If you have not had a conversation with your staff about AI tools, you already have shadow AI. It is not a hypothetical. Someone in your office has pasted a contract, a patient note, or a client’s financial statement into a free chatbot to get a faster summary, a cleaner rewrite, or a quick answer. They did not do it to cause a problem. They did it because it worked and nobody told them not to.

This post covers three things: what that actually risks, how to find out how much of it is happening in your office, and what a workable policy looks like, including a safe tool your staff can use instead of quietly working around you.

What “shadow AI” actually means

Shadow AI is any AI tool an employee uses for work without IT or ownership signing off on it. It is the AI version of shadow IT: personal Dropbox accounts, unsanctioned Slack workspaces, spreadsheets emailed to a personal address so someone can work on them at home. The pattern is the same. Staff find a tool that saves them time, and they use it, policy or no policy.

The difference with AI is what gets typed into the box. People do not think of a chatbot as “sending data somewhere.” It feels like using a smarter search engine. So they paste in:

• A draft contract or NDA to get a plain-language summary
• A patient intake note to help write a referral letter
• A spreadsheet of vendor pricing to build a comparison
• An email thread with a client’s account numbers to draft a reply
• HR complaint details to help phrase a response

None of that is malicious. All of it can leave your control the moment it is submitted to a free, consumer-grade AI tool, depending on that tool’s terms of service and data retention settings.

What it actually risks

For a law firm, accounting practice, medical or dental office, or manufacturer handling customer specs, the exposure breaks down into a few concrete categories:

• Confidentiality and privilege. Client and patient data pasted into a public tool may be logged, used to improve the underlying model, or reviewed by the vendor’s staff, depending on the plan and settings. For attorneys and healthcare providers, this raises real questions about confidentiality obligations and, for healthcare, HIPAA.
• Contractual exposure. Many client and vendor contracts include confidentiality or data handling clauses. Pasting that data into a third-party AI tool without checking those terms can put you in breach without anyone realizing it happened.
• No audit trail. If a free consumer chatbot is where sensitive data goes, you have no record of what was shared, by whom, or when. If something goes wrong later, you cannot reconstruct it.
• Inconsistent output quality. Beyond the security question, ungoverned AI use means nobody is checking whether the tool’s answers are accurate. A hallucinated citation or a wrong number in a client-facing document is a different problem, but it compounds the first one.

This is not a reason to ban AI. Banning it just pushes the behavior further underground, onto personal phones and personal accounts where you have even less visibility. The goal is to find out what is happening, and replace it with something that works and is actually safe.

How to find out how much of this is happening

You will not get a straight answer by sending an email that says “please don’t use AI tools.” Most people will not read it as applying to them, because they do not think of what they are doing as “using an AI tool” in the way you mean it. Two approaches work better, and they work best together.

1. Ask directly, in person, without blame. A short conversation with team leads or department heads gets more honest answers than a survey. Frame it as “I want to know what’s actually being used so I can make it safe, not to get anyone in trouble.” Ask specifically: “Has anyone used ChatGPT, Copilot, Gemini, or similar tools for client work? What did you use it for?” People will tell you once they understand the goal is a safe option, not a crackdown.

2. Check what your systems already show you. If you use Microsoft 365 Business Premium or E3, check Entra ID under Enterprise applications for AI tools staff have already signed into with their work account (an easy way to spot ChatGPT, Gemini, or similar showing up as a consented app), plus your firewall or DNS filter logs for traffic to consumer AI domains. Microsoft 365 E5 includes Defender for Cloud Apps, which adds a dedicated Generative AI app category that shows this activity directly, but you do not need E5 to get a useful picture. If you are not sure what your current setup shows you, that is a five-minute conversation with whoever manages your Microsoft 365 tenant.

Between the interviews and the admin console check, you will get a realistic picture: which tools, how often, and for what kind of work.

What a one-page acceptable use policy covers

You do not need a 12-page document nobody reads. A single page that staff will actually remember covers:

• What is approved. Name the specific tool(s) staff are allowed to use, and note that the approved tool is different from the free public version of the same product.
• What is off-limits. Client names, patient information, account numbers, contract terms, anything covered by an NDA, and anything you would not want to see forwarded outside the company.
• Where approved AI use is allowed. Internal drafting, general research, formatting help, brainstorming, non-sensitive summarization.
• Who to ask when unsure. One named person or role. “If you are not sure whether something is safe to paste into an AI tool, ask [role] first.”
• What happens with mistakes. Make clear that reporting an accidental paste of sensitive data is not a punishable offense. You want people to tell you, not hide it.

Keep it to one page. Post it somewhere staff will actually see it, not buried in an employee handbook nobody opens after onboarding.

Giving staff a safe option

A policy without an approved alternative just tells people no. Two paths work, depending on your size and sensitivity level:

• Licensed tools with the right settings. Microsoft 365 Copilot and similar enterprise AI tools, configured correctly, keep data inside your existing Microsoft 365 tenant boundary rather than sending it to a public model with consumer terms of service. The setup and permission boundaries matter here. Default settings are not always the safe settings.
• Private, self-hosted AI. For firms handling especially sensitive data, or that want data to never leave infrastructure they control, a private AI deployment removes the third-party data handling question entirely. This is a bigger project than flipping on a licensed tool, but it is the strongest answer to “where does our data actually go.”

We covered the tradeoffs between these two paths in more detail in Microsoft 365 Copilot vs Private AI, and if you have not started thinking about AI at all yet, AI for Small Business: Where to Actually Start is the right starting point. For what Business Premium includes versus Standard, see Microsoft 365 Business Premium vs Standard.

Next step

If you want a straight read on where your office actually stands, an AI Readiness Assessment gives you a clear picture of current AI use, exposure points, and what a safe rollout looks like for your size and industry. Visit the AI Integrations page to see the options, or call 224-531-1000 to talk it through.