What can an AI policy cover?

An AI policy can cover anything an AI writes or an agent does. In practice policies fall into six domains: safety, security, compliance, reliability, output quality and integrity.

The six domains

DomainWhat it is aboutAn example policy
SafetyHarm to peopleThe assistant does not give instructions for self-harm
SecurityAttacks on or through the systemThe agent does not send credentials to an outside server
ComplianceLaw, regulation and contractThe assistant does not give legal advice to users in this region
ReliabilityDoing the job without breaking thingsThe agent does not delete files outside its workspace
Output qualityAnswers that are right and on briefThe assistant does not state a price that is not in the catalogue
IntegrityHonest use of the productUsers do not use the support bot to write their homework

What they have in common

Each is worded as something the AI writes or its agent does: a reply, a tool call, a call placed by a voice agent. That is what a guardrail can see and stop. "Prevent fraud" is a goal. "The agent does not issue a refund above the order's value" is a policy.

Why not a fixed list

Every product has rules of its own. A bank, a game and a coding tool share almost none. A fixed list of ten categories fits nobody exactly, which is why policies are written, not chosen: see natural-language policies.

With White Circle

Control enforces unlimited custom policies across all six domains, on every prompt, response and tool call.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL