Expense reimbursements: what to automate, what to keep human, and the data it depends on

Software can read receipts, check claims against policy, pay approved amounts and chase overdue advances. People still need to write the policy, judge unusual claims, and decide what to do about money an employee owes. None of it works until card feeds, cost centres and approver records are clean.

Setting policy and approval limits

This stays with people. Someone has to decide what the organisation will pay for, at what level, and who signs off. Those are choices about culture and risk. No model can make them for the business.

What software does well here is enforcement. Once the rules exist as structured fields in the expense tool, such as per diem rates by location, hotel caps, or which categories need a second approver, the system applies them every time. The weak point is translation. Policies written as long prose documents tend to contain phrases like "reasonable" or "where appropriate" that a machine cannot test. Before automating, rewrite each rule so it can be checked against a field. Keep the judgement calls visibly separate, and route them to a named person.

Communication of the policy can be largely automated too. Inline prompts at the point of entry teach more than an annual email ever did.

Capturing tax data

Receipt capture is the most mature piece. Optical character recognition and current AI models extract merchant, date, amount and tax lines reliably from clear images and e-receipts. They struggle with handwritten slips, crumpled paper, foreign invoices and receipts that mix personal and business items.

Recoverable tax is where errors cost real money. A model can suggest the tax treatment, but a tax specialist should own the mapping rules: which expense types allow recovery, which countries require a full invoice and not a till receipt, how mileage is treated. Review a sample of extracted tax lines regularly. Drift shows up quietly.

Approving claims and advances

Most claims are routine. Automated checks can clear them: amount within limit, receipt attached, category matches merchant, no duplicate submission. Duplicate and anomaly detection is a genuine strength of software, because it compares every claim against every other one, which no manager does.

A person is still needed for anything that falls outside a rule. Entertainment with clients, travel that breaks policy for a good reason, advances for unusual trips. The manager approving should know the work, not just the budget line.

Watch for rubber stamping. When approvers receive a stream of claims the system has already passed, they stop reading. Sending them only exceptions keeps their attention where it matters.

Paying out

Once a claim is approved, payment is mechanical. Systems batch approved items, post them to the ledger with the right cost centre and tax code, and send payment through the payroll run or a bank file. This is safe to automate fully, provided the general ledger coding is reliable and bank details are verified on change.

Advances need slightly more care. The system should record each one against the employee and expect it to be settled by a later claim or a repayment.

Managing personal accounts

Every employee carries a running balance: advances not yet cleared, personal spend on a corporate card, overpayments. Software can track these balances, send reminders and escalate on a schedule.

Deciding what happens next is human work. Deducting from salary, writing off a small balance, or raising a sensitive matter with a departing employee involves employment law and goodwill. Keep those decisions with finance and HR together.

What the data has to look like first

Automation amplifies whatever is already in the system. Before switching anything on, check the following.

• The employee master matches HR records, including who has left and who reports to whom.
• Approval hierarchies reflect the real organisation, with delegates set for absences.
• Corporate card feeds arrive complete and map to the right cardholder.
• Expense categories link cleanly to general ledger accounts and tax codes, with no catch-all "miscellaneous" doing heavy lifting.
• Cost centres and project codes are current, and closed ones are blocked.
• Bank details have a verified change process.

If approvers are wrong in the system, automated routing sends claims to the wrong desk faster than paper ever did.

Questions to ask the people who run it

The written process and the lived one usually differ. Ask the people doing the work:

• Which claims get sent back most often, and what is missing from them?
• Are there approvals that happen by email or in conversation before anything is entered in the tool?
• Which policy rules does everyone quietly ignore, and who allowed that?
• How are claims handled when the approving manager is on leave?
• Where do personal charges on corporate cards end up, and who notices?
• Which receipts get re-keyed by hand, and why does the scan fail on them?
• What happens to an advance when the trip is cancelled?
• When an employee leaves owing money, who finds out, and how?
• Which reports does tax or audit ask for that take manual work to produce?

The answers show where the documented rules are fiction. Fix those gaps before teaching a system to follow them.

APQC's Process Classification Framework® (PCF) is an open standard developed by APQC, a nonprofit that promotes benchmarking and best practices worldwide. To download the full PCF or to view definitions and measures, please visit www.apqc.org/pcf.