LLM Governance Checklist for Production Deployments

Production Readiness Checklist

These controls are organized by lifecycle stage: verification before go-live, enforcement at runtime, and monitoring after deployment.

Pre-Deployment: Identity, Permissions, and Governance Mapping

• Agent identity is provisioned separately from human user accounts and integrated with existing enterprise IAM
• Permission scope is defined per tool or API call and mapped to the agent's approved use case
• Governance policies, roles, and accountability structures are documented, consistent with the NIST AI RMF Govern function
• Input and output validation controls are specified to address prompt injection and insecure output handling
• Third-party models, tools, and plugins in the value chain have been reviewed for integration risk
• The deployed model and configuration version are recorded against the corresponding governance approval

Deployment: Runtime Enforcement and Audit Logging

• Tool-call restrictions are enforced at runtime, not only validated at design time
• Permission checks execute before tool or function invocation, not after
• Enforcement is applied through a centralized policy point rather than per-application logic
• Audit logs capture agent identity, invoked tools, and decision context at the time of execution
• Policy definition is separated from policy enforcement to support independent review and change control
• An escalation path exists for policy violations detected during runtime execution

Post-Deployment: Monitoring, Drift Detection, and Incident Response

• Periodic review cycles are scheduled to detect behavioral drift from the originally approved use case
• Incident response procedures are tested against actual runtime log output, not only tabletop scenarios
• The full checklist is re-verified whenever models, tools, or permission scopes change
• Audit log integrity is maintained so records can be relied on for later review
• Runtime controls are periodically tested against known adversarial technique references, such as those cataloged in MITRE ATLAS
• Ongoing monitoring activity is mapped to the NIST AI RMF Manage function for accountability