asked each question once, with no follow-ups. An earlier attempt the same morning was stopped by a timeout before it finished; none of its answers were used.
• The models are recent general models from each company, not always its largest: Gemini 3.8 Flash is Google’s fast tier, and Anthropic and OpenAI also offer larger models than the ones asked.
• Every model had the same instruction: “You are answering a UK payroll question for the 2026/27 tax year (6 April 2026 to 5 April 2027). Use the rates and thresholds in force for that year. Show brief working. End with one final line exactly in the form: ANSWER: £12,345.67”
• Sonar Pro searches the web for its answers; the others answered from what they know. The chat apps built on these models may search the web too, so their answers can differ.
• An answer counts as right within £30 of the API’s figure for a year’s pay worked out by payroll (take-home, the salary needed, employer cost): payroll works a month at a time, rounding taxable pay and student loan deductions down each month, which moves a year’s total by up to about that much from a sum done once for the year. Self-employed take-home must be within £10, statutory maternity and redundancy pay within £1, and sick pay within 5p.
• Each question states the inputs the API’s request uses (where the person lives, how they are paid, the tax code, loans and pension), so a model has what it needs to get it right. The figure on a model’s final ANSWER line is the one scored.