What is an AI AR agent and how does it work

An AI AR agent is autonomous software that does the work of accounts receivable and is accountable for the result. It reads the state of each account, decides the next action, takes that action across your email and ERP, and follows through until the invoice is paid or a person needs to weigh in. Where a dashboard shows you the work and a chatbot answers questions about it, an agent does the work, continuously, across the whole ledger.

The simplest way to picture it is a capable collector or AR analyst who never runs out of hours. They open an account, see it is 25 days late with an open promise to pay, decide to follow up, send the right message, log the result, and move to the next account. An AI AR agent runs that same loop, on every account, every day, escalating the cases that need a human decision.

What an AI AR agent is

An AI AR agent is not a feature bolted onto a reporting tool. It is a worker. Three properties define it, and a tool missing any of them is something else.

• It acts, not just informs. It sends the reminder, applies the payment, and opens the deduction case in your real systems, rather than surfacing a task for a person to do.
• It decides from context. It chooses the next action based on the actual state of the account, not a fixed schedule, so it can wait for a promised date, escalate a dispute, or pause outreach when appropriate.
• It owns an outcome. You measure it on cash recovered and DSO reduced, the same goals you would give a person, not on emails sent.

That last point is the heart of it. An agent is accountable. This is the difference between generative AI that drafts and an agent that does: the agent is the one holding the outcome.

The perceive-decide-act loop in AR

An agent works by running a loop, the same one a person runs without naming it. Understanding the loop is how you understand that autonomy is engineered and auditable, not magic.

Perceive. The agent builds a current picture of each account. It reads the open invoices and their ages, the payment history, any promises to pay, open disputes, and the recent email thread. When a new customer email arrives, it reads that too and updates its picture, recognizing a payment promise, a dispute, or a request for a copy of the invoice.

Decide. From that picture, the agent chooses the next-best action. A 40-day-late invoice with no contact gets a firmer reminder. An invoice with a promise-to-pay date next week gets left alone until then. A short pay gets a deduction case opened. The decision comes from the account's actual state, which is why the agent can do the right thing across thousands of accounts that are each in a different situation.

Act. The agent takes the action in the real system. It sends the email, applies the cash against the matching invoices, opens or routes the dispute, or schedules the next follow-up. Then it observes what happens and loops back to perceive, so a customer reply or a payment immediately changes what it does next.

This loop is what lets an agent carry a task across days and replies instead of one turn at a time. It is also the mechanism behind faster collections, because the agent never forgets to follow up and never lets an account go quiet, which is the discipline that brings DSO down.

What the agent connects to and controls

An agent is only as capable as the systems it can read and act in. A real AR agent connects to the stack where receivables actually live.

• The ERP or accounting system. To read open invoices and customer master data, and to write back applied cash, credit memos, and dispute codes.
• Email and messaging. To send reminders and, crucially, to read and act on customer replies, which is where most of the real AR conversation happens.
• Payment and remittance data. To match incoming payments to invoices and parse the remittance that explains short pays.
• Customer and contract data. Price lists, terms, and proof of delivery, so it can validate deductions and answer pricing questions correctly.

The breadth matters because AR work crosses systems constantly. A payment lands in the bank feed, has to be matched to an invoice in the ERP, and clears a cash application exception that was holding up the account. An agent that only sends email cannot do that. An agent wired into the whole stack can.

Where humans set policy and approve

Autonomy and control are not opposites. The agent runs the work; people set the boundaries and own the judgment calls. In practice that means a few clear layers of human authority.

You set policy up front: the collections cadence, the tone, which accounts are sensitive, and the dollar thresholds above which the agent must ask first. You hold approval gates on high-stakes moves, like writing off a balance, agreeing a payment plan, or escalating a strategic account. And you review after the fact, because the agent logs every decision and the reasoning behind it, so you can audit what it did and why.

The result is bounded autonomy. The agent handles the routine volume freely, which is most of the work, and brings you the exceptions with a complete picture rather than a blank investigation. Oversight scales because you supervise by policy and exception, not by checking every action.

Agent vs chatbot vs RPA bot

The label "AI" gets applied to three very different things, and conflating them is how buyers get burned. An agent is none of the other two.

A chatbot answers questions in natural language. Ask it the balance on an account and it tells you. Useful, but it does not act on the ledger, follow up with a customer, or apply a payment. It describes work; it does not do it.

An RPA bot automates a fixed sequence of clicks. It is fast and reliable on a stable, structured task, and it breaks the moment something is unexpected, a remittance in a new format, an email phrased differently, an exception the script did not anticipate. It follows rules; it does not reason, so the messy cases get kicked back to a person. That is most of AR.

An AI AR agent reasons over real, messy context and takes accountable action across systems. It handles the unstructured email and the ambiguous short pay that an RPA bot cannot, and it acts on the ledger where a chatbot only talks. The test is simple: does it do the work and own the outcome, or just describe it and hand it back?

What to expect in the first 90 days

A credible agent earns its authority rather than demanding it on day one. A realistic rollout starts supervised and widens as trust is established.

In the first few weeks, you connect the systems and let the agent run in a supervised mode where it proposes and you approve, so you can see its decisions against accounts you know. As the decisions prove sound, you raise the thresholds and let it act directly on the routine cadence, keeping approval gates on the high-stakes moves. By the end of the period, the agent is working the full ledger continuously, you are reviewing exceptions and audit logs rather than initiating outreach, and you should be able to see the early movement in DSO and recovered cash that tells you it is doing the job.

The goal is not to hand over the function blind. It is to reach the point where the agent runs the volume, the team handles judgment and relationships, and the cash outcome is measurably better than manual work could produce.

How Rex works as an AI AR agent

Rex is an agentic AI AR agent built on exactly this loop. It perceives the state of every account by reading your ERP and the live email thread, decides the next-best action, and acts, sending collections messages, applying cash, and working deductions across the whole ledger. It follows each case through to a result and escalates the calls that need a person, with every decision logged for you to review.

Rex is accountable for the outcome, not the activity. You hold it to cash recovered and DSO down, set the policy and the approval gates, and let it run the work a team would otherwise do by hand. See how Rex runs accounts receivable end to end as an agent.