Expense tools expose each function separately, so grant them one by one. The agent may read claims and card transactions, attach and read receipts, categorise lines, flag policy breaches and return a claim to the employee for a fix. Approving or rejecting a claim, changing amounts, card controls and reimbursement stay with people. Every action is recorded against the claim.
What does the connection involve, and who owns it?
An expense management tool collects card transactions and out-of-pocket claims, holds receipts, applies policy rules, routes claims to managers and sends approved amounts for reimbursement. Each of those is a function exposed through the tool's integration interface, and each is a separate capability for the agent. The general pattern of an agent calling a system's functions is explained in what is tool use.
The expenses lead owns the tool, the policy rules and the agent's grants. Line managers approve their teams' claims. Payroll or accounts payable handles reimbursement. The data protection owner cares about what the agent can see, because expense data reveals where employees go and with whom. The wider finance position is in AI agents for finance: books you can audit.
Which permissions should you grant, and at what depth?
| Action in the expense tool | Depth | Condition | Stays with a person |
|---|---|---|---|
| Read claims and card transactions | 1 | Within the entities in scope | Nothing |
| Read receipts | 1 | Same | Nothing |
| Attach a receipt to a transaction | 3 | Amount and date match | Mismatches |
| Categorise lines | 3 | From the category list | New categories |
| Add a policy flag or comment | 3 | Rule cited | Nothing |
| Return a claim to the employee for correction | 3 | Factual errors only | Disputed returns |
| Send reminders | 3 | On the approved schedule | Changing the schedule |
| Approve a claim | None | Refused by your limits | Line manager |
| Reject a claim | 2, held | Always held | Line manager or expenses lead |
| Edit claimed amounts | None | Refused by your limits | Employee, on return |
| Freeze or change card limits | None | Refused by your limits | Expenses lead |
| Export for reimbursement | None | Refused by your limits | Payroll or payables |
| Edit policy rules | None | Refused by your limits | Expenses lead |
Returning a claim is different from rejecting it. A return says "this needs a fix", such as a missing receipt or a wrong category, and the employee can resubmit. A rejection says "we will not pay this". The first is routine and reversible; the second affects someone's money and is a judgement. Whether an agent could ever take those decisions is discussed in can an AI agent approve or reject expense claims.
How do you set it up, step by step?
- Person, tool administrator: creates a service account for the agent, clearly named, with only the functions in the table.
- Person, expenses lead: writes the policy as checkable rules. Limits per category, receipt thresholds, approval routes. The agent cites rule numbers in every flag; how well agents can apply a policy is discussed in can an AI agent enforce an expense policy.
- Person: sets the agent's own limits on the platform to match the table.
- Agent, depth 1: runs read-only for a cycle, proposing flags and categories without writing them.
- Attest: replays past claims, scoring categorisation, receipt matching and policy flags against the floor.
- Person: enables writes per capability once the agent's certificate is issued.
What stays with a person?
Every decision about whether money is paid, and anything that affects an employee's card. A rejected claim may leave someone out of pocket. A frozen card may leave them stranded on a trip. Those decisions need someone who knows the circumstances and can talk to the employee.
Policy also stays with people. If the agent could edit rules, it could make its own flags disappear.
What is recorded for each action?
Each action is filed as a verdict in Pryme: the claim, the line, the action, the rule cited, the bounds checked and the one that decided. The expense tool's own history shows the change under the agent's service account. Flags and comments are visible to the employee, so they must be factual and cite the rule, not speculate about intent.
Two kinds of entry need regular review. Rejections drafted and held, because they show where policy and practice diverge. And refused attempts, such as an approval the agent tried and the platform refused, which suggest the agent's brief needs correcting.
How do you revoke access?
Remove a function from the service account and it stops in the tool. Withdraw a capability on the platform and the agent stops attempting it. Disable the service account to stop everything. A certification run that clears no depth drops the agent to depth 1 the same day: it can still read, but no longer attaches, categorises or flags. Test each level before go-live.
Where does it usually go wrong?
Flags read as accusations. "Possible personal expense" on a colleague's dinner claim reads as an allegation. Word flags as rule checks: "Rule 4.2: meals above the limit need a manager note."
Receipts are matched loosely. A receipt for a similar amount on another day closes the wrong gap. Require date and amount to match within tight bounds.
Location data is over-used. Receipts reveal where people were. Use them to support claims, not to build pictures of employees' movements.
The connection overlaps with others. Card spend can appear in the expense tool and the AP tool. Keep grants consistent with connecting an AI agent to your accounts payable automation tool, and report through the governed route in connecting an AI agent to your BI tool.
Questions people ask
Can the agent pre-approve claims for managers?
It can mark a claim as having passed every policy check, which speeds up the manager's review. The approval itself stays with the manager.
Should employees know an agent is reviewing their claims?
Yes. Tell them in the expense policy and in the agent's comments. Under the EU AI Act, people must be told when they interact with an AI system, and transparency with staff is good practice anyway.
Can the agent split a claim that mixes business and personal costs?
It can propose a split and return the claim to the employee to confirm. It does not change the claim itself.