Separate recommendation from authority.
An agent can classify, retrieve evidence and propose an action without being allowed to execute it. Keeping those boundaries explicit lets teams gain automation value before granting broad autonomy.
Engineering insight · agentic AI
Enterprise AI agents need explicit authority, evidence and human approval.
Agentic systems become useful when they can take actions, but that is exactly where governance matters. The architecture should state what an agent may read, decide, propose and execute.
Tool permissions by role
Evidence-grounded retrieval
Human approval for consequential actions
Audit and fallback paths
A practical agent control model
- Router: classify intent and risk
- Retriever: gather approved evidence
- Worker: draft the action or answer
- Verifier: check constraints, evidence and confidence
- Approval gate: require a person when the action is sensitive
- Executor: call a narrowly permitted tool
- Audit layer: record the evidence, tool call and decision
Fail closed when evidence is weak.
If a required source is unavailable, a policy conflicts or model confidence is insufficient, the system should stop or escalate instead of inventing certainty.
Measure operational quality.
- Task success rate
- Escalation rate and reasons
- False-positive/false-negative cost
- Evidence coverage
- Human correction rate
- Latency and per-task cost
Start with the operating problem
Want an agent architecture for one of your workflows?
Describe the users, data and decisions involved. I can map the agent roles, tool permissions, validation and approval points.