Prescriptive process monitoring

What is prescriptive process monitoring?

Prescriptive process monitoring goes one step past prediction. While a case is still running, it recommends the action most likely to improve the outcome and weighs that action against its timing, cost, and side effects.

Predictive process monitoring might estimate the chance that a case misses its deadline. Prescriptive process monitoring asks the follow-up question: is it better to speed the case up, reassign it, request more information, or do nothing at all? The recommendation takes account of capacity, policies, and knock-on effects, not just the raw risk score.

From prediction to intervention

A high risk score does not mean an action will help. Some cases run late for a root cause that can no longer be changed, so flagging them and intervening only wastes effort.

On top of a risk model, the system needs to know which actions are available and what effect each one tends to have on similar cases. A sound decision compares the likely outcome under each possible action, including doing nothing. That is a question about cause and effect, not only a forecast.

An example from service cases

A model predicts that a complex case will probably overrun its resolution time. The options might be to assign a specialist, ask the customer for more detail, or raise the priority.

A specialist tends to help most once the type of problem is known. Escalating too early instead creates an extra handover and more waiting. So the advice is held back until there is enough information and a specialist is actually free. The person handling the case sees the estimated effect, the uncertainty, and the reason behind the suggestion.

Treatment effect and uplift

A treatment effect is the difference between the expected outcome with an action and the expected outcome without it, for a single case or a group.

Uplift models try to find the cases that improve because of the action. They separate those from cases that would have turned out fine anyway and cases that end badly whatever you do. Acting only where the uplift is real is what makes the effort pay off.

The catch is that historical data rarely contains actions that were assigned at random. Selection bias and confounding factors make the effect hard to estimate from records alone.

Actions and constraints

Start from a short catalogue of actions the organisation can actually carry out, each with its conditions, cost, and owner. There is no point recommending something the process cannot reliably deliver.

Constraints include capacity, separation of duties, customer preference, regulation, and a ceiling on how often you intervene. When several cases compete for the same scarce resource, an optimiser can spread it where the expected gain is largest, for example assigning five available specialists to the cases where they will help most.

Timing

An early intervention has more time to work but rests on less information. A late one can be aimed more precisely yet may arrive too late to matter. Model the decision points and the stopping rules, and do not repeat the same advice after every event without recording that the action has already been taken. Factor in the cost and disruption of the action itself, since an extra call or email can add friction for the customer.

How it differs from predictive monitoring and business rules

Predictive monitoring is judged on how well it forecasts the next event, the outcome, or the remaining time. Prescriptive monitoring is judged on the decisions it drives and the effects they produce. A slightly less accurate predictor can lead to better actions if it is good at spotting the risks that are actually treatable, so measure realised uplift, cost, and unwanted effects, not only accuracy.

A business rule prescribes a fixed action under known conditions. It is easy to understand and fine for stable policies. A prescriptive model uses data to estimate context-dependent effects, which asks for more evaluation and monitoring in return. The two work well together: rules mark out which actions are allowed, and the model ranks the options inside that boundary. This is also where prescriptive monitoring overlaps with an AI agent that proposes next steps under human oversight.

Evaluating before go-live

A model trained on history can tell you what happened under the old way of working, but not directly what a new decision policy would cause. Off-policy evaluation tries to estimate the value of a different decision rule under assumptions about how actions were chosen in the past. Where it is responsible to do so, a randomised or phased pilot gives stronger evidence of real effect. Set stopping criteria in advance and protect vulnerable groups.

What to watch out for with prescriptive process monitoring

Feedback loops distort the next model. Advice changes the process and therefore the data future models learn from. A successful intervention can make a high-risk case look as if it was never at risk. Log the prediction, the information available, the advice, the action chosen, any override, and the result as separate facts, and do not train blindly on the final status.

Allocation can be unfair. A resource limit can quietly disadvantage certain customers or regions. Analyse how actions and outcomes break down by relevant group, and let people override a recommendation with a reason.

Human oversight needs teeth. A review is only meaningful with enough context, time, and authority to act on it. A mandatory click box over a score nobody understands is not real control.

Prove it against a baseline. Prescriptive monitoring earns its place only when the recommended action produces measurably better outcomes than a simple, explainable rule. Keep a safe fallback for when the model or an integration fails, and version the model, the policy, and the action catalogue together.