A flag is only as good as the rule behind it. Turn the expense policy into numbered, checkable rules, then let the agent test every claim line against them and attach a flag that cites the rule, the amount and the evidence. The employee can respond. The manager approves or declines the line with a reason. The expenses lead owns the rules and reviews patterns.
What is the process, and who owns it?
Expense policy sets what the company will pay for: limits on hotels and meals, which travel class, what needs a receipt, what is never claimable. Traditional checks are a manager's glance and a finance sample. Both miss a lot, and both apply the policy unevenly. An agent can read every line and apply the same test every time; the case for that consistency is in can an AI agent enforce an expense policy.
The expenses lead owns the policy rules, the flag wording and the review of patterns. Line managers own approval of their teams' claims, including flagged lines. Employees own their claims and their explanations. The wider finance position is in AI agents for finance: books you can audit.
What does the agent prepare, step by step?
- Person, expenses lead: writes the policy as rules. Each rule numbered, with a test the agent can apply: "Rule 3.1: hotel per night, by city band, up to the band limit." Rules the agent cannot test are marked "manager judgement". Evidence: the rule set and its version.
- Agent, depth 1: reads each claim line with its receipt, category, date, location and any attendees listed. Evidence: the inputs.
- Agent, depth 1: tests the line against every applicable rule. Evidence: each rule tested and the result.
- Agent, depth 3: attaches flags to lines that fail a hard rule, citing the rule, the limit and the amount over. Evidence: the flag on the claim.
- Agent, depth 2: notes softer signals for the manager. A weekend claim for a role that does not usually travel, a meal for one claimed as client entertaining, a mileage claim longer than the usual route. These are worded as questions. Evidence: the note.
- Person, employee: responds to any flag with an explanation, or corrects the claim.
- Person, manager: approves or declines each flagged line, with a reason if approving despite the flag. Evidence: the decision and reason.
- Agent, depth 2: drafts the pattern review for the expenses lead: rules most often breached, managers who approve every flag, rules that may need changing.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Rules | Nothing | None | Expenses lead writes them |
| Read | Reads line, receipt, context | 1 | Nothing |
| Test | Applies every applicable rule | 1 | Nothing |
| Hard flags | Attaches cited flags | 3 | Nothing |
| Soft notes | Raises questions for managers | 2 | Interpreting them |
| Respond | Nothing | None | Employee |
| Decide | Nothing | None | Manager |
| Patterns | Drafts the review | 2 | Policy changes |
What does a named person approve, and why?
The manager approves every flagged line. A flag says the line breaks a rule; it does not say the line should not be paid. A hotel over the limit during a conference when every nearby room was over the limit is a sensible exception. Only the manager, who knows the trip, can judge that, and their reason is recorded so the pattern review can see how often each rule is set aside.
The expenses lead approves the rules and every change to them. If a rule is overridden most of the time, the rule is wrong, and changing it is a policy decision. The general principle, that agents apply policy evenly while people handle exceptions, runs through all expense work.
Where does it usually go wrong?
Vague policy. "Reasonable" hotel costs cannot be tested. Every such phrase becomes either no flag or a flag on everything. Rewrite the policy before switching the agent on.
Flag fatigue. Soft signals flagged as hard breaches flood managers, who start approving everything without reading. Keep hard flags for hard rules.
Managers approve every flag. The record will show it. The pattern review should name managers whose override rate is far above their peers, for the expenses lead to discuss.
Local rules are ignored. Allowances and costs differ by country and city. Rules need location bands, and the agent must read location from the receipt, not the employee's home office.
Attendee details are over-collected. Client entertaining rules need attendee names, but flags should not expose them more widely than the claim already does.
Everything is switched on at once. Start with a few rules and a few teams, then widen. Switching the agent's flagging on per team is a natural use of a feature flag, and the rollout approach is set out in how do you roll out an AI agent gradually.
Which controls matter, and what evidence do they leave?
Every flag cites a rule. Evidence: rule number and version on each flag.
Flags are factual. "Rule 4.2: meal above limit by 18" rather than "possible abuse". Evidence: flag text, sampled by the expenses lead.
Manager reasons on overrides. Evidence: the reason field on every approved flagged line.
Agent cannot approve, decline or edit. Evidence: refused verdicts on any attempt.
Pattern review each period. Evidence: the review and the expenses lead's actions.
What should you measure?
- Flags per hundred lines, by rule.
- Override rate by rule and by manager.
- Employee responses that led to a correction, against those that explained a valid exception.
- Time from submission to decision on flagged claims.
- Rules changed after the pattern review.
Questions people ask
Is a policy flag the same as a feature flag?
No. A policy flag marks an expense line that breaks a rule. A feature flag is a software switch for turning a capability on or off, which you might use to control where the agent's flagging is live.
Can the agent decline a claim that breaks a hard rule?
It flags; the manager decides. Some organisations auto-decline a short list of never-claimable items, but that is a policy choice with its own approval and record.
Should employees see flags before submitting?
Yes, if your tool allows it. A flag shown at entry time lets the employee correct or explain before the manager sees it, which saves everyone a round trip.