Can an AI agent chase your unpaid invoices?

Can an AI agent do accounts receivable?

It can do the chasing; it should not decide what a customer owes. An AI agent for accounts receivable can watch your aging report, send reminders from templates you approve, call customers, log promises to pay and propose which payment belongs to which invoice. Agents do this in production today, according to the vendors that sell them. Stuut reports that its agent handles 95% of collections outreach for EZG Manufacturing, whose two-person AR team works on collections part-time. Abivo reports that its voice agent let OFS Group go from four callers to one. Both figures come from the vendors' own case studies.

What stays with you is every decision that changes what the customer owes or how you treat them: disputes, short payments, payment plans, credits, write-offs and handing an account to an agency or lawyer. Treat this as a supervised hire. The agent works the list; a person owns the exceptions.

This page covers collections and cash application for invoices you have already sent. For the supplier side see accounts payable, and for the wider role see bookkeeper.

What the agent does and what stays with you

Agent

3of 10

• Monitor open invoices and aging
• Send due-date and overdue reminders from approved templates
• Log replies and promises to pay

Agent with review

4of 10

• Write non-routine collection emails
• Call customers about overdue invoices
• Propose matches between payments and invoices (cash application)
• Prioritize accounts and recommend escalations

Human

3of 10

• Resolve disputes and short payments
• Agree payment plans and credit extensions
• Approve credits, write-offs and handover to an agency or lawyer

The split follows one rule: the agent can do anything that follows an approved script and can be checked against the ledger, and a person decides anything that changes a balance or the terms. Reminders on a cadence are the clearest case. Stuut describes EZG rolling out in phases, starting with due-date reminders, then weekly emails, then automated phone calls. That order is a sensible one to copy: let the agent prove itself on low-risk messages before it speaks to customers on the phone.

Free-form writing is where review matters. Microsoft's Copilot in Dynamics 365 Finance can draft a reminder email from a customer's overdue invoices and payment history, but Microsoft calls the draft a suggestion and says it is your responsibility to check that it is complete, accurate and appropriate before you send it. Apply the same standard to any agent that writes its own wording.

Cash application sits in the middle. Invoiced by Flywire says its CashMatch AI attaches a confidence score and a short-pay flag to every match and posts above a threshold you set. Treat everything below that threshold as your team's work. In our view, a short payment deserves a person's look too, because it is often the first sign of a dispute.

A worked example: one overdue invoice

This is an illustration, not a customer story. Assume, for this example, that your cadence sends a reminder 7 days after the due date and a call at 21 days.

Invoice 2207 for $6,400 to a wholesale customer is 7 days past due. The agent sees it in the aging report, checks that the account is not on the do-not-contact list and that the invoice is not marked disputed, and sends the 7-day template from the AR inbox with the invoice attached. Nothing comes back. At 21 days the agent calls the billing contact on file. The customer's AP clerk says they paid $5,200 last week and are holding $1,200 because one pallet arrived damaged.

The agent logs the call, finds the $5,200 payment sitting unmatched, proposes matching it to invoice 2207 as a short payment and marks the remaining $1,200 as disputed, which stops further reminders. Your AR owner gets the dispute with the call summary and the delivery note. She checks with the warehouse, agrees a $1,200 credit and approves it in the accounting system. The agent did the chasing and the paperwork; the person decided the credit.

What you need to provide

The agent touches customer data (names, emails, phone numbers) and financial data (balances, payment history, bank remittances). Give it the least access that does the job:

• Read access to open invoices, contacts and payment history, and the right to log activity, but not to post credits, write off balances or change terms.
• A shared AR mailbox to send from and receive replies in, so your team sees every conversation.
• Your cadence, templates and escalation rules in writing, plus a do-not-contact list. Intras, which builds collections agents, describes the same pattern: reminders go out only on templates the team approved, and disputes go to a named owner.
• If the agent calls customers, a phone number and a decision on consent (see the rules below).

Two sets of US rules are worth raising with counsel before you start. First, the FDCPA and the CFPB's Regulation F govern debt collectors collecting consumer debts. Regulation F defines a debt as an obligation arising from a transaction primarily for personal, family or household purposes, and excludes from "debt collector" a creditor's own officers and employees collecting in the creditor's name. It also says a creditor that collects its own debts under a name suggesting a third party is a debt collector. Whether your agent's outreach falls under these rules depends on who your customers are and whose name the agent uses. Second, the FCC ruled in February 2024 that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" cover AI technologies that generate human voices, so such calls require the prior express consent of the called party absent an emergency purpose or exemption. Ask counsel which obligations apply to your customers and jurisdiction.

Job brief template

Use this as the brief for a vendor, an operator or an internal team setting up the agent. Replace the bracketed values with your own.

The acceptance criteria to keep are the zero-tolerance ones: no wrong balance, no message to a held account and no dispute left unrouted. Even a high overall accuracy rate can leave individual customers with an incorrect amount; check sampled messages against the ledger.

Risks and human checkpoints

The main risk is not that the agent fails to collect. It is that it chases the wrong thing politely and often: a paid invoice, a disputed amount, a customer you were handling carefully. We recommend sampling sent messages against the ledger each week and counting "we already paid" replies; neither check catches every error, but both can help surface possible cases of the agent chasing the wrong invoices for review.

The second risk is that disputes stop reaching people. A good agent logs everything, which makes it easy for a dispute to sit in a log nobody reads. Make the routing to a named owner part of the acceptance criteria, and read the inbox yourself during the first weeks.

How to run a trial

Run the agent on a slice for one month before you hand it the whole ledger.

1. Pick 20–50 customers with routine payment behavior. Leave out key accounts and anyone in a dispute.
2. Start with emails only, on your existing templates. Add calls later, after counsel has confirmed the consent position.
3. Keep your current process running for customers outside the slice, so you can compare.
4. Each week, check 50 sent messages against the ledger, and list every reply, promise to pay and dispute with the date it reached a person.
5. If cash application is in scope, review every proposed match before it posts for the first two weeks, and count corrections.
6. Track the numbers you already report: overdue balance in the slice, invoices over 60 days and hours your team spent on the slice.

1. 1

   Pick 20–50 customers

   Routine payment behavior; leave out key accounts and anyone in a dispute

2. 2

   Start with emails only

   On your existing templates; add calls later, after counsel has confirmed the consent position

3. 3

   Keep your current process running

   For customers outside the slice, so you can compare

4. 4

   Each week, check 50 sent messages

   Against the ledger; list every reply, promise to pay and dispute with the date it reached a person

5. 5

   Review every proposed match

   If cash application is in scope, for the first two weeks, and count corrections

6. 6

   Track the numbers you already report

   Overdue balance in the slice, invoices over 60 days and hours your team spent on the slice

If the messages are error-free for two weeks and every dispute was routed within a day, widen the slice. For more on structuring the test, see how to run a paid trial with an AI agent, and for the brief itself, how to write a job description for an AI agent. Related finance jobs: bookkeeping and reconciliation, month-end close and the finance overview.

Questions buyers ask

What does it cost compared with a collector? The human side is published: O*NET OnLine lists a 2025 median wage of $47,030 a year ($22.61 an hour) for bill and account collectors (SOC 43-3011), from BLS data, wage only. The agent side is not: Abivo describes its plans as a flat monthly rate based on active contacts and amount in collections and asks you to get in touch, and Stuut and Invoiced by Flywire did not show prices on the pages checked. So there is no cost comparison on this page. Ask vendors for the price at your volume and compare it with the hours your team spends chasing.

Will customers know they are dealing with an AI? That depends on how you set it up and on the rules where you operate. For AI-voice calls in the US, see the FCC ruling above.

Does this reduce days sales outstanding? Vendors say so: Stuut reports a 5-day DSO improvement at EZG, Abivo reports more than 30 days at OFS Group, and Invoiced by Flywire says its clients see an average 14-day improvement. These are vendor figures without independent checks. Measure your own in the trial.

This page is not financial, tax or accounting advice.

Sources

1. 1
   Collections coordinator summary

   Microsoft LearnPrimaryAccessed Nov 14, 2026

2. 2
   43-3011.00 - Bill and Account Collectors

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

3. 3
   12 CFR 1006.2 Definitions (Regulation F)

   Electronic Code of Federal Regulations (eCFR)PrimaryAccessed Nov 14, 2026

4. 4
   Declaratory Ruling, FCC 24-17: Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts

   Federal Communications CommissionPrimaryAccessed Nov 14, 2026

5. 5
   EZG Manufacturing case study

   StuutVendorAccessed Nov 14, 2026

6. 6
   How OFS Group Recovered $842K with AI

   AbivoVendorAccessed Nov 14, 2026

7. 7
   Agentic AI for accounts receivable automation

   Invoiced by FlywireVendorAccessed Nov 14, 2026

8. 8
   AI collections agent for accounts receivable

   Intras AIVendorAccessed Nov 14, 2026

9. 9
   Abivo pricing

   AbivoVendorAccessed Nov 14, 2026