Guardrails
Controls placed around an AI system — input and output checks, permission limits, review steps and escalation rules — that keep its behaviour within acceptable bounds.
Guardrails are the controls that keep an AI system's behaviour within acceptable limits. They can sit before the model (screening inputs), after it (checking outputs for policy, accuracy or sensitive data), or around its actions (restricting which tools, data and operations it can use, and when a human must approve).
The word is often used as reassurance: "we have guardrails". A more useful habit is to make them specific and testable: what exactly is blocked, what is reviewed, who is alerted, and how often does each control fire or fail?
What good guardrails include
- Permission boundaries that reflect that capability is not authority
- Review and approval steps where a person's judgment is genuinely needed (human in the loop)
- Checks against known failure modes such as hallucination
- A route for uncertain cases to be handed back
- A plan for failure: how it is detected, contained and reversed, which is the substance of a recovery contract
Guardrails reduce risk; they do not remove it. Ask what happens when one fails.
Read more in OpenAI Decisions API turns probability into policy.
Related terms
Capability is not authority
A design principle for AI agents: being technically able to perform an action does not mean the agent should be permitted to perform it. Delegation needs gradients of authority.
Recovery contract
A short written artifact in an AI vendor evaluation that states how failures are detected, contained, reversed and handed back to people — and who is responsible at each step.
Hallucination
When an AI model produces fluent, confident output that is false or unsupported by its sources. What matters is which errors are unacceptable in a given workflow and how they are detected.
Human in the loop
A design where a person reviews, approves or corrects an AI system's work at defined points. In practice it can mean a person with real judgment, or a person who has become the loop.
Used in these essays
AI is removing the coordination tax — the next bottleneck is authority
As AI reduces the cost of moving work between teams and systems, many organizations will discover that the remaining delay is not coordination. It is permission to decide.
An AI agent should get authority per task, not permissions per role
A person's role describes standing access across many situations. An AI agent should receive a narrower authority envelope for the task it is executing now.
Design systems need an interaction policy layer for AI
Components can keep an AI experience visually on-system while the agent is behaviorally off-policy. Design systems now need reusable rules for what AI may present, decide, execute, defer, expose as uncertain, confirm, reverse and escalate.
Your AI agent can follow the objective and still break the company
AI agents can satisfy a local objective while violating the organization’s broader constraints. The control problem is not only model alignment; it is management design.
OpenAI Decisions API turns probability into policy
The provocative part of OpenAI's Decisions API is not that AI can make choices. It is that a model score can quietly become an action — and an action can quietly become policy.
The agent is not the operating model
Giving AI more autonomy does not remove organizational complexity. It gives that complexity permission to act.