During the Second World War, the US Navy analysed which parts of returning bombers had the most bullet holes. The conclusion seemed obvious: reinforce those spots. The mathematician Abraham Wald saw it differently. The planes that came back had survived the damage. You needed to look at the spots without bullet holes. Those were where the fatal hits landed, on the planes that never returned.
Survivorship bias. You draw conclusions from what you see and miss what is invisible.
The same thing plays out in AI adoption inside organisations. The visible AI projects, the official pilots, the approved tools, those are the planes that came back. The real story is in what you cannot see.
The invisible adoption
In virtually every organisation we work with, the same thing is happening. Employees are using ChatGPT, Claude, Copilot, on their own phones, with their own accounts, without anyone knowing. They write emails with it, summarise meetings, generate first drafts of reports. Sometimes at work, sometimes at home in the evening.
The reflex of most organisations: this is a compliance problem. Data is leaking out. There is no policy. We need to ban it or regulate it.
That reflex is understandable. It is also short-sighted.
Because those employees are telling you something valuable. They are pointing precisely to where the friction is. Where the work is so frustrating or time-consuming that people are searching for solutions on their own initiative. That is strategic information.
From offence to compass
The smartest approach: amnesty. No punishment for unofficial AI use. Bring it into the open. Map which tools people are using, for what, and why.
The answers are revealing. "I use ChatGPT for customer emails because our template system doesn't work." "I have Claude summarise my meetings because otherwise I spend two hours on minutes." "I generate first drafts of proposals because the quoting process has six steps that nobody understands."
Every unofficial AI tool is a pointer to a broken process. And those pointers are more reliable than any consultancy report. People vote with their feet. Or in this case: with their prompts. It is Follow the Friction in its purest form, the friction finds itself.
What to do with it
Three steps.
Inventory. Ask each team: which AI tools are you using, for what, and what value does it deliver? Map it out. Not to audit, but to understand. The places with the most shadow usage are your best starting points. An AI workshop is an effective way to have this conversation in a structured setting.
Professionalise. The tools that deliver value, make them official. With the right security, the right data agreements, the right boundaries. That shifts you from unmanaged sprawl to focused adoption without losing the energy behind it.
Learn. The patterns in shadow usage tell you more about your organisation than many a strategy session. Where are the process breaks? Where is friction greatest? Where are people finding their own solutions because the official route is too slow or too cumbersome?
The difference
Organisations that treat Shadow AI as a threat lose the information it contains. Organisations that treat it as radar get a free diagnosis of where the value is.
Abraham Wald saved aircraft by looking at what was invisible. The best AI strategy begins in the same place: with the work your team is already doing when nobody is watching. The question then is whether your organisation is ready to act on it, or whether something else needs to happen first.