The number that surprises leadership teams is rarely the number of licences. It's the number of people who've been using AI for a while already, without anyone knowing. Not secretly in the sense of bad faith – but because nothing official exists and the work still has to get done.
That's given rise to the term shadow AI, echoing shadow IT. And as with shadow IT: the problem isn't staff disloyalty. The problem is that a need exists which the organisation isn't meeting.
How Shadow AI Comes About
The typical path is unremarkable. Someone has too much to do, tries an AI tool privately, notices a task now takes ten minutes instead of sixty, and keeps doing it that way from then on. They mention it to a colleague, who does the same. Within six months, part of the workforce is working with tools that were never approved.
The motivation is almost always productivity, not rule-breaking. That matters for how you respond.
What's Actually Risky About It
Four things, in descending order of significance.
The contract terms are different. If someone uses a private consumer account, personal-user terms apply. Under commercial agreements – Team, Enterprise or API use – established providers don't use company content to train their models by default. With private accounts, that depends entirely on settings nobody in the business controls. Same technology, completely different legal footing.
There's no trail. If nobody knows who put which content into which tool, nothing can be reconstructed if it matters. If a question comes from audit, a client or a regulator, there's no solid answer. With a company solution and audit logs, there is.
The knowledge stays with individuals. Someone who's spent months training themselves into a good way of working takes it with them when they change jobs. In a company solution, recurring workflows can be captured as reusable skills that everyone uses. This is the most underrated point: shadow AI makes individual staff faster, but never the business.
Quality can't be checked. Without a defined process, everyone writes their own way. For client communication, that means five tones of voice, five quality levels, no shared baseline.
Why a Ban Is the Wrong Move
The reflexive response is a policy banning private AI use. That has three effects, two of which are unwanted.
First, the usage doesn't disappear – it just becomes less visible. Second, you lose the people who are furthest along internally – precisely the ones you'd need as champions for a later rollout. Third, you buy yourselves peace of mind at the cost of a lag your competitors don't have.
A ban is also hard to enforce when every private phone has a browser on it. It doesn't stop the usage – it just stops people talking about it.
What Works Instead
The most effective step is unspectacular: create an official, approved option that's at least as good as what people already use privately. If the company solution is worse than the private account, the private account wins – every time.
That means four things. Company accounts with sign-in through your existing system, so access ends automatically when someone leaves. Clear rules on which data classes are allowed and which aren't – in language someone without legal training understands. Access to your own systems via controlled connections, so nobody has to copy data in by hand anymore. And an amnesty: say openly, once, that past usage isn't an issue and that you want to know what's actually needed.
That last point costs nothing and pays off the most. We've repeatedly seen the best use cases come up in exactly that conversation – from people nobody had thought to ask before.
A Worthwhile Interim Step
Before you decide on licences, clarify one question: what tasks are your staff already doing with AI today? Not as a check-up, but as a stocktake. In almost every case, the result is the most realistic use-case list you'll get – it doesn't come from a slide deck, but from actual day-to-day work.
That's exactly what our readiness check builds on. We look at what's already there, instead of asking what could be.
.svg)