Insights
Treating Shadow AI as R&D Data
Most organisations are treating shadow AI as a security problem. I think it is the most honest research and development data they have ever been given, and they are about to delete it.
The pattern is now measurable. Prodoscore’s August 2026 analysis of activity across 72 companies found employees actively using nearly 50 distinct AI tools. Three of the five most-used platforms were consumer products rather than the ones IT had provisioned. ChatGPT reached roughly three times as many employees as the sanctioned enterprise assistant, and accounted for close to five times the hours.
The standard reading is non-compliance
Policy failed, procurement was too slow, people leaked data into systems nobody approved. All of that is true, and the risk is real, and none of it is the interesting part.
Consider what those numbers actually are. Thousands of people, unpaid and unasked, ran an evaluation of the tools available for their own work. They compared. They abandoned things that did not help. They adopted things that did, at personal cost, without training budget or a change-management deck. That is a distributed procurement study conducted by the only people who know what the work really requires.
Organisations spend serious money trying to obtain exactly this information — vendor bake-offs, pilots, steering committees — then produce a decision optimised for contract terms, and are surprised when adoption stalls. Meanwhile they already hold a revealed answer about which tools fit which tasks, and their instinct is to shut it off.
Two things make this expensive rather than merely ironic
The first is that the workaround usually contains a diagnosis. Someone using an unapproved tool is telling you, precisely, where the approved system fails. Not as an opinion in an engagement survey, but as behaviour, with effort behind it. Workday’s January 2026 research found that nearly 40 percent of AI time savings are consumed by rework — correcting, rewriting, verifying output — and that only 14 percent of employees consistently get clearly positive net results. People do not go around a system that works.
The second is what the response teaches. When the honest reply to “how are you doing this?” is career-limiting, the practice does not stop. It goes quiet. And once it is quiet, the organisation loses the ability to see either the risk or the invention. You end up governing a fiction: a policy that describes work nobody is doing, sitting on top of work nobody will describe.
I do not think the answer is permissiveness. Data going into unmanaged systems is a real liability, and the answer to a real liability is not to admire it.
An amnesty with a ledger
The answer is to separate the two questions the organisation keeps collapsing into one. Where the data went is a control problem, and it should be handled strictly. Why the person went there is an intelligence problem, and it should be handled greedily.
In practice that means an amnesty with a ledger. Ask what people are actually using and what for; guarantee that the answer is never used against them; publish the results internally so the discoveries stop being private property. Then treat the top items as a backlog, not a violation list. Most organisations will find they have a working map of their own friction, assembled for free, that no consultant could have produced.
There is a broader principle underneath. When execution becomes abundant, the scarce thing an organisation holds is what its people have noticed. Noticing shows up first as deviation — someone doing it differently because the official way is worse. An organisation that punishes all deviation is not safer. It is simply blind in the one direction where its future is arriving.
The question I would ask a leadership team this month is not “how do we stop this?” It is “what did our people learn that we have no way of hearing?”