What Is Shadow AI?

What counts as shadow AI

Shadow AI covers any AI use at work that sits outside what the organization has approved or can see. Common forms include:

• Personal accounts on chatbots like ChatGPT, Claude, or Gemini, used for work tasks.
• Browser extensions that rewrite, summarize, or translate the web pages you have open, which usually means they can read those pages, often on every site you visit.
• Free online tools for transcription, PDF summaries, slide decks, or images.
• AI features switched on inside other apps, such as a meeting note-taker or a writing assistant in software the team already uses, without anyone reviewing them.

The label depends on approval and visibility, not on the tool. The same assistant can be approved at one company and shadow AI at another. Even within one company, the business version of a tool may be approved while personal accounts on the same tool are not, because the data terms and admin controls can differ.

Why it happens

Shadow AI is rarely a sign of bad intent. Most of it starts with someone trying to get work done faster. A few forces drive it:

• The tools are useful and easy to get. Many are free or low cost, need only an email address, and run in a browser tab or on a phone.
• Approval is slow or unclear. If requesting a tool takes weeks, or nobody knows who to ask, people skip the step.
• There is no policy. Without written guidance, staff make their own calls about what is safe to paste in.
• The approved option falls short. If the sanctioned tool is harder to use or less capable than a free one, people quietly go back to the free one.
• AI arrives on its own. Software vendors keep adding AI features to products teams already use, sometimes turned on by default.

Each of these causes is something an organization can address directly, which is why the most workable responses target the causes rather than the tools.

Risks for your organization

The core problem is that an organization can’t manage what it can’t see. Key risks:

• Data exposure. Client details, employee records, financials, or plans may go into tools whose data terms nobody has read. Depending on the service and settings, conversations may be stored or used to improve the provider’s models. Protections arranged for approved accounts, such as contractual limits on training and retention or admin controls, don’t extend to personal accounts. Neither does zero data retention (a provider agreeing not to store prompts and responses), which some providers offer business customers.
• Confidentiality and compliance. Pasting in material covered by a nondisclosure agreement (NDA), a client contract, or privacy and industry rules may break commitments the organization has made.
• Intellectual property. Proprietary material, such as product designs, code, or a firm’s own methods, may be shared under terms nobody checked. Those terms set what the provider may do with what you enter and what rights you have to the output. Output can also occasionally resemble someone else’s protected work.
• Inconsistent quality. People use different tools in different ways, and unchecked errors, including hallucinations (confident but made-up details), can slip into client-facing work.
• No record of how work was produced. If a report, clause, or analysis is questioned later, the organization may have no record it can access of which tool shaped it or what was shared with it.

How to respond without a blanket ban

Banning AI outright tends to push use further out of sight, especially when anyone can open a chatbot on a personal phone. A more workable response makes the safe path easier than the workaround:

1. Find out what is already in use. Ask openly, without blame, which tools people rely on and for what.
2. Approve tools that are good enough. Offer at least one capable assistant on a company-managed account with clear data terms and admin controls (protections often sold under an enterprise-grade label), so the approved route is also the convenient one.
3. Write a short, clear policy. Cover which tools are approved, what data must never go in, when a person must review AI output, and how to request a new tool.
4. Train people. Show staff how to use the approved tools well and how to spot weak or wrong output.
5. Build in visibility. Organizations that build their own AI apps can route those requests through an AI gateway to log usage and add technical AI guardrails, such as screening for sensitive data. These controls see only the traffic that passes through them, so they work alongside policy and training, not in place of them. Smaller teams can start with a simple list of approved tools and a fast way to request new ones.

Revisit the policy and tool list regularly, since new tools and features appear often.

Further reading

• AI Fluency AssessmentA free assessment of where your business stands on AI adoption, a useful starting point before you write an AI policy.
• AI Training & SpeakingCovers SN3’s executive sessions, team training, and custom workshops, relevant if training is part of your response to shadow AI.