Policies: your rules as code, enforced at runtime
Policies turn your acceptable-use guidelines, data-handling requirements, and regulator commitments into executable rules. Applied on every message, every tool call, every agent step. Before the model, not after.
Five gates, one direction: safe
A request doesn't reach a model unless it cleared every gate. A response doesn't reach a user unless it did too.
Identity & role
authn/authzWho is asking? Verified via SSO, scoped to their Space and role.
Content policy
contentBlock-lists, personal-data detection, redaction, category filters applied both to input and output.
Data residency
residencyEU data stays in EU-region models. NL-only datasets route to NL-only providers.
Model whitelist
modelsOnly pre-approved models allowed per Space. Fine-tuned judgments for high-stakes work.
Audit seal
evidenceEvery allow or deny is logged with inputs, policy version, and the matching rule.
Guardrails aren't a training thing
Enforced at the edge
Rules fire before the request leaves your governed boundary. Not in a post-hoc classifier that might have already leaked data.
Versioned and reviewable
Policies live in Git-like history. Every change is reviewed, dated, and tied to the person who made it.
Fast enough to not matter
Sub-50 ms in the hot path. Policy checks never become the reason people "just use ChatGPT" instead.
Common policies we've shipped
Strip personal data before model call
Replace BSN, IBAN, emails, and phone numbers with tokens. Restore them only for approved recipients.
Block external-only AI for HR
HR Spaces can only use AI deployed in-region with a signed DPA.
Require approval for mass email
An agent wanting to send to >50 recipients must clear a human approval gate first.
Flag prompts that risk unreliable answers
Low-confidence answers on regulated topics get a "citation required" flag and a reviewer.
Bring your policy, see it enforced in 30 minutes.
Send us one rule from your AI acceptable-use policy. We'll encode it live on the call and show you the audit entry.