Enterprise AI Governance That Unblocks Delivery
A working governance model for AI systems: an inventory, risk tiers, named owners and controls that run in the delivery pipeline instead of in a quarterly committee.
Answer first
Enterprise AI governance is the set of decisions, owners and controls that decide which AI systems may run, with what data and permissions, and what evidence keeps them running. It unblocks delivery when it is small and specific: one inventory, three or four risk tiers, a named owner per system, and checks that run in the release pipeline. It slows delivery when it is a committee that reviews slide decks.
Why most governance programs stall
Governance usually starts as a policy document. The document says AI must be "fair, transparent and accountable", and nobody can tell from it whether a specific release is allowed to ship. Teams then route every request to a committee, the committee becomes the bottleneck, and people start using tools outside the process.
The fix is to turn principles into decisions a delivery team can make on a Tuesday afternoon:
- Is this system in the inventory, and who owns it?
- Which risk tier is it in?
- What evidence does that tier require before release?
- Who can stop it, and how fast?
If a team can answer those four questions without a meeting, governance is working.
The minimum model
1. One inventory
Keep a single register of AI systems in use or in development: purpose, owner, model or vendor, data it can read, actions it can take, users affected, risk tier and last review date. Banking supervisors have long treated an inventory as the base of model risk management; the April 2026 revision of US model risk guidance, SR 26-2, still describes a comprehensive model inventory as common industry practice. The same logic applies outside banking. You cannot govern what you have not listed.
2. Risk tiers based on what the system can do
Tier on consequences, not on the model brand. A practical starting point:
- Tier 1, assistive: drafts or summarizes for a human who decides. Example: an internal writing assistant.
- Tier 2, decision support: output feeds a decision about a customer, employee or money, and a human approves each one.
- Tier 3, acting: the system writes to records, sends messages or moves money without per-item approval.
- Tier 4, regulated or high impact: uses covered by law or sector rules, such as the EU AI Act's high-risk categories (employment, education, credit and others).
Each step up adds evidence requirements. Tool access and write permissions move a system up a tier faster than model choice does. See access control patterns for enterprise agents.
3. A named owner per system
Every inventory entry names a business owner who answers for outcomes and a technical owner who answers for behaviour. "The AI team" is not an owner. Owners approve changes, accept residual risk in writing and are paged when the system misbehaves.
4. Evidence gates in the pipeline
Write each tier's requirements as checks the release pipeline can run or verify:
- an evaluation set with pass thresholds, rerun on every model, prompt or tool change;
- permission scopes declared in configuration and reviewed when they change;
- logging of inputs, outputs and actions at the level the tier needs;
- a tested rollback or kill switch for Tier 3 and above;
- for Tier 4, the documentation your regulator or standard expects.
Policy-as-code makes these checks enforceable rather than advisory. See policy-as-code for AI systems.
Step-by-step setup
- List every AI system in use, including vendor features switched on inside existing SaaS tools, and record an owner for each.
- Assign a risk tier to each system using what it can read and do, not which model it runs.
- Write the evidence each tier needs before release, and turn as much of it as possible into automated checks.
- Set a review cadence by tier, plus mandatory review on material change (new data source, new tool, new user group).
- Give one small group authority to approve Tier 3 and Tier 4 releases, with a published turnaround target.
- Report inventory coverage, open exceptions and incidents to leadership on a fixed schedule.
Mapping to the frameworks people ask about
You do not need to pick one framework before starting. The NIST AI RMF, released in January 2023 for voluntary use, organizes the work into four functions: Govern, Map, Measure and Manage. NIST says the framework is being revised. ISO/IEC 42001:2023 specifies an AI management system that an organization can be certified against. The minimum model above maps cleanly to both: the inventory, owners and approvals sit under "Govern" (the AI RMF asks for mechanisms to inventory AI systems in GOVERN 1.6), context and risk tiers under "Map", evaluation gates under "Measure", and monitoring and rollback under "Manage". For a side-by-side view, read ISO 42001 vs NIST AI RMF.
For the security side of agent and LLM systems, the OWASP Top 10 for LLM Applications (2025) is a useful checklist. In practice, "Excessive Agency" and "Prompt Injection" are the two items most likely to move a system up a tier.
How to measure whether it is working
Track a few numbers and review them monthly:
- Inventory coverage: systems found by discovery (SSO logs, expense data, vendor feature flags) that were not in the register.
- Turnaround: days from request to decision, by tier.
- Exceptions: count and age of releases shipped with an accepted risk instead of passing evidence.
- Incidents: count, severity and time to contain, by tier.
Falling turnaround with flat or falling incidents is the signal that governance is enabling delivery. Rising exceptions mean the requirements are unrealistic or the tooling is missing.
Decision framing
Governance is worth the effort when it lets you say yes faster and with evidence. Frame the business case around decisions it enables: which systems can expand to more users, which can take actions instead of only drafting, and which vendor features can be switched on safely. Avoid promising cost savings from governance itself; the value shows up as releases that would otherwise have stalled.
When to go lighter
If you run only Tier 1 assistive tools, a short acceptable use policy and an inventory may be enough. See what to put in an AI acceptable use policy. Add tiers and gates when the first system starts to act on records or customers.
Related reading
- AI risk-tier classifier: a free tool that applies these tiers to one system
- AI governance FAQ
- AI governance frameworks and standards, verified
- Enterprise MCP connector governance after the stateless pivot
- Procurement intake for models, agents and MCP connectors
- AI reliability and operations services
Sources
- NIST AI Risk Management Framework (AI RMF 1.0)
- NIST AI 100-1: Artificial Intelligence Risk Management Framework (PDF)
- ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system
- European Commission — AI Act regulatory framework and timeline
- Federal Reserve SR 26-2: Revised Guidance on Model Risk Management
- OWASP Top 10 for LLM Applications (2025)