Photo by Kaboompics.com on Pexels
It’s the third of the month. Your operations manager opens the expense queue and finds 23 reports waiting: hotel folios, Uber receipts, a client dinner, two Home Depot runs, a conference badge someone photographed instead of the registration invoice. She has a vendor call in twenty minutes. She opens the first four carefully, skims the next ten, and approves the rest because the totals look normal.
Nobody decided to stop reviewing expenses. It just happens, one busy month at a time. This post covers what expense review is supposed to catch, why it turns into rubber-stamping, what to fix in your policy before you automate anything, and what it looks like when an agent checks every line so the approver only sees the items that are actually out of policy.
What expense report review is supposed to catch
A good review answers a short list of questions for every line, not just every report:
- Is there a receipt, and does it match? Amount, date and merchant on the receipt agree with the card transaction or the reimbursement claim.
- Is it within policy? Meal limits, lodging caps, mileage rates, approved categories, pre-approval for anything above a threshold.
- Is it a duplicate? The same dinner submitted by two people, or the same receipt on this month’s report and last month’s.
- Is it coded right? The correct expense account, client or job, so job costing and billable expenses aren’t wrong at month-end.
- Is the documentation complete? Who attended the client dinner, what the trip was for, an itemized hotel folio instead of a card slip.
That last point isn’t just internal tidiness. Under the IRS’s general rule in Publication 463, you need documentary evidence such as receipts, canceled checks or bills to support travel and business expenses, and the exception for missing receipts applies only to expenses other than lodging that are under $75. If you reimburse employees under an accountable plan, the same publication describes the reasonable time periods the IRS accepts, such as accounting for expenses within 60 days. Rules change and your situation may differ, so confirm the details with your accountant before you build a policy around them.
Why manual review turns into rubber-stamping
The problem isn’t that approvers are careless. It’s that the job is shaped wrong for a person.
- Almost everything is fine. When most lines are routine, attention fades. The one hotel night above the cap looks exactly like the twelve that weren’t.
- The checks need context the approver doesn’t have open. Spotting a duplicate means remembering what someone else submitted last month. Nobody keeps that in their head.
- It arrives in a batch, at the worst time. Reports pile up at month-end, right when the same people are busy closing the books.
- Saying no is awkward. Sending a report back to a senior salesperson over a $14 overage feels petty, so it slides, and the policy quietly stops meaning anything.
Most expense software helps with capture: snap a photo, the receipt lands in the report. That removes typing. It doesn’t remove the reviewing, because the tool still waits for someone to open each report and decide.
Fix the policy before you automate the review
Automating a vague policy gets you vague answers faster. Before any agent touches an expense report, the policy has to be something a careful new hire could apply without asking you. Look for these gaps:
- Limits written as feelings. “Reasonable meals” can’t be checked. “Up to $60 per person for client meals; attendees required” can.
- Exceptions that live in someone’s memory. If the owner always allows the higher hotel rate in New York or Toronto, write it down.
- Categories nobody agrees on. If three people code the same software subscription three different ways, fix the chart of accounts mapping first.
- No owner for the gray zone. Decide who rules on exceptions, and how fast, before a system starts sending them anywhere.
This is the kAIzen part of the work: process first, then AI. The rewrite usually takes a few working sessions, and it pays off even if you never automate anything.
What a proactive agent does with every expense line
A proactive agent doesn’t wait for the approver to open the queue. It reads the signal that already exists, the receipt images, emailed e-receipts and the corporate card feed, as soon as they show up, and works each line:
- Extracts the receipt. Merchant, date, amount, tax and line items, the same way AI invoice data extraction turns supplier PDFs into structured data.
- Matches it to the transaction. Card charge to receipt, or reimbursement claim to receipt, and flags card charges that never got a receipt.
- Checks it against the written policy. Limits, categories, pre-approvals, required attendees, lodging documentation.
- Looks across reports. Same receipt twice, same dinner from two employees, a mileage claim on a day the same person rented a car.
- Suggests the coding. Account, job or client, based on how your team has coded similar lines before.
- Asks the employee first when something is missing. A missing folio or attendee list goes back to the person who submitted it, before it ever reaches an approver.
Then it checks itself. Every time an approver overrides a flag (“that client dinner was pre-approved by phone”), the agent records why and proposes a rule or policy change for a person to approve. Over time the flags get fewer and more accurate.
What the approver sees: only the exceptions
Instead of 23 reports, the approver gets a short list: the hotel night $80 over the cap, with the policy line it breaks; the dinner that appears on two reports; the $400 software charge with no receipt after two reminders. Each item comes with the reason and the evidence, so the decision takes a minute, not a hunt.
Lines that passed every check stay visible for audit, but nobody has to read them. That’s the same idea behind accounts payable automation done well: the routine flows through, the exception comes to you.
What stays human, and how trust is earned
Reimbursing an employee is money going out, so it never leaves a person’s view. The agent prepares the decision; a person approves the payment. A few other things stay human too:
- Judgment calls on exceptions, including the ones the policy didn’t anticipate.
- Any conversation with an employee about a pattern, a repeat issue or possible misuse.
- Changes to the policy itself.
And the agent doesn’t start out trusted. In the first weeks, a person reviews everything alongside it, and every disagreement is a calibration point. Autonomy goes up with evidence: first the agent suggests, then it clears clean lines with spot checks, and only then does it clear routine lines on its own. We cover that ladder in more detail in human-in-the-loop AI approvals.
Where to start
Pull last quarter’s expense reports and count three things: how many lines broke the written policy, how many were approved anyway, and how many hours approvers spent in the queue. If the first two numbers are close, your review has already turned into a formality. If you want to put a dollar figure on those hours before deciding anything, our ROI calculator is a quick way to size it.
Want to see this in your company?
Schedule a free discovery call with RubikSoft and learn how an AI Agent can transform your operations in weeks.
Schedule Your Free Consultation