Sector notes
Where AI is actually safe to deploy
Deployment safety is no longer a yes or no question. It varies by industry, application and regulatory tier, and the organisations getting a return are matching scope to data readiness and the right level of human oversight.
Executive summary
4
consequence tiers that decide the oversight a deployment actually needs
Core conclusions
- Two systems on the identical model can sit at opposite ends of the risk register depending on what happens when they're wrong and how fast anyone finds out.
- The question has shifted from whether to deploy to where deployment is operationally viable — a question about consequence, not capability.
- Organisations capturing returns aren't deploying most widely; they're the ones whose scope matches their data readiness and regulatory tier.
For an infrastructure operator the question has shifted from whether to deploy to where deployment is operationally viable — which is a question about consequence, not about capability.
Two systems using the identical model can sit at opposite ends of the risk register depending on what happens when they are wrong, and how quickly anyone finds out.
Sorting by consequence, not by technology
| 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 |
Matching scope to readiness
The organisations capturing returns are not the ones deploying most widely. They are the ones whose deployment scope matches their data readiness and their regulatory tier, and who narrowed the first project until those three agreed.
That is the judgement we are actually paid for at the design stage — including saying which decisions in scope should not be automated yet.
