Executive summary
4
consequence tiers that decide how much oversight a deployment needs
Key conclusions
- Two systems built on the same model can sit at opposite ends of the risk register. Their position depends on what happens when they are wrong and how quickly anyone finds out.
- The question has moved from whether to deploy to where deployment is operationally viable. The answer depends on the consequences of a wrong output.
- Organisations that capture returns match their scope to their data readiness and regulatory tier. They narrowed the first project until all three aligned.
For an infrastructure operator, the question has moved from whether to deploy AI to where deployment is operationally viable. The answer depends on the consequences of a wrong output and has little to do with the model's capability.
Two systems using the same model can sit at opposite ends of the risk register. Their position depends on what happens when they are wrong, and how quickly anyone finds out.
The consequence of a wrong answer sets the oversight required
| Consequence of a wrong answer | Oversight this implies |
|---|---|
| Cosmetic or easily reversed by the user | Sampling review, published error rate |
| Financial, bounded, recoverable | Threshold-based escalation with a named reviewer |
| Regulatory record or personal data | Full logging, retention, and an auditable decision trail |
| Physical action on plant, vessel or vehicle | Abstention path, hard interlocks, human authority to proceed |
Match the project's scope to your data readiness and regulatory tier
The organisations that capture returns match their deployment scope to their data readiness and their regulatory tier. They narrowed the first project until scope, readiness and tier all aligned.
Clients pay us for that judgement at the design stage. It includes telling them which decisions in scope should not yet be automated.
