Shadow AI: your team already voted.

Nobody asked permission. A sales rep pastes the pipeline into a public tool to draft follow-ups. Finance uploads a P&L to summarize it. An engineer drops in source code to debug faster. Each of them is solving a real problem with the best tool they can reach. None of them read the terms of service.

The ban already failed

Prohibition assumes people use AI to defy you. They don’t. They use it because the sanctioned alternative is missing or worse than the public tool. A ban with no viable replacement does not reduce exposure — it reduces visibility. The data still leaves. You just stop hearing about it.

The uncomfortable truth: your employees have already run the cost-benefit analysis. The public tool won. Your move is not to relitigate their decision. It is to change the options on the table.

Know what is actually at stake

“Data exposure” is too abstract to act on. The concrete risks:

• Retention you don’t control. You don’t know what was pasted, where it is stored, or whether it trains someone else’s model.
• No audit trail. When something goes wrong — a leaked contract, a wrong number in a client report — there is no record of which tool touched the data.
• Offboarding gaps. Work done in a personal AI account walks out the door with the employee. You cannot revoke it, wipe it, or review it.
• Permission sprawl. The AI sees whatever the user pastes, including data that user was never supposed to have. Your access-control mess becomes an AI problem.

Make the governed path the easy path

People don’t circumvent good tools. They circumvent bad ones. The playbook:

• Ship one approved tool that is genuinely good. A governed assistant connected to company data, easier to reach than the public alternative. If the sanctioned path is the path of least resistance, adoption follows.
• Fix permissions first. AI should only ever see what the user is already allowed to see. If your data access is a mess, that is project one — the AI project comes second.
• Write a one-page policy. What is approved. What never gets pasted anywhere. Who to ask. Anything longer won’t be read, which is the same as not having one.
• Measure the program, not the people. Track adoption, data flows, and output quality at the system level. The goal is a better program, not catching individuals.

Start with an honest inventory

Before any policy: find out what is actually in use. Survey the teams, check network and endpoint signals, and ask directly — with no punishment attached. An honest inventory is the foundation everything else stands on. Most leadership teams are surprised by the results. That surprise is the business case for doing this properly.

Governance is not the brake on AI adoption. It is what makes AI safe enough to accelerate. Companies that treat it as infrastructure pull ahead. Companies that treat it as paperwork fall behind.