Shadow AI: The Unapproved Tools Your Staff Already Use at Work

It is already happening in your business

If your business has more than a handful of staff, someone is already pasting work into an AI tool you have never approved. A customer complaint dropped into free ChatGPT to draft a reply. A supplier contract summarised by a browser extension. A meeting recorded by a transcription bot that nobody in charge ever invited. Surveys of workplace AI use consistently find that a large share of employees use these tools without telling their employer, and small businesses, with thinner IT oversight, are more exposed than most.

The instinct is to ban it. That instinct is understandable and usually wrong. Bans do not stop the behaviour; they move it onto personal phones and home laptops, where you have no visibility at all. The widely reported 2023 incident in which Samsung engineers pasted confidential source code into ChatGPT happened at a company with far more controls than a typical SMB. The lesson is not that staff are reckless; it is that the tools are useful enough that people will find a way.

Why staff hide it

• Fear of looking lazy, or of revealing that a task takes less time than everyone believed
• Fear that admitting AI use invites a redundancy conversation
• A blanket ban, which turns every admission into a confession
• No approved alternative that is as good as the free tool they found
• Nobody ever asked

Notice that every item on that list is a management problem, not a technology problem. That is good news, because management problems can be fixed without buying anything.

The risks actually worth worrying about

Not all shadow AI use is dangerous. The risk lives almost entirely in what data goes in and what output goes out unchecked.

• Data protection: pasting customer or staff personal data into a consumer AI tool is a disclosure to a third party. Without a proper agreement in place, that sits badly with UK GDPR, and the ICO expects you to know where personal data goes
• Training on your inputs: consumer tiers of many AI tools may use conversations to improve models unless the user opts out, while business tiers typically exclude this by default. Staff on personal accounts get the consumer default
• Confidentiality: client NDAs and your own commercial secrets do not stop applying because the recipient is a chatbot
• Unchecked output: quotes, legal-sounding letters and technical answers going to customers without review
• Orphaned accounts: work products accumulating in personal logins that walk out of the door with the employee

Note what is not on the list: the mere fact of AI use. Drafting, summarising and brainstorming with non-sensitive material is mostly upside.

Finding it without a witch-hunt

The goal of discovery is a map, not a charge sheet. If people believe the survey is a trap, you will get a map of nothing.

• 1. Run an anonymous survey asking which tools people use and for what, with an explicit written promise of no disciplinary consequences
• 2. Scan expense claims and company card statements for AI subscriptions
• 3. Check data you already have: Microsoft 365 and Google Workspace admin reports show third-party app sign-ins and browser extensions
• 4. Ask team leads a positive question: what has saved your team the most time this year?

Expect to find more than you thought, concentrated in writing-heavy roles: sales follow-ups, job adverts, social posts, reports and proposals.

The amnesty-then-approve framework

Once you can see the landscape, run this sequence rather than reaching for a policy template:

• 1. Declare an amnesty window: four weeks in which anyone can declare any tool with zero consequences
• 2. Catalogue what surfaces: the tool, the use case, and what data goes into it
• 3. Triage by data sensitivity, not by tool name: green for public and non-sensitive material, amber for internal data, red for personal data, client-confidential material and financial records
• 4. Approve fast: for the popular tools, buy the business tier, which typically adds admin controls, single sign-on and no-training-by-default terms, then migrate people onto company accounts
• 5. Publish a one-page policy: the approved list, the red-data rules, and a request route for new tools with an answer promised in days, not months
• 6. Review quarterly; the tool landscape moves fast enough that an annual review is a dead letter

Make the sanctioned path the easy path

Shadow AI persists when the official route is worse than the unofficial one, so make it better. Paid accounts are faster and more capable than free tiers. A shared prompt library saves everyone reinventing the same wheel. A named AI lead gives questions somewhere to go. A short training session on what must never go into a prompt does more than a policy PDF nobody reads.

Handled this way, shadow AI is not a scandal; it is free user research. Your staff have already run the pilots and found the use cases that work. Your job is to keep the gains and strip out the risk. If you want help with the tooling, the policy or the rollout, our team works with UK SMBs on exactly this.