What is being predicted
Case-level prediction targets are concrete. For an open invoice: will it be paid late, and by how many days? Will it be disputed? For an order line: will it miss the promised date? For a ticket: will it breach its SLA, be reopened, escalate? Alongside these outcome targets, models predict the next activity and the remaining time, and flag cases whose trajectory looks unlike anything normal — anomalies that no rule yet describes.
How the model sees a case
A running case is a prefix: the events so far, their timing, their attributes, and the objects involved — the supplier's history, the customer's payment pattern, the approver's current load. Models are trained on thousands of completed cases where the outcome is known, then applied to open cases as new events arrive. Gradient-boosted models on well-engineered prefix features are a strong, explainable starting point for tabular ERP data; sequence models can add lift where the order of events carries more signal. The choice should be made by back-testing on the organization's own history, not by fashion.
Calibration: making 70% mean 70%
A probability is only useful if it is honest. If the cases scored at 70% risk go wrong 70% of the time, the number can drive decisions — how many cases to work today, which to escalate, where a threshold belongs. Calibration is measured, monitored for drift as the process changes, and reported. A model that is confidently wrong is worse than no model.
Explanations in process language
A score without a reason is hard to act on and easy to distrust. Explanations translate model attributions into the vocabulary of the process: second approval loop; supplier disputed three of the last five invoices; amount above approver limit; no promise-to-pay recorded. These are the drivers a collections specialist or an AP lead already recognizes — and they point directly at the intervention.
The intervention window
Prediction is only valuable while something can still be done. Each prediction should carry the time remaining in which an action can change the outcome — days until the discount window closes, until the promised ship date, until the SLA clock runs out. Ranking by risk alone sends teams to cases that are already lost; ranking by risk, value exposed and remaining window sends them where the head start is worth the most.
About accuracy claims
Prediction quality depends on the process, the data and the horizon. A published “accuracy” percentage without those three is marketing, not evidence. The honest approach is to agree the target and the quality threshold at the start of an engagement, back-test on history before anything reaches a queue, and report quality continuously in the product afterwards.
Rank by risk, value and window — not by risk alone. The head start is the product.