Shadow AI, meaning employees using AI tools that IT never approved or does not know about, is already the norm in most organisations, not the exception. It is happening at the top of the organisation chart as much as at the bottom.
A steering committee I sat with was 40 minutes into drafting a stricter Shadow AI policy when the CIO said the quiet part out loud:
“You all know we are describing ourselves.”
Nobody disagreed. The meeting became more useful after that. It is a pattern we have seen before: employees using AI without approval almost always turns out to be widespread once anyone actually looks.
The numbers say that room is every room. Seventy-eight per cent of AI users bring their own tools to work, according to the Microsoft Work Trend Index 2024. Meanwhile, 49% of employees have hidden their AI use from a manager, according to WalkMe's 2025 survey. Ethan Mollick gave the hiders a name that stuck: secret cyborgs.
of AI users bring their own AI tools to work.
of employees have hidden their AI use from a manager.
of workers use personal AI tools daily, according to MIT's 2025 data.
The standard corporate response is a policy document and a firewall rule. Here is why that fails, and the 90-day path that works instead.
Why bans fail, mechanically
Bans do not remove demand; they relocate it. When Samsung banned ChatGPT in 2023 after a source-code leak, usage did not stop. It moved to personal devices, where the company had even less visibility than before. Substitution, not compliance, is the reliable outcome, and it comes with a bill.
Breaches involving high levels of Shadow AI cost an average of $670,000 more, according to IBM's 2025 analysis, and Shadow AI detections are rising fourfold year on year, according to the Verizon 2026 DBIR.
There is also an enforcement problem nobody likes stating plainly. The people who would enforce the ban are its heaviest violators. A rule the leadership breaks daily is not a rule; it is décor.
Reframing Shadow AI as a demand signal
Read Shadow AI as a demand signal and the whole picture reorganises. Your people have already found where AI creates value in their jobs, quietly redesigned their workflows around it, and are running the result on personal accounts.
MIT's 2025 data made the scale plain. Roughly 90% of workers use personal AI tools daily, against about 40% of companies with official subscriptions.
That is not only a compliance crisis. It is the most detailed and honest requirements document your AI programme will ever receive. It is simply written in browser histories instead of workshops.
The 90-day path
Amnesty and inventory
Announce a no-consequences disclosure window. Ask employees which AI tools they use and what problems those tools solve.
Outcome: A practical map of real AI demand, not a confession booth.
Build the fast lane
Establish an approved-tool route with a real service commitment. Answer requests in days, not quarters, and cover the five most common use cases.
Outcome: A sanctioned route that is faster than the workaround.
Make it stick
Start a tool registry, instrument approved usage and publicly recognise teams that disclose clever unsanctioned workflows and help migrate them.
Outcome: Better visibility, adoption and disclosure.
Keep the data rules to one page
Pair the amnesty with three clear data tiers. If the classification system requires a workshop to explain, employees will ignore it.
Green data
Public and low-risk information that may be used with approved general-purpose tools.
Amber data
Internal or controlled information that requires an approved, governed AI platform.
Red data
Sensitive information that must never leave designated governed systems.
The fast lane applies the same principle as adopting AI without disrupting operations: the sanctioned path only wins if it is faster than the workaround.
The honest edge
Amnesty is not absolution. Some data can never touch a personal tool. If the inventory surfaces genuine exfiltration of red-tier data, that is a security incident, not an amnesty item.
The tiers are what make the generosity safe. Governance without teeth is theatre; teeth without a fast lane is simply a slower way to fail.
Frequently asked questions
What is Shadow AI?
Shadow AI is employees using AI tools that IT has not approved or does not know about, usually because those tools solve a real problem faster than the official options do.
Why do bans on unsanctioned AI tools fail?
Bans do not remove the demand. They relocate it, often onto personal devices where IT has even less visibility. That is what happened when Samsung banned ChatGPT in 2023.
What does the 90-day governance plan involve?
It has three phases: 30 days of no-consequences amnesty and inventory, 30 days building a fast approved-tool route, and 30 days making the approach stick through a tool registry and usage tracking.
Does amnesty mean no consequences for past AI use?
Amnesty covers disclosure, not a blanket pass. If the inventory reveals an actual exposure of sensitive data, it should be treated as a security incident, not an amnesty item.
One more honest note
I do not know a reliable way to reach the last few per cent who will hide usage no matter what. I have stopped designing for them. Design for the 90% who would happily use the sanctioned path if it were good.
The Monday move fits into three sentences at your next town hall: “We know most of you use AI tools we have not approved. For the next 30 days, we simply want to know what and why, with no consequences. Then we will build you a faster legal way to do the same thing.”
Say that, mean it, and you will learn more about your organisation's AI readiness in a month than any maturity assessment will tell you.
Related insights
