Guardrails

Guardrails are technical and organisational safeguards. They define what an AI system may do, which data it may process and when human review or approval is required.

Three layers of technical guardrails

For practical planning, guardrails can be grouped by where they act: on inputs, on system actions and on generated outputs.

Inputs: Controls determine which data and requests a model may receive. These include data classification, access restrictions, filters and validation of file types or request content.

Actions: Permissions limit which tools and systems an AI application may use. Critical or difficult-to-reverse steps may also require confirmation.

This layer is particularly relevant to agentic AI, where several tool calls may follow one another.

Outputs: Checks validate format, source grounding, sensitive content or other domain criteria before a result is used further.

Guardrails in production

Clear guardrails show which steps may run automatically and where approval is required. The permitted level of automation should follow from the risk and potential impact of an error.

Typical building blocks in implementation

Typical components include permission systems, approval steps, logging and automated checks against defined rules or reference data.

Where a potential high-risk trigger under the EU AI Act exists, human oversight and documented control points may be among the legal requirements to assess.

The required guardrails should therefore be defined during architecture and operating-model design according to the specific risk.