There's a lot of noise about AI in AR right now. We break down which AI applications have real evidence behind them and which are pattern-matching rebrands.
Virtually every AR software vendor updated their marketing in 2024-2025 to include "AI" somewhere in the headline. The result is a landscape where it's genuinely difficult for finance teams to evaluate what's real, what's overstated, and what's just a rules engine with better copy. We built Cashvyne to do a specific thing — predict invoice-level 60 DPD risk before invoices go overdue — and we've had to think carefully about where machine learning actually adds value in this domain versus where it's being added for marketing reasons. This is our honest take.
What Actually Works: Three High-Confidence Applications
Payment timing prediction at the invoice level
This is the application with the strongest evidence base, and it's the one we've built around. The core task: given an open invoice and a customer's payment history, predict whether that invoice will be paid on time, 1-30 DPD, 31-60 DPD, or 60+ DPD. This is a classification problem with structured time-series data — historical payment records, invoice amounts, payment terms, days-to-pay across the last 12-24 months — and machine learning genuinely outperforms simple rules in this context.
Why does ML outperform rules here? Because the relevant signals are non-linear. A customer who paid late twice in the last 12 months is not simply "twice as risky" as one who paid late once. The pattern matters: did both late payments cluster at quarter-end? Did they follow large order volumes? Did they coincide with a period when the customer was also late on other suppliers' invoices (which you can infer from their payment timing drift relative to their own baseline)? A well-trained classification model can detect these interaction effects. A rules engine can't, unless someone has hand-coded every possible combination.
The honest caveat: this works best with at least 12-18 months of payment history per customer. For new accounts with no payment history, prediction reverts to base rates for the customer segment — which is still useful, just less personalized. For portfolios with very few accounts (under 50 customers), the training signal may be insufficient to outperform a well-calibrated rules-based system.
Cash application and remittance matching
Matching incoming payments to open invoices is a high-volume, low-judgment task — exactly the type of work where machine learning applied to historical matching patterns reduces both manual processing time and error rate. The AI learns which invoice fields to use as matching signals for each customer (some remit by invoice number, some by PO number, some by amount, some by a customer-specific reference code), and applies that learned mapping to new remittances automatically.
This application is well-established at the high end of the market (platforms like HighRadius have offered it for years). It's now reaching mid-market at more accessible price points. The improvement over rules-based matching is meaningful: training on historical matches handles exceptions and variations that rigid rules miss, which reduces manual exception handling. For a company processing 800-1,200 invoices per month, this is genuinely worth evaluating.
Dispute categorization and routing
When a customer emails in about an invoice dispute, the first step is classifying the dispute type (pricing error, quantity discrepancy, wrong billing address, duplicate invoice, service delivery dispute) and routing it to the right person. Natural language classification applied to dispute email text can do this accurately enough to reduce manual triage time. It's not a dramatic improvement, but it's real and measurable: average time from dispute receipt to first substantive response drops by 20-40% when classification happens automatically rather than requiring a person to read and route each email.
What Is Overstated: Three Applications That Underdeliver
"AI-generated" personalized dunning emails
The pitch: large language models generate a custom dunning email for each customer, personalized to their payment history and relationship context, that outperforms templated emails. The reality: most AR dunning emails are read by an AP clerk who processes dozens of vendor invoices per day. The marginal value of a more personalized email is low in this context — what matters is whether the email arrives on schedule, contains the right invoice reference, and asks clearly for payment. A well-designed template with dynamic fields (invoice number, amount, due date, days overdue) outperforms an LLM-generated email most of the time, because LLM generation introduces variability and occasional errors into a process where consistency and accuracy are what matter. We don't use generative AI for dunning messages at Cashvyne for this reason.
Credit risk scoring for new customers from "alternative data"
Several vendors claim to use alternative data sources — social media signals, news sentiment, business registry data — to score creditworthiness for new B2B customers who have no payment history with you. This sounds compelling but is difficult to validate independently. The signal quality from alternative data for mid-market B2B companies (where the "customer" is often a 20-person regional distributor with minimal social footprint) is weak. Traditional credit data — trade credit reports, D&B scores, credit bureau pulls — remains the more reliable input for initial credit decisions on unknown accounts. The alternative data story is more compelling for consumer credit and for large public companies where news flow is meaningful.
Predicted cash collection amounts with high precision
A common vendor claim: "our AI predicts exactly how much cash you'll collect each week." The mechanics here are sound in theory — aggregate payment probability scores across open invoices to get a probability-weighted expected collection amount. But the precision implied in these pitches often exceeds what the models can actually deliver, particularly for portfolios with high concentration in a few large accounts where the behavior of one customer swings the aggregate forecast significantly. A payment probability model gives you a distribution, not a point estimate. Vendors who present this as a precise weekly cash prediction are overstating what the underlying model can do. The useful output is a probability-weighted range with an explicit confidence interval — not "you will collect $X this week."
How to Evaluate a Vendor's AI Claims
When an AR vendor tells you they use machine learning or AI for a specific function, ask four questions:
What is the training data? A model trained on generic B2B payment data is different from a model trained on your payment history. Generic models have more data but may not capture the idiosyncrasies of your customer base. Per-account models need enough transaction history to train. Know which type you're being shown.
What is the evaluation metric and how was it measured? "Our model is 87% accurate" means nothing without knowing what the accuracy metric is (precision, recall, F1, AUC) and how it was measured (holdout test set, cross-validation, live A/B comparison). Be skeptical of accuracy claims that aren't accompanied by this context.
What is the fallback when the model doesn't have enough data? Every ML model degrades gracefully — or doesn't — when applied to thin data (new accounts, seasonal accounts, accounts with volatile payment history). How a vendor handles this edge case tells you a lot about the robustness of the implementation.
What does the model actually output, and how does it integrate into the workflow? A prediction that sits in a separate analytics tab nobody checks is not adding value. The output needs to be integrated into the AR team's daily work queue — a risk flag, a recommended action, a modified dunning sequence — not a separate report.
Where We Sit on This Spectrum
We're going to be direct: Cashvyne does one ML task well. We take your historical payment data, build a per-customer payment behavior model, and score open invoices for 60 DPD probability while they're still in the net-30 window. That score then drives which dunning sequence the invoice enters. We don't do generative dunning emails, alternative data credit scoring, or multi-week cash collection forecasting with point-estimate precision. The scope is intentional — we would rather do one prediction accurately than five predictions superficially.
The questions above are the same ones you should use to evaluate us. Our model requires at least 12 months of payment history to produce reliable per-customer scores. For accounts with thin history, we default to segment-level base rates. We evaluate model performance against holdout test periods from your own data in the initial setup — not against an aggregate benchmark from other customers. And the model output is a risk score that routes invoices into work queues, not a separate report that requires a separate workflow to consume.
We are not arguing that all AI claims in AR are false. The applications with real evidence — payment timing prediction, cash application, dispute classification — are genuine improvements over rules-based alternatives for teams with sufficient data. The test is whether the vendor can show you the evidence with appropriate rigor, not whether they can tell a compelling story about what the model does.