AI collections agent: how to use AI to chase unpaid invoices

It helps to compare three things that often get lumped together.

Reminder rules are the scheduled emails built into accounting systems, covered in QuickBooks invoice reminders and Xero invoice reminders. They send the same message whatever the customer last said.

A chatbot can write a good reminder if you ask, but it does not know an invoice is overdue unless you tell it.

An AI collections agent starts from the invoice. It knows the due date from your accounting system, writes to the right contact, reads the reply, and works out what that reply means for the invoice. For how this fits into the wider receivables process, see our article on how AI is changing accounts receivable work.

These tasks are repetitive and easy to check, which makes them good work for AI.

• Draft and send follow-ups. A short, specific email with the invoice number, amount and due date, in a tone that changes as the invoice ages. Our collections email sequence shows the usual stages.
• Read and classify replies. Most replies fall into a few groups: already paid, disputing the invoice, promising a date, or "I'm not the right person". Each needs a different next step.
• Track promises. "We'll pay on the 15th" becomes a date to watch, with a polite follow-up only if the payment does not arrive.
• Keep the record. Every message, promise and change logged against the customer and invoice.

• Disputes. When a customer says the work was wrong or the amount is off, chasing should stop and someone who can fix it should reply. See our guide to invoice disputes.
• Payment plans and discounts. Agreeing to split a balance or waive a fee is a commercial decision. Our payment plan agreement guide covers how to put one in writing.
• Sensitive relationships. Your largest customer, or one in difficulty.
• Anything unusual. If the AI is unsure what a reply means, it should ask rather than guess.

A general assistant is a fast way to write better reminders, and you stay in control because you send every message yourself.

A warning about data. Do not paste in anything your contracts, privacy policy or customers would not expect you to share with an outside service. Check your company's rules and the assistant's data settings first. The prompts below use placeholders, so you can fill in names and details yourself before you send.

Draft a first reminder:

Write a short, polite email reminding a business customer that an invoice is now overdue.
Invoice: [Invoice number], amount [Amount], due [Due date], now [Number] days late.
Keep it under 90 words, include the invoice number and amount, and ask them to confirm a payment date.
No exclamation marks and no threats. Sign off as [Your name].

Read a reply and suggest the next step:

A customer replied to an overdue invoice reminder. Here is the reply with names removed: [Paste reply].
Tell me whether this is: already paid, a dispute, a promise to pay (and the date), a request for something (such as a PO or a copy), or a wrong contact.
Then draft a short response that fits.

The limits are practical. You still have to notice which invoices are overdue, remember who promised what, and check nobody has paid since you last looked. For a few late invoices a month, that is fine. If you would rather ask ChatGPT about your real invoices than paste them in, see ChatGPT for invoices, which covers connecting it through a plugin. Our guide to asking for payment politely has more wording to work from.

An agent connects to your accounting system for due dates and payments, and to your inbox so it can send and read email in your name. It follows up, reads replies, tracks promises and brings hard cases to a person. Some services pair the software with a team who oversee it, which suits businesses without a dedicated credit controller (as the role is called in the UK). The trade-off is setup and trust: you agree what the agent may do alone and what needs approval.

Use this checklist on any tool, including your own prompts.

1. Approval. Can you review messages before they go, and choose which situations always come to a person?
2. Audit trail. Is every email, reply, promise and change recorded against the customer and invoice?
3. Your voice. Can you edit the wording, and does it sound like your business rather than a template?
4. Stops when paid. Does it check for payment before every message, so nobody is chased for money already sent?
5. Stops on disputes. Does routine chasing pause when a customer raises a problem?
6. Data handling. Do you know where your customer data goes?

If a tool cannot answer the first four, keep it to drafting. Our page on collections software has a longer list of questions for vendors.