AI Agent Governance

Enterprise AI Agent Governance: An Operating Model

For the governance lead moving from one controlled workflow to agents across teams: a federated operating model, decision-rights table and monthly evidence packet.

Uli PrantzBuilds and operates all-agents
Published
Key takeaways

Key takeaways

  • Central governance should set minimum controls and own exceptions; process teams should own each agent's purpose, procedure and outcomes. Sending every decision to one committee creates a queue, not control.
  • Keep one inventory that joins business owner, technical identity, tools, data, risk tier, policy version and lifecycle state. A catalog of agent names alone cannot answer who has authority.
  • Tier the deployed use, not the model brand. The same model summarizing an internal document and releasing a payment belongs under different controls.
  • Review evidence by exception: new permissions, policy denials, human overrides, failed runs, stale owners and overdue reviews. Dashboard totals hide the items that require a decision.
  • Treat material changes as new authorization events. A new tool, data class, external audience or autonomy boundary can invalidate the approval even when the agent name stays the same.
Enterprise AI agent governance is the operating model that applies common minimum controls across an organization while leaving each business process with a named owner. Central governance owns the policy, tier definitions, inventory standard, evidence reviews and exceptions. Process teams own purpose, procedure, outcomes and day-to-day response. Security and platform teams enforce identity, access, logging and lifecycle controls. The join between those roles is the agent inventory and its evidence packet.

What changes when one governed process becomes a portfolio

The process-level framework asks five questions: access, autonomy, record, escalation and review. At enterprise scale, the questions stay the same. The difficulty is discovering every deployment, applying consistent tiers, keeping owners current and reviewing exceptions without a committee reading every run.

NIST's AI RMF assigns governance across the organization rather than to a single technical team. Current Microsoft maturity guidance likewise connects enterprise standards with identity, data access, logs, human oversight and lifecycle operations. A workable model therefore has a small central rule set and distributed ownership, with evidence flowing back.

Set decision rights before forming the committee

A federated decision-rights model. The final column prevents shared accountability from becoming no accountability.
DecisionRecommendsDecidesProvides evidence
Approve a new low-impact processProcess ownerDepartment control ownerBuilder or platform owner
Approve higher autonomy or external actionProcess and security ownersNamed risk owner or governance bodyPlatform, testing and approval owners
Grant production accessSystem and process ownersIdentity or data ownerAuthorization service
Accept a time-limited exceptionControl ownerNamed exception authorityProcess owner until expiry
Pause after an incidentAny monitoring or response ownerPre-authorized incident roleRun and incident records
Retire an agentBusiness ownerPortfolio or platform ownerIdentity, system and records owners

Reserve the central body for policy, high-tier approvals, cross-team conflicts and exceptions. If it approves each ordinary workflow edit, teams will either wait or route around it. Local owners may decide within a published boundary; movement across the boundary comes back centrally.

Build an inventory that can answer operational questions

An inventory entry should join the business deployment to the technical actor. Start with purpose, department, business owner, technical owner, lifecycle state and risk tier. Add the exact identities, environments, tools, data categories, external audiences, maximum autonomous action, policy version, last review and next review.

This is not a model catalog. One model may support several uses with different consequences; one process may switch models without changing its control tier. Inventory the deployed purpose and its permission envelope, then link the current model and procedure as versioned components.

Questions the inventory should answer without interviewing the original builder.
QuestionFields that answer it
Who can stop this today?business owner, incident role, lifecycle state
What can act outside the company?action tier, tools, external audiences, approval boundary
Which access survives a departed owner?owner directory ID, agent identities, last recertification
What changed since approval?procedure, model, tool, data and policy versions
Which review is overdue?last review, interval, next due date, exception expiry

Review an exception packet, not a wall of totals

NIST's 2026 monitoring report describes a real scaling constraint: human review does not grow at the speed or complexity of production systems. Use human attention on deviations that require a decision. Totals belong in context; exceptions belong on the agenda.

Monthly enterprise evidence packetSimulated example data

The cover page lists active agents by tier and lifecycle state, then identifies changes since the prior packet. The decision queue contains new tools or data classes, autonomy increases, repeated policy denials, human overrides, incidents, failed rollback tests, stale owners, expired exceptions and overdue access reviews.

Each item carries one requested decision, one accountable owner, supporting run or configuration evidence, and a due date. Clean low-tier agents do not receive a slide each; they remain available in the inventory and sampled control report.

Use the same gates from proposal through retirement

  1. 01Propose and classifyHuman approval
    Name the process, owner, affected people, data, tools, action types and intended autonomy.
  2. 02Test under supervisionHuman approval
    Run normal, boundary and failure cases; capture approvals, denials, overrides and recovery evidence.
  3. 03Authorize productionHuman approval
    Approve the procedure, permission envelope, autonomy line, monitoring and incident owner together.
  4. 04Monitor and changeCode
    Route material changes, control failures and expired reviews back to a named decision state.
  5. 05Suspend or retireHuman approval
    Disable schedules and credentials, settle queued work, preserve required records and mark the inventory state.

Limitations and when not to use this

  • This operating model needs adaptation to the organization's legal, security and sector obligations. It does not assign a regulatory classification.
  • An inventory does not discover unsanctioned agents by itself. Pair attestation with technical discovery in identity, network, procurement and platform data where available.
  • Central reporting can create false confidence when source systems omit events. Test evidence completeness and cold replay rather than relying on dashboard availability alone.

Sources

  1. AI Risk Management Framework CoreNIST Accessed 29 July 2026
  2. Agentic AI maturity model: governance and securityMicrosoft Learn Accessed 29 July 2026
  3. Challenges to the monitoring of deployed AI systemsNIST Accessed 29 July 2026

Build the first evidence packet

Pick five active agents, fill the inventory fields, and bring only missing owners, new permissions, overrides and overdue reviews to the next governance meeting.

Build the first evidence packet
About the author

Uli Prantz

Builds and operates all-agents

Uli Prantz builds all-agents, the process-automation platform this site documents. He writes about the operational side of automating recurring business work: where deterministic code beats model judgment, where it does not, and where a human still has to approve.

Bring one process. We will scope it in 30 minutes.

You leave the call knowing whether it is a fit, what can become code and what still needs a person.

Book a discovery call