Why isn't approving AI tools the same as governing AI?

Because approval only covers the tools that came through procurement and got a policy, and that is rarely where AI adoption actually happens. New AI tools are adopted in a browser tab, not a purchasing process, so the set you have approved is a subset of the set people use. Governing the approved tools well can still leave most of your real AI activity, on personal accounts and tools you never sanctioned, completely outside your controls. Approval governs intent; only measuring actual usage governs reality.

How much of our AI use are we probably not seeing?

More than most leaders expect. When organisations move from "the tools we approved" to measuring what staff actually use, a shortlist of a handful of tools routinely turns into dozens, and a large share of AI activity turns out to be happening on personal accounts rather than the enterprise tiers provided. The precise numbers vary, but the pattern is consistent: the governed set is the minority, and the invisible set is bigger than the strategy assumed.

Why are personal AI accounts a governance problem?

Because your enterprise controls only reach the accounts and tools they know about. Single sign-on, CASB and DLP work on managed, sanctioned tools; they are blind to someone using a personal ChatGPT account, a free image generator, an unmanaged meeting note-taker or a browser extension. So a policy that is perfect for your managed tools does nothing about the AI activity happening on personal accounts, which is often where the sensitive data and the real exposure sit.

Isn't AI risk mostly about a few careless employees?

No, and that assumption misreads where the exposure is. AI usage concentrates heavily: the median employee barely uses it, while a small group of power users generate an order of magnitude more activity. Those power users are usually not reckless, they are among your most productive people, finding ways to work faster. So your biggest concentration of AI exposure often sits with your best people doing their jobs, on tools you cannot see, which is a very different problem from policing a careless few.

Where should AI governance actually start?

With visibility, not with a policy. You cannot write meaningful rules, build a register, or assess risk for AI you have not seen. The productive order is: measure what AI is genuinely in use, build the register from that reality rather than from your approved list, classify each system by risk, and only then let the policy govern something real. A strategy written before you have looked governs an imagined organisation, not the one you have.

Can't we just block personal AI tools?

Blocking helps at the edges but rarely solves it, and on its own it usually makes things worse. Prohibition pushes usage further underground, onto personal devices and accounts you can see even less, and it removes value your best people are getting from the tools. The more durable fix is to make the sanctioned path better than the workaround: provide enterprise tools that are genuinely good, with clear rules people will follow, so there is little reason to reach for a personal account in the first place.

You're governing the AI you approved, not the AI your people use

The dangerous moment in AI governance is not having no strategy. It is having one that looks complete. You have blocked the consumer version of ChatGPT and rolled out an enterprise tier. Copilot is going out through Microsoft 365. A couple of teams have approved access to something else, and there are one or two niche tools besides. There is a policy for each. On paper, you are governed.

Then someone asks the only question that really tests it: how do you know that is all of them? And if you are honest, you do not, because you have been governing the tools you approved, not the tools your people actually use. Those are two very different lists, and the gap between them is where most AI governance quietly fails.

The confident answer that falls apart

Picture the board asking a reasonable question: what AI are our people using, and how is it governed? You have a tidy answer ready. Five tools, give or take, each with a policy, consumer versions blocked, enterprise tiers provided. It feels like control.

The follow-up is the one that does the damage. How do we know we have covered everything? Because the tidy answer describes approved usage. It was never built from a measurement of actual usage. Nobody looked at what staff were really doing; they documented what the organisation had sanctioned and assumed the two matched. They almost never do.

Approved is not the same as actual

AI adoption does not arrive through procurement. It arrives through a browser tab. A new writing assistant, an image generator, a coding tool, a meeting summariser, a browser extension: any of them is one sign-up away, no purchase order required. So the tools you have approved are the subset that happened to come through a process, and the tools in use are whatever people found genuinely helpful.

When organisations finally measure the difference, it is not a rounding error. A shortlist of a handful of approved tools turns into dozens of real ones, most of which leadership had never heard of. You can have immaculate policies for five tools and still be blind to the majority of your AI footprint. Governing the approved list is not the same as governing your AI, and mistaking one for the other is the core error.

The half you cannot see

The most uncomfortable part is where the activity lives. Your enterprise controls, single sign-on, CASB, DLP, only work on the accounts and tools they know about. They are built around the sanctioned, managed environment. They are blind to anything outside it.

And a large share of real AI activity happens outside it: on personal accounts. Someone pasting a document into their own ChatGPT login, a free image tool, an unmanaged note-taker sitting in meetings, an extension quietly sending content to a third-party model. None of it touches your identity provider, your proxy or your data-loss controls. So the policy that is perfect for your managed tools governs nothing that is happening on personal accounts, and that is frequently where the sensitive data and the real exposure are. A governance model that only reaches sanctioned accounts is, by definition, governing the minority of your AI use.

This is the practical meaning of a phrase that sounds like a cliche until you live it: you cannot govern what you cannot see. The tools you have not inventoried are not covered by your policy, your risk assessment or your controls, no matter how good those are for the tools you have.

The risk is not evenly spread

There is a second surprise waiting in the numbers, and it changes where you should look. AI usage does not distribute evenly across a workforce. It concentrates, hard. The median employee touches AI lightly. A small group of power users, often the top few per cent, generate many times more AI activity than everyone else, sometimes an order of magnitude more.

The instinct is to assume those heavy users are the problem. They are usually the opposite: they tend to be your most capable and productive people, the ones who have found how to use AI to move faster. Which leads to an awkward truth. Your largest concentration of AI exposure often sits with your best people, doing their jobs well, on tools your controls cannot see. That is not a discipline problem to be stamped out; it is a governance and enablement problem to be met. A model that assumes risk lives with a careless few, and that usage is uniform, is looking in the wrong place.

Why “we have policies” is false comfort

A strategy that governs the approved tools gives you the feeling of coverage without the substance, and that is arguably worse than knowing you have a gap, because it stops you looking. The board hears “governed.” The risk register says “covered.” The board’s own oversight rests on an answer that was never measured. And because everything appears fine, nobody runs the check that would show otherwise, until an incident, an audit, a data-protection complaint or one pointed question does it for you.

The exposure is not hypothetical while it is invisible. The EU AI Act and data protection law apply to the AI you use, not just the AI you approved; obligations attach whether or not you knew a tool was in play. “We did not know that was being used” is not a defence that survives contact with a regulator.

Visibility comes first

The correction is a reordering. Governance does not start with a policy; it starts with a truthful picture of what is actually in use. The sequence that works:

Measure actual usage, not approved usage. Find the tools people really use, including the ones on personal accounts. No single method catches everything, so triangulate: ask, in a genuinely no-blame way; comb subscriptions and expenses; audit the AI features switched on inside tools you already own; and use network, SSO and managed-browser telemetry where you have it. We cover the how in finding the shadow AI you do not know you are using.
Build the register from what you find, not from your approved list. Every tool goes into one AI system register, each with a named owner. This is the artefact your governance actually attaches to.
Classify each one by risk, by the data it touches and what it decides, so effort follows exposure: how to classify your AI systems by risk.
Make the sanctioned path better than the workaround. The reason people use personal accounts is that they help and are frictionless. Provide enterprise tools that are genuinely good, with rules people will actually follow, so there is little reason to route around you. Bans alone push usage further out of sight. See writing an AI policy people follow.
Bring the power users in. Learn what your heaviest users need and give them safe capability, rather than clamping down on the people delivering the most value.
Re-measure. The tool list changes monthly. Visibility is not a one-off audit; it is a standing capability, which is precisely why governance has to be a system, not a document.

Why this is what a management system is for

If the pattern above sounds like a lot to keep running, that is the point, and it is exactly why a serious approach is a management system rather than a strategy deck. An AI Management System aligned to ISO 42001 starts with a complete inventory and keeps it current, ties each system to an owner and a risk assessment, and builds the evidence that shows it is all actually operating. It forces the measurement and the maintenance that a strategy on paper skips. The register, the classification, the review cycle are not bureaucracy; they are what stops the gap between approved and actual from reopening the moment you look away. That is what a working AIMS contains.

The short version

The trap is not having no AI governance. It is having governance that covers the tools you approved and mistaking it for coverage of the AI your people actually use. When you measure the difference it is usually large: far more tools than you listed, a big share of activity on personal accounts your controls never see, and the heaviest use concentrated among your most productive people. You cannot govern what you cannot see, so visibility comes first. Measure real usage, build the register from reality, classify the risk, and make the sanctioned path good enough that people do not route around it. Then, and only then, does your policy govern something real.

Want a fast, honest read on where your AI governance actually stands? The free AI governance readiness assessment takes about ten minutes, no email required, and surfaces exactly these gaps. When you want to turn what you find into a register and a working management system, talk to us.

For the full picture, see the AI governance guide.