← All articles
AI Strategy5 min read

Your employees are already using AI you didn't approve

Employees are already using public AI tools with company data. Survey what is really happening, provide an approved path, and draw a boundary people can remember.

Your employees are already using AI you didn't approve

If you have not set an AI policy yet, your employees have already set one for you — in browser tabs, with whatever tool was fastest. A salesperson pastes a customer list into a public chatbot to draft follow-ups. An analyst uploads a contract to summarise it. A recruiter feeds CVs into a free tool to rank candidates. They are just getting the work done.

That is shadow AI: AI use outside any approved tool, any data agreement, and any record of what went where. It is what happens when useful technology arrives faster than governance — and the longer leadership pretends it is not happening, the more company data ends up in systems the company has no contract with and no way to delete from.

Start by finding out what is actually happening

Do not start with a ban. Start with a survey, and make it safe to answer honestly. Ask teams which tools they use, for what, and what kind of data goes in. You will usually find three groups: people using approved tools well, people using public tools for harmless tasks like drafting emails, and a smaller group pasting genuinely sensitive material — customer data, contracts, source code, HR records — into tools never meant to hold it.

That map is the whole strategy: harmless uses tell you which approved tools to provide, sensitive uses tell you where the real risk sits. A blanket ban treats both groups the same and drives risky use underground — people do not abandon a tool that saves them an hour a day because a policy told them to.

Make the approved path as good as the shadow one

The fastest way to shrink shadow AI is to make the sanctioned option competitive: an approved tool with business-grade terms, no training on your inputs, and access controls matching your existing systems — as easy to reach as the public alternative. If it takes three logins and a ticket to IT, it will lose to the tool that takes one click, every time.

Then draw the boundary in plain language. Green: drafting, brainstorming, summarising public material. Amber: internal documents, approved tool only, human check before anything leaves the company. Red, always: customer personal data, credentials, source code, financial records, and anything under NDA in any public tool. A ten-page policy nobody reads is not a boundary; it is plausible deniability.

You also cannot un-paste data. If sensitive material has already gone into a public tool, record the exposure, check its retention terms, rotate anything credential-like, and tell affected customers if it warrants it. Going forward, the controls that work are boring: an approved-tool list tied to real data agreements, blocking consumer AI tools on company devices where the risk justifies it, and a channel where employees can ask "can I use this for that?" and get an answer in hours.

Shadow AI is not a sign that your people are careless. It is a sign that they found value faster than the organisation provided a safe way to capture it. We help companies turn that quiet adoption into a governed AI setup — approved tools and clear boundaries, without giving up the productivity they already found.

Field note: In one company we advised, the survey found a support lead pasting ticket histories — customer names and order details included — into a public chatbot to draft replies. Nobody had told her not to; the approved-tool list lived on an intranet page she had never seen. The fix was an approved tool on her desktop the same week and a one-page boundary sheet. Within a month, public-tool use in that team had stopped — not because anyone was policed, but because the approved path finally did the job.

Related service AI products

Building something worth shipping?

We take on a small number of AI product engagements. Tell us what you are building — we reply quickly.