Expense and reimbursement agent

Background

The client is an IT services company of 120 employees whose frequent travel and client meetings produce 1,800 expense items a month. Employees photographed receipts and typed the date, vendor, account category, amount, and reason into the expense system. It took twelve minutes per item, and most people did it in a month-end batch.

At month end, two finance staff compared the corporate card statement with submitted receipts line by line, chased missing items, and re-keyed the expense sheet into the accounting system by account. Close took three days. Managers opened and approved each of their team’s items, but policy violations such as over-limit meals, weekend spending, and the same receipt submitted twice were often discovered only after close. How much a department had spent this month was known only when the spreadsheet totals arrived in the middle of the next.

What we built

The goal was to remove typing and matching and put only items that need judgment in front of a person. Approving a payment stays with a person, always. We built five things.

1. Reading receipts and card charges

When an employee photographs a receipt or forwards a corporate card authorization, the agent reads the date, merchant, amount, and tax and assigns an account category. It picks the category from the merchant type and the employee’s past submissions, and when confidence is low it shows two candidates to choose from. The employee only confirms a one-line reason.

2. Card reconciliation

Daily card authorizations from the card company are matched to submitted receipts by amount, date, and merchant. An authorization with no receipt prompts the employee to upload one; a receipt with no authorization (a personal card payment) is marked separately.

3. Policy review

The company’s expense policy was turned into a rules table: per-person meal limits, time conditions for overtime meals, attendee lists for entertainment, prohibited merchant types, duplicate receipts, and payments outside travel dates. A violation is flagged with the clause it breaks; everything else passes as “no issue.” Finance edits the rules table themselves.

4. Approval screen

Managers approve clean items in bulk from a summary (count, total, breakdown by category) and open only flagged items. Approvals and rejections are logged with reasons, and rejected items go back to the employee with the reason. No expense is paid without approval.

5. Journal entries and dashboard

Approved sheets become journal entries in the accounting system with account, department, and project codes. Spending by department and category appears on a dashboard the moment it is approved, and a department that passes 80% of its budget triggers a notice to its manager.

How it was rolled out

Weeks	Work
1–2	Analyzed 10,000 expense items and rejection history from the previous six months to draft category rules and the policy table. Aligned policy interpretation with finance
3–5	Built receipt reading, card reconciliation, and policy review. Ran a month of expenses in shadow mode alongside human processing
6–8	Added the approval screen, accounting integration, and dashboard. Engineering and sales teams used it first
9–10	Company-wide rollout. Employee guidance and a rules-table editing guide for finance

In shadow mode, automatic category assignment agreed with people 92% of the time, and the agent found 1.4 times as many policy violations as people had found after close. After finance revised the policy table twice, false positives (clean items flagged) fell below 3%.

How the day changed

Employees upload one receipt photo right after paying, confirm the reason, and are done. The month-end evening with a pile of receipts is gone. Instead of opening 150 items one by one, managers check about 20 flagged ones and approve the rest from a summary.

Finance’s close dropped from three days to half a day. Card matching and re-keying disappeared; what remains is checking flagged items and refining the policy table. Department spending is on the dashboard without waiting for close.

The table below compares each task before and after.