What Makes a Proper Decision Support System?

Another tutorial from the AI tutorial series, and this time on a question very close to home:

📄 What Makes a Proper Decision Support System?

Its core claim is one I keep repeating to students and colleagues: a machine learning model is not, by itself, a decision support system — and neither is an optimisation model. A prediction has no decision maker, no alternatives, no trade-offs, and no accountability. An “optimal” solution often assumes the objective is known and fixed, when in practice it rarely is.

The tutorial goes back to the DSS classics (Simon, Gorry and Scott Morton, Keen, Little, Sprague) and forward to modern work on human–AI decision making, and synthesises twelve properties of a proper DSS in four layers: decision focus, human agency, epistemic transparency, and institutional embedding. It also turns them into a rubric you can use to score a real system. As with the others in the series, it is AI-generated: a good map, but go to the sources for anything you build on.

How it connects to my research

This is essentially the backbone of much of what I work on:

• From ML models to DSS — Transforming Machine Learning Models to Decision Support Systems argues for exactly this shift, and A comparison between DSS and ML models for churn prediction compares the two against DSS quality criteria.
• Interpretable, consistent decision models from data — DIDEX and DEX-LL learn DEX models with explicit, monotone, auditable decision rules.
• What-if analysis for policy makers — the Balkan Peace Index and the internal migration DSS let users explore how outcomes change under different scenarios.
• Accountability and fairness — my algorithmic fairness work, from FairDEA to FairAW, is about making sure decisions can be justified to the people they affect.

Enjoy, and let me know what you think.