No. Separation of duties means whoever prepares a payment does not approve it, and that applies to agents as it does to people. A second agent built on the same brief and model does not count as independent either. The agent may prepare and pre-check the payment; a named person approves it, and another person releases it.
Why doesn't "the agent checked it" count as approval?
Approval exists so that a second mind looks at a payment with its own information and its own accountability. If the preparer approves its own work, the second look never happens. Mistakes in preparation pass straight through, and so does anything deliberate.
An agent checking its own payment is the same mind looking twice. It will make the same reading of an ambiguous invoice, trust the same sources and miss the same signals. Its checks are useful, and they should run, but they are part of preparation, not approval.
What about two different agents?
It is tempting to build a "preparer agent" and an "approver agent" and call that separation. It is weaker than it looks. If both agents use the same model, see the same data and follow similar briefs, their errors are correlated: the invoice one misreads, the other probably will too. And neither can be held accountable for approving a payment.
A second agent can add value as a reviewer, checking for duplicates, bank detail changes and policy breaches. Its findings go to the person who approves. But the approval itself, the act of saying "this payment should go", belongs to someone who can answer for it.
What may the agent prepare, and what stays with people?
The agent, at depth 2: - selects approved, due invoices; - runs duplicate, bank detail and amount checks; - builds the payment file; - writes a summary of exceptions for the approver; - drafts the bank instructions.
People: - the invoice approver confirms the business owes the money; - the payment run approver reviews the run and its exceptions; - two releasers send the payment at the bank.
The step-by-step process is in how to prepare supplier payment runs without releasing them, and the bank file stage in how to prepare bank payment files for release.
An illustrative example
Take a mid-sized software company that wants to speed up supplier payments. Its first design has one agent processing invoices, approving those under 5,000 and adding them to the payment run.
The finance controller rejects it. Under that design, the agent would read an invoice, decide it is valid and authorise payment, all on its own reading. A spoofed invoice under 5,000 from a lookalike supplier would be paid without any person seeing it.
The revised design has the agent process invoices and route each to the budget holder, who approves in the AP tool. The agent then builds the payment run, runs its checks and writes a one-page exception summary. The payments manager approves the run. Two finance staff release it at the bank. The agent still does most of the work, and the run is faster than before, but every payment has a named approver who is not the preparer.
What controls make it safe?
| Control | Evidence it leaves |
|---|---|
| Preparer and approver are different identities | Both recorded on each payment; the system refuses a match |
| Agents cannot hold approval rights | Role scope on file; approval permissions held only by people |
| Approver independent of the invoice requester | Authority matrix check recorded |
| Release by two people | Bank audit log |
| Exception summary reviewed | Approver's sign-off on the summary |
| Every verdict filed | Allowed, held and refused payments on record |
Where Pryme fits
In Pryme, the minimum chain checks every action against role scope, certificate, sources, the customer's limits and the platform guardrails. An agent's role scope can include preparing a payment; approving a payment it prepared is refused, and the verdict names role scope as the bound that decided.
The full rules for run approval are in whether an AI agent can approve a payment run, with the review steps in the payment run review checklist. The wider controls are in AI agents for finance: books you can audit.
Questions people ask
Can the agent approve small payments if a person approves the invoice?
The invoice approval is the person's decision that the money is owed. The payment still needs approval and release by people, though many teams use lighter review for small, routine runs.
Does this apply to refunds and payroll too?
Yes. Any money leaving the business needs a preparer and a separate approver. The same agent cannot be both for a refund, a payroll run or a transfer.
What if we only have one finance person?
Use a second person outside finance, such as a director, as approver. Small teams often find this the hardest control to maintain, which is a reason to keep it, not drop it.