Your organization claims to have a mature AI governance program. An auditor asks for evidence of a decision the program influenced last quarter. Can you provide it?
If you're facing a CISO who insists their AI governance is audit-ready, but you can't find the owner, decision trail, or controls matching their deployed agents, you're seeing a compliance gap that will matter in 2025. While 74% of organizations believe they could pass an AI compliance audit today, only 27% have a fully mature AI governance program. That 47-point gap is where your audit findings will emerge.
This playbook guides you through assessing AI governance programs, from scoping to validation. It's designed for external auditors and assessors who need to move beyond policy review to test whether governance truly influences behavior.
The Problem: Governance Theater vs. Governance Function
The issue isn't the absence of AI policies. Most organizations have them. The problem is that these policies often don't connect to decisions, controls, or accountability.
When auditing AI governance, determine if the program has done any of these four things:
- Changed a deployment decision
- Eliminated a feature or use case
- Forced a model retrain
- Delayed approval
If the governance program can't point to a specific instance of these actions, you're auditing paperwork, not a control environment. Paperwork doesn't prevent AI incidents. 65% of organizations have experienced an AI incident or near-miss, but fewer than half have controls matching their deployed agents.
Another red flag: 54% of organizations can't identify who owns AI governance. When governance resides in a committee instead of with a specific accountable party, defining your audit scope becomes much harder.
What You Need Before Starting
Before starting fieldwork, ensure you have:
Documentation access:
- AI governance policy and procedures
- AI use case inventory (first-party and third-party)
- Risk assessment or risk register for AI systems
- Incident response procedures (general and AI-specific)
- Meeting minutes from governance committees or working groups
- Decision logs or approval records for AI deployments
Stakeholder access:
- The named owner of AI governance (if one exists)
- Security leadership (CISO or equivalent)
- Data governance or data privacy lead
- Engineering or product teams deploying AI agents
- Vendor management or procurement for third-party AI tools
Technical access (if testing controls):
- Model validation reports or testing documentation
- Access logs for AI systems or platforms
- Configuration settings for AI agents in production
- Training data lineage or data governance records
If the organization can't produce a named owner in the first conversation, document that as a potential finding immediately. Governance without ownership fails the accountability test.
Step-by-Step Assessment
1. Map the AI Footprint
Start with visibility. Request the inventory of AI systems, agents, and tools. Don't accept "we use ChatGPT" as an answer. You need:
- First-party AI: models or agents the organization built and deployed
- Third-party AI: vendor platforms integrated into workflows
- Shadow AI: tools employees use without formal approval
Test the inventory by sampling recent procurement records, cloud spend reports, or API usage logs. If the inventory doesn't match what's in the environment, you've found a visibility gap.
2. Identify the Governance Owner
Ask directly: "Who owns AI governance?" If the answer is "the AI governance committee," ask who chairs it and who has final decision authority. If it's still a group, that's a control deficiency.
Document:
- The owner's title and reporting line
- Whether they have budget authority
- Whether they have veto power over deployments
If ownership is unclear, governance can't function. You can't audit accountability that doesn't exist.
3. Test the Decision Trail
This is the fingerprints test. Request evidence of governance decisions made in the last 90 days. Specifically:
- A deployment that was delayed or denied
- A feature that was removed or modified based on governance review
- A model that was retrained due to governance findings
- A vendor AI tool that was rejected or required additional controls
If the team can't produce at least one example, the governance program isn't influencing behavior. It's documentation theater.
For each decision, validate:
- Who made it
- What criteria they used
- Whether it's documented
- Whether it was enforced
4. Validate Controls Against Deployed Agents
86% of organizations have AI agents live in production. Ask for the control mapping for each high-risk agent. You're testing whether controls exist and whether they match the risk tier.
For each agent, verify:
- Risk classification (high, medium, low)
- Assigned controls (input validation, output review, access restrictions)
- Monitoring or logging in place
- Incident response procedures
Sample a high-risk agent (anything touching PII, financial data, or automated decisions). Test whether the controls are configured and operating. If the organization can't demonstrate that the payroll agent has different controls than the meeting summarizer, they're not risk-tiering. That's a finding.
5. Review Data Governance Foundation
Strong AI governance starts with strong data governance. If the organization doesn't have data classification, data lineage, or data quality controls in place, their AI governance is built on sand.
Test:
- Whether training data is classified and tracked
- Whether data provenance is documented
- Whether data quality is measured
- Whether data access controls apply to AI systems
If data governance is weak, AI governance can't compensate. Note this as a foundational gap.
6. Assess Incident Response Readiness
44% of organizations have written AI-specific incident response procedures. Ask to see them. Then test whether the procedures cover:
- Model drift or performance degradation
- Prompt injection or adversarial inputs
- Unintended outputs or hallucinations
- Data leakage through AI systems
If the procedures exist but haven't been tested (tabletop exercise, simulation, or actual incident), they're untested controls. If they don't exist at all, and the organization has agents in production, that's a material gap.
Validation: How to Verify It Works
Your validation step is confirming that governance isn't just policy. Test these outcomes:
- Decision authority is clear: You can name the person who approves or denies AI deployments.
- Decisions have consequences: You found at least one deployment that was delayed, modified, or rejected in the last quarter.
- Controls match risk: High-risk agents have demonstrably different controls than low-risk ones.
- Incidents are tracked: The organization can produce a log of AI-related incidents or near-misses and show how governance responded.
- Data governance is intact: Training data is classified, tracked, and governed.
If you can validate all five, the governance program is functional. If you can't validate even one, you're looking at a major nonconformity.
Maintenance: Ongoing Audit Considerations
AI governance isn't static. Your follow-up audits should test:
- Inventory drift: Is the AI footprint growing faster than governance can track?
- Control decay: Are controls that existed at the last audit still operating?
- Ownership changes: If the governance owner left, was accountability reassigned?
- Policy updates: Has the governance policy been updated to reflect new risks or deployments?
Set a review cadence tied to the organization's deployment velocity. If they're shipping new agents monthly, governance should be reviewed quarterly. If deployment is slower, annual reviews may suffice.
The U.S. spent $700 billion on AI infrastructure in a single year. Governance programs need to move at the same speed as the infrastructure they're meant to control. Your job as an auditor is to test whether they actually do.
