Starting question
A use case begins with observable friction.
Delays, rework, duplicated steps and decisions made without information are more useful starting points than a generic automation request. They allow the current process to be compared with one bounded change.
Efficiency is not the only outcome: quality, worker autonomy, verification burden, safety and routes for challenge can improve or deteriorate.
Operational relations
Five steps that must remain connected.
Activity
Identify real work, exceptions and the expected result.
Friction
Expose delays, errors, rework and loss of context.
Assistance
Define what the system prepares, proposes or performs.
Decision
Assign review, approval, stopping and challenge.
Effect
Observe quality, time, workload, risk and effects on people.
Boundaries and responsibility
A metric must not hide the work required to produce it.
A faster response can transfer hours of checking to other people. More uniform classification can hide important exceptions. Evaluation should include the new work created by the system, not only the work it appears to remove.
When AI monitors performance or informs decisions about people, transparency, proportionality, representation and responsibility require specific examination.
Practical object
AI Process Decision Map
Complete the AI Process Decision Map. Each field exposes a relationship to verify before extending the system.
Process
Which event starts and which result closes the work?
People
Who operates, checks, decides and experiences the effect?
Friction
Where are time, quality or context lost?
AI intervention
Which step changes in practice?
Information
Which data and sources support that step?
Decision
Who can correct, stop or challenge it?
Measure
Which signals expose benefit and harm?
Fallback
How does work return to a known process if the system fails?
First test
A short test should produce knowledge, not merely an output.
- 01Observe the process
Collect ordinary cases, exceptions and informal steps with the people doing the work.
- 02Choose one friction
Define the problem and outcome before selecting a technical solution.
- 03Simulate the new step
Use controlled data and keep the final decision outside automation.
- 04Measure the whole work
Include correction, checking, escalation and newly created tasks.
- 05Decide with people
Review outcome, impact and conditions for extension with operators and accountable owners.
Public sources
References for verification and further work.
- ILO, Generative AI and Jobs: A Refined Global Index.
- ILO, The impact of GenAI on jobs, productivity and work organization.
- NIST AI RMF Core.
These sources support initial design. Legal, professional, ethical and organisational requirements depend on the case and the responsible functions.
Frequently asked questions
Manager or process owner: questions to clarify before the project.
How do you choose a process suitable for AI?
Start from observable friction, available sources, a verifiable result and manageable consequences. Frequency alone is insufficient when errors and responsibility remain opaque.
Which metrics are needed?
They depend on the outcome and may include quality, errors, total time, verification burden, exceptions, safety and effects on people. Baselines must be measured rather than invented.
Does AI replace a process step?
It may assist or transform specific tasks. Coordination, checking and exceptions created by the new step must also be observed.
Who should approve the project?
The owner of the outcome together with the relevant people, data, security, technology and compliance functions.
How is an alternative preserved?
Document the known process, retain necessary data and access, and define stopping and fallback conditions before extension.