What is a tool risk rating?
A tool risk rating is a low, medium or high classification given to each tool an AI agent can call, based on write access, reversibility, permissions and financial impact, used to decide which calls need extra checks.
Not every tool call deserves the same scrutiny. Looking up an order is harmless if it goes wrong; issuing a refund or deleting a record is not. OpenAI's A practical guide to building agents (2025) turns this into a simple rule: assess each tool, give it a rating of low, medium or high, and let the rating trigger automated actions, such as pausing for guardrail checks before a high-risk function runs or escalating to a person.
A rating moves the safety decision from the moment of the call, when a person or classifier has seconds to judge it, to design time, when the team can think through the worst case calmly.
The factors OpenAI lists
Read-only or write access
A tool that only reads cannot change anything, though it can still expose data.
Reversibility
A draft can be discarded and a soft delete restored; a sent email or a paid invoice cannot be taken back.
Required account permissions
A tool running with an administrator's rights can do far more damage than one scoped to a single user.
Financial impact
Payments, refunds and orders carry direct cost when they go wrong.
What a rating triggers
Low
Runs without interruption and is logged.
Medium
Runs after automated checks, such as argument validation, recipient allowlists, amount caps or rate limits.
High
Pauses for guardrail checks and, where needed, a person. OpenAI's guide names cancelling orders, authorising large refunds and making payments as actions that should trigger human oversight until confidence in the agent grows.
Ratings declared by the tool itself
The Model Context Protocol lets a server attach annotations to a tool, including hints that the tool does not modify its environment, that it may make destructive updates, that repeated calls with the same arguments have no extra effect, or that it reaches an open world of external systems. The specification says clients must treat annotations as untrusted unless they come from a trusted server.
Annotations are a useful input to a rating, especially for third-party MCP servers, but the rating belongs to the team deploying the agent, because the same tool can be low risk in one workflow and high in another.
Common mistakes
Rating the tool, not the arguments
A refund of 5 and a refund of 5,000 go through the same tool. Where impact varies with the arguments, the rating should too.
Ignoring combinations
A tool that reads private data and a tool that fetches a URL can each look low risk, yet together form an exfiltration path. Source-sink analysis catches this.
Ratings that never change
OpenAI ties human oversight to confidence in the agent's reliability, so ratings should be revisited as eval and production evidence builds up.
Too many highs
If most tools are rated high, approvals become constant and stop being read.
How Motiontic's engineers apply it
Motiontic's engineers rate every tool in the scope document before it is built, with the client's owner of the underlying system, and keep the rating next to the tool's definition in code. The rating sets the control: logged, guarded by argument checks and limits, or paused for approval. Ratings are reviewed whenever a tool is added or its permissions change.
Example tool risk ratings for a support agent
| Tool | Access | Reversible | Rating | Control |
|---|---|---|---|---|
| Look up order | Read | Not applicable | Low | Logged |
| Draft reply | Write, draft only | Yes | Low | Logged; a person sends it |
| Update shipping address | Write | Yes, with history | Medium | Customer verified first, argument checks |
| Issue refund | Write, financial | Partly | High | Amount cap, approval above a threshold |
| Cancel order | Write | Often not | High | Human approval |
Common questions
What is a tool risk rating?
A tool risk rating classes each tool an AI agent can use as low, medium or high risk, based on factors such as write access, reversibility, permissions and financial impact, so riskier calls get more checks.
Where does the idea of tool risk ratings come from?
The term is used in OpenAI's A practical guide to building agents (2025), which lists tool safeguards among layered guardrails. OWASP's LLM06 Excessive Agency risk covers the same ground from the attacker's side: too much functionality, permission or autonomy.
How do tool risk ratings relate to human-in-the-loop?
The rating decides where a person is needed. Low-risk tools run freely, while high-risk tools pause for approval, which keeps human review for the few actions where it matters.
Are MCP tool annotations a risk rating?
Not on their own. They are hints from the server about how a tool behaves, and the MCP specification says clients must treat them as untrusted unless the server is trusted.
Related
Blast radius (AI agents)
The worst an agent could do, measured by what it can reach.
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
AI guardrails
Checks on what goes into a model, what comes out, and what an agent is allowed to do.
Approval fatigue
Too many approval prompts, and people stop reading them.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.