Ask ten AR managers what agentic AI means and you’ll get ten answers. Ask them what it changed in their department last quarter and the room gets quieter.
That gap matters. Deloitte’s Q4 2025 CFO Signals survey found that 87% of North American CFOs expect AI to be extremely or very important to finance operations in 2026, and 54% named integrating AI agents as a transformation priority. At roughly the same time, Gartner predicted that more than 40% of agentic AI projects will be canceled before the end of 2027, largely because of escalating costs, unclear business value, and inadequate risk controls.
Both findings can hold at once. And receivables is a reasonable place to find out which side of that split you land on, because the work runs at high volume, most decisions have a verifiable right answer, and the results show up in days and dollars rather than in sentiment.
What agentic AI means in accounts receivable
Agentic AI in accounts receivable is software that takes a goal, breaks it into steps, pulls the context it needs from your ERP and banking data, chooses an action within limits you’ve defined, and carries the work through to a result. The shift from earlier automation is the decision. A rules engine walks the path you drew for it. A predictive model tells you what it expects to happen. An agent picks a course of action and performs it.
That distinction gets blurred in a lot of vendor material, and Gartner has a name for the blurring: agent washing, where an existing chatbot or scripted bot gets relabeled. Worth keeping in mind when you’re reading a datasheet.
How agents differ from rules and predictive models
Rules-based automation | Predictive AI | Agentic AI | |
|---|---|---|---|
| What it does | Follows a path you defined | Scores or forecasts an outcome | Chooses an action and performs it |
| When something unexpected appears | Stops, or routes to a queue | Flags it for a person | Reasons through it, or escalates by policy |
| What your team does | Builds and maintains the rules | Interprets the score and acts | Sets policy, reviews exceptions, audits |
| An AR example | Send dunning notice on day 31 | Predict this customer pays around day 47 | Match a short payment to a trade promotion, open a dispute case, route it to the right owner |
None of this makes the first two columns obsolete. Rules still handle the clean, repeatable majority of receivables work perfectly well, and they’re cheaper to run. Agents earn their place on the messy remainder, which is usually where the aging and the overtime live.
Where agents are earning their keep in AR right now
Cash application
Cash application is the clearest early win, because a match is either right or it isn’t. Remittance data arrives as PDFs, portal downloads, emails, and EDI files, often in a different format for every large customer. An agent can retrieve the remittance, read it, reconcile it against open items, handle a partial payment that reflects a deduction, and post the result. Serrala’s cash application solution runs at automation rates of up to 99% in customer environments, which gives you a sense of the ceiling on well-structured payment data.
Unapplied cash is the metric to watch here, more than headcount.
Collections prioritization and outreach
Most collections teams are working an aging report that treats a 45-day balance from a reliable customer the same as a 45-day balance from an account that has slipped three quarters running. Agents change the sequencing. They rank accounts by a combination of value at risk and likely payment behavior, draft the outreach, log promises to pay, and follow up when a promise lapses. Serrala’s collections management capabilities apply that prioritization inside SAP or in the cloud, depending on where the process lives.
The human part stays human. Deciding whether to press a key account in the middle of a contract renewal is not a task you hand to software.
Disputes and deductions
Deductions are where cash quietly leaks. A short payment arrives, someone has to work out whether it reflects a valid promotion, a pricing error, a delivery shortage, or an unauthorized chargeback, and by the time that’s resolved the trail has gone cold. Serrala launched a Disputes and Deductions Cloud solution in June 2026 built around centralizing that lifecycle and surfacing the root causes behind repeat claims. Root cause is the interesting part: resolving a dispute faster is useful, but preventing the next forty of the same type is where the working capital shows up.
Credit and risk monitoring
Credit reviews tend to happen on a calendar rather than in response to events. Agents can watch payment behavior, external risk signals, and exposure continuously, then propose limit changes or flag deterioration between formal reviews. Serrala’s credit and risk management approach keeps decisions within limits your credit policy defines, so a person is approving the policy rather than approving every individual case.
What agentic AI in AR still doesn’t do well
This part gets skipped in most articles on the topic, which is a shame, because it’s the part that determines whether your project joins Gartner’s 40%.
Agents inherit your data problems
They don’t fix duplicate customer masters, inconsistent payment terms across entities, or a chart of accounts that three acquisitions have made unrecognizable. If your matching rate is poor today because remittance data never arrives in usable form, an agent will be poor at matching too. Fixing the upstream data is unglamorous and it’s usually the highest-return work available.
Contract ambiguity is still a human problem
Rebate terms buried in a 2019 master agreement, verbal side arrangements a regional sales director made, pricing exceptions that were never written down anywhere. An agent can only reason over what it can read.
Relationship judgment doesn’t automate
Your five largest customers account for a disproportionate share of receivables and an even larger share of the consequences if the collections tone is wrong. Agents should handle the long tail and prepare the ground for the top accounts, not run them.
Novel situations have no precedent to learn from
A customer entering administration, a sanctions change affecting a payment corridor, a currency control that appeared last week. These need a person, and the useful design question is how quickly the system recognizes it’s out of its depth and hands over.
Being candid about these limits is not a reason to wait. It’s how you scope a project that survives its first quarter.
Governance decides whether this works
The reason so many agentic projects stall isn’t model quality. It’s that nobody could answer the audit question.
Deloitte’s Q2 2026 CFO Signals survey, published in July 2026, found 43% of CFOs citing potential litigation from the use of protected or private content as their leading external AI concern, followed by cybersecurity at 41% and regulatory complexity at 36%. Those aren’t abstractions in receivables, where you’re handling customer payment data, credit files, and contractual terms.
Four questions are worth settling before anything goes live:
- What is the agent allowed to do without asking? Post a match under a set variance threshold, yes. Write off a balance, probably not. The threshold should be a policy decision, documented, and adjustable.
- Can you reconstruct any decision afterward? Every action needs an audit trail showing what data the agent saw, which rule or policy applied, and what it did. If your auditors can’t follow it, you’ll end up reviewing everything manually and losing the benefit.
- Who owns the exception queue? Agents produce fewer exceptions but harder ones. Someone senior enough to resolve them needs the time.
- How does autonomy expand? Sensible programs start narrow and widen the mandate as results hold up, rather than switching everything on at once.
Serrala built its governed AI agents, launched in April 2026, around exactly this: human-in-the-loop oversight, role-based governance, and a fully auditable record of every action across AR, AP, and payments. You can see how that fits into the wider architecture in our write-up of the Serrala Finance Platform.
A starting sequence that doesn’t bet the quarter
Pick a process with a verifiable right answer
Cash application matching, remittance retrieval, and dispute intake all qualify. You can check the output. Start where success is unambiguous, because that’s what builds the internal case for the next step.
Baseline before you automate
Record your current straight-through matching rate, unapplied cash balance, DSO, average dispute resolution time, and hours spent on manual follow-up. Do it now, in writing. Six months from now nobody will remember what the starting point was, and a program without a baseline is very hard to defend in a budget review. Our AR automation ROI calculator is a quick way to frame the numbers if you need a starting model.
Write the escalation rules before the automation rules
Decide what the agent does when it isn’t confident, and who receives the handoff. Teams that define this first tend to move faster later, because expanding autonomy becomes a matter of adjusting a threshold rather than rewriting the design.
Measure working capital, not activity
Task counts and hours saved are easy to report and easy to dismiss. DSO, unapplied cash, bad debt provision, and dispute cycle time are the numbers your CFO carries into a board meeting. Connecting AR performance to those measures is largely a reporting problem, and Serrala Analytics exists to close that gap.
Where this is heading, with the appropriate caution
Speculating openly: the next stage looks less like better individual agents and more like coordination between them. A credit agent that notices deterioration, a collections agent that adjusts sequencing accordingly, and a disputes agent that spots a pricing error causing the short payments in the first place, all working from shared data rather than in separate applications. Serrala has described this as multi-agent orchestration, and elements of it are running in production today.
What’s genuinely unclear is how fast the trust threshold moves. Finance teams have been comfortable with automated cash matching for years. Comfort with an agent adjusting a credit limit or negotiating a payment plan will take longer, and honestly it probably should. Anyone who tells you they know the timeline is guessing.
Frequently asked questions about agentic AI in AR
What is the difference between AR automation and agentic AI in AR?
AR automation follows predefined workflows: an invoice goes out on a schedule, a reminder fires on a set day, a payment matches when the reference number is clean. Agentic AI adds decision-making. It handles cases the workflow didn’t anticipate, choosing an action within policy limits instead of stopping and queuing the item for a person.
Can AI agents make financial decisions on their own?
Within boundaries you set, yes. Most implementations let agents act autonomously on low-risk, high-volume decisions such as matching a payment inside a defined variance, and require human approval for anything with balance-sheet consequences. The governance framework, not the technology, defines where that line sits.
How long does it take to see results from agentic AI in AR?
It depends heavily on data readiness. Teams with clean customer master data and structured remittance flows often see movement in matching rates within a quarter. Organizations working across multiple ERPs and years of inconsistent data usually spend the first months on the data itself. That work isn’t wasted, but it should be in the plan and the timeline.
Does agentic AI in AR require replacing our ERP?
No. Agents run alongside existing systems, whether SAP-embedded or cloud-native. Serrala works across SAP, Microsoft, NetSuite, Sage, and other ERP environments, and the practical constraint is usually data access rather than the ERP itself.
Which AR process should we automate with agents first?
Cash application, for most organizations. Volume is high, the correct answer is verifiable, the effect on unapplied cash is immediate, and success there makes the case for collections and disputes considerably easier to argue.
If you’re weighing where agentic AI fits in your receivables operation, our overview of AR automation covers the process view, and Serrala AI covers how the agents work across finance. For a wider picture of how these tools are changing the finance function, AI in accounting is a useful next read.
