Policy should attach controls to an AI system in context, not to a model name in isolation. The same model may support low-stakes brainstorming in one setting and rank access to an essential service in another. A useful policy unit therefore includes the model, interface, input data, operator, affected population, decision authority, fallback process, and scale of deployment. Risk tiering uses that complete unit to reserve the strongest evidence, oversight, and reporting duties for situations where failure matters most.
Start with a scenario register. For each plausible harmful scenario $s$, estimate conditional probability $p_s$, severity $h_s$, and exposure $e_s$. A simple prioritization score is $$R=\sum_{s=1}^{m}p_s h_s e_s.$$ Exposure may represent users, decisions, or time at risk; severity may be a normalized impact scale rather than money. The units and time horizon must be stated. This score ranks attention, but it is not a claim that unlike harms have become morally interchangeable.
Average risk can hide concentrated or tail harm. Compute a score for each affected group, $R_g$, and inspect $R_{\max}=\max_g R_g$ alongside the aggregate. For rare severe outcomes, add a tail measure such as $$T_\alpha=\mathbb{E}[H\mid H\ge q_\alpha],$$ where $q_\alpha$ is a high quantile of harm $H$. Also record reversibility, ability to opt out, and dependence on the decision. An error that can be corrected before action is different from an irreversible denial discovered months later.
Map scores and qualitative factors to explicit tiers. For example, tier one may require documentation and basic testing; tier two may add subgroup evaluation, human review, and an appeal path; tier three may require independent evaluation, controlled deployment, incident reporting, and a demonstrated fallback. Thresholds $T_1<T_2<T_3$ should be written before reviewing a favored system. A critical-rights flag or catastrophic scenario should be able to raise the tier even when the aggregate score is below a numerical threshold.
Uncertainty belongs in the decision rather than in a footnote. If probability estimate $p_s$ has uncertainty allowance $u_s$, a conservative score can use $$R^{+}=\sum_s \min(1,p_s+u_s)h_s e_s.$$ Sparse subgroup data, distribution shift, automation bias, and correlated failures should widen uncertainty or trigger a higher tier. Teams should compare the AI system with a relevant human or non-AI baseline, but a weak baseline does not excuse avoidable harm. The decision record must identify residual risks that remain after controls.
Operationalize the tier with a control matrix. Record the intended use, prohibited uses, accountable owner, evidence required before release, monitoring metrics, review frequency, appeal route, and shutdown authority. Reassess when the model, data, interface, user population, decision consequence, or deployment scale changes. Useful diagnostics include unresolved high-severity scenarios, missing subgroup estimates, controls without owners, and risk assumptions that have not been tested. A tier is valuable only when it changes concrete obligations and can be audited later.