How AI Agents Can Automate Accounts Receivable

Accounts receivable is one of the few business functions where the work literally never stops. Invoices post at any hour. Payments land in bank feeds overnight. Reminders fall due on a schedule that does not care whether anyone is in the office. The AR inbox fills with remittance advices, copy requests, disputes and promises to pay.

That makes it an obvious candidate for automation — and also one of the easiest places to automate badly. This article is about how to do it well: what the work actually is, where AI genuinely helps, and why the most important design decision is not which model you use but what you refuse to let it decide.

What AR work actually consists of

Strip away the job titles and receivables work is a handful of repeating activities:

• Getting invoices out — making sure every invoice posted in the accounting system actually reaches the right person at the customer, complete and correct, and confirming it arrived.
• Following up — reminders before and after due dates, escalating in tone and seniority, pausing when a customer disputes or pays, and recording promises to pay.
• Applying cash — matching money that arrives against the invoices it was meant for, which is simple when the reference is clean and tedious when it is a partial payment covering four invoices with a deduction nobody explained.
• Handling disputes and deductions — working out why a customer short-paid, classifying the reason, getting it to the person who can resolve it, and tracking it until it closes.
• Watching credit risk — noticing that a customer’s payment behaviour is slipping before the exposure becomes a problem.
• Answering the inbox — the constant stream of “please resend invoice 1042” and “we paid this last week” and “we are disputing the freight charge.”
• Reporting — aging, days sales outstanding, collection effectiveness, cash forecasts; the numbers the rest of the business runs on.

Look at that list and a pattern appears. Some of it is reading, sorting, drafting and prioritising: classifying an email, drafting a reminder, ranking which accounts to call today, summarising a dispute. Some of it is calculating, deciding and recording: matching a payment, posting it, applying a late fee, changing a credit limit, issuing a credit memo.

The first kind is what current AI is genuinely good at. The second kind is where it must not be trusted on its own.

Why “let the AI decide” is the wrong design

It is tempting to hand the whole function to a capable model and let it act. The reasoning is usually that the model is right most of the time, and the exceptions can be caught later.

With money, that reasoning fails in three ways.

First, the cost of a wrong action is asymmetric. A reminder sent to the wrong contact is embarrassing. A payment posted against the wrong invoice corrupts your ledger. A credit memo issued without authority is a financial loss and, in many organizations, a control failure that has to be reported.

Second, “right most of the time” is not auditable. When a controller asks why a figure in the aging report is what it is, “the model computed it” is not an answer. Every number in a finance function needs to be reproducible from inputs and rules.

Third, customers are not test data. A dunning email with a hallucinated invoice number, or a tone that reads as threatening, damages a relationship that took years to build.

The design that works is a split: AI reads, drafts, classifies and prioritises. Deterministic policy code calculates, decides and executes. Humans approve high-risk actions. Everything is logged. The model’s job is to make the humans faster and better informed, not to replace the controls.

The sub-agent model

Rather than one agent trying to do all of receivables, the approach we have found robust is a coordinating agent and a set of specialists, each responsible for one of the activities above:

The coordinating agent receives work (scheduled jobs, events from the accounting system, requests from the console), routes each task to the right specialist, checks every proposed action against the organization’s policies — thresholds, roles, approvers — and sends anything gated by those policies to a human.

Two things about this table are worth noticing. The right-hand column is never empty for anything that touches money. And the reporting row is the only one with “none” — because reporting reads; it never acts.

The approval queue

The approval queue is where the design becomes real for the finance team. It is a single list of actions the system wants to take but is not allowed to take alone: a proposed cash match below the tolerance, a credit memo for a dispute, a dunning sequence about to start on a customer for the first time, a reminder that deviates from the template.

Each item shows what the system proposes, why, and the evidence. A person approves, amends or rejects. The decision — and who made it — is logged against the item.

Done well, the approval queue is not a bottleneck. It is the opposite: it is the only place the team has to look, because everything that did not need judgement has already been handled. The team stops chasing and starts deciding.

The thresholds themselves — which amounts need a controller, which need the CFO, which reminders can go automatically — belong to the organization, not to the software. They should be set when the system is installed and editable afterwards, with the history kept.

What changes for the AR team

The honest answer is: the job gets smaller and more senior.

The volume work — sending, reminding, matching the clean cases, answering “please resend” — disappears from the daily list. What remains is the work that needed a person all along: the awkward partial payment, the dispute that is really a sales problem, the customer whose behaviour has changed, the decision to escalate.

The team also gets something it rarely had before: a complete record. Every reminder sent, every promise made, every match proposed and accepted or rejected, is in one place, timestamped, attributable. When a customer says “nobody told us,” there is an answer.

How the AR Automation Console implements this

Qumaxa’s AR Automation Console is built on exactly this model — the same split, the same seven specialists, the same approval queue.

A small connector is installed where the accounting system lives (today, QuickBooks Online or Microsoft Dynamics GP) and talks to it with API-based, least-privilege access — never a user’s password. The cloud console holds the agents, the organization’s policies and the approval queue. Every figure in the reporting is computed deterministically; a language model never produces a number.

It is early-access software, and its product page says plainly what is current and what is on the roadmap — additional ERP adapters, a reconciliation agent, further finance agents on the same platform. We think that is the right way to introduce automation into a function that handles money: clearly, with the controls visible, and with people in charge of the decisions that matter.

Written by Altaf Hussain Khan, Director, Qumaxa. Qumaxa builds custom software, automates business processes and applies AI for organizations in Pakistan and internationally.