How should AI systems be risk-tiered so that governance is proportionate?

What determines the tier

Not model capability or size, but the consequence of being wrong and the degree of autonomy.

• Consequence: does an error affect someone's access to a service, their money, their safety, their rights? A recommendation that is wrong wastes attention; an eligibility decision that is wrong causes real harm.
• Reversibility: can the affected person notice and challenge the outcome, and can it be undone?
• Autonomy: does the system decide, or does it inform a person who decides? An automated decision with no human in the path is categorically higher risk than the same model producing a suggestion.
• Population and scale: how many people, and are any of them in a position of particular vulnerability?
• Contestability: is there a path for someone to seek review?

What each tier should require

• Low: documented purpose, basic monitoring, an owner.
• Medium: independent validation, performance monitoring segmented by group, human review of a sample, documented limitations.
• High: pre-deployment evaluation including adversarial testing, meaningful human oversight, continuous monitoring with defined intervention thresholds, an appeal route, and a documented impact assessment.

The error that recurs

Tiering by technology rather than by use. A large model producing marketing copy is lower risk than a simple scoring rule denying credit, and a tiering scheme that ranks by model sophistication gets both wrong.

What makes it work

Tier assigned at design time and reassessed on material change, with the tier determining the controls automatically rather than through negotiation. And a documented escalation route for genuinely novel uses, because a tiering scheme fixed at design time will not anticipate what teams build next.