What an AR agent genuinely finishes today, why the weight sits in cash application rather than in chasing, and where you are better off keeping it out.
What is an AI agent for accounts receivable?
An agent has agency. It does not follow a path you mapped out in advance, but works out within set bounds which steps are needed. The difference with rule-based automation and with workflows is covered in AI agents for finance.
Two points from there apply here without qualification.
The label tells you nothing. AI-first and AI-native are about how something was built. Agent is about what it can do. What you want to know is what happens to the payment that does not fit an invoice.
The AI reads and investigates, the rule decides. In receivables that distinction is sharper than in payables, because here it concerns money you already have. Which invoice you clear against which receipt is a posting, and it has to come out the same way every time. The investigation needed to know which invoice it is, is exactly where agency belongs.
Where the time actually goes in accounts receivable
The standard claim in this category is that the bottleneck sits in the conversations: customers asking questions, raising disputes, going quiet. That holds for large organisations with their own credit control team.
In the Dutch mid-market we see something else. The chasing is already handled. There is a tool, or it runs off the ERP's standard function, and the reminders go out fine. What is missing is a working way to put the money that came in back against the open items.
The stories are strikingly consistent. At a travel services company on NetSuite the system refuses a match on a one-cent difference, so €90 against €90.01 blocks the entire bank reconciliation. There is no live bank feed, so statements are exported and uploaded by hand every day. A receipt in euros cannot be cleared against an invoice in Canadian dollars, so that runs through a suspense account with manual journal entries behind it. Someone there called it suspense account gymnastics, which is a better description than we could have come up with.
This is not a matter of having bought the wrong tool. That same organisation had one: a specialised reconciliation product advertising a 99 percent match rate that in practice stayed under 20 percent. Remittance advices could not even be loaded into it. On large files of more than a thousand lines it failed outright. The premium tier ran to roughly €1,300 to €1,400 a month, at which point the conclusion was: then we will find something else.
What this means for your decision: if your reminders already go out automatically and your team still loses days a month, your problem is not in the chasing. Buying more dunning functionality will not fix it.
Six things an AR agent takes over
1. Matching payments to open items
The heaviest part, which is why it comes first. The bank transaction arrives and has to go against the invoices underneath it.
Where it goes wrong is the same every time. One payment covers forty invoices, with a remittance advice sitting as a PDF in a separate email or never sent at all. Something was partly paid. A credit note was offset. The customer withheld a bank charge or a factoring fee. The amount is a few cents off through rounding or exchange rate.
An agent reads the remittance advice out of the attachment or out of the transaction description, links the lines to the open items, and clears the difference according to the rule you set: exchange difference, bank charges, settlement discount, or too small to bother with. What it cannot close reaches your team with the candidates attached, instead of as an empty line on a statement.
In more depth: reconciliation software and matching and bank reconciliation in NetSuite.
2. Handling the receivables inbox
Copy invoices, statements of account, questions about a specific line, a note that payment is coming next week. It all lands in the same mailbox and it is nearly always routine.
An agent reads what is being asked, pulls the right document and answers in the thread. The difference with an auto-reply is that it understands the question: someone asking why invoice 20419 is still open when there is a credit note against it does not want an acknowledgement of receipt.
3. Chasing, dosed on what the customer actually does
This is the part you probably already have, so the question is what an agent adds. The answer is the dosing. A customer who always pays on day 45 against agreed terms of 30 needs a different conversation from one who is late for the first time. A customer with an open dispute should not be chased at all, because you lose goodwill and solve nothing.
An agent looks at payment behaviour, at what was answered before and at which questions are still open, and adapts the timing and the tone to that. Not because it is friendlier, but because a reminder sent to someone who just raised a dispute does not shorten your average payment term.
4. Spotting disputes before they are called disputes
A customer who writes that something is wrong with an invoice rarely files a formal dispute. They ask a question. If nobody recognises that question as a dispute, the invoice stays open while reminders keep going out over it.
An agent classifies the message, links it to the invoice and the line in question, pauses the reminders on that invoice, and puts it to the colleague who owns it. Usually that is someone in sales or operations rather than finance, and that is exactly why it stalls today.
5. Capturing payment promises and following them up
"We will pay next Friday." That currently lives in someone's head or in a mailbox, and it gets followed up when somebody remembers.
An agent takes the commitment out of the message, records the date, and comes back to it if the money is not there on that date. Referring to what was promised, which is a different and more effective conversation than a standard reminder.
6. Invoicing from what was actually agreed
At the start of the chain: raising invoices from contracts, subscriptions, delivered hours or orders, including the deviations agreed per customer. How that goes wrong and why it pollutes the rest of the chain is covered in AI agents in order to cash.
Traditional receivables management versus AI agents
| Task | Traditional receivables management | AI agent |
|---|---|---|
| Chasing | Schedule based on the due date | Timing and tone based on payment behaviour, earlier replies and open questions |
| Receivables inbox | Someone reads and answers, or there is an auto-reply | Reads the question, pulls the document, answers in the thread |
| Dispute | Only a dispute once someone logs it as one | Recognises it in the message, pauses the reminders, puts it to the right colleague |
| Payment promise | In someone's head or in a mailbox | Recorded with a date, and followed up if the money does not arrive |
| Cash application | Only the exact match; the rest is manual | Reads the remittance advice, links partial payments and offsets, clears the difference by your rule |
| Foreign currency | Suspense account plus manual journal entries | Matches across currencies and clears the exchange difference where you want it |
| A one-cent difference | Blocks the match | Cleared within your tolerance, with the reason recorded |
| The posting in your ERP | Fixed rules | Fixed rules here too: deterministic code you described in plain language, not a model's judgment call |
Where you do not want agency
In receivables it concerns money that is already in and customers you want to keep. That makes two boundaries sharper than on the payables side.
The clearing itself is deterministic. Which receipt against which invoice, which difference to which account, up to what amount without intervention. That is code you set. A system that clears a three-cent difference to exchange results today and to bank charges next month is unusable, however defensible each individual choice is.
The tone towards a customer is no place for full freedom. An agent allowed to decide for itself how firm a reminder gets can damage a relationship you spent years building. The practical split: it decides when and on what basis, you decide the range it writes within. And on the first few rounds you read along.
Do the rest of the arithmetic too. A system that clears correctly 95 percent of the time leaves a hundred wrongly linked items on two thousand lines a month. Your customer finds those, not you, and the conversation that follows costs more than the clearing itself.
Where it stops
Without a remittance advice there is sometimes nothing to match. If a customer transfers one amount with no description and no advice, and the amount matches no combination of open items, the only way out is to ask. An agent can ask that question and process the answer, but it cannot guess. Anyone promising this always runs automatically is selling you something other than what you get.
What your ERP does not offer, an agent cannot write into it. On Exact, writing reconciliations back does not run through the ordinary REST connection, which means it has to go another way. You only hit limits like that during the build. Ask about it per action you want to automate.
An agent does not create a bank connection that does not exist. If your ERP runs without a live bank feed, an import remains. An agent can take that import over and process it, but it cannot make a connection the bank or the ERP does not offer.
How to start
- Count the lines that stay open. Not how many transactions you process, but how many fail to link automatically each month and why. Split that into missing advice, partial payment, offset, currency and cent differences. That split decides what you need to solve.
- Write down your clearing rules. Up to what amount can a difference be cleared, and to which account per type. Those rules already exist, they are just not written anywhere.
- Check whether chasing really is your problem. If reminders go out automatically and it still costs days, it sits elsewhere.
- Run it supervised first. Every link as a proposal, your team approves, then you move the threshold.
- Measure the number of manually linked lines, not your DSO. DSO moves for ten reasons at once and says little about whether this works.
Frequently asked questions
Agentic software that works out for itself which steps are needed to get a sales invoice from sent to reconciled: invoicing, chasing, answering queries, spotting disputes, following up payment promises and matching payments to open items. It works inside your existing ERP and records every action in an audit trail.
Those are built to handle chasing, and they do it well. They do not solve linking incoming money to open items, and at most mid-sized companies that is where the work sits. You do not have to replace them; an agent can run alongside on the same ERP data.
Partly, and not in the way usually promised. Faster cash application means your open items list is correct, so you stop chasing customers who already paid and start chasing the ones who did not. That saves cycle time and it saves irritation. The biggest gain is in hours, not in days.
Within the rules you set, yes. Up to what amount a difference clears by itself and to which account per type is your call. What falls outside that reaches your team, with the investigation already attached.
Yes, and it is one of the places with the most to gain. The usual workaround is booking a receipt through a suspense account and clearing the exchange difference with a manual journal entry. That is exactly the kind of repeat work that can run on fixed rules.
Yes, those are the three Claridy runs on. What is available through the connection differs per ERP, and on reconciliation you notice that more than on invoice processing. Ask per action whether it genuinely works.
Processing inside the EU, read-only where possible, and an exportable audit trail per action. The SOC 2 programme is under way and certification is coming. In more depth: is AI safe for your books.
What Claridy does on this side sits on the order to cash and reconciliation product pages. The chain before it: AI agents in order to cash. The payables side: AI agents for accounts payable. On the distinction between automation, workflows and agents: AI agents for finance.
Last checked: 2026-08.