AI agents for accounts receivable

What an agent can own on the receivables line

Receivables is a line, not a department. An invoice is raised, sent, reminded, perhaps disputed, paid in whole or in part, allocated, and reported on. Most of the line is routine and follows rules. A little of it is judgement, and the judgement is concentrated at three points: when the customer pushes back, when cash arrives that does not match, and when it is time to escalate. The table sets out where an agent can own the step outright and where it should prepare and hand over.

The economic case does not depend on any single step. In the United Kingdom the Office of the Small Business Commissioner puts the cost of late payment to the economy at around eleven billion pounds a year, and most of that is not bad debt; it is money that arrived late because nobody asked for it on time, or asked the wrong person, or asked for the wrong amount. An agent that never forgets to ask, and only asks for what is actually owed, moves the receivable towards cash without anyone having to be diligent on a Friday afternoon.

Where it starts: the invoice has to be right

The largest lever on receivables is not collections, it is invoice accuracy. An invoice with the wrong purchase order number, the wrong contact, the wrong tax treatment or a line the customer does not recognise will not be paid on time whatever you do afterwards, because it goes into the customer's query pile rather than their payment run. An agent helps here because it raises the invoice from the record rather than from memory: the project or order it bills from, the contact marked as the accounts contact, the customer's reference, the tax rule for the customer's country.

This is also where a person's rules should be explicit. If invoices above a certain value need a second look before they go, tell the agent so and it will create the invoice, hold it as a draft and ask. If a customer requires a purchase order on every invoice, put that on the customer record and the agent will refuse to send without one. Rules like these are cheap to state and expensive to discover after the fact.

Collections without a collections department

For a small finance team the collections work is the part that slips, because it competes with everything else and because chasing people is unpleasant. An agent removes both problems for the routine stages. The design that works is a written chase policy and a standing request: each morning, find what is outstanding, send the message for the stage each invoice has reached, skip anything paused or disputed, and hand over the exceptions. The mechanics are covered in the companion piece on automating invoice chasing; here is what the receivables view of the same run looks like when the finance lead asks a broader question.

Notice the promise-to-pay entry. A customer who says they will pay on the twelfth is not an overdue account for the next week; they are a commitment to track. The agent records the promise, pauses the chase until the date, and resumes if the date passes with nothing received. That single behaviour is the difference between a chase that feels professional and one that feels automated.

Cash allocation: the unglamorous step that decides your numbers

Everything upstream is wasted if cash arrives and sits unallocated. Unallocated receipts make the ageing report wrong, which makes the chasing wrong, which annoys customers who have paid. The work is matching: a receipt for $6,160.00 from Acme with a reference of INV-1042 is a clear match and the agent should record it without asking. A receipt for $12,320.00 with no reference, from a customer with three open invoices, is not clear, and the agent should propose the most likely allocation and wait.

The rule to give the agent is a confidence rule, not a value rule. Record the payment when the amount matches one open invoice exactly, or the remittance names the invoice, or the customer has exactly one open invoice. Otherwise propose and ask. Part payments follow the same logic: record what arrived against the invoice named, leave the balance outstanding, and let the chase continue for the balance rather than the whole. Multi-currency receipts need the invoice currency and the received currency both recorded so the difference lands where the accountant expects it rather than as a phantom underpayment.

The reporting a person actually needs

Receivables reporting in small firms tends to be an ageing report nobody reads, because it lists everything and prioritises nothing. An agent that has been running the line can answer better questions. Which customers are habitually late, and by how much. How much cash is promised for this week and how reliable those promises have been. What share of invoices were paid by the due date this month compared with last. How many invoices needed a person, and how overdue they were when they did.

That last measure is the one to watch. Count exceptions, not reminders. If the agent is doing its job, the number of reminders is boring and steady while the number of invoices that reach a person falls, and the ones that do reach a person arrive earlier with their history attached. If instead the exception count rises, the policy is wrong or the record is wrong, and either is worth an hour of attention.

What has to be true of the software underneath

None of the above is a property of the agent. Claude, ChatGPT and their peers are all capable of reasoning about an overdue invoice. What decides whether the work can be handed over is whether the system holding the receivables exposes the right actions, with the right controls, to an agent acting for a named person.

• Balance outstanding per invoice, not just a status. Part payments and credit notes have to reduce the amount chased.
• The accounts contact on the customer record, so the agent sends to the person who pays rather than the person who bought.
• Email in the same system as the invoice, so a reply lands against the record and the agent can see that a customer has responded before it sends the next reminder.
• Tasks and pauses as data, so a promise to pay or a colleague saying leave it is something the agent reads, not something it is told in a chat and forgets.
• Per-user permissions enforced on every call, so the agent of someone who cannot record payments cannot record payments either.
• A log of every action, with inputs and result, so the month-end review of what the agent did is the same review you would do of a colleague.
• A path back to the ledger, such as a sync to the accounting package your accountant works in, so the receivables view and the books agree.

Sois is one system built this way. Invoicing, bills, multi-currency, regional tax and a Xero sync sit in the accounting module alongside contacts, the inbox and tasks, and the whole workspace is an MCP server. You connect the agent you already use with a single OAuth sign-in, it sees the tools your role allows and nothing else, and every call is checked again when it runs and written to the log. When your own agent does the reasoning, Sois performs no AI on your behalf and charges nothing for it; when you have no agent, the Chat Agent Gateway takes one plain-language message and does the work. Whichever system you choose, run the list above against it before you hand over the receivables line, because an agent given a system that fails any one of those conditions will produce exactly the embarrassing chase you were trying to avoid.

Questions people ask

This article is reviewed when the products it describes change. Next scheduled review: December 4, 2026.