AI and agents
Guardrails
Guardrails are the checks placed around an AI system that limit what it may receive, say or do, such as input filters, output validation, tool permissions and spending limits.
How it works in practice
A model’s instructions are a request; guardrails are what enforce it. Some run before the model, such as removing personal data or blocking off-topic requests. Some run after, such as checking that an extracted invoice total matches the line items or that a reply contains no prices the system did not supply. Others sit around the tools: allow lists, per-run limits on refunds or emails, and approval steps.
Good guardrails are specific to the job and tested. A generic "be safe" instruction does little. A rule that blocks any refund over $200 without a person’s approval is easy to check and easy to explain to an auditor.
A worked example
A retailer’s support agent can offer a discount code, but code creation is capped at 15% and five codes a day per agent. Anything above that goes to the support lead’s queue with the conversation attached.
Examples in this glossary describe typical builds for companies of 1 to 1,000 people. They are not named client results.















