Your team already uses AI. Make that use safe.

“Half the team is pasting client work into some chatbot, and nobody asked.” Once Artificial Intelligence (AI) tools got good, most agencies heard a version of that sentence. Quiet use of tools nobody approved has a name: shadow AI. It is common and usually well meant. The fix is a short set of rules that makes the safe way the easy way.

What is shadow AI, and how common is it?

Shadow AI is any AI tool used for work that the company has not reviewed or approved: a personal chatbot account, a browser extension, a free transcription app. Microsoft and LinkedIn surveyed 31,000 knowledge workers in 31 markets in early 2024 and reported that 78% of AI users are bringing their own AI tools to work¹, rising to 80% at small and medium-sized companies. Microsoft's own wording is that this puts company data at risk¹. It is a vendor survey, so read the exact number loosely. The direction is hard to argue with. People reach for a tool that helps, and where no approved one exists they pick their own.

What goes wrong when someone pastes in client work?

The risk sits in what goes in. The Open Worldwide Application Security Project (OWASP) lists sensitive information disclosure among the top risks for Large Language Model (LLM) applications, and much of its list of protected information reads like an agency's shared drive: personal and health data, financial data, credentials and access keys, and trade secrets². An unreleased product brief pasted into a personal account has left your control the moment you press enter.

What happens next depends on the plan, not just the tool. Anthropic, for example, updated its consumer terms to give Free, Pro and Max users the choice to allow their data to be used to improve Claude³, and states that the update does not apply to services under its Commercial Terms, including Claude for Work and the API. Other vendors draw their own lines between personal and business plans. Two people can use the same chatbot on the same afternoon and be covered by different rules.

Should we just ban AI tools?

A ban is the instinctive answer, and it tends to backfire. If most people who use AI at work already bring their own tools, a ban mostly moves the habit out of sight, where you can no longer ask what is going in. A better start is to find out what people already use, then give them something safer to use instead.

You do not need a binder to do that. The U.S. National Institute of Standards and Technology (NIST) publishes an AI Risk Management Framework built around four functions: Govern, Map, Measure, and Manage⁴. It describes the framework as intended for voluntary use⁴, and a separate Generative AI Profile helps organizations identify the risks specific to tools like these. A ten-person team can borrow the shape of it on a single page.

What's Real?

• •Use is already widespread. Microsoft's survey found 78% of AI users bringing their own AI tools to work¹. It is self-reported and vendor-run, so treat it as a strong signal rather than a measurement.
• •The risk is the input. OWASP's list of sensitive data includes financial data, credentials and trade secrets², which is exactly what lands in a rushed prompt.
• •Plan terms differ. Consumer and business plans can be covered by different data rules, as Anthropic's split between consumer and Commercial Terms³ shows. Read the terms of the plan your team uses.
• •Help exists and is voluntary. NIST's framework is intended for voluntary use⁴. In the European Union (EU), the AI Act goes further: Article 4 requires providers and deployers to take measures to support the development of AI literacy⁵ among their staff.

The “so what” — for anyone serving tech clients

• •Run the amnesty first. Ask the team what they use and promise no blame. You cannot write rules for tools you have not seen.
• •Write the red lines on one page. Client-confidential material, credentials and personal data stay out of any tool that is not approved. Everything else can move.
• •Check the contract and the plan. Your client agreements may already say something about AI tools or subcontractors; read them before a client asks. Then check which plan each approved tool is on. The same product can carry different data terms.
• •Pair the rules with the brand guardrails. What goes into a tool and what comes out of it are two halves of one job; the brand-safety brief⁶ covers the second half.