Decision Model

A decision model is a formal description of how a decision comes about. It specifies which information feeds into a decision, how that information is weighted or linked, and what result emerges from it. That sounds abstract, but it is very concrete: when a bank automatically decides on a loan application, it follows such a model. It specifies above which income, which credit score — that is, a creditworthiness value indicating how reliably someone repays debt — and which further conditions a loan is approved or rejected. The model makes the decision independent of any single person and, in principle, auditable by others.

Decision models as a foundation for explainable AI

In discussions about AI systems, one question keeps coming up: why did the system decide this way? Decision models are one answer to that. They make the logic of a decision visible and verifiable — both for people and for regulators.

This is especially important when a decision directly affects someone. If a job application is rejected, an insurance application denied, or a social benefits claim turned down, those affected have a legal right to an explanation in many countries. An explicit decision model is what makes this explanation possible in the first place. Systems without such a model — such as deep neural networks that do not disclose their internal computational steps — are considered “black boxes” and are therefore facing increasing regulatory pressure.

A common misconception: decision model and AI model are not the same thing. An AI model can be part of a decision model — for example, an image recognition system that provides an input. The decision model, however, is the overarching structure that determines what happens with that result afterward.

Structure of a decision model

A decision model typically consists of three elements: inputs, decision logic, and output. The inputs are all the information the model takes into account — for example, age, income, purchase history. The decision logic describes how this information is linked together. That can be a simple if-then rule, a table of point values, or a more complex statistical formula. The output, finally, is the result: approved or rejected, risk class A or B, price X or Y.

A widely used standard for representing such models is DMN — short for Decision Model and Notation. DMN is a graphical language that companies can use to write down decision logic in standardized tables and diagrams. A DMN decision table looks like a spreadsheet: the rows contain possible combinations of inputs, and the last column shows the corresponding result. This allows even someone without programming knowledge to read and check the logic.

The key difference from a simple algorithm — a sequence of computational steps — is that a decision model deliberately separates the decision logic from the technical code. Business experts can adjust the rules without developers having to change the source code. This speeds up adjustments and reduces errors.

Decision models in products and current debates

Decision models are embedded in many systems people use every day without noticing it. Credit card fraud is detected in real time using a model that checks whether a transaction matches previous purchasing behavior. Streaming services decide which content to recommend based on a model. Navigation apps calculate the best route according to a model that weighs distances, traffic conditions, and road types.

In the political debate over the EU AI Act — the first comprehensive law regulating AI in Europe — decision models play a central role. The law requires so-called high-risk AI, meaning systems used in sensitive areas such as job placement, lending, or law enforcement, to have their decision logic documented and explainable. Companies that cannot present clear decision models risk heavy fines.

The topic is also gaining importance in AI research. Researchers are working to extract decision models retroactively from existing neural networks — that is, to translate the implicit logic of a black box into explicit rules. This is technically difficult, but pressure from regulation and the public is making it increasingly necessary.