What AI actually does in collections
Take away the marketing and there are two jobs where AI models are useful in accounts receivable today:
- Writing. Producing a reminder that fits this invoice, this customer and this conversation, instead of filling the blanks in a template.
- Reading. Working out what a customer’s reply means (a promise, a dispute, a question, “already paid”) and pulling out the useful facts, like the promised date.
Some vendors go further, with AI voice agents that place calls or agents that log in to customers’ AP portals. Those can be useful at scale, but they’re newer and harder to supervise. For most small and mid-sized businesses, the two jobs above are where the time goes.
Writing: templates vs context
Template reminders fail in a familiar way. Day 30 says the same thing as day 14 with a firmer adjective. They ignore the reply the customer sent last week. They thank a customer for their business while chasing them for the third time. Customers learn to skim them.
A reminder written from context reads differently. Compare:
Template: “This is a friendly reminder that invoice #1052 for $4,410.00 is now 7 days overdue. Please arrange payment at your earliest convenience.”
From context: “Hi Dana, thanks for sorting out 1047 last week. Invoice 1052 for $4,410.00 was due on October 30 and I don’t see it yet. Is it in a payment run, or does it need a PO like the last one?”
The second one uses what’s known: the contact’s name, the last invoice, the fact that a PO caused the last delay. That’s what a careful person would write, and it’s easy to answer because it asks a real question.
Reading replies
Replies are where manual collections really costs time, and where AI helps most. Most replies to a reminder fall into a handful of types:
| Reply | Type | Sensible next step |
|---|---|---|
| “It’s in Friday’s payment run.” | Promise to pay | Pause until Friday; resume if nothing arrives. |
| “The hours don’t match our log.” | Dispute | Stop chasing that invoice; send it to a person. |
| “Can you resend with our PO?” | Question | Send to a person to fix and resend. |
| “We paid this on the 20th.” | Already paid | Hold reminders; check the books. |
| “I left in August, try AP.” | Wrong contact | Hold; ask for the right person. |
| “Out until the 10th.” | Out of office | Wait and retry. |
Notice that “Friday’s payment run” has to become an actual date. That’s the kind of small inference language models are good at and rule-based filters are bad at.
Where AI fails
Language models make mistakes, and in collections some mistakes are expensive. The common ones:
- Misreading mixed replies. “We’ll pay half Friday, but the second line item is wrong” is a promise and a dispute at once.
- Confident wrong facts. A model can state an amount or a date that isn’t in the invoice if it isn’t tightly constrained to the real data.
- Tone drift. Too casual with a formal client, too stiff with a friendly one.
- Deciding things it shouldn’t. Offering a discount or a payment plan nobody authorized.
None of these is a reason to avoid AI. They’re reasons to design around it.
Guardrails that matter
Before you let any AI tool write to your customers, check for these:
- Rules decide timing. The cadence should be something you set and can read. The AI writes; it doesn’t choose when.
- Approval mode. You see and approve every draft until you choose otherwise.
- A fallback to a person. Replies the model isn’t sure about should come to you, not get a best guess.
- Facts come from the books. Amounts, dates and invoice numbers should be pulled from your accounting data, not generated.
- No unauthorized offers. The AI shouldn’t promise discounts, extensions or plans you haven’t set up.
- A full log. Every message sent and every reply’s classification, so you can see what happened and why.
Get those right and AI does what it’s good at: the writing nobody has time for, and the reading that otherwise sits in an inbox until Friday.