Where the Enterprise Brain becomes governed — authoring the workflows agents follow, the policies that constrain them, and the gates where humans must approve.
Builds on: Phase 3 — Semantic Linking & Binding
Phase 3 makes the graph queryable. Phase 4 makes it governed — every agent action is scoped, every decision is policy-checked, every critical step has a human gate.
OKF Workflow Authoring
Author the multi-step workflows that agents execute — periodic KYC reviews, SAR narrative generation, regulatory impact assessments. Each workflow is an OKF-conformant definition with named steps, decision points, and explicit human checkpoints. Workflows compile to version-pinned graph nodes.
OPA/Rego Policy Editor
Write and test the policies that govern every Brain decision. Data access policies (who can see what), tool-gating policies (which agent steps can use which tools), output validation policies (what shape an answer must have). Authored in Rego, tested against sample data, deployed through the four OPA enforcement gates.
Approval Gates & Risk Tiers
Define where agents must stop and wait for human approval. Risk-tier configuration determines which decisions auto-proceed, which pause for review, and which require senior sign-off. The gates are architectural — not optional hints the model can override.
HITL Interrupt Points
Design the human-in-the-loop moments: what the human sees when an agent pauses, what evidence is pre-assembled for their decision, what actions they can take, and what happens after they approve or reject. Every interrupt point is a first-class authored surface, not an afterthought.
Agent & Access Registry
Register what plugs into the Enterprise Brain — which agents, which tools, which external systems. Each registered consumer gets a scoped identity, a named owner, and explicit permission boundaries. No anonymous consumption. No unscoped access.
Confidence & Abstention Thresholds
Configure when the Brain should answer and when it should abstain. Per-topic, per-query-type confidence thresholds determine the boundary between a governed answer and a transparent "I don't have enough evidence to answer this." Abstention is a feature, not a failure.
Key personas for this phase
RM
Raj Mehta
Brain Administrator — manages the agent registry and access
GO
Governance OfficerNew in Phase 4
Owns policy authoring, approval gates, and audit governance
PN
Priya Nair
Knowledge Engineer — authors workflows and HITL interrupt designs