Organizations need a way to distinguish an AI feature that drafts internal notes from a system that influences a consequential business decision. Without classification, governance tends to fail in one of two directions: everything receives the same heavy review, or teams decide informally that their own use is low risk.
A practical tiering model gives teams a consistent starting point while still allowing specialists to escalate unusual cases.
Classify the use, not just the model
The same underlying model can support very different activities. Risk depends on what the system is doing, who is affected, what data it uses, how much autonomy it has, and what happens when it is wrong. Classify the specific use case rather than assigning one permanent risk label to a vendor or model.
Consider consequence
Ask what a plausible failure could affect. Internal convenience is different from access to employment, money, essential services, legal rights, safety, or health. Higher-consequence uses should generally receive stronger review, testing, oversight, and monitoring.
Consider autonomy and human control
A system that suggests a draft for a knowledgeable employee to review is different from one that takes an external action automatically. Examine whether a person can meaningfully review the result before the consequence occurs, whether they have enough information to disagree, and whether the action can be stopped.
Consider data sensitivity and exposure
Personal, confidential, regulated, security-sensitive, or proprietary information can increase the impact of misuse or failure. Identify what data enters the system, what providers receive it, what is retained, and what other systems the workflow can reach.
Consider uncertainty and reversibility
Some errors are easy to detect and undo. Others may be difficult to identify, affect many people quickly, or create consequences that cannot be fully reversed. Novel uses, weak testing evidence, rapidly changing dependencies, and poorly understood failure modes are reasons to increase scrutiny.
Use tiers to determine controls
A simple internal model might use low, moderate, high, and prohibited or restricted categories. Low-risk uses can follow standard controls and lightweight registration. Moderate uses may need documented testing and owner approval. High-risk uses may require security, privacy, legal, compliance, domain, and executive review. Some uses may be outside the organization’s risk appetite altogether.
The exact tiers should reflect the organization’s obligations and operating context. Classification is a governance tool, not a substitute for applicable law or specialist advice.
Reclassify when the use changes
Risk classification is not permanent. New data, expanded permissions, new users, increased autonomy, different decisions, model changes, or deployment at greater scale can move a system into another tier.
The purpose of classification is proportionality: spend the most governance effort where failure matters most while keeping ordinary low-risk work understandable and manageable.