An AI agent risk appetite statement says how much risk the board accepts from agents acting on the firm's behalf, written as actions and limits rather than adjectives. It sets, for each class of action, the highest depth any agent may hold, the limits that apply, the metrics that signal a breach, and who must act when one occurs.
Why write appetite in actions?
Most risk appetite statements for AI read like "low appetite for AI-related conduct risk". Nobody can test that against what an agent does at 3pm on a Thursday. An appetite expressed as actions can be: "no agent may release a payment above 10,000 without a named person" is either true or not.
That is the argument in risk appetite for AI agents should be written in actions. The template follows it. It keeps a short qualitative statement for the board, then translates it into action classes with ceilings and triggers that operations can enforce.
When should you use it?
Draft it before the first agent moves past a pilot, and have the board approve it alongside the AI agent policy. Revisit it annually and whenever the firm wants to allow a new class of action. Each pilot then works inside it: the pilot scope statement should cite the appetite lines it relies on.
Individual agents are then assessed against it. The AI agent risk assessment template compares an agent's residual risk with the appetite, line by line.
How do you fill in each field?
- Action class groups actions by effect: read, draft, change customer records, communicate externally, move money, make decisions about individuals.
- Depth ceiling is the highest depth any agent may reach for that class. Evidence may justify less; nothing justifies more.
- Limits add amount, volume or population caps.
- Key risk indicators are measures that show appetite is being approached: override rate, held items aged past target, complaints mentioning an agent, failed certification runs.
- Trigger and tolerance gives the level at which the indicator requires action, and the level at which agents in that class are suspended.
- Owner is the executive who answers for the class.
The template
Part A: board statement
Two to four sentences on the firm's overall stance, naming the outcomes it will not accept.
Part B: action classes
| Field | Guidance |
|---|---|
| Action class | Group of actions by effect |
| Appetite | Low, moderate, or none |
| Depth ceiling | 1 to 4 |
| Limits | Amount, volume, population |
| Always held for a person | Actions within the class that are never autonomous |
| Key risk indicator | What is measured |
| Trigger | Level that requires review |
| Tolerance | Level that requires suspension |
| Owner | Accountable executive |
A filled example
A hypothetical regional bank.
Board statement. The bank accepts agents preparing and carrying out routine servicing within tested limits. It does not accept any agent making a credit decision about an individual, or moving customer money without a named person's approval above small limits. It will suspend any class of agent activity that produces customer detriment.
| Action class | Appetite | Ceiling | Limits | Always held | KRI | Trigger | Tolerance | Owner |
|---|---|---|---|---|---|---|---|---|
| Read and summarise | Moderate | 1 | Data classes per DPIA | None | Data access outside role | Any | Any repeated | COO |
| Draft for a person | Moderate | 2 | None | All external sends | Drafts edited heavily | Rising for 2 weeks | n/a | COO |
| Change customer records | Low | 3 | Address and contact only | Bank details, name | Corrections after agent change | 5 per month | 20 per month | Head of operations |
| Communicate externally | Low | 3 | Approved templates only | Complaint responses, arrears letters | Complaints citing agent | 3 per month | 10 per month | Chief customer officer |
| Move money | Low | 4 | 500 per payment, own-customer transfers | All above limit; all to new payees | Payments recalled | 1 | 3 | CFO |
| Decisions about individuals | None | 2 | n/a | All | Any final decision by agent | Any | Any | Chief risk officer |
The last row matters most. Appetite of "none" with a ceiling of 2 means agents may prepare credit cases for underwriters but never decide.
Where Pryme fits
Pryme enforces the depth ceilings and limits in this statement as part of the minimum chain: the customer's own limits sit beside role scope, certificate, sources and platform guardrails, and the lowest wins. Breaches of tolerance can be traced through filed verdicts. More in agent governance is a bound, not a model.
Questions people ask
Should each business line have its own appetite?
Keep one statement at board level and let business lines set narrower limits beneath it. A business line may be stricter than the board, never looser.
How often should key risk indicators be reported?
Monthly to the executive risk committee, quarterly to the board, and immediately when a tolerance is breached.
What if an agent's evidence supports a higher depth than the ceiling?
The ceiling wins. Raising it is a board decision, made on the evidence, not something an agent owner can do.