Predictive Process Monitoring

Traditional process monitoring tells you what has happened: which cases completed on time, where the bottlenecks were, how many exceptions occurred last month. It is valuable, but by definition retrospective. By the time a problem appears in a monitoring dashboard, the affected case has already been delayed, the compliance deadline has already been missed, the customer has already received a late shipment.

Predictive process monitoring shifts the time horizon. Instead of reporting on completed cases, it scores running cases in real time — estimating the probability that each in-flight case will result in a late delivery, a payment default, a control violation or any other outcome that the organization wants to prevent. With this information, process owners can intervene while intervention is still possible.

How Predictive Process Monitoring Works

The core of predictive process monitoring is a machine learning model trained on historical event log data. The training data consists of completed cases with known outcomes. The model learns to associate patterns in the partial execution of a case — the activities completed so far, the time elapsed, the case attributes — with the probability of a given outcome.

At runtime, the model is applied to cases currently in progress. Each case receives a prediction score that is updated as new events are recorded. When the score crosses a defined threshold, an alert is triggered — routing the case to the appropriate person or automated action.

Three Prediction Types

• Outcome prediction: Will this case end in a specific outcome? (late payment, approval rejection, delivery failure)
• Remaining time prediction: How long will this case take to complete from its current state?
• Next activity prediction: Which activity is most likely to occur next in this case? Useful for resource planning and exception detection.

The Intervention Window: Why Timing Matters

The value of a prediction is directly related to the available intervention window — the time between when the prediction is made and when the outcome becomes irreversible. A prediction made 30 minutes before a deadline has little practical value. A prediction made 5 days before a payment due date, while there is still time to contact the customer and arrange payment, has significant commercial value.

Designing predictive monitoring around intervention windows requires understanding the process:

• What is the typical duration between the predictable early warning signal and the eventual outcome?
• What action can be taken in that window, and by whom?
• What is the cost of a false positive (unnecessarily triggering an alert) vs. a false negative (missing a real risk)?

Calibrating prediction thresholds to optimize this trade-off is a key part of implementing predictive monitoring in practice.

High-Value Use Cases

Accounts Receivable

Predict which open invoices are at risk of late or non-payment based on customer payment history, invoice age, dispute patterns and order characteristics. Route high-risk invoices to collections teams before the due date — enabling proactive follow-up that retrospective aging reports do not support.

Purchase-to-Pay

Predict which purchase orders are at risk of late delivery based on supplier history, lead time variance, order complexity and current order volume at the supplier. Trigger sourcing alternatives before stock-outs occur rather than after.

Approval Workflows

Predict which approval cases will breach their SLA based on approver workload, historical approval times for similar cases and time already elapsed. Escalate approaching-SLA cases before the breach rather than after.

Compliance Monitoring

Predict which transactions are likely to result in a control violation — a posting without an approval, a three-way match exception, a segregation-of-duties conflict — while the transaction is still in progress and can be stopped or redirected.

Data Requirements and Model Accuracy

Predictive models are only as accurate as the data they learn from. Three factors most affect prediction quality:

• Historical volume: Models need enough completed cases to learn reliable patterns. As a practical minimum, several hundred examples of each outcome class are needed for supervised classification models.
• Data completeness: Missing events, imprecise timestamps or incorrectly recorded case attributes degrade model accuracy. ERP systems typically provide high-quality event data, but data cleaning is almost always required before model training.
• Concept drift: Process behavior changes over time — new suppliers, new products, economic shifts, regulatory changes. Models trained on older data may lose accuracy as the process evolves. Periodic retraining is a routine operational requirement.

Standard Business Central reporting is retrospective by design. Monitoring in-flight cases and scoring them against predictive models requires a dedicated process intelligence layer that reads live transaction data continuously — capabilities that are not part of standard BC functionality.

Related Concepts

• Process Monitoring
• Process Mining
• AI Process Mining
• Event Log
• Key Risk Indicators
• Conformance Checking
• Process Analytics