Yes for clean matches; no for decisions after a mismatch. An AI agent can match bank receipts to open invoices and post them when amount, customer and reference agree. Short payments, overpayments, unidentified receipts, write-offs and credit notes are drafted by the agent and decided by a person, so applying cash and forgiving debt never share an actor.
Why is cash application suited to an agent?
Because most of the work is matching, and most matching is evidence-based. A receipt arrives with a payer name, an amount and a reference; a remittance may list the invoices it covers. The agent compares them with open items in the ledger. When everything lines up, there is no judgement involved, only care, and agents are careful at volume. The process is defined in what cash application is, and its place in the wider cycle in what order-to-cash is.
The hard part is the remainder: receipts with no reference, payments that cover some invoices but not others, amounts that are short by a deduction the customer took. These need context and sometimes a conversation with the customer.
Where is the separation of duties?
The risk in cash application is not mostly about the cash coming in. It is about what can be hidden in the difference. If one actor can apply a short payment and also write off the shortfall, it can make an unpaid balance disappear. The same goes for issuing a credit note to close a gap. So:
| Step | The agent | A person |
|---|---|---|
| Match receipt to invoice where amount, customer and reference agree | Applies it at depth 3, within limits | Reviews a sample |
| Match where the remittance lists several invoices | Applies it if the total agrees | Reviews a sample |
| Short payment | Applies what was paid; drafts a note on the shortfall | Decides: chase, dispute or write off |
| Overpayment | Holds it on account; drafts a refund or allocation | Approves the refund or allocation |
| Unidentified receipt | Suggests likely matches with reasons | Confirms the match |
| Write-off of a small difference | Drafts it with the evidence | Approves it, under a write-off limit |
| Credit note | Drafts it | Approves and issues |
The person approving write-offs and credit notes should not be the person who maintains the agent's matching rules either, so nobody can change both what is matched and what is forgiven.
An illustrative example
Take a wholesale distributor that receives a few hundred customer payments a week.
A receipt of 12,400 arrives from a retailer with a remittance listing four invoices that total 12,400. The agent applies the payment to all four and files the match with the remittance. A second receipt of 9,750 arrives with a remittance listing three invoices totalling 10,000 and a note: "less 250 for damaged goods". The agent applies 9,750 across the three invoices and leaves 250 open on one of them. It drafts a note for the credit control team with the remittance and the customer's deduction reason, and checks whether a goods return was recorded. None was.
A credit controller reviews it, contacts the customer, and agrees a credit note of 250 after confirming the damage. She approves the credit note, which the agent drafted. The agent then closes the item. Every step is in the record, with two different actors on the application and the credit.
What can go wrong, and which controls stop it?
Matching the wrong customer. Two customers with similar names, one paying for the other. The agent applies only where customer identity agrees; otherwise it suggests and holds.
Applying to the wrong invoice. A payment that matches an amount but not a reference may belong to a different invoice of the same value. Tolerance rules decide when amount alone is enough, and they should be strict.
Duplicate invoices. A customer may pay what looks like a single invoice twice because it was issued twice. The agent should flag the possible duplicate rather than apply the second payment against a different item; detection is covered in whether an AI agent can detect duplicate invoices.
Rule drift. Matching rules changed quietly to increase auto-match rates. Rule changes need approval and a change record.
Other controls:
- Write access limited to applying receipts, not to customer master data, credit notes or write-offs.
- A daily summary of applied, held and unmatched receipts for credit control.
- Sample review of auto-applied matches each week.
The full process, step by step, is in how to apply cash receipts to open invoices with an AI agent.
Where Pryme fits
In Pryme, bank receipts arrive as sealed facts from the bank feed and remittances as external or signed, depending on their source. Applying a matched receipt and approving a write-off are separate capabilities with separate holders, and every verdict is filed. More is in AI agents for finance: books you can audit.
Questions people ask
Can the agent read remittances from emails and portals?
Yes. It should treat their contents as information to match against the bank receipt, not as proof of payment.
What tolerance should auto-matching allow?
Set it low and specific: exact amounts by default, with small tolerances only for known causes like bank charges, approved by the finance lead.
Should the agent contact customers about short payments?
It can draft the query for credit control to send. Disputes and negotiations stay with people.