Skip to content

AI governance that runs, not a binder on a shelf

9 min read by John Bagnall

AI governance sticks after the audit when it runs like an operating rhythm, not a document set. That means five things: a named owner with the time to do it, an intake gate new AI cannot route around, a fixed review cadence, clear triggers that force a re-review when something changes, and evidence produced as a by-product of the work rather than assembled the week before an auditor arrives. Miss any of them and governance quietly turns into a binder on a shelf, accurate on the day it was written and less true every week after.

Most organisations do not decide to let governance lapse. It just stops, because nothing was built to keep it going. Here is why that happens and what a version that keeps running looks like.

Why governance stops after the audit

Governance usually arrives as a project: a deadline, a project lead, maybe a consultant, a budget, and a clear finish line, the audit, the certificate or the customer’s due diligence. Projects are good at getting things built. They are bad at keeping things running, because when the finish line is crossed, the deadline, the lead and the budget all go away.

What is left is a set of good documents with three problems:

  • Nobody owns them. The person who drove the project has gone back to their day job, or the consultant has left. Questions have no one to go to.
  • Nothing triggers a review. A new AI tool, a vendor swapping its model, a team using an approved tool for something new: none of it prompts anyone to update anything.
  • They describe a moment, not a system. The register lists the AI you had in March. The risk assessments describe systems that have since been changed.

And AI changes faster than almost anything else a business governs. New tools appear weekly, vendors update models without asking, and the AI your people actually use drifts away from the AI you approved. We covered that gap in governing the AI you approved, not the AI your people use. A static document set cannot keep up with that. Only a process can.

The signs your governance is a binder

Be honest about these. Two or three of them is common; most of them means the system has stopped.

  • The AI register has not changed in more than a quarter, but you know new tools have arrived since.
  • New AI tools get bought or switched on without passing through any review.
  • The last management review was the one held for the audit.
  • Asked “who owns AI governance?”, people give different names, or none.
  • Answering a customer’s AI questionnaire means a scramble to find or recreate evidence.
  • The policy still references tools you no longer use, or misses ones you do.
  • Risk assessments exist for systems, but not for how those systems are used now.

What “running” actually means

Governance that runs is not more paperwork. It is a small amount of the right work, done regularly, by someone who owns it. Five mechanics make the difference.

1. An owner with time, not a committee

One named person is accountable for keeping the system running, with protected time to do it, and a senior sponsor who backs them when something needs deciding. A committee can advise and approve; it cannot own. Ownership by committee tends to mean ownership by no one, which is how most binders end up on shelves. The board’s role is to make sure that owner exists and to look at what they report.

2. An intake gate new AI cannot skip

Most governance drift happens at the front door. If new AI tools, features and use cases can go live without a review, the register is out of date the moment it is finished. The fix is to wire the gate into processes people already use: procurement and expense approval, security review, the product release checklist, SSO and app approval. A short request form, a quick risk classification, a proportionate review, and a decision logged. Low-risk tools should clear it in a day, or people will route around it.

3. A fixed cadence

Governance survives on a calendar. A proportionate rhythm for a small or mid-sized organisation looks like this:

How oftenWhat happensWhat it produces
As requests arriveNew AI tools and use cases go through the intake gateRisk classification, decision, register entry
MonthlyOwner reviews the register, open risks and overdue actions; checks for unregistered AIUpdated register and action log
QuarterlyLeadership review: status, risks, decisions needed, regulatory changesMinuted management review, priorities
AnnuallyInternal audit, policy refresh, literacy and training refreshAudit report, updated policy, training records

None of these meetings needs to be long. The monthly review can be an hour. What matters is that it happens every time, so nothing is ever more than a few weeks out of date. This is what ISO 42001 means by monitoring, internal audit and management review; we unpack those requirements in Clause 9 and the improvement loop in Clause 10.

4. Triggers that override the calendar

A cadence keeps the system alive. Triggers keep it accurate. Certain events should force a review of a system regardless of where you are in the calendar:

  • The vendor changes the underlying model, or you switch providers.
  • The system starts using new data, especially personal or sensitive data.
  • It is used for a new purpose, or by a new group of users.
  • Its output starts reaching customers or affecting decisions about people.
  • An incident, complaint or near miss involving it.
  • A regulatory change that affects it, such as the EU AI Act’s moved deadlines.

Write the triggers down, tell system owners about them, and add them to vendor contracts where you can, so you hear about model changes rather than discover them.

5. Evidence as a by-product

The binder problem is partly an evidence problem. If evidence only gets assembled for audits, it is always out of date and always a scramble. If it is created by doing the work, it is always current. The intake form is the risk record. The monthly review’s notes are the monitoring evidence. The decision log is the accountability trail. Keep them in one place, the register as the spine, and the answer to “can you prove it?” is a link, not a project.

Keep it proportionate, or it will not survive

The fastest way to kill governance is to make it heavier than the risk it manages. If every low-risk tool needs a two-week review, people will stop asking. Match the depth of review to the risk tier: a quick check and a register entry for low-risk tools, a full impact assessment only where AI affects customers, people’s rights or significant decisions. Then make the compliant path the easy path, the same principle that makes an AI policy people actually follow. Governance that is light where it can be and serious where it needs to be is governance people keep doing.

Measure whether it is running

A few simple signals tell you, and your leadership, whether governance is alive:

  • Coverage: AI systems in the register versus AI systems you find when you look (SSO logs, expense data, browser extensions, vendor feature releases).
  • Gate compliance: new tools that came through intake versus around it.
  • Speed: how long a typical review takes. If it is creeping up, people will start skipping it.
  • Actions: open and overdue governance actions.
  • Time to evidence: how long it takes to answer a customer or auditor question with proof. Days means you are running; weeks means you are rebuilding.

Put these in the quarterly review. Numbers leadership looks at are numbers that get attention.

Who runs it

Someone has to own the rhythm. For larger organisations that is an internal AI governance lead, often sitting in risk, compliance or the CTO’s office. Many small and mid-sized organisations do not need, or cannot yet justify, a full-time role, but still need the work done every month. That is the gap a fractional governance lead fills: someone who runs the cadence, keeps the register and action log current, handles reviews, and reports to leadership, working alongside your technology, legal, security and HR teams rather than replacing them.

Either way, the test is the same: if the person who built your governance left tomorrow, would it still run next month? If the answer is no, it was a project, not a system. That is also why implementation should be done by someone who leaves you with something you can operate, a point we make in why ISO 42001 takes a builder, not just an auditor.

The short version

Governance does not fail at the audit; it fails in the months after, when the project ends and nothing keeps it going. Governance that runs has a named owner with time, an intake gate wired into how people actually buy and build AI, a monthly and quarterly cadence, triggers that force a review when something changes, and evidence that falls out of the work. Keep it proportionate so people keep doing it, measure a few simple signals so leadership can see it, and ask honestly whether it would survive the person who built it leaving. A binder describes your AI on the day it was written. A running system describes it today.

Want to know whether your governance is running or sitting on a shelf? The free AI governance readiness assessment gives you an honest read in about ten minutes, no email required. If you need someone to own the rhythm, our AI Governance as a Service runs it for you as a fractional governance function, and our AI governance consulting builds the system in the first place. For the full picture, see the AI governance guide.

Frequently asked questions

Why does AI governance fall apart after the audit?

Because it was built as a project, not a process. The project had a deadline, a lead and a budget; when the audit passed, all three went away. What is left is a set of documents describing the organisation on the day they were written, with no owner, no schedule and nothing that forces a review when the AI changes. AI use changes faster than almost anything else a business governs, so that gap opens within months.

How often should AI governance be reviewed?

On two clocks. A fixed cadence: new AI requests handled as they arrive, a monthly review of the register, risks and open actions, a quarterly leadership review, and an annual internal audit and policy refresh. And event triggers that force a review regardless of the calendar: a model or vendor change, new data, a new use or user group, an incident, or a change in regulation. The calendar keeps it alive; the triggers keep it accurate.

Who should own AI governance day to day?

One named person with the time to do it, backed by a senior sponsor. In smaller organisations that is often a CTO, head of risk or compliance lead with protected hours, not a side task. Ownership by committee tends to mean ownership by no one. Organisations that do not need a full-time role often use a fractional governance lead to run the cadence alongside their own teams.

How do you know if your AI governance is actually working?

Track a handful of signals. How many AI systems are in the register versus how many you find when you look; how many new tools came through the intake gate versus around it; how long reviews take; how many actions are overdue; and how long it takes to answer a customer or auditor question with evidence. If those numbers are moving in the right direction and leadership sees them, it is working.

Does ISO 42001 require ongoing governance, or just certification?

Ongoing. ISO 42001 is a management system standard, so it expects you to monitor and measure the system, run internal audits, hold management reviews and continually improve (Clauses 9 and 10). Certification bodies run surveillance audits during the certificate cycle precisely to check the system is still operating, not just that it existed on audit day.


Building something you need to govern?

Start with a fixed-scope AI Opportunity & Risk Audit.

Meet an Expert