What Is an AI Agent for Accounts Receivable?

An AI agent for accounts receivable is software that reads the state of an invoice or account, decides the next action under finance policy, and takes that action in your real systems (email, ERP, payment tools). It follows through until cash posts or a person needs to weigh in. It is not a chatbot that answers questions about your aging report. It is not RPA that replays a fixed click-path until the UI changes and the bot breaks. An AR agent is measured on cash recovered, exceptions handled, and DSO pressure relieved, not on how many reminders it sent.

What it actually does in AR

On a working book, agent work looks like jobs, not demos:

• Prioritize who to chase: value at risk and payment behavior, not “oldest invoice first” by default.
• Draft and send the next touch: tone and channel matched to the account, with a human gate where policy requires it.
• Capture promises to pay: reschedule the next action to the promised date instead of hammering the cadence.
• Pause on disputes: stop collections when the customer is fighting the invoice and route the case with context.
• Help cash application: match remittance to open items and escalate the ambiguous ones.
• Hand remittance detail to payment rails: once a customer pays or commits, the agent’s job includes getting settlement data into the systems that post cash, not stranding it in an inbox.

If a vendor’s “AR agent” only drafts nicer emails, you bought an Assist tool with a new label. Drafting is one of the six jobs above, not all of them.

Four kinds of AR agents (same invoice, different jobs)

Most teams deploy one kind of AI and wonder why results feel incremental. In finance, agents fall into four categories, Assist, Automate, Advise, and Configure, and AR needs more than the first. (I wrote the full frame in The 75% of AI Value You’re Leaving on the Table.)

Take one overdue invoice:

• Assist drafts the collection email, and a collector reviews and sends it.
• Automate runs the reminder cadence, stops on payment, and escalates only real exceptions.
• Advise notices this customer’s payment pattern predicts a miss six weeks out and flags the account before it ages into crisis.
• Configure sees that variance thresholds for this vendor tier create noise and proposes tightening the rule so humans stop reviewing false positives.

Four agents can touch one invoice. None of them is “the AR chatbot.” Treating them as one capability is how boards get Assist theater and call it transformation.

Agent vs automation vs RPA vs “AI-powered AR”

Human-in-the-loop belongs in policy for high-stakes writes, credit holds, and legal tone. It does not belong as a placebo on every low-risk step. (More on that in human-in-the-loop placebo and the autonomy-tax framing on the frameworks page.)

How to evaluate a vendor claim

Before you buy “agentic AR,” ask:

1. What can it write to the ERP or billing system without a human click, and what is blocked by design?
2. Is every action auditable (who/what/why) for controllers and auditors?
3. Does a dispute or promise automatically change the next action, or does the cadence keep firing?
4. How does remittance move from customer reply → match → payment rails / cash app?
5. What happens when confidence is low? Does it escalate with context, or silently guess?
6. Can you set approval thresholds by amount, segment, or action type?
7. Are you buying Assist, Automate, Advise, and Configure, or a slide that says “agents” once?

If the answers are vague, you are not evaluating an agent. You are evaluating marketing.

Read next: four stages of agent evolution, invoice as agent.