An organisation deploys many AI systems with very different risk profiles. How should they be tiered, and which controls attach to each tier?

Why tiering is the first move

Applying uniform governance to every model produces one of two failures: controls heavy enough for a credit decision applied to an internal summarisation tool, which stops useful work and is routed around; or controls light enough for the summariser applied to the credit decision, which is the actual risk.

Tiering is what makes proportionate governance possible, and the tier must be determined by consequence rather than by technique — the model's architecture is irrelevant to the risk assessment.

The dimensions that determine tier

• Consequence of a wrong output. A misranked recommendation, a delayed ticket, a denied loan, a missed diagnosis — these differ by orders of magnitude and are the primary axis.
• Autonomy. Does the output inform a human, or does it act? A system that takes action without review is in a materially higher tier than one that suggests, regardless of what it does.
• Reversibility. Can the effect be undone, and how quickly?
• Scale and population. How many people are affected, and are any of them in a vulnerable group?
• Regulatory exposure — credit, employment, healthcare, insurance and law enforcement carry specific obligations independent of the organisation's own assessment.
• Contestability. Can a person affected by the decision challenge it, and is there a route?

Controls by tier

Low tier — internal, advisory, reversible: basic evaluation, monitoring for degradation, an owner, and a documented purpose. Proportionality means genuinely light, or the framework is discredited by its first application.

Medium tier — customer-facing, advisory or reversible: pre-deployment evaluation against a held-out set, drift and quality monitoring in production, a documented fallback for model unavailability, human review of edge cases, and a defined escalation path.

High tier — consequential decisions about people, or autonomous action: mandatory human review before consequential effect · bias and fairness testing across relevant groups, repeated in production rather than only pre-deployment · full documentation of training data, evaluation methodology and known limitations · adversarial testing and red-teaming · explainability sufficient to support a challenge · a contestation route with a human decision-maker · continuous monitoring for disparate impact · and an approval gate with accountable sign-off.

Highest tier — safety-critical or legally restricted: everything above, plus independent validation, regulatory engagement, and a deliberate decision about whether to deploy at all.

The controls people skip that matter most

• Production monitoring for fairness and drift. Pre-deployment testing establishes behaviour on the evaluation set; the population changes, and a model fair at launch may not be fair a year later — which nothing detects without ongoing measurement.
• The fallback. What happens when the model is unavailable, and has that path been load-tested?
• A tested contestation route, rather than a documented one.
• Retirement. A model with no owner and no review still running three years later is a common and serious finding, and it is invisible without an inventory.

And the inventory itself is the precondition: an organisation that cannot enumerate its deployed models cannot tier them, and most cannot.