An AI agent risk assessment answers one question before launch: if this agent does the wrong thing, how bad is it — and what stops it? Traditional AI assessments focus on model quality and data. Agents add a more important dimension: the actions they can take. This template keeps the review short, concrete and repeatable, and fits into a wider AI agent governance process.

Start with actions, not the model

List every tool the agent can call and what each one does in the real world. The model matters, but the tools define the blast radius. An excellent model with permission to delete production is riskier than an average model that can only read help articles.

The AI agent risk assessment template

1. Purpose and owner

  • What is the agent for, in one paragraph?
  • Who is the accountable owner?
  • Who are its users, and who is affected by its actions?

2. Tools and actions

For each tool: what it does, which systems it touches, whether it reads or writes, and whether its effects are reversible.

3. Impact rating per action

Rate each action on three questions:

Question Low High
Can it be undone? Easily Not at all
Who is affected? Internal only Customers or the public
What's at stake? Convenience Money, personal data, safety, legal

4. Threats

Consider how each high-impact action could go wrong: prompt injection, wrong arguments, over-broad access, loops and retries, or a model change. The agentic AI risks guide lists the common ones.

5. Controls

For each high-impact action, record the control:

  • Allowed automatically, within which limits?
  • Held for approval — by whom, with what timeout?
  • Denied outright?
  • How is it logged?

These become your AI agent policy.

6. Residual risk and decision

What risk remains with controls in place, and who accepted it? Record the decision, the date and the conditions — for example "re-review if the agent gains write access to billing".

Assign a risk tier

A simple three-tier model keeps reviews proportionate:

  • Tier 1 — read-only, internal. Light review; logging on.
  • Tier 2 — writes to internal systems or contacts customers. Policy rules and approvals for defined actions.
  • Tier 3 — moves money, touches sensitive data or production. Strict rules, named approvers, dry-run before launch, weekly review.

Keep it current

An assessment describes the agent at a point in time. New tools, a new model or a new use case should trigger a quick re-review of the affected actions — not a full restart.

Where Gaten fits

Gaten turns the controls in your risk assessment into enforced rules: allowed limits, approvals and denials on every action, with a record you can attach to the assessment. See risk & compliance.

Frequently asked questions

What is an AI agent risk assessment?

A structured review of what an agent can do, how its actions could go wrong, and which controls limit the impact — completed before launch and updated when it changes.

How is it different from a model evaluation?

Model evaluations measure output quality. An agent risk assessment focuses on the real-world actions the agent can take and the controls around them.

How long should it take?

For most agents, an hour or two with the owner, an engineer and a risk reviewer — longer for agents that move money or handle sensitive data.