Start with the action, not the algorithm.
Before a model is selected, the work needs a decision that is specific enough to inspect. The Decision Test turns a broad ambition into an operational question: what is being decided, by whom, at what moment, and with what consequence if the signal is wrong?
A vague request such as “predict which claims will close” becomes a narrower support question: “Can the team prioritise open files for review when capacity is limited, without automating the closure decision itself?” The difference sets the target, the acceptable error and the role of human judgement.
A prediction task is only useful when its output changes a defined action without hiding who remains accountable for the final decision.
- 01Name the action
Describe the next action a person or team could take after seeing the result.
- 02Define the timing
Fix the moment at which the information must be available to be useful.
- 03Make the error concrete
Ask what false positives and false negatives change in the real workflow.
- 04Set the boundary
Write down what the system must never decide on its own.
Sources & rationale
- AI Risk Management Framework 1.0National Institute of Standards and Technology · 2023
Useful for framing intended use, context, impact and accountable human oversight before a system is measured.
- Reproducibility and Replicability in ScienceNational Academies of Sciences, Engineering, and Medicine · 2019
Supports explicit questions, transparent decisions and documentation that make a study inspectable.