Guardrails

Guardrails are the rules, filters, and checks that constrain what an AI employee is allowed to do — keeping it on-task, on-brand, and out of irreversible trouble.

Guardrails span the whole stack. At the prompt layer: 'never give legal advice'. At the tool layer: 'this employee can read the database but not write to it'. At the output layer: a classifier that blocks PII, profanity, or off-brand language before send. At the action layer: spend caps, rate limits, approval gates. Good guardrails are invisible until needed. Heavy-handed guardrails make the employee useless. The skill is picking the smallest set that covers the actual failure modes you care about.

Example

AI SDR guardrails: never claim a feature the product doesn't have, never quote prices not in the official sheet, never message a domain on the do-not-contact list, max 50 emails per hour, escalate any reply with the words 'lawyer' or 'unsubscribe was ignored'.

How OpenLabor uses it

Every OpenLabor employee ships with role-appropriate guardrails. You add company-specific ones in plain English.

Are guardrails the same as content filters?

Content filters are one type of guardrail. Guardrails also include tool permissions, spend caps, approval gates, and behavioral rules.

Can guardrails be bypassed by prompt injection?

Some can. The robust ones live outside the model — at the tool/permission layer — where the model can ask but the system says no.

Related: approval-gate, hallucination, audit-trail.

AI Labor Glossary