How can AI help us check staff expenses without treating everyone as suspicious?

AI can read a receipt, compare its details with an expense claim and highlight something that needs explaining. That is useful administration. Asking it to decide whether an employee is dishonest is a different proposition, and not one you should buy on the strength of a confident demonstration. A duplicate-looking receipt is a question, not a finding of fraud.

Consider an illustrative example: two £28.50 taxi claims have the same date and supplier. They might be duplicates. They might also be separate journeys taken by colleagues after the same meeting. A sensible system presents the original documents, explains the match and asks a finance colleague to check. An unsuitable system labels both people suspicious or automatically refuses reimbursement.

The purpose should be consistent application of a published expenses policy, fewer missing details and faster repayment. Those benefits help employees as well as finance. Do not expand that purpose into monitoring where people spend their evenings, guessing their health or comparing their lifestyles. The fact that a receipt contains information does not make every possible inference a legitimate business requirement.

Before buying, ask the supplier to demonstrate a legitimate exception as well as an obvious duplicate. Can a reviewer correct the extracted amount, record an explanation and close the flag without leaving an accusation attached to the employee? If the product cannot distinguish a query from misconduct, it is a poor fit for fair expense checking.

The strongest candidates are receipt extraction and explanations of clear policy mismatches: a missing business purpose, an amount that differs between receipt and claim, an absent receipt or a possible duplicate. Exact limits and arithmetic usually belong in ordinary rules. AI may help interpret messy documents, but it should not invent the policy against which the employee is judged.

Separate three decisions. First, did the software read the document correctly? Second, does the claim meet the company's policy or need an authorised exception? Third, what is the correct tax and accounting treatment? A language model's explanation does not settle any of those decisions. Finance must retain responsibility, especially where VAT, mileage, overseas receipts or mixed personal and business spending are involved.

Ask what the product cannot reliably read. Faded receipts, refunds, tips, foreign currencies and several receipts in one photograph are sensible demonstration cases. The answer should include uncertainty and an accessible original, not just a confidence percentage with no business meaning.

You may not need a custom AI project. Xero's UK expense product already describes receipt capture that fills supplier, date and amount, alongside human claim approval. Its page lists Grow at £39 per month excluding VAT after the introductory promotion, with expenses and mileage for one user. That is an example checked on 14 November 2026, not a quotation for your team; verify current plans and additional-user charges. Compare an existing feature with a custom build before paying for integration you may already have.

Employees should understand what information is examined, why it is needed, who sees it and what happens when a claim is flagged. Explain this before rollout, in ordinary language. A useful message is: we are checking submitted claims against the same policy, and a person will handle questions. A vague announcement about intelligent fraud detection creates a very different expectation.

The ICO's monitoring-workers guidance explicitly links workplace monitoring with data protection and worker trust. It is marked as under review following the Data (Use and Access) Act, so check current detailed guidance rather than treating an old checklist as permanent legal advice. Establish an appropriate lawful basis and assess necessity and proportionality. Assess whether a data protection impact assessment is required, particularly if the proposal involves systematic monitoring or consequential profiling.

Ordinary expense administration does not justify continuous location tracking or access to a personal bank account. Ask for the evidence necessary to support the claim, not a complete picture of someone's private life. Where a submitted receipt reveals health information or another sensitive detail, restrict access and assess the additional requirements before processing it with AI. The ICO's worker-health guidance is a relevant starting point.

A supplier should explain retention, subprocessors, model-training use and access controls. A business subscription alone is not proof that these arrangements meet your needs. Keep the expense-checking purpose narrow, and do not quietly reuse the output for performance rankings.

The person reviewing an exception needs enough information to disagree with the system. Showing a red risk score and asking someone to approve it is not meaningful review. Show the source receipt, the extracted values, the actual policy provision and the precise reason for the flag. Keep language factual: possible duplicate or missing evidence, not dishonest employee.

Use the same published rules for directors and junior staff. Different jobs can legitimately involve different spending, but a higher expense total is not evidence of abuse. A travelling engineer and an office-based administrator should not be compared as though they have the same working day. Check whether the product disadvantages people whose receipts are harder to read or whose work requires unusual purchases.

Give employees a straightforward correction route. They should be able to explain a legitimate claim, supply another image or challenge an incorrect interpretation without feeling they are appealing a disciplinary verdict. Record the final human decision and its reason. Do not let an AI-generated suspicion remain in an HR record after the underlying issue has been resolved.

If there is a real misconduct concern, move it into the established investigation process. The tool can preserve relevant evidence, but it cannot substitute for a fair investigation or qualified employment advice. In particular, do not let a flag automatically trigger disciplinary action, payroll deductions or an indefinite reimbursement hold. Legal duties and the circumstances of the payment need separate consideration.

Measure both financial usefulness and the burden imposed on staff. Ask how many flags were correct, how many legitimate claims needed extra explanation, how long reimbursement took and how much reviewer time was actually saved. A vendor's accuracy percentage is not enough unless you know what was measured and whether it resembles your documents.

Here is hypothetical arithmetic, not a predicted result. If checking 200 monthly claims takes four minutes each, that is 800 minutes. Reducing initial checking to two minutes saves 400 minutes. But 50 unnecessary queries taking eight minutes each consume the entire saving. A system that produces impressive alerts can still make the business slower and less trusting.

Include subscription charges, integration, reviewer training, support and exception handling in the comparison. Do not count an unproven fraud saving as guaranteed cash recovered. A trial should use representative, appropriately protected claims and a human-reviewed reference set. Include genuine exceptions and ordinary honest mistakes, not just examples designed to make the supplier look good.

Record keeping also matters. HMRC says employers must retain expense and benefit records for three years from the end of the relevant tax year. Its example requires a record of when and why an employee travelled, with receipts where possible. Other accounting or tax obligations may require longer retention. An AI summary is not a reason to discard original supporting evidence or keep unrelated personal information indefinitely.

Do not buy AI expense checking to repair an unclear expenses policy. If managers disagree about what is allowed, the software will scale that disagreement. Clarify reimbursement rules and approval authority first. Nor is a custom build sensible simply because a supplier can demonstrate one: a modest claims volume, a standard form and existing accounting software may already solve the problem.

Reject products that cannot explain individual flags, allow corrections or separate finance queries from allegations. Be cautious if the proposed value comes mainly from secretly watching staff, linking personal accounts or producing behavioural risk scores. Those are materially different buying decisions from reading submitted receipts, with greater legal and cultural consequences.

This can be a good fit where repetitive document checking is genuinely consuming time, existing tools fall short and a named finance owner can oversee exceptions. Staff should gain clearer decisions and quicker repayment, not just another system asking them to prove themselves. If you cannot name that employee benefit, reconsider the proposal.

Before signing, ask for a demonstration using your difficult examples and a written explanation of access, retention, correction and human approval. If you want help assessing whether the proposal is proportionate, discuss the workflow with us. The right answer may be to use the software you already have rather than commission anything new.