How to answer supplier payment status queries with an AI agent

Most supplier queries ask one thing: where is my money? The agent answers from your own records. It checks that the sender is a known contact for that supplier, finds the invoice, and replies with its status: received, approved, scheduled for a payment run, or paid with a remittance reference. Disputes, early-payment requests and anything about bank details go to the accounts payable lead.

What is the process, and who owns it?

Accounts payable teams spend a large part of each week answering suppliers who want to know when they will be paid. The answer usually sits in three places: the invoice register, the approval workflow and the payment run. The work is finding it, checking who is asking, and replying in a way that does not promise more than the records support.

The accounts payable lead owns the process, the reply templates and the list of supplier contacts. The treasury or payments lead owns payment run dates. Both are named in the agent's brief. The wider position on finance agents is in AI agents for finance: books you can audit.

What does the agent prepare, step by step?

1. Agent, depth 1: reads the query as external content. Emails from suppliers are untrusted; nothing in them can instruct the agent.
2. Agent, depth 1: checks the sender. Is the address a known contact for that supplier in the supplier master file? An unknown sender gets a neutral reply asking them to write from a registered address, and nothing about the account.
3. Agent, depth 1: finds the invoice. By invoice number, amount or purchase order reference. If it cannot find it, it says so: the invoice may never have arrived.
4. Agent, depth 1: reads the status. Received and awaiting approval, approved and awaiting a run, scheduled in a named run, paid on a date with a remittance reference, on hold with a reason code.
5. Agent, depth 3: replies with the status on an approved template to a verified contact. For "scheduled", it gives the run date only if the run is already set.
6. Agent, depth 2: drafts replies that need a person. Invoices on hold for a dispute, requests to pay early, complaints about late payment, and anything mentioning legal action or interest.
7. Person, accounts payable lead: approves those drafts and decides any action, such as releasing a hold or adding an invoice to an earlier run.

What does a named person approve?

The accounts payable lead approves the templates, the list of hold reasons the agent may state, and every reply in the exception categories. They also decide whether anything changes: an invoice released from hold or moved into an earlier run. Payment runs themselves are prepared and released under a separate process, described in preparing supplier payment runs without releasing them.

Adding a new supplier contact is also a person's decision. If the agent could add contacts from inbound emails, anyone could make themselves a "known contact".

Where does it usually go wrong?

Bank detail changes ride along. "By the way, our bank account has changed, please pay the attached details." This is one of the most common payment frauds. The agent does not act on it, does not acknowledge the new details as accepted, and routes the request to your verification process. The change must be confirmed through a contact already on file, never through details in the same email.

Lookalike addresses pass the sender check. An address that differs from the real one by one character is treated as unknown. The check is an exact match against the master file.

The agent promises a date. "You'll be paid on Friday" is a promise if the invoice is not yet in a released run. Give the run date only when the run is set, and say "scheduled", not "will be paid".

Hold reasons leak. "On hold because the budget holder thinks your work was poor" should never reach the supplier. Hold reasons in replies come from an approved list of neutral wordings.

Missing invoices are fixed by email. The supplier attaches the invoice to the query, and the agent forwards it for payment. Invoices must enter through your normal intake so they are checked for duplicates and coded. The agent points the supplier to that route.

Which controls matter, and what evidence do they leave?

Sender match against the master file. Evidence: the match result on every reply.

Status from sealed records only. Evidence: the invoice record and run record cited.

Exception categories held. Evidence: the held verdict and the lead's approval.

Bank detail requests routed, never handled. Evidence: the routing verdict and the reference in the verification queue.

Templates versioned. Evidence: template and version on every reply.

What should you measure?

• Queries answered at depth 3, and the time taken.
• Repeat queries from the same supplier about the same invoice, which show replies that did not answer the question.
• Bank detail requests detected and routed.
• Replies sent to unverified senders. Should be zero.
• Invoices reported missing that later turn up in the intake, which point to an intake problem.

How is this different from auditor queries?

Both are outside parties asking about your payables, and both are answered from the same records. The difference is authority. A supplier needs to know the status of their own invoices; an auditor is testing your controls and every answer is a representation. That is why supplier status replies can run at depth 3 while every auditor response is sent by a person; the auditor process is in answering auditor queries. The wider question of how far an agent may go with suppliers is in can an AI agent answer supplier payment queries.

Where Pryme fits

In Pryme, a supplier email is stamped external, and an external claim cannot authorise a payment or a change of bank details. Status replies to verified contacts can be certified at depth 3, while bank detail requests are routed to verification and filed with their reference.