Set them per capability, below the delegated authority your people already hold. A finance agent needs one limit for each thing it does, such as coding invoices, proposing payments or drafting accruals, with a value cap, a cumulative cap, a counterparty rule and a named approver above the cap. The limit is checked at the moment the agent acts.
Why can't the agent borrow a person's limit?
Your delegation of authority was written for people. A payables clerk with a limit of 5,000 also has judgement and a manager. An agent can repeat the same action hundreds of times an hour, so borrowing the clerk's limit gives it the clerk's authority at a machine's pace. The agent's limits sit inside your delegation matrix, lower and narrower, and every amount above them goes to a person who already holds that authority. The wider design is in AI agents for finance: books you can audit.
What should an approval limit contain?
An amount on its own is not a limit. For each capability, write down seven things:
- The action. "Propose a supplier payment", not "payments".
- The value cap per item. The largest single amount the agent may act on.
- A cumulative cap. A daily total per supplier and overall, so ten items of 900 cannot pass where one of 9,000 would not.
- The counterparty rule. Only suppliers on the approved master file, with bank details unchanged for a set period.
- The evidence required. A matched purchase order and goods receipt, for example.
- The approver above the cap. A named role from the delegation matrix, not "the finance team".
- The review date. When the limit is next looked at, and by whom.
The cumulative cap does the most work. Splitting is the oldest way round a per-item limit, and an agent can be steered into it by a supplier's emails. Our guide to detecting split purchases below approval limits covers the patterns to watch for.
What can the agent do, and what stays with a person?
| Capability | The agent | Stays with a person |
|---|---|---|
| Code supplier invoices | Codes matched invoices under the cap (depth 3) | Unmatched invoices, new cost centres |
| Propose a payment run | Drafts the run (depth 2) | Releasing the run |
| Accruals | Drafts the journal with support (depth 2) | Posting anything above materiality |
| Customer refunds | Drafts the refund (depth 2) | Approving it and anything to a new account |
| Payroll changes | Prepares the change with evidence (depth 2) | Approving and releasing it |
Preparation is agent work and release is human work. The approval matrix for agent-prepared payments sets out who signs at each value band. Refunds and payroll follow the same shape; see how to set refund limits for an AI agent and how to let an AI agent prepare payroll changes for approval.
How do you choose the numbers?
Start from your own transactions, not a round figure that feels safe. Pull a quarter of history and set the first cap at the top of the routine, well-evidenced band. Raise it only on evidence: the agent's results on your own past cases at the higher value, reviewed by the process owner. A cap that rises because the team is busy will be breached.
A worked example
Take a mid-sized wholesaler. The agent may propose supplier payments up to 2,000 per item and 10,000 per supplier per day, to suppliers whose bank details have not changed in 60 days.
It proposes 1,800 to a regular haulier with a matched order and receipt, and the payables manager releases it. An hour later a second invoice from the same haulier, for 1,900, quotes the same delivery reference. It is under the item cap, but the duplicate reference trips the counterparty rule. The verdict is held, the record names the customer's own limit as the deciding bound, and the manager sees both invoices side by side.
Which controls make the limits hold?
- Enforce at the moment of action. A limit written in the agent's instructions is a request. A limit checked by the platform before the call reaches the ledger is a control.
- The lowest bound wins. Your limit is one of several. If the agent's certificate allows less than your limit, the certificate decides.
- The agent cannot change its own limits. Changes go through change control, with the process owner and a second approver.
- Every verdict is filed. Allowed, held and refused decisions are recorded with the bound that decided, so an auditor can test the limit, not just read it.
- Limits expire. Review each one at least quarterly, and lower it the same day if a test fails.
Where Pryme fits
In Pryme, your approval limits are one of five bounds every action meets, alongside role scope, the agent's certificate, its sources and the platform guardrails. The lowest bound decides. Attest evidence from your own cases sets how deep each capability may go, and every verdict names the bound that decided it in an append-only record.
Questions people ask
Should the agent's limit match the person it supports?
No. The agent's limit should be lower, because it acts faster and without judgement. Treat the person's limit as a ceiling.
Can limits differ by supplier or entity?
Yes, and they often should. A long-standing supplier with stable bank details can carry a higher cap than a new one, and a subsidiary with weaker controls may need a lower cap than the parent.
What happens if we lower a limit in the middle of a run?
The next action is checked against the new limit. Anything already drafted but not released stays with the approver, who sees it against the lower figure.