TurfAI User GuideGovernance
How TurfAI keeps AI automation safe, private, and auditable.
Governance is the set of controls that wrap every workflow, agent, and squad: who can do what (access), what data leaves the boundary (privacy), and what happened (audit). For an enterprise platform this is not a feature bolted on the side — it surrounds the whole loop.
The three pillars
1. Access control (RBAC + ownership)
Role-based access control governs who can see and do what. On top of roles, an ownership model separates system resources (shared, read-only to most users) from user resources (private to their owner). Users own, edit, and delete their own workflows, agents, prompts, and chatbots. See Administration for the full model.
2. Privacy (Data Shield)
Data Shield tokenises PII before content reaches an external
LLM — the model sees EMAIL_84d0a133, never the real address — and restores the originals in
the response. Tokenisation is deterministic within a request and isolated across requests.
3. Audit & analytics
Every shielded LLM call and key platform action writes an immutable audit row — entity counts by type (never raw values), layers invoked, latencies, retention timestamp, and a correlation ID linking back to the run. On top of audit sit analytics: token usage with USD cost, jobs executed, per-workflow success rates.
Fail-closed where it counts
| Situation | Behaviour | Why |
|---|---|---|
| PII detector errors at ingress | Fail-closed — block the call | A false negative is a breach; fail loudly |
| Audit emit fails | Fail-open — call succeeds, warning logged | Audit is observability, not correctness |
| Policy requires shield, node didn't opt in | Workflow fails terminal (DATA_SHIELD_POLICY_VIOLATION) | The author declared the contract; the executor enforces it |
Data Shield v0.5 covers 8 PII types across three LLM entry points. The RAG/chatbot paths, free-text names/locations (NER), and customer-held KMS are v1.0. Always check the Data Shield limitations before relying on coverage.