Your organization is shipping AI features whether or not you have a process for it. Someone in marketing is drafting copy with a model. Someone in support has wired a chatbot to your knowledge base. Someone in engineering just gave an agent write access to a production system because it made a demo look great. The question is not whether AI is entering your environment. It is whether each use case gets scrutiny before it goes live, or after something goes wrong.
The answer is not a heavyweight review board that teams route around. It is a lightweight risk-tiering step: a few minutes of structured questions at the start that sort each AI use case into a tier, and that tier decides how much review the use case gets and which controls apply. Low-risk uses sail through. High-risk uses get the scrutiny they deserve. Nobody wastes time, and nothing dangerous ships unexamined.
Why Tiering, And Why Now
AI risk is not evenly distributed. A model that suggests email subject lines and a model that autonomously approves refunds are not the same risk, and treating them alike is the mistake in both directions. Review everything heavily and teams stop telling you about their AI projects. Review nothing and you learn about the autonomous refund agent from your finance team's variance report.
This is exactly the problem the NIST AI Risk Management Framework (AI RMF) is built to address. Its core functions (Govern, Map, Measure, and Manage) are about matching your effort to the actual risk of a system. Tiering is how you operationalize the Map function: you cannot manage what you have not first characterized. On the technical side, the OWASP Top 10 for LLM Applications names the concrete failure modes to watch for, including prompt injection, insecure output handling, sensitive information disclosure, and excessive agency. Tiering is how you decide which of those matter for a given use case.
Tiering is not bureaucracy. It is the step that lets you skip bureaucracy for the ninety percent of AI uses that do not need it, so you can concentrate on the ten percent that do.
The Four Questions That Set The Tier
You can determine an AI use case's tier with four questions. They are deliberately simple, because a tiering step nobody can answer in five minutes is one nobody uses.
- Data sensitivity. What data does the model see or produce? Public marketing copy is one thing; customer records, source code, health or financial data, or anything covered by a regulation is another. The more sensitive the data, the higher the tier, and the more the OWASP concern of sensitive information disclosure applies.
- Autonomy. Does the model only suggest, with a human acting on the output, or does it take action on its own? A model that drafts a reply is low autonomy. An agent that sends the reply, calls an API, moves money, or changes a system is high autonomy, and directly triggers the OWASP excessive agency risk.
- Audience. Who is exposed to the output? An internal tool used by ten trained employees is lower risk than a feature facing your entire customer base or the public, where prompt injection and insecure output handling become live threats that adversaries will probe.
- Human oversight. Is there a human in the loop who can catch and correct a bad output before it causes harm, or does the output flow straight through to real-world effect? Strong oversight lowers the tier; no oversight raises it.
Notice how these four combine. High sensitivity plus high autonomy plus a public audience plus no oversight is your top tier, because that is exactly the combination that turns a model error into an incident. Low sensitivity, suggestion-only, internal, with a human reviewing every output is your bottom tier, and it should move fast.
How The Tier Routes The Use Case
The point of a tier is that it decides what happens next automatically. The team does not negotiate for a review path; the tier assigns it.
- Low tier: self-service. The team documents the use case, confirms a short baseline of controls (no sensitive data, logging on, a clear owner), and ships. No meeting required.
- Medium tier: a lightweight review. A security reviewer checks input and output handling, confirms the data flows, and validates the relevant subset of the OWASP Top 10 for LLMs against the design.
- High tier: full review with named controls. Threat modeling, prompt-injection testing, output validation, least-privilege scoping of any tools the model can use, human-in-the-loop checkpoints, and monitoring, all mapped to NIST AI RMF Measure and Manage before launch.
This routing is the whole payoff. It gives fast-moving teams a clear path to ship low-risk AI without friction, and it guarantees that the genuinely risky use cases cannot quietly skip the review they need. The tier does the sorting, not a person, so it scales as your AI footprint grows from three use cases to three hundred.
Make It Part Of The Intake, Not An Afterthought
Tiering only works if it happens before the build. The moment to tier an AI use case is when it is proposed, as part of whatever intake your teams already use for new projects. Bolt it onto the existing request, keep it to the four questions, and make the output a tier plus a routing decision the requester sees immediately.
Done this way, tiering is not a gate that teams resent. It is a quick sorting step that most use cases pass in minutes, and that quietly catches the dangerous ones before they reach production. Over time it also becomes your inventory: a running record of every AI use case, what data each touches, how autonomous each is, and what controls each carries. That inventory is something the NIST AI RMF Govern function asks for and most organizations lack.
If you want to stand this up without designing it from scratch, start with the AI Use Case Risk Tiering tool, which walks a use case through these questions and returns a tier and a recommended review path. Pair it with the plain-English guide to AI security so your team understands the NIST AI RMF and the OWASP failure modes behind the tiers. Tier your AI use cases now, before an incident does the tiering for you.