
An AI accounts receivable agent is software that reads your open invoices, works out who is overdue and what the sensible next step is, and then either drafts that step or carries it out. The pitch is attractive because chasing payment is repetitive work that people skip when they are busy. The risk is that a system able to message a customer is also able to message the wrong person at the wrong time. This guide sets out what to hand over, what to keep, and how to tell the difference before the first reminder goes out.
What the agent is actually doing
Strip away the branding and most accounts receivable agents do four jobs. They read the ledger, classify each open balance by age and history, choose the next action from a set of rules, and then execute or queue that action. Each job can be done well or badly, and each fails differently. A misread ledger produces a reminder for an invoice that was paid last week. A badly drawn rule sends a firm notice to a customer who is in a dispute. An agent that acts without leaving a record produces an argument you cannot settle later.
- Read: pull invoices, payments and credit notes, and reconcile them before anything else happens.
- Classify: group balances by days past due, customer and amount, and flag anything with an open dispute or a recent part payment.
- Decide: map each group to a step on your escalation sequence, using the terms that apply to that customer.
- Act: draft a message for review, send it on schedule, or hold it and record why.
Where an agent earns its keep
The work an agent does well is the dull, high-frequency part: noticing that an invoice crossed its due date overnight, drafting a reminder in the right tone, and making sure nothing is forgotten on day 31. A person is slow at this and tends to let it slide during a busy week. It is also the process most payment problems turn out to lack, as the piece on payment terms explained argues. An agent does not fix unclear terms, and it cannot make a customer with no intention of paying start paying. What it does is make a clear process run without depending on anyone's memory.
Set the limits before the agent acts
Write the boundaries down before switching anything on. The agent should inherit your policy, not invent one of its own.
- Amount thresholds. Balances above a value you choose go to a person for review, however old they are.
- Stop conditions. An open dispute, a recent part payment or a customer on a payment plan pauses the sequence automatically.
- A tone ceiling. Define the firmest message the agent may send unattended. Anything beyond it, such as a final notice that cites a statute, waits for a human to approve it.
- Legal limits. Late charges and interest depend on the contract and on the state. Check the rules in the late fee library and the approach in how to charge late fees on invoices before any automated charge is allowed.
- Contact rules. Name who receives reminders, whether copies go to your own team, and which contacts must never be emailed automatically.
What a person should still decide
Three decisions should stay with a person: whether to write off a balance, whether to refer a debt to an outside collector, and whether to offer a concession on terms to a valued customer. Each carries commercial and sometimes legal consequences, and each depends on context the ledger does not hold. A useful test is reversibility. If a message cannot be recalled once it is sent, a person approves it first. The escalation ladder that gets invoices paid is a sensible place to draw that line: the early rungs can run automatically, and the final rungs are reviewed.
Connecting an agent without losing control
Many teams will connect an agent to their invoicing tools through an API or the Model Context Protocol. Duebay's MCP developer page lists the tools its server exposes, including reading the aging report, drafting a dunning sequence, computing a late fee against a cited statute, and sending a reminder. The MCP tools specification is explicit that there should always be a human in the loop with the ability to deny tool invocations, and that clients should make clear which tools a model can reach. Read it before you grant any agent a tool that sends mail.
In practice that means giving each key the narrowest access the job needs, keeping read permissions separate from send permissions, and logging every call. Whether a particular send needs a person's confirmation is a decision for your own setup. Do not assume that a tool will ask a person before it acts.
What a good audit trail records
When a customer asks why they received a reminder, you should be able to answer in a sentence. For every automated action, record the invoice, the balance at that moment, the rule that selected the step, the message sent and the time it went. Commentary on governed receivables agents makes the same point, that the system should show what data it saw, which policy applied and what it did. Serrala's piece on what agentic AI does and does not do in accounts receivable is a useful counterweight to vendor claims of full autonomy.
A Monday queue, worked through
This is an illustration, not a benchmark. Imagine a ledger of forty open invoices on a Monday morning. Under a sensible setup, most sit within terms and need nothing. A handful are a few days late and receive the scheduled friendly reminder without anyone touching them. A few have passed the point where a firmer notice is due, and those wait in a queue for a person to approve. One is flagged as disputed and is paused until someone resolves it. The agent's job is to produce that queue accurately and quickly. The person's job is to clear the approvals and settle the dispute.
How to tell whether it is working
An agent is only worth keeping if the outcome improves, and the outcome to watch is cash arriving on time. Measure it before you switch the agent on and again after a month, using the same ledger and the same method. Days sales outstanding is the simplest summary: if it falls while your terms stay the same, the process is doing its job. If it does not move, the problem is more likely the terms or the customer mix than the speed of reminders.
Look at the exceptions too. Count how many automated messages were held, how many were corrected by a person, and how many customers replied to say the reminder was wrong. A rising number of corrections means the classification rules need tightening before the agent is trusted with a firmer step. A steady, low number means the policy you wrote is matching what actually happens.
Questions people ask
Can an AI agent collect payment on its own? It can prompt the customer and follow through on the steps you have approved. Whether money moves depends on the payment methods you connect, not on the agent.
Is an agent safe for a small team? Often it is safer than a busy person skipping reminders, provided the thresholds and stop conditions are set. Start with reminders only, and add escalations once you have reviewed a few weeks of the log.
What should I review first? Check the classification before the wording. A message in the wrong tone is easy to fix. A reminder sent for an invoice that is already paid damages trust.
Do I need a special product to do this? No. The written policy and the audit trail matter more than the brand. A spreadsheet with a clear policy beats an agent with none.
The tools behind Duebay's developer surface run on the same engine as the app. Grant only the permissions a job needs, log every call, and widen access when the record shows it is safe.
Read the MCP developer docs