The problem
PSA (PAYE Settlement Agreement) and P11D reporting depend on consistent, accurate classification of employee expenses and benefits. In most organisations this work is done once a year, under time pressure, using exports from the expense system, payroll data and accounts payable ledgers pulled into spreadsheets.
Finance and tax teams then spend weeks reading through thousands of expense lines, applying judgement about whether items are taxable, allowable, trivial, entertainment, staff welfare or benefit-in-kind. Descriptions are inconsistent, receipts are missing, category fields are misused and the same type of expense often ends up coded differently depending on who entered it.
The result is heavy manual review, inconsistent treatment between years, and the constant risk that something is missed or misclassified before submission to HMRC.
Why it matters
Misclassified expenses create direct financial and reputational risk. Under-reporting on a PSA or P11D can lead to HMRC enquiries, interest, penalties and potential reassessment of prior years. Over-reporting means the business pays more tax than it needs to. Both outcomes erode trust between finance, tax, HR and the wider business.
From an operational view, the annual scramble pulls senior finance and tax resource away from higher-value work. And because the process is largely undocumented and spreadsheet-based, knowledge sits with individuals rather than the organisation, making it hard to scale, audit or hand over.
The opportunity
Expense classification is an ideal candidate for a governed, AI-assisted workflow. The rules are well understood, the data is structured enough to standardise, and large language models are very capable of suggesting a tax treatment based on a description, merchant, amount and category.
By combining no-code automation, embedded AI classification and a clear review layer, the team can move from a once-a-year manual exercise to a continuous, controlled process. AI proposes the classification, humans review the exceptions, and every decision is logged for audit.
Example workflow
1. Connect the source data
Pull expense transactions from the expense management system, corporate card feeds, accounts payable and payroll. Include descriptions, merchant names, GL codes, amounts, dates, employee details and any attached receipt metadata.
2. Standardise and prepare the data
Normalise merchant names, clean category fields, deduplicate, and enrich each line with employee role, location and cost centre. Flag missing receipts, missing descriptions and unusual amounts.
3. Apply business logic
Apply deterministic rules first: known merchants, known GL codes, trivial benefit thresholds, mileage rules and staff entertainment limits. Route anything that falls outside clear rules to AI-assisted classification.
4. Run checks and controls
Use AI to propose a PSA, P11D, taxable, non-taxable or trivial benefit classification with a short rationale and a confidence score. Flag low-confidence items, unusual patterns and items that contradict prior treatment.
5. Produce outputs
Generate a clean, structured dataset ready for PSA calculation and P11D population, with full traceability back to the source transaction. Produce summary packs by category, cost centre and employee group.
6. Review exceptions
The tax and finance team review only the exceptions and low-confidence items in a structured queue, rather than every line. Decisions are recorded and feed back into the rule set and AI prompts.
7. Move to governed operation
Run the workflow monthly or quarterly rather than annually. Maintain version control over rules, prompts and treatments, with clear ownership and an audit log of every classification decision.
What good looks like
- A single, standardised dataset of expenses across all source systems.
- Deterministic rules applied first, with AI only used where judgement is needed.
- Every classification has a rationale, a confidence score and an audit trail.
- Exceptions are reviewed in a structured queue, not a spreadsheet.
- Prior-year treatments are visible and used to drive consistency.
- The PSA and P11D submission is a by-product of a controlled process, not a project.
Benefits
For the finance and tax team
- Far less manual review at year end.
- Consistent treatment of similar expenses across the business.
- Clear documentation of why each item was classified a certain way.
- More time for genuine judgement calls and planning.
For leadership
- Reduced risk of HMRC challenge, penalties or restated submissions.
- Predictable PSA cost and clearer benefit-in-kind exposure.
- Confidence that tax compliance is governed, not improvised.
For the wider business
- Faster, cleaner expense processes for employees and managers.
- Better data on staff entertainment, welfare and travel spending.
- A repeatable model that scales as the business grows or acquires.
Where to start
A good first version focuses on one population of expenses, typically corporate card transactions or staff entertainment, over a recent twelve-month period. Build the standardised dataset, apply the deterministic rules, layer in AI classification for the grey areas, and compare the output to the existing PSA or P11D treatment. The differences become the basis for both immediate corrections and ongoing improvements to the workflow.
How 4th Revolution can help
4th Revolution is finance-led and data-led. We specialise in no-code automation and embedded AI for processes where governance, controls and tax compliance matter. We understand PSA, P11D, benefit-in-kind rules and the realities of finance and HR data.
Our goal is not just to build a workflow. It is to create a governed, repeatable process with clear ownership, documented rules, a full audit trail and AI used where it genuinely adds value. The output is a tax compliance process you can defend to auditors, HMRC and your board.
Example outcome
Before: the tax team spends six weeks each year manually reviewing tens of thousands of expense lines across spreadsheets, with inconsistent treatment between reviewers and limited audit trail.
After: expenses are classified continuously through the year, AI handles the routine grey areas, the team reviews only flagged exceptions, and the PSA and P11D submissions are produced from a single governed dataset with full traceability.