AI Agents in Expense Management: Where It Breaks

The loop, as it actually runs

1. Expense incurred
2. receipt captured
3. coded to category and project
4. checked against policy
5. approved
6. reimbursed and posted

Where it breaks

Not in the steps. In the joins between them:

• Receipt chasing is entirely manual
• Policy checks are inconsistent between reviewers
• Coding errors surface at close
• Reimbursement is slow enough that people notice

What that actually costs

This one is paid in two currencies. The first is the finance team's — receipts chased for a fortnight after every trip, coding corrected at close, policy applied differently depending on who reviewed it. The second is morale, and it is the one people underestimate: reimbursement slow enough that employees notice is a small, regular reminder that the company is disorganised about their money. Neither shows up as a cost anywhere. The first is buried in close timelines and the second is buried in how people feel about expense reports, which is a real thing even though no report measures it.

What we would build first

Receipt capture and chasing goes first, because everything downstream is blocked on it. Coding, policy checks and reimbursement all wait on a document that a person has to remember to send. An agent that watches the card feed, notices a charge with no receipt attached, and asks the cardholder directly — then again, on a schedule — removes the dependency the whole loop hangs on. It is also the least contentious thing to automate, which matters when you are introducing software into how people get paid back.

It runs inside Ramp, QuickBooks Online, Gusto, Slack, Gmail — the tools already in use.

How this one goes wrong

Policy checking is where an agent oversteps. Expense policy is written to be applied with some judgement — the client dinner that ran long, the flight rebooked because a meeting moved — and an agent applying the written rule strictly will reject things every human reviewer would have approved, at volume, to people's faces. The guard is that the agent flags rather than rejects: it can say which rule a claim appears to breach and route it to a reviewer, but the decline is a person's to make. Automating the judgement here buys you a small efficiency and a large amount of resentment.

The individual jobs inside this loop — what starts each one and where it has to hand back to a person — are covered under finance & bookkeeping.

Where it lands hardest: staffing & recruiting agencies, property management, insurance brokerages, managed service providers and mortgage brokerage. Each of those pages carries the sector’s own economics — the three roles that absorb the work, and the season that decides when a build can start.

In most companies this work sits with executive assistants, administrative assistants and office managers. Each of those pages carries the wage arithmetic for that role, and says plainly where the numbers do not justify the spend.