You could be forgiven for thinking PAYE Settlement Agreements are well-trodden ground for in-house tax teams. After all, the rules have been around for decades, the deadlines don’t move, and most employers run through the cycle every year. So why does the work seem to get harder, not easier?
A PSA lets an employer settle, in a single annual payment to HMRC, the income tax and Class 1B NIC due on certain expenses and benefits provided to employees. It’s used for items that are minor, irregular, or impractical to put through payroll or the P11D – staff entertaining, small gifts, working meals and the like. The 2025/26 tax year is closed, we’re in the midst of P11Ds and the PSA window is coming fast.
Key dates for 2025/26
- 5 July 2026 – apply for, or amend, your PSA with HMRC
- 31 July 2026 – calculation ‘normally’ expected to be submitted
- 22 October 2026 – payment of tax and Class 1B NIC due (19 October if paying by post)
Late payment penalties run at 5% of the outstanding amount after 30 days, with further 5% charges at six months and twelve.
The rules look simple. They aren’t.
Take the £150 annual function exemption. This is the rule that lets employers cover staff Christmas parties, summer events etc. without a tax charge.
Most people in employment tax know this one. £150 per head, per year, no PAYE or NIC due. Simple enough. Except…
- Per head – each attendee, including any guest brought. Two heads if an employee brings a partner.
- Inclusive of VAT and ancillary costs (transport, accommodation), not just food and drink.
- Conditional on the event being open to all employees, or to all employees at a location.
- And – this is the one that trips people up most often – if the spend exceeds £150 per head for an event, the WHOLE cost becomes taxable, not just the excess. A Christmas party at £155 per head can’t exempt £150 with £5 on the PSA. The whole £155 has to be reported and grossed up.
Similar traps sit in trivial benefits (£50 per occasion, with conditions on cash-equivalence and not being given in recognition of services), in qualifying long service awards, and along the subsistence-versus-entertainment line that tax teams argue about every year.
Whilst the headline rules are widely known, applying them correctly is harder than it looks – and that’s before we get to the data.
Isn’t it all about THE DATA…?
Beneath the rules, preparing a PSA is essentially a classification exercise across thousands of rows of expense and benefit data (often hundreds of thousands!). HMRC defines the categories. Your finance and HR systems define the rows. Funnily enough those systems weren’t built with tax classification in mind – and I know many that seem to purposely work against it, especially global systems!
Three things show up in almost every PSA cycle.
- Finance categories don’t map to tax categories. A typical ledger carries headings like “Office expenses,” “Team welfare,” “Hospitality,” “Training & development” – none of which tells you whether an item is staff entertainment, client entertainment, subsistence, or a trivial benefit. The team has to re-classify, and the size of that task is set by the volume of expense data, not by the number of items that end up on the PSA.
- The same supplier appears under different categories across the year. One employee codes a Pizza Express invoice as “Hospitality.” Another codes the same restaurant on a different day as “Team event.” A third uses “Client meeting.” This is what happens when categorisation is left to the moment of claim, but it means you can’t take the existing category as the answer.
- Recipient identity is rarely captured cleanly. A team-only lunch, a client lunch and a mixed lunch each attract different treatment – and the expense record almost never tells you which it was. Consistency in interpretation is what separates a defensible position from one that quietly accumulates risk.
The combined effect: tax decisions are being made against data that needs heavy preparation first, and the preparation itself is judgement heavy. In most teams that work happens in spreadsheets, on a tight calendar, with two or three people at best (some outsource this overseas). Classification ends up less consistent than anyone would like, and the reasoning behind each call rarely gets captured anywhere you could point HMRC to next year.
Machine learning
Many years ago we tried to use machine learning – it wasn’t ready. However, with the increase in small and large language models, the improvement in connections, server speed, ability to train models through the use of cleverly designed prompts. All these mean we’ve now been able to build a solution to help. PSA.I – a UK-built tool, powered by AI, aimed at the data side of the PSA cycle. It reads expense and benefit data in whatever form your systems export, applies treatment rules across the population, flags borderline cases for human review, and keeps the reasoning alongside the row.
What I particularly like is that it remembers what the team has done before. Corrections, overrides and exceptions made during one cycle are retained and applied in the next, so the tool aligns with your policies over time and year-on-year inconsistencies fade.
For employers running high-volume expense data and inconsistent category coding (which is pretty much all of you reading this!), it makes the classification step substantially less painful, and much more auditable.
Summary
PSAs aren’t going anywhere. The rules will stay hard to apply, the data messy, and the calendar tight. The teams that handle the cycle well treat the data – and the consistency of rule-application across it – as the work itself, rather than as a set-up step before the work begins.
Get the data right, and the rest follows.
If you’d like to talk through your 2025/26 PSA – scope, new items, the data side, or whether PSA.I might fit – please get in touch.