AI for UK Payroll: RTI, Auto-Enrolment and Holiday Pay Checks

AI can help with UK payroll before and after the pay run: spotting anomalies in an exported pay run before the Full Payment Submission, explaining holiday-pay and auto-enrolment rules and pointing to the GOV.UK page, and drafting staff replies. The calculations, the RTI submissions and each auto-enrolment decision need a manual or software check, because small errors compound across a pay run.

Think of AI as a reviewer sitting next to your payroll software, not inside it. Payroll software calculates and files. A reviewer reads the numbers before they go and asks "does this look right?" That second pair of eyes is cheap with AI, and its limits are clear: it reads what you export, it cannot see what you did not give it, and it can state a rule with total confidence when the rule has changed.

Follow me on Instagram@sagnikteaches

What is safe to automate and what is not

Keep personal data out of the prompt

Payroll data is among the most sensitive a small business holds: names, National Insurance numbers, bank details, pay, tax codes, absences, sometimes health information in statutory sick pay records. You hold it under data-protection law, so decide what leaves your payroll system. A workable rule: nothing identifiable goes into an AI tool, ever. For every check below, export only the columns the check needs and replace the employee name with a code such as E017. Delete National Insurance numbers, bank details and addresses from the export. Use a business plan with training off by default and a firm account rather than a personal one; the tutorial on safe ChatGPT prompts for payroll teams shows how to phrase checks that never include pay data.

Connect on LinkedInSagnik Bhattacharya

An even safer route for numeric checks is to ask AI to write a spreadsheet formula or an Excel check sheet, and then run your data through it locally. The data never leaves your machine. The ChatGPT and Excel guide explains how to ask for formulas and test them.

Subscribe on YouTube@sagnikteaches

Check 1: Pre-submission anomalies before the FPS

GOV.UK's guidance on reporting through RTI says you send a Full Payment Submission (FPS) from your payroll software on or before your employees' payday, including everyone you pay even if they get less than £96 a week (the figure the page gave on 4 November 2026). You report the usual payday even if you pay early or late. If information changes after you have sent it, you send a corrected FPS. The practical lesson is that mistakes should be caught before the submission, because corrections cost time and can mean a second submission and employee confusion.

An anomaly check compares this period with the last one and looks for things that are unlikely rather than wrong. Export the pay run as a table (employee code, gross pay, tax, employee NI, employer NI, pension, net pay, hours, holiday pay, any statutory pay) and the previous run in the same format.

Compare the two pay runs below (this period and last period). Flag:
1. Any employee whose gross pay changed by more than 20%.
2. Any employee with zero or negative tax or NI when last period
   had a value.
3. Any employee in this run who was not in the last, and vice versa.
4. Any net pay greater than gross pay.
5. Any identical gross amounts for 5 or more employees (possible
   copy-down error).
6. Any employee with pension contributions last period and none now,
   or the reverse.
For each flag, give the employee code, the field, and both values.
Do not explain the cause. List only.

A realistic sample result (illustrative): out of 38 employees, six flags. Two are expected (a new starter and a leaver, both known). One is a 41% jump for E012, which turns out to be a back-pay adjustment entered twice. One is a zero NI line for E025 whose pay fell below the threshold in a short week. Two are noise. You investigate each in the payroll software and correct before submission. The value is the E012 catch: it would have sent an overpayment, and a corrected FPS afterwards.

Set the thresholds for your own business. A restaurant with variable hours might need 40% for gross pay swings; an office with fixed salaries might use 5%. Over time, tune the list of checks by recording which flags turned out to matter.

Check 2: The deadline calendar around the submission

GOV.UK describes the cycle that follows an FPS. In the next tax month, which starts on the 6th, you can see what you owe from the 10th in your online account, send an Employer Payment Summary (EPS) by the 19th to claim reductions such as statutory pay, and pay the balance by the 22nd (the 19th if paying by post). Those dates are on the GOV.UK pages; confirm them on the current page whenever you set up a calendar, because they can change.

AI can turn those rules into a monthly calendar of reminders and a short checklist for the person running payroll. That is simple, repetitive and low risk, and a missed date is the kind of thing a reminder prevents. Ask for a table of dates for the coming twelve months based on your pay date, then check the table yourself against GOV.UK and the calendar before using it. Do not let AI compute due dates around weekends and bank holidays without that check.

Check 3: Holiday pay, rules first and arithmetic second

Holiday pay is where chatbots most often give a plausible, wrong answer, because there are several regimes inside one topic. GOV.UK's pages on holiday entitlement and holiday pay set out the basics:

• Most workers are entitled to 5.6 weeks' paid holiday a year; for a five-day week that is 28 days, and statutory entitlement is capped at 28 days.
• Part-time workers with regular hours get 5.6 weeks, which for three days a week means 16.8 days.
• Irregular-hours and part-year workers accrue leave at 12.07% of the hours worked in each pay period (rounded up to the nearest hour from 0.5), and a week's pay is worked out from their average pay over the previous 52 weeks, counting only weeks in which they were paid and going back as far as 104 weeks if needed.
• For regular-hours workers, at least four weeks of the statutory leave must be paid at the "normal" rate of pay (which includes things such as commission and regular overtime), and the remaining 1.6 weeks at the "basic" rate. Irregular-hours workers are paid all their leave at the normal rate.
• Rolled-up holiday pay (adding an amount to the hourly rate) is not allowed for regular-hours workers; employers can use it for irregular-hours and part-year workers unless their leave year began on or before 31 March 2024.
• From 6 April 2026, employers must keep detailed records of annual leave and holiday pay for at least six years.

Those points came from the GOV.UK pages when I read them on 4 November 2026. Treat them as a starting position, and always re-read the page before you build anything, because holiday pay is an area that changes.

A quick sum shows how small the numbers get. GOV.UK's own example: a weekly-paid worker who works 30 hours accrues 30 × 12.07% = 3.621 hours, which rounds up to 4 hours of leave, available from the next pay period. A chat answer that says "about 3.6 hours" is wrong on the rounding rule, and over a year of weekly pay runs those part-hours add up to a real difference in what the worker can take.

Using AI for rule look-ups, not figures

A good prompt asks the tool to name the category of worker, state the rule in plain English, and tell you which GOV.UK page to open to confirm. Then you open it.

I run payroll for a small business. A worker has irregular hours
(between 8 and 30 a week), is paid hourly, and her leave year
started on 1 June 2025. In plain English, which rules decide how
her statutory holiday is accrued and how a week's pay is worked
out? Name the GOV.UK pages I should check. Do not calculate any
amounts. If the answer depends on facts I have not given, list them.

A sensible answer names the 52-week average, notes that only weeks with pay count, mentions the maximum 104-week look-back, flags the leave-year start date as relevant to the accrual and rolled-up pay rules, and lists missing facts such as contract terms and whether pay is regular. That is a good primer. You still read the GOV.UK page because the tool may have summarised a rule that has changed.

A spreadsheet check for the average

For irregular hours, a short sheet can check the week's-pay figure from your payroll software. Ask AI to write the formulas for a 52-row table (week start, hours, pay, paid? yes/no), an average over paid weeks only and a note for the 104-week limit. Test it with two or three of the worked examples in GOV.UK's guidance on calculating holiday pay for workers without fixed hours or pay, and only then use it on live data. A realistic bug caught in testing: the first formula averaged across all 52 weeks including unpaid ones, understating the week's pay. The official example showed it at once. That is the reason for testing against published examples before anyone relies on the sheet.

Check 4: Auto-enrolment status

GOV.UK says an employer must automatically enrol a worker into a pension scheme and make contributions if the person is classed as a worker, is aged between 22 and State Pension age, earns at least £10,000 a year, and usually works in the UK. There are exceptions (for example, a worker who has given notice or has already left a scheme within the stated period), and different arrangements for some company directors without employment contracts. Read the current GOV.UK and Pensions Regulator pages before you rely on any figure, since earnings thresholds are reviewed.

The same GOV.UK guide sets the minimum contributions for most automatic-enrolment schemes at 8% in total, of which the employer pays at least 3%, usually worked out on earnings between £6,240 and £50,270 a year. It also says an employer need not contribute for a worker who opted in and earns £520 a month (£120 a week, £480 over four weeks) or less. Scheme rules can use a different earnings basis or pay more, so type in the figures for your own scheme. They turn a vague "check pensions" into a test AI can run on an export: list every eligible jobholder whose employer contribution is below 3% of the qualifying earnings for the period.

Payroll software normally assesses status automatically at each pay run. The manual check is a sample of the people for whom the answer is least obvious:

• New starters in their first pay run.
• Staff whose pay in the period was near the earnings trigger.
• Staff who recently turned 22 or reached State Pension age.
• Anyone who opted out, with the date and the evidence.
• Workers with two jobs in the business, or a change of hours.

Export a list with employee code, date of birth (or just age band), pay in period, status assigned by the software and pension contribution taken. Ask AI to compare that list with the criteria you have typed in from GOV.UK, not from its memory.

Criteria (from GOV.UK, as I have read them today): worker; aged
22 to State Pension age; earnings at or above the annual trigger of
[figure]; ordinarily works in the UK. Below is a table with age
band, annualised earnings, status assigned and contribution taken.
List every row where the assigned status does not match the
criteria, every row where status is "eligible" but no
contribution was taken, and every row where the employer
contribution is below [your scheme's employer rate] of the
earnings it is based on. Do not correct anything.

Illustrative result: of 22 staff, two rows flagged. A worker who turned 22 last month, earns above the trigger and is still marked "non-eligible jobholder", and a worker marked "eligible" with no contribution because the pension start date was entered after the pay date. Both go back into the software for a human to settle. The software could have caught the first at the next assessment; the second would have been an underpayment you find later and have to put right.

Check 5: Replies to staff queries

Employees ask the same questions: "why is my pay lower this month?", "how much holiday have I got left?", "why has my tax code changed?" Drafting the reply with AI is safe if you give it your own policy text and the facts from the payslip with names removed, then read the answer before sending. The risky part is the figure. The tutorial on how payroll bureaus use AI to cut errors and queries has more reply patterns.

A before and after (illustrative): the draft says "Your holiday pay was calculated at your basic rate". You know this worker's regular overtime counts towards normal pay for at least four weeks of statutory leave, so you correct it to "your holiday pay for the first four weeks reflects your normal pay, including regular overtime, as the rules require". The tool applied a plain-vanilla rule. You applied the correct one.

A small-business example: a twenty-person pay run

Say a recruitment agency with twenty employees runs monthly payroll in cloud software. On the Wednesday before payday the payroll lead exports the draft run and last month's run, replaces names with codes, deletes National Insurance numbers and bank details, and runs the anomaly prompt. Four flags come back. She finds that one flag is a commission payment entered twice, one is a leaver who should have been marked with a leaving date, and two are expected. She corrects the first two in the software, and the final run is submitted on payday.

Two days later, a part-time worker on irregular hours asks how her holiday pay was worked out. The lead asks AI for the rule outline, opens the GOV.UK page, checks the 52-week average in her tested spreadsheet against the software's figure, and replies using the practice's holiday-pay policy text. Because the leave year in the contract starts in June, she confirms the accrual rule with the page rather than with the chat. At the end of the quarter she picks six staff and checks their auto-enrolment status against the criteria. Nothing is wrong this time. She notes the date and the sample in the payroll file, which is the kind of evidence an enquiry about records will look for.

Where it goes wrong

• Stale rules. A tool quotes last year's threshold. Show-up: the figure does not match the GOV.UK page. Fix: type the criteria in yourself from the current page, with a date.
• Wrong regime. A tool applies the regular-hours rule to an irregular-hours worker. Show-up: the average differs from the official example. Fix: name the worker category in the prompt and test against the examples.
• Silent arithmetic error. A chat calculation is off by a small amount. Show-up: your software disagrees. Fix: never use a chat figure for pay; formulas in a spreadsheet only, tested.
• Identifiers in the export. A National Insurance number left in a column. Show-up: you see it when you re-read the prompt. Fix: delete the chat, tell whoever owns data protection, and use a template export that has no such columns.
• Too many flags. A threshold that is too tight and buries real errors. Show-up: the team stops reading them. Fix: tune thresholds after each run based on which flags mattered.

Year-end and corrections: where a second look pays

Most payroll pain concentrates at the edges of the tax year and in the corrections that follow a mistake. GOV.UK says that if you find an error in an FPS you should send a corrected FPS as soon as possible, and that the last report of the tax year carries extra information. Those are software tasks. What AI can do is prepare you for them: draft a year-end checklist from the steps you tell it your software requires, or read an export of the final pay run for the usual year-end oddities.

This is the final pay run export of the tax year (names removed).
Check for: employees with a cumulative gross that is lower than in
the previous run (should not happen); any employee marked as
leaver who is still being paid; any employee with a tax code ending
in a letter that differs from the one last period; any employee
without a pay-frequency marker. List findings by employee code.
Do not suggest corrections.

A leaver who is still on the run is the classic find. It usually means someone forgot to enter the leaving date, which affects the final pay, the P45 and the pension record. Catching it before the final submission saves an awkward correction in April.

A weekly and monthly rhythm that fits a small team

A small business gets the most from these checks when they are part of a fixed rhythm, not a heroic effort. One shape that works for a twenty-person monthly payroll: five working days before payday, collect changes (starters, leavers, hours, bonuses, absences); three days before, run the draft and the anomaly comparison; two days before, resolve flags and check the sample of holiday and auto-enrolment cases; one day before, approve; on payday, submit the FPS and keep the confirmation; the following month, check the liability against the figure in the online account from the 10th, send the EPS if needed by the 19th and pay by the 22nd. Each step has a named owner and a place to tick it off, and the AI-assisted checks are two lines in that list, not the whole process.

If you have a part-time bookkeeper or an external bureau, agree which of these checks they run and which you do, so nothing is assumed to be someone else's job. Ask them what they already check; many bureaus run exactly this kind of comparison already.

Setting it up and keeping it honest

Start with one check, the anomaly comparison, and run it alongside your normal process for three pay runs without changing how you work. Count the flags that turned out to be real errors. If none did, the check is not worth keeping in its current form; adjust the thresholds or drop it. Then add the deadline calendar, then the holiday and auto-enrolment sample checks. Keep a one-page procedure with the prompts, the data columns allowed in an export, the account to use, and the manual checks that remain. For timesheet-side errors that feed into payroll, the timesheet checks tutorial is the natural companion, and holiday request tracking helps keep the leave records that the rules now require. If you want to map the whole pay-run process with someone, that is what my AI implementation consultation covers, and you can begin at the contact page.

UK payroll and AI: questions answered

Can AI submit my RTI return?

No, and it should not. FPS and EPS submissions go from your payroll software to the tax authority. AI is useful beforehand, reviewing an exported pay run for oddities, but the submission itself, the figures and the correction of any mistake stay inside the payroll software under your control.

Is it safe to paste payslip data into a chatbot?

Not with names, National Insurance numbers or bank details. Export only the columns a check needs, replace names with employee codes, and use a business plan. Better still, run the check on a spreadsheet with formulas you can see, using AI to write the formulas rather than to read the data.

Can AI calculate holiday pay for irregular-hours staff?

It can explain the method and write a spreadsheet formula, but the figure should come from your payroll software or a calculation you have tested against the official guidance. Irregular hours bring a 52-week average, skipped unpaid weeks and leave-year start dates, which is exactly where chat answers go wrong.

How often should I check auto-enrolment status?

At every pay run, because status can change when pay, age or hours change. Most payroll software assesses automatically, so the manual check is a sample: take new starters, people near the earnings trigger, and anyone who opted out, and confirm the software's decision agrees with your reading of the rules.

Further reads

• Employment Contracts With AI: Drafting Steps and Legal Checks — Contracts that set the pay terms your payroll applies.
• How to Write an Employee Handbook With AI and What to Check — Writing the holiday and pay policies staff will read.
• How to Anonymise Client Data Before You Paste It Into AI — Masking identifiers before any prompt.
• Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare on tax work.
• AI Literacy Requirements: What Your Staff Need to Know — What staff running payroll should know about AI.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in an AI tool's privacy terms.
• Payroll Year-End Checklist: Where AI Helps and Where It Misleads — A year-end checklist showing where AI helps and misleads.
• Can AI Draft Pension Auto-Enrolment Letters? Check Regulator Wording — Drafting auto-enrolment letters and checking the wording.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: GOV.UK guides Running payroll (FPS and EPS parts), Holiday entitlement (including holiday pay) and Workplace pensions, read on 4 Nov 2026; check the current pages and your payroll software's guidance before acting.