Shadow AI is what happens when employees use AI tools — ChatGPT, Gemini, Claude, AI meeting note-takers, browser plugins — for work without IT knowing about it. It’s the AI-era version of “shadow IT,” and according to Gartner research published in May 2025, 69% of organizations either suspect or have evidence their employees are doing it. For a small business, the risk isn’t the chatbot itself — it’s the company data being pasted into it.
What Is Shadow AI in Plain English?
Think of it as any AI tool touching your business data that you didn’t choose, license, or configure. Common examples we see across Western WA small businesses:
- An employee pasting a customer list into the free ChatGPT tab to “clean it up.”
- A paralegal feeding a draft motion into a consumer LLM to “summarize for the partner.”
- An admin running an AI meeting recorder during a client consultation.
- A salesperson installing a browser AI summarizer that has full access to every webpage they visit, including your CRM.
- A developer using a free-tier code assistant that may train on the snippets it sees.
None of these tools are inherently malicious. The problem is that data leaving your environment through an unmanaged tool is data you no longer control.
Why Are Employees Using AI Without Asking?
Mostly because it works, it’s free, and they’re trying to do a good job. The same survey research consistently finds that employees adopt AI faster than employers can write policies — and the gap is widening, not closing. IBM’s 2025 Cost of a Data Breach report found that only 37% of organizations have shadow-AI policies in place. The rest are running on hope.
People reach for shadow tools when sanctioned ones are slow, missing, or perceived as worse. If your team has no licensed AI option, every individual will pick their own. If your team has a licensed option that’s harder to use than the free alternative, they will still pick the free one. The pattern is familiar from the shadow-IT era — the answer isn’t more rules, it’s making the approved path the path of least resistance.
What Could Actually Go Wrong?
The realistic risks aren’t theoretical. Three short stories make the point.
Samsung, April 2023. Within 20 days of allowing internal ChatGPT use, Samsung experienced three separate leaks — engineers pasted in source code, an employee dropped in internal meeting transcripts, another shared semiconductor test data. Samsung banned generative AI tools company-wide shortly after. The lesson isn’t “ban AI”; it’s that even sophisticated companies underestimate how fast confidential data flows out once a chatbot is on the desktop.
Air Canada, February 2024. A British Columbia Civil Resolution Tribunal held Air Canada liable for its chatbot’s incorrect advice about bereavement fares — CAD $812.02 to the customer. Small award, big precedent: companies are responsible for what their AI tells customers, even when employees set the bot up unofficially. (Note: this was a Canadian tribunal, not binding U.S. law, but cited widely as the first clear ruling.)
Regulated-data leakage. What we typically see in healthcare and legal environments — an employee summarizing a PHI-laden chart note or a draft pleading in a free chatbot — is usually a one-keystroke compliance event with no audit trail. We’ll cover the regulated-industry specifics in a separate post on AI risks for healthcare and legal practices.
How Bad Is the Problem?
Two data points worth knowing:
- Gartner forecasts that by 2030, 40% of organizations will experience a security or compliance incident tied directly to shadow AI.
- IBM’s 2025 breach data shows shadow-AI-related breaches averaging around $4.63 million in cost — roughly $670K more than typical breaches — and taking 247 days to detect versus 241. (That figure is across organizations of all sizes; for a small business the dollar impact is smaller in absolute terms but proportionally worse.)
The detection lag is the part most owners underestimate. Shadow AI leaks tend to be silent — there’s no ransomware note, no encrypted drive, just an employee whose ChatGPT history quietly contains your last three quarterly reports.
How Do I Get Shadow AI Under Control Without Becoming the Bad Guy?
A reasonable sequence for a 10-to-50-person business, ordered from cheapest to most involved:
- Take inventory first, write rules second. Ask your team — by survey, not surveillance — what AI tools they’re already using. You’ll learn more in one Friday than from any audit tool. This also signals you care about their productivity, not just policing.
- Pick a sanctioned alternative. A single licensed AI option with enterprise data protection (Copilot Chat free tier with company sign-in, ChatGPT Enterprise, or Claude for Work) covers most use cases. Make it as easy to reach as the free option.
- Write a one-page AI acceptable use policy. Cover approved tools, prohibited inputs (client PII, source code, regulated data), and a no-blame reporting line. We cover this in detail in how to write an AI acceptable use policy for your business.
- Add technical guardrails. DNS filtering categories that flag AI tools (most modern MSP stacks have this), Microsoft Defender for Cloud Apps to surface unsanctioned SaaS, Entra ID conditional access to gate consumer logins on company devices. This is the same playbook we recommend in our remote work security guide.
- Train, then re-train. Twenty minutes during onboarding, twenty minutes annually. Bundle it with phishing training since the failure modes overlap.
- Have an incident playbook ready. If a leak happens — and it eventually will — your incident response plan should already cover it.
Most importantly: don’t lead with a ban. Bans push usage to personal phones and personal accounts, which is worse than what you started with.
ROI Technology Inc. helps Western Washington small businesses build practical AI usage guardrails without slowing their teams down. Contact us or call (888) 707-3652 to schedule a shadow-AI exposure review.