Your AI governance standard tiers each system by its use case - who is harmed if it is wrong and how badly - and it works for the twelve internal applications. It produces nothing usable for the general-purpose model your platform team is about to expose to every team through an internal API, because the use is whatever the caller does. OpenAI's Preparedness Framework, published in December 2023 and updated to version 2 in April 2025, tiers by capability instead: it tracks biological and chemical capability, cybersecurity and AI self-improvement against two thresholds named High and Critical, and requires safeguards that sufficiently minimise the risk of severe harm before a system that reaches High is deployed. What forced that shape, and where would copying it be a mistake?

The situation that forces it

Use-case tiering requires knowing the use, and a provider of a general capability does not know it. A model served through an API is used tomorrow for something nobody wrote down today, so the consequence of being wrong is unbounded at design time and the tier cannot be computed. The only properties the provider controls are what the model can do and what safeguards sit around it.

What they chose

Capability as the tiered object. Version 2 dropped the earlier four-level low-to-critical scale to two decision thresholds, which matters more than it looks: a threshold only earns its place if something specific happens when it is crossed. High capability requires safeguards that sufficiently minimise the risk before deployment; Critical requires them during development as well. A named internal body, the Safety Advisory Group, reviews whether the safeguards are sufficient and recommends an outcome to leadership rather than deciding alone.

The structural choice is tiering the thing you can measure, and gating the thing you can withhold - the deployment.

Why it fits their constraints

Three properties of the provider's position make it the only workable shape:

• Evaluations can be run by the provider on the model itself, repeatedly, before anyone deploys it.
• Thresholds set in advance convert a judgement call under launch pressure into a pre-registered commitment, which is the same reason a release gate beats a release review.
• A small number of categories can be resourced. A category list that grows with every new risk idea becomes a document nobody runs.

What it costs

The tier says nothing about the harms that dominate most organisations' actual exposure. A model can sit far below any frontier capability threshold and still approve the wrong refund, cite a hallucinated policy to a customer or leak one tenant's ticket into another's answer. A governance scheme whose thresholds never fire is indistinguishable from no scheme, and it costs the credibility you need for the controls that do matter.

Where copying it would be a mistake

Do not import capability thresholds into an enterprise that controls its own uses. If you know the use, tier by consequence - who is affected, how reversible the decision is, whether a human decides. If you provide a capability to others, tier by capability and publish the limits. If you do both, run both and bind them in the gateway: the calling application presents a use class and a tier, and the gateway refuses a use class the model has not been cleared for. That admission check is the only place the two schemes meet in code rather than in a policy document.

The second copying error is organisational. A standing advisory group of nine people reviewing twelve internal applications is the design-authority queue failure with new vocabulary. At that volume the review belongs in the pipeline, with the group seeing only what the rules flag.

Common weak answers

• "Tier by model size or by whether it is generative." Neither predicts harm. A small classifier deciding eligibility outranks a large model summarising meeting notes.
• "One process for everything." It produces a questionnaire that the twelve applications pass and the platform cannot answer.
• "Wait for the regulation to tell us the tiers." Tiering is how you decide where your own controls go. External classifications land on top of it.