AI Agent Accounts Receivable: What It Can Actually Collect

AI agent accounts receivable work splits into two very different jobs, and that split matters more than any accuracy number a vendor quotes you. One job is reading an invoice correctly, matching it to the right account, and drafting a reminder that sounds like it came from your office. The other job is deciding whether a slow-paying customer gets a firm call or a grace period, which is a relationship call, not a data lookup. An agent is already good at the first job. It should not be trusted alone with the second, and knowing exactly where that line sits is what a weekly owner review is built to check, because the alternative is finding out where the line was after a customer complains about a reminder that landed wrong.

Can an AI agent follow up on unpaid invoices?

Yes, for the mechanical half of the job: watching for overdue invoices, pulling history, and drafting a reminder on a schedule.

An agent can watch your invoice system for anything past its due date, pull the account's payment history, draft a reminder in your own voice, and send it on a schedule you set, without a person opening the aging report first. That is exactly the kind of repeatable, rule-based task an agent handles well, because the inputs, an invoice, a due date, a contact, rarely change shape from one account to the next.

In practice that looks like a short, specific routine running in the background: an invoice crosses 15 days past due and the agent sends a plain reminder with the amount and the original due date; it crosses 30 days and the tone firms up slightly, still polite, now naming a specific next step; it crosses 45 or 60 days and the agent stops sending anything on its own and instead puts the account on a list for a person to call. That staged structure, not a single blanket reminder, is what keeps the mechanical half of the job from accidentally doing the relationship half's work.

The scale of the problem is why this matters for a small business specifically. A September 2026 survey of more than 600 companies across the United States, Canada, and Mexico by the trade credit insurer Atradius found that late payment affects seven in ten companies in the region, and overdue invoices average 23% of a company's B2B receivables at any given time. A fifth of the money a business is owed is sitting past due at any moment, which is exactly the kind of backlog that slides when a person is chasing it by hand between everything else on their plate. For a business with, say, $40,000 in outstanding invoices on the books, that 23% figure would put roughly $9,200 sitting overdue right now; that specific dollar amount is an illustrative example built on the survey's percentage, not a client result, but it shows why a backlog this size is worth automating the chasing part of, even before touching the harder judgment calls.

What can an agent actually do inside accounts receivable follow-up?

Four tasks: flagging overdue invoices, pulling account history, drafting the reminder, and logging what went out.

None of these four require judgment about the customer relationship. They require consistency, which is the part a person juggling five other jobs tends to lose first. Flagging catches an invoice the moment it crosses its due date without waiting for a person to notice. Pulling history means the reminder references the right amount and the right prior conversation, not a generic notice. Drafting the reminder means writing it in a tone that matches how your business actually talks to customers. Logging means writing down exactly what went out and when, so nothing gets double-chased or forgotten.

The logging step matters more than it sounds like it should. A person chasing invoices by hand, especially one doing it between other responsibilities, tends to lose track of which account already got a reminder this week and which one is still waiting its turn. A customer who gets two reminders four hours apart because two different people, or one distracted person, both acted on the same aging report is a worse outcome than a customer who waits an extra day for the first one. An agent that writes down exactly what it sent and when removes that specific failure mode entirely, independent of whatever else it gets right or wrong.

What this duplicates, and what it does not

A collections specialist spends real time on exactly those four steps before they ever pick up a phone: checking who is overdue, pulling the history, writing the note, and tracking the outcome. An agent doing that work does not replace the judgment a specialist brings to a strained customer relationship. It replaces the clerical half of the job so the specialist's attention goes to the handful of accounts that actually need a human voice, rather than being split thin across every account on the aging report regardless of whether that account needs a human voice at all.

How accurately does an agent read the invoice before it decides who to chase?

It varies by 51 points between models on the same invoice, which is the question most collections pitches skip entirely.

A 2025 benchmark tested eight multi-modal AI models, from the GPT-5, Gemini 2.5, and Gemma 3 families, against three invoice datasets totaling 1,850 documents: 500 clean, computer-generated invoices, 1,000 scanned receipts, and 350 scanned invoices. The researchers compared two different strategies for reading each document: having the model look at the image directly, versus first converting the document to structured text and handing the model that instead. The best result, Gemini 2.5 Pro reading document images directly, hit 96.50% field-extraction accuracy on the clean, computer-generated invoices. The weakest model tested managed only 45.69% on those same clean invoices, a 51-point gap on the easiest document type in the study, and that gap alone is the clearest argument for testing your own agent against your own paperwork instead of trusting a single reported number.

Why a scanned invoice is a harder test than a clean one

Accuracy dropped further once the documents got messier. Even the top-performing setup fell to 87.46% on scanned receipts, and the paper flagged one field in particular, the IBAN or account number, as consistently worse than the rest because scanning confuses a zero with the letters O or U. That single-field weakness matters more in accounts receivable than it might elsewhere, because a misread account number does not just produce a wrong-looking record. It can route a payment confirmation, or a reminder, to the wrong place entirely. A small business rarely sends clean, computer-generated invoices exclusively. If your own paperwork includes scans, faxed copies, or a vendor's handwritten adjustment, the agent reading it is working closer to that lower number than the headline one, which is a real reason to check its output on your specific documents before trusting it to prioritize a list unsupervised.

Three fields worth a human spot-check before the agent prioritizes a list

• The exact amount owed, since a single misread digit sends a reminder for the wrong balance.
• The due date, since a confused date can flag a current account as overdue or miss one that actually is.
• The account match itself, since two customers with similar names are exactly where an agent's confidence outruns its accuracy.

Can an agent decide which overdue account to chase first?

Partially: a model can rank accounts by how likely they are to pay soon, but this is older ground than it looks.

IBM Research built a prototype with a multinational bank specifically to rank which overdue accounts a collector should call first, because a collector's own experience and list-building decisions were never recorded anywhere the system could learn from directly. The team's point, stated plainly in the paper, is that invoice data is genuinely complex to model: a given account's payment pattern depends on its history, its size, and context the raw numbers do not capture on their own. Working around that, the prototype reached up to 81% prediction accuracy on which accounts were likely to pay soon versus stay overdue, and the authors' simulations suggested prioritizing collector calls this way could save on the order of $1.75 million a month, a simulated estimate from their own modeling, not an independently audited result.

The useful part of that finding is not the dollar figure. It is that ranking accounts by likelihood of payment, a sorting problem, is something a model can meaningfully help with, while deciding what to actually say to the one account that is 90 days overdue and arguing about a disputed charge is not a sorting problem at all. The two look similar from a distance, since both involve an overdue account and a decision about what happens next. They are not similar once you look at what the decision actually requires.

Which accounts still need a person, not an agent?

Three situations: a disputed invoice, a strategic account, and anything heading toward a collections agency or legal language.

A dispute over the invoice itself needs a person because the customer says the amount or the work is wrong, and an automated reminder would just escalate a misunderstanding. A strategic account needs a person because the relationship matters more than this one invoice, and a tone-deaf reminder can cost more than the payment is worth. Anything heading toward a collections agency, a payment plan, or legal language needs a person because the business is making a real decision about the relationship, not just asking for money again.

Why these are judgment calls, not data lookups

Each of those three situations depends on context an invoicing system does not hold: history with this specific customer, how much the account is worth beyond this invoice, and what the business is actually willing to do if the money does not come. Writing down, in advance, which accounts route straight to a person and which ones the agent can work alone is the same kind of boundary covered in a closer look at agent handoff rules, and it is worth setting before the first reminder goes out, not after an account relationship gets strained.

What should an owner check weekly to know the follow-up is actually working?

One number: the dollar amount that actually moved from overdue to paid, not how many reminders went out.

An agent that sends twenty reminders and collects nothing is not working, even if its activity log looks busy. Pair that weekly check with a spot-check of the invoice data behind one flagged account, confirming the amount, date, and account match are correct, the same three fields worth checking before the agent ever prioritizes a list. The pattern is close to what shows up when tracking an agent's performance over its first 90 days: completion and accuracy are different numbers, and only one of them tells you the money actually came in.

A second thing worth watching is the accounts the agent escalated versus the accounts it should have escalated. If a dispute or a strategic account slipped through and got a routine reminder instead of landing on a person's desk, that is the handoff rule failing, not the invoice-reading accuracy, and it is worth fixing that boundary before it happens twice. A weekly ten-minute check against both numbers, money collected and accounts correctly escalated, catches the two failure modes that actually cost a business something: money left on the table, and a relationship handled badly.

What does it cost to keep chasing invoices by hand, against running it through an agent?

The clearest baseline for the human side is public: a median wage of $47,030 a year for the role most small businesses already pay for in some form.

The US Bureau of Labor Statistics counted 158,830 bill and account collectors employed nationally as of May 2025, earning a median wage of $22.61 an hour, or about $47,030 a year. That is the role most small businesses are either paying for directly, or asking an office manager or bookkeeper to absorb on top of their actual job. Weighing a dedicated collections hire, a slice of someone's existing time, or an agent against that number is the real decision, and it depends on how many accounts actually go overdue each month, which is specific to your own books, not a number any blog post can hand you. If you want that number for your own invoice volume, your free Systems Plan walks through it against what you are collecting today.

A similar comparison involving a different client-facing task, drafting the updates a business sends customers rather than chasing them for payment, is in a look at whether an agent can write client status updates, including the same kind of field-level accuracy check that matters here.

Methodology

This post on ai agent accounts receivable draws on four sources. Atradius's Payment Practices Barometer for North America (September 2026, 600+ businesses surveyed) supplies the scale of the late-payment problem. A 2020 IBM Research paper (arXiv:2008.07363) supplies the prioritization-accuracy and simulated-savings figures, built and tested with a multinational bank; its dollar estimate is explicitly a simulation in the original paper, not an audited outcome, and this post treats it that way. A 2025 benchmark of eight multi-modal AI models (arXiv:2509.04469) supplies the invoice-reading accuracy figures across clean, scanned, and receipt-style documents. The US Bureau of Labor Statistics' Occupational Employment and Wage Statistics for bill and account collectors (May 2025) supplies the cost baseline for the role most businesses already pay for, in some form, to do this work by hand.