An action risk tier is the category an action falls into based on how much harm it can do and how hard it is to undo. It decides the minimum depth an AI agent needs before it may take that action, whether a person must release it, and how many approvers are required. Tiers are set per action, before any agent is allowed near it.
Why do actions need risk tiers?
Because agents do many different things, and treating them all alike fails one way or the other. Treat every action as high-risk and the agent is useless, waiting for approval to read a policy. Treat every action as low-risk and the same agent can send a payment as freely as it drafts a note.
Tiers let you be strict where it matters and quick where it does not. They also make autonomy decisions consistent across agents: two agents that can both change a customer's address face the same rules, because the action has one tier.
How are action risk tiers set in practice?
List every action each agent could take, including side effects such as notifications. For each, ask two questions: what is the worst plausible harm, and can it be undone, by whom, and until when? The action reversibility register template records the answers. Then place each action in a tier.
A common set of tiers lines up with depth:
| Tier | Typical actions | Minimum depth | Default treatment |
|---|---|---|---|
| Read | Look up a record, search policy | 1 | Allowed within role |
| Draft | Prepare a letter, a credit memo, a payment file | 2 | A person sends or submits |
| Routine, reversible | Update a contact preference, reopen a case | 3 | Allowed within bounds |
| Hard to undo | Send a payment, close an account, file a report | 4 | Held for a named person; second approver for the highest amounts |
Borderline actions go up a tier, not down. Whether sending an email is an irreversible action is a good test case: once it reaches the recipient, it cannot be unread.
What does that look like in an example?
A wealth platform's servicing agent can do four things: view holdings, draft a switch request, update a client's communication preferences, and place the switch. Viewing is read. Drafting is draft. Preferences are routine and reversible. Placing the switch is hard to undo, because once it reaches the market, reversing it means a second trade at a different price. The agent may earn depth 3 for preferences while the switch stays held for an adviser.
How is an action risk tier different from depth?
The tier belongs to the action; depth belongs to the agent. A tier says "this action needs at least depth 4 and a person's release". Depth says "this agent has earned depth 3 for this capability". The action proceeds only if the agent's depth meets the tier and every other bound agrees. The two are set by different people for different reasons: process owners tier actions; certification evidence sets depth.
Which terms relate to action risk tiers?
- Staged action: how higher-tier actions are prepared and held for release.
- Compensating action: the remedy when a lower-tier action turns out to be wrong.
- The irreversible action checklist helps decide which tier an action belongs in.
- Regulated autonomy explains why reversal is the question that sets the tier.
Questions people ask
Who decides an action's tier?
The process owner proposes it and the risk owner signs it off. Disputed cases default to the higher tier until there is evidence to lower it.
Can an action's tier change?
Yes, if the action changes. Adding a reversal window, so a payment can be cancelled before submission, may move it down a tier. Removing one moves it back up.