Finance teams evaluating AR technology get pitched a lot of vague claims about what AI will do for their collections process. "Optimize your outreach." "Recover more cash." "Reduce DSO automatically." These statements are not technically false, but they are not informative either. They do not tell you which specific decisions the AI is making, what inputs it requires, where it fails, or what it leaves entirely to humans.
I want to describe what AI in AR collections actually looks like at the decision level, because that is the level at which a finance team needs to evaluate it. Not capability categories. Actual decisions.
Decision 1: When to Send a Reminder
This is the timing prediction problem. The input is historical payment data for a specific customer: when they have received invoices, when they have responded to reminders, when they have actually paid, and what day-of-week and time-of-day patterns appear in those responses. The output is a prediction: for this customer, the next scheduled contact should be delivered at this specific time.
The model here is not complex by ML standards. It is essentially a behavioral pattern-matching problem. The signal it is looking for is consistent: do certain time windows produce faster responses from this customer than others? If a customer has responded to 8 of their last 10 reminders within 24 hours when those reminders were sent on Tuesday mornings, and ignored 5 of 7 reminders sent on Friday afternoons, the prediction is straightforward.
What this does not do: it does not predict whether the customer will pay. It predicts when they will be responsive to a contact. Those are different problems. A customer with a cash flow issue will not respond faster to a better-timed reminder. Timing prediction helps with the large share of late payments that are process delays or attention gaps, not financial inability.
Decision 2: Which Channel to Use
Channel routing uses the same behavioral signals in a different dimension. If a customer has a pattern of opening emails but not acting on them, while responding quickly to direct messages, that pattern informs the channel selection. If a customer has historically responded to email through day 30 but ignored email after that point, the model routes them to a different channel at day 31.
Channel routing in AccordX currently covers email, SMS, and phone-coaching prompts (structured talking point guidance for a human to place a call). We do not fully automate phone calls. The rationale is that phone calls in a collections context require real-time judgment that a model cannot reliably provide. The phone-coaching prompt is a middle path: the AI identifies that this account needs a phone contact and prepares the talking points, but the conversation itself happens with a human.
Decision 3: What to Write
Message generation is the most visible AI capability and the one that requires the most nuance to evaluate correctly. A language model generating collections messages is working from a set of inputs: the invoice details, the customer's response history, the aging stage, the account relationship context, and the tone guidelines set by the finance team. The output is a draft message calibrated to those inputs.
The critical limit here is that the model does not know what it does not know. If a customer is dealing with an internal restructuring that has slowed their AP process, a relationship manager might know this from a recent conversation. The model does not. If a customer has raised a pricing concern that is currently being resolved by the sales team, the model does not know unless that information has been explicitly flagged in the account context.
This is why exception routing matters so much in the architecture. Message generation works well when the account context is complete and the situation is standard. It should exit to a human queue when the context is incomplete or when behavioral signals suggest something unusual is happening.
What AI in AR Collections Does Not Do
It does not manage disputes. Dispute resolution requires understanding what the disagreement is, gathering documentation, coordinating with internal teams, and making judgment calls about whether to credit, adjust, or escalate. This is a human workflow with document management and cross-functional coordination. An AI can flag that a dispute is in progress and route the account out of the automated follow-up queue, but it does not resolve the dispute.
It does not make credit decisions. Evaluating whether to extend new credit, put an account on hold, or modify payment terms involves financial analysis and relationship judgment that goes beyond collections operations. Collections AI is not a substitute for credit management.
It does not guarantee payment. This should be obvious but is worth stating explicitly: timing a reminder better and choosing a more responsive channel improves the probability of getting a response and accelerating payment. It does not change the underlying financial situation of a customer who genuinely cannot pay. DSO improvement from better collections operations reflects improved process efficiency, not a change in your customer's payment capacity.
What Accurate Expectations Look Like
The teams that get the most value from AR automation are the ones with accurate expectations about what problem they are solving. If your DSO problem is primarily a process problem, meaning reminders go out at inconsistent times, use generic templates, ignore channel preferences, and miss accounts that fall below the priority threshold, then AR automation addresses that directly. The improvement is in process consistency and coverage across your portfolio.
If your DSO problem is primarily a credit quality problem, meaning a significant portion of your overdue receivables are from customers with genuine financial difficulties, better reminder timing and message drafting will not move the number much. That is a different problem requiring different interventions.
The honest framing: AR automation is a process improvement tool. It makes your collections process run at consistent quality across every account, every week, without the coverage gaps that come from manual prioritization. The ceiling on what it can accomplish is your portfolio's underlying collectability. Within that ceiling, the improvement is real and measurable.
