Playbook

Keep the right person in control

Clear permissions and review steps make an AI workflow easier to operate. Use this guide to decide who can access information, approve actions and handle exceptions.

Get the Control Templates Read the Production Guide
Governance map
Control stack
5 Required control families
Review
Target evidence coverage
3
Risk-based gate types
24 hrs Incident review target
Sources Action logging
Least Privilege default
Direct answer

The core controls every production AI workflow needs

Start with the information the workflow can access, the actions it can take and the person responsible for each decision. Then test the approval, correction and recovery steps.

Least privilege

Give each tool only the access needed for its task. Separate reading, drafting and changing records, and test that permission checks hold at the point of action.

Policy and allowlists

Define what the agent can access, which actions are permitted, and what content or destinations are blocked before the agent can call tools.

Evidence and review

Capture prompts, retrieved context, tool actions, confidence signals, overrides, and final outcomes so reviewers can reconstruct a decision path.

Control system

How to place approval gates without stalling operations

Gate 1

Pre-action review

Use before a payment, message send, policy exception, or system update. The reviewer sees the recommendation, rationale, source evidence, and intended action.

Gate 2

Threshold review

Use when confidence scores, retrieval quality, or classification certainty fall below the release threshold. The agent routes to manual handling instead of guessing.

Gate 3

Exception review

Use when the workflow encounters a novel case, policy conflict, or missing data pattern. This gate prevents silent drift from becoming a production habit.

Framework mapping

Use security frameworks to organize the review

Use OWASP-informed control language

  • Prompt and tool abuse risk belongs in pre-execution policy checks.
  • Data leakage risk belongs in retrieval boundaries, redaction, and role-based access.
  • Excessive agency risk belongs in approval design, action scopes, and rollback planning.

Use NIST AI RMF as an operating map

  • Govern: name accountability, ownership, and risk review cadence.
  • Map: document workflow boundaries, actors, and action surfaces.
  • Measure: define evidence, thresholds, and monitoring signals.
  • Manage: specify mitigations, approvals, and remediation steps.
Evidence package

What security and risk reviewers need to inspect before launch

  • Workflow architecture summary with data sources, tools, identities, and operators
  • Control matrix that maps risk points to gates, policies, and evidence
  • Audit log examples showing input, context, action, reviewer, and final state
  • Evaluation results for happy path, edge cases, and prohibited actions
  • Incident and rollback procedure for failed tool calls or unsafe outputs
Templates

Control matrix and approval artifacts

Use the templates page for reusable control categories, gate logic, and review prompts.

 Insight

Human-in-the-loop approval gates

Read the operational pattern for placing approval boundaries without creating a queue disaster.
FAQs

Common governance design questions

The minimum set is least-privilege access, policy and allowlist checks, risk-based approvals, complete audit evidence, and evaluation gates tied to release management.
Place them before financial, legal, customer-impacting, or irreversible actions, and whenever confidence or retrieval quality drops below the release threshold.
Use it to structure accountability, measurement, and risk treatment decisions. It is most useful as an operating map for reviewers and builders, not as a one-line badge claim.

Need controls your security team can actually inspect?

Use the questions on this page to describe your task, review the available information and decide what a useful test should show.

Talk through your task See Our Approach