How an AI Agent Calls Customers About Overdue Invoices
How an AI agent calls customers about overdue invoices: the data it needs, what it says on the call, how a promise to pay is captured, and when a human takes over.
Pratheek Adi
Co-Founder & CTO
Table of contents
Share
More than half of US small businesses are owed money on unpaid invoices at any given time. The Intuit QuickBooks 2025 Small Business Late Payments Report found that 56% of US small businesses carry unpaid invoices, averaging about $17,500 each, and 47% have invoices more than 30 days overdue. Atradius puts the wider picture at 43% of credit-based B2B sales in the United States were overdue in 2025. Every one of those invoices needs a follow-up, and most finance teams do not have the hours to make the calls.
That is the job an AI agent that calls customers about overdue invoices is built for. This page explains, plainly, how it works: the data it needs, how it picks who to contact, what it says on the phone, what happens after the call, when it hands the account to a person, and which systems it connects to.
An AI agent that calls customers about overdue invoices is software that reads your accounts receivable aging, phones (and emails or texts) the right contact at each overdue account, references the specific invoices, asks for a payment date, records the answer as a promise to pay, a dispute, or a request, and hands anything that needs judgment to your team. It works the routine follow-up list every day so people only touch the exceptions.
What Is an AI Agent That Calls Customers About Overdue Invoices?
An AI collections agent is a voice and messaging program that runs the follow-up step of accounts receivable on its own. It is not a robocall that plays a recording. It holds a two-way conversation: it introduces itself, names the invoices, listens to the answer, responds to the common replies, and writes the outcome back to your system.
The category exists because follow-up is the part of collections that fails first: a small team is expected to chase hundreds of accounts by hand. Gartner reports that 59% of finance leaders now use AI somewhere in the finance function, and receivables follow-up is one of the first places it earns its keep, because the work is repetitive, scripted, and measurable.
Still asking whether an AI can call your customers at all? Start there. This page assumes the answer is yes and explains the mechanics.
What Data Does the Agent Need Before It Can Call?
The agent needs four things: an aging report, contact details, your payment terms, and a list of accounts it must leave alone. Everything else is configuration.
The aging report
The accounts receivable aging is the source of truth. It lists every open invoice by customer with the invoice number, date, due date, amount, and balance. Most teams already produce it; the agent just needs it daily. If you export the aging report from QuickBooks, from Xero, or from NetSuite, that file is enough to start.
Contact details
A phone number and an email address for the person who pays the bill, per customer. This is the constraint most teams underestimate: an accounting system often stores the project manager or the buyer, not accounts payable. The agent can only reach who it has, so a contact cleanup in the first two weeks matters more than any setting.
Terms and a do-not-contact list
The agent needs to know the terms (net 30, net 60) so it can calculate days overdue correctly, and it needs a list of exclusions: accounts in dispute, accounts with a payment plan, holdbacks that are not yet due, and any customer you have decided to handle personally. A good agent honors both a per-customer pause and a per-invoice exclusion, because the two are different problems.
How Does the Agent Decide Who to Contact and When?
The agent works from rules you set, not from a mood. The typical rule set has five parts:
Days overdue: contact starts at a threshold you choose (many teams start at 30 or 60 days past due, some sooner).
Minimum balance: invoices under a floor you set are skipped, so nobody gets a call about a few dollars.
Cadence: how many touches per customer per week, and how many days between them.
Channel order: email first and call second, call first, or all channels, per account or per segment.
Calling window: business hours in the customer's time zone, never outside them.
Credit Pulse's 2025 benchmark data found that automated reminders that begin before the due date outperform post-due follow-up by 12 to 18 days of DSO. The cadence should be steady and early, not aggressive and late; consistency, not tone, is what moves the aging.
Each morning the agent re-reads the aging file, drops anything paid or excluded overnight, and builds the day's list. A customer that paid yesterday is never called today.
What Does the AI Agent Actually Say on the Call?
The agent speaks as a named member of your team, states the reason for the call, references the specific invoices, and asks for a date. Here is a plain version of a first call, the kind a well-configured agent makes:
The opening
"Hi, this is Kate calling from General Contracting Co. I'm reaching out about invoice 4471 from June 12th for $4,250, which is now 45 days past due. Do you have a moment?"
The opening is short and specific: company, invoice, amount, age. No threat, no pretending to be something it is not.
The ask
"Can you tell me when we can expect payment on that invoice?"
Then it listens. Most answers fall into a handful of buckets: it is scheduled, it was already paid, we never got the invoice, there is a problem with it, the wrong person is on the line, or please stop calling. The agent has a response for each. If the customer gives a date, the agent confirms it back: "Great, I'll note that for Friday the 18th. If anything changes, you can reach our accounts team at 555-0100."
What a good agent never says
It never claims a customer is on credit hold or that service has been cut unless your team has actually done that. It never argues about a dispute; it records it. It never calls a number that has been removed. And it never invents a discount, a payment plan, or a deadline you have not approved. Every consequence the agent mentions is one you configured, in words you wrote or approved.
If you want to hear it rather than read it, the standard sample is on our homepage. Curious what this sounds like in practice? Here's a 98-second sample call: https://collectionshub.net/#live-demo
What Happens After the Call?
After every conversation the agent writes an outcome, updates the account, and schedules or cancels the next touch. Your team sees the result in a dashboard and an end-of-day summary; the customer gets a short email confirming what was agreed. The common outcomes:
Outcome | What the agent does | What your team sees |
|---|---|---|
Promise to pay | Confirms the date, sends a written recap, pauses outreach until the date passes, resumes if unpaid | Promise date on the account, flagged if it slips |
Already paid | Asks for the date and method, marks the invoice for reconciliation, stops outreach | A reconciliation task with the customer's payment details |
Needs an invoice copy | Sends or requests the PDF, resets the follow-up clock | A "copy sent" note, or a task if your ERP cannot export it |
Disputed invoice | Records the reason in the customer's words, stops calling about that invoice, escalates | A dispute ticket with the transcript, assigned to a person |
Wrong contact | Asks for the right name and number, updates the record, retries | A corrected contact, visible in the log |
Voicemail or no answer | Leaves a short message with your callback number, retries later in the cadence | Attempt count and next scheduled touch |
The recap email matters as much as the call. A customer who agreed to Friday now has it in writing, from the same name that phoned them, with the invoice attached. That is the step busy collectors skip, and the step that turns a verbal promise into a paid invoice.
When Does the Agent Hand Off to Your Team?
The agent hands off whenever judgment is needed, whenever the customer asks for a person, and whenever a limit you set is reached. Those three triggers cover almost every case:
Judgment calls
A disputed amount, a request for a payment plan, a customer who says the work was not finished, a large account that is suddenly silent. The agent records what it heard and stops; a person decides what happens next. The agent is allowed to ask for money. It is not allowed to negotiate, threaten, or make exceptions.
An explicit request for a human
"I need to speak to a real person" ends the automated conversation. Depending on configuration, the agent gives the customer a direct line, schedules a callback from a named colleague, or, where live transfer is set up, moves the call to your desk. It never pushes the customer to stay on the line.
Hard limits you set
Maximum attempts per invoice, a ceiling on the age of invoices it may chase, a dollar threshold above which every account is human-first, and a one-line pause command ("pause this customer") your team can send at any time. If you are weighing whether an AI caller will hurt your customer relationships, those limits are the honest answer: the agent is only as pushy as the rules you give it.
Which Systems Does It Connect To?
Most AI collections agents connect in one of three ways: a native integration with the accounting or billing system, a daily export of the aging report, or a spreadsheet you maintain. All three work; they differ in how much reconciliation your team does by hand.
Native integration: the agent reads open invoices and payments directly, so a payment posted this morning stops a call this afternoon. Typical targets are QuickBooks Online, Xero, NetSuite, Sage Intacct, Dynamics 365 Business Central, and subscription billing such as Chargebee.
Daily aging export: your team or a scheduled report sends the aging file each morning; the agent reconciles it against yesterday and builds the day's list. Common where the ERP is on-premise or has no API.
Spreadsheet feed: a normalized sheet with one row per invoice, kept by finance, sometimes fed by a connector. The most flexible route and the one that needs the most discipline.
Whichever route you take, the agent should write back: outcomes, promise dates, updated contacts, and disputes belong in your system of record or your dashboard, not in a silo. If you are comparing vendors on exactly this, our buyer's guide to AI voice agents lists the questions to ask, and our roundup of the best AI collections software covers where each tool fits.
How Does Abivo's Kate Do This?
Abivo's agent, Kate (renamed per client), runs every step above as one worker: she calls, emails, and texts from a name and an email domain branded as part of your team, references the specific overdue invoices, captures promises to pay and disputes, sends the written recap, and routes anything that needs a decision to your inbox. Your collectors are copied on her emails from day one, and any of them can pause a customer by replying to Kate or clicking Disable in the platform.
Kate connects to QuickBooks, Xero, NetSuite, Sage Intacct, Dynamics 365, Chargebee, and Flywire, and she accepts a daily aging export or a normalized spreadsheet when a native connection is not available. Each day ends with a summary of every conversation and what was agreed. For an example of what this looks like on a real receivables book in fire protection and commercial services, read the OFS Group case study.
The first two weeks are contact cleanup and rule tuning; the aging starts to move after that.
Practical Takeaways Before You Switch One On
Fix the contact list first. The agent can only reach the accounts payable person you give it; a project manager's number is not a collections contact.
Start with one segment. A single branch, one invoice type, or invoices over a certain age. Prove the cadence before you widen it.
Write the exclusions down. Disputes, payment plans, holdbacks, and the accounts you handle yourself, in a list the agent reads every morning.
Approve every consequence sentence. If the agent may mention credit holds or service interruption, you decide the wording and the day it may say it.
Copy your team on everything for the first month. It is the fastest way to catch a wrong invoice reference or a tone you dislike.
Measure one number. Percentage of the book over your overdue threshold, weekly. If it is not falling after six weeks, the rules are wrong, not the customers.
Frequently Asked Questions
Does the AI agent tell customers it is an AI?
Yes, when asked, and the best configurations say so plainly. The agent introduces itself by name as part of your team, and if a customer asks whether they are speaking to a person, it answers honestly and offers a human contact.
What if a customer says the invoice was already paid?
The agent asks for the payment date and method, thanks the customer, marks the invoice for reconciliation, and stops outreach on it. Your team confirms against the bank or the payment processor. If the payment cannot be found, a person follows up, not the agent.
Can the agent call customers in a different time zone?
Yes. Calling windows are set per customer or per region, so a customer in Vancouver is called during Vancouver business hours even when your office is in Toronto.
How is a promise to pay tracked?
The agent records the promised date on the account, sends the customer a written recap, and pauses outreach until that date. If the invoice is still open the day after the promise, the agent resumes with a reference to the earlier conversation, and your team sees the slip in the daily summary.
Want to see how this would run on your receivables? Head to Get Started and we will walk through your aging report with you.
