01 the business
Training and education provider
150 staff, many travelling trainers
Managers approve claims on their phones and rarely open receipts. Finance finds the problems later: missing receipts, alcohol on client lunches, hotels above the cap and the same taxi claimed twice.
02 before
The problem
Managers approve expense claims on their phones between meetings and rarely open the receipts. Finance finds the problems later: missing receipts, alcohol on a client lunch, a hotel above the cap, the same taxi claimed twice.
- Approvals happen without reading receipts.
- Policy breaches are found weeks later.
- Duplicate claims are hard to spot across months.
03 the build
What the AI does
- Step 1: Picks up new claims from Expensify, Ramp or Xero Expenses as soon as they are submitted.
- Step 2: Reads each receipt and checks that the merchant, date, amount and tax match what the employee entered.
- Step 3: Compares every line against your expense policy, such as meal caps, travel class, approved vendors and receipt rules.
- Step 4: Looks for repeat claims across employees and past months, including the same receipt submitted twice.
- Step 5: Adds a plain note to the claim that lists each issue and quotes the policy clause it relates to.
paused: waiting for a person to approve
How it was put together
- Picked up new claims from the expense tool as they were submitted.
- Checked receipts against what the employee entered.
- Compared every line with the expense policy.
- Looked for repeat claims across people and months, and added notes quoting the policy clause.
04 human in the loop
Where a person stays in charge
The employee's manager still approves or rejects every claim, now with the policy notes in front of them. Finance makes the final call on anything flagged as a possible duplicate.
- Managers still approve or reject every claim.
- Finance decides on possible duplicates.
- Notes are factual and cite the policy, never accusing anyone.
05 rollout
From kickoff to live in 3 weeks
- Week 1Turn the policy into clear, checkable rules.
- Week 2Check last quarter's claims to calibrate.
- Week 3Run on all new claims.
06 after
What changes
Managers approve with the facts in front of them, and finance stops reopening claims weeks later. Employees get a clear reason when something is sent back.
Assume 1,500 claim lines a month and 1 minute of finance review per line, with checks clearing 80% automatically.
1,500 × 1 minute = 25 hours a month of review. With 80% cleared, finance looks at 300 lines, about 5 hours.
What to measure
- Claims with policy issues flagged before approval.
- Issues found after payment.
- Finance time on expense review.
- Duplicate claims caught.
07 evidence
What the research says
In 1,921 fraud cases, expense reimbursement schemes made up 13% of cases with a median loss of $50,000, and frauds at organisations with under 100 staff had a median loss of $141,000.
NIST's framework organises AI risk work into four functions, Govern, Map, Measure and Manage, with a profile for generative AI added in July 2024.
NIST, AI Risk Management Framework 1.0 and Generative AI Profile (opens in a new tab) · 2023
08 lessons
Mistakes to avoid
- A vague policy. The checks can only apply clear rules.
- Treating flags as wrongdoing. Most are honest mistakes.
- Checking only new claims. Duplicates often span months.
09 questions
Common questions
- How common is expense fraud?
- In the ACFE's 2024 study of 1,921 fraud cases, expense reimbursement schemes made up 13% of cases, with a median loss of $50,000.
- Which expense tools does it support?
- Expensify, Ramp and Xero Expenses are common.