A reference architecture for decision intelligence

This paper describes a layered architecture for decision-intelligence systems. It is intended for architects and analytics leaders who need a common vocabulary for planning such systems.

The foundation provides governed access to internal and external data, along with the context needed to interpret it: definitions, lineage and freshness. Without this layer, higher layers inherit its ambiguity.

Forecasts, classifications, simulations and optimisation results are produced here. Each output should carry an estimate of uncertainty and a pointer to the model version that produced it.

Policies, constraints and objective functions are expressed explicitly. Separating this layer from the models lets business owners change policy without retraining anything.

Recommendations reach people or systems through workflow with approval, override and logging. Outcomes are captured and compared with expectations, completing the learning loop.