Shadow AI is employees using AI tools their organization hasn't approved, and IBM's 2026 breach study says security incidents involving it more than doubled in a year. Most breached organizations still had no AI governance to manage AI or detect shadow AI. Good governance has to do two jobs at once: control what AI can touch, and give people a sanctioned option that's useful enough that they don't go around it.
What is shadow AI?
IBM defines it simply: shadow AI is "where workers use unapproved AI." (IBM, 2026)
It rarely looks like a security event. It looks like a manager pasting a draft contract into a personal chatbot account to tighten the language. An analyst uploading a customer spreadsheet to a free tool to get a quick summary. A browser plug-in that reads every page an employee opens. Each one feels like getting work done. None of them goes through your access rules or leaves a record you can review.
How common is shadow AI at work?
Very common, by every measure we found.
- In Microsoft and LinkedIn's 2024 Work Trend Index, "78% of AI users are bringing their own AI tools to work," and 52% of people who use AI at work "are reluctant to admit to using it for their most important tasks." The survey covered 31,000 knowledge workers in 31 markets. (Microsoft, 2024)
- MIT NANDA's 2025 report found that "only 40% of companies say they purchased an official LLM subscription," while "workers from over 90% of the companies we surveyed reported regular use of personal AI tools for work tasks." (MIT NANDA, 2025)
- An IBM-sponsored 2025 survey of 1,000 U.S. office workers found that 80% use AI in their roles, but "only 22% rely exclusively on tools provided by their employers." (IBM, 2025)
The second number in the Microsoft finding matters most. When half your AI users won't admit what they use it for, you can't govern it by asking.
What does shadow AI cost when something goes wrong?
IBM's Cost of a Data Breach Report 2026, conducted by Ponemon Institute, studied 602 organizations that had a breach between March 2025 and February 2026. Here's what it found on shadow AI (IBM, 2026):
- Security incidents involving shadow AI "more than doubled to 43% this year from 20% last year."
- Those incidents cost an average of USD 5.39 million, up from USD 4.63 million the year before. The study's global average for all breaches was USD 4.99 million.
- The most reported impacts were data loss or compromise (49%) and disrupted operations (42%).
- In about one in five shadow AI incidents, organizations reported paying a regulatory fine.
Keep the scope in mind. These are organizations that were breached, not all organizations, and the costs are averages. But the direction is clear, and it matches what employees say they're doing.
Why is AI governance falling behind?
IBM's answer: "AI adoption is outpacing oversight. That trend has only increased." The numbers (IBM, 2026):
- 68% of breached organizations "lacked AI governance to manage AI or detect shadow AI," up from 63% the year before. A third (33%) said they were developing policies, up from 22%.
- 38% of organizations that enforced AI governance required IT approval for AI deployments, the most common control IBM tracked, down from 45%.
- 19% reported that their governance and security teams coordinate their efforts.
- 21% reported a breach involving an AI model or application, up from 13%. Among them, 92% "lacked proper AI access controls," which IBM says include role-based access and multifactor authentication.
- Only 40% reported using access controls on AI models and data.
IBM also notes that the root causes of AI-related breaches were "often structural: compromise of connected APIs, applications and cloud misconfigurations, indicating governance failures—not model risk."
That last line is the key. The model usually isn't the weak point. The weak points are the connections around it and the rules nobody set.
Why isn't a ban on AI tools enough?
Because people use shadow AI for a reason. In IBM's U.S. survey, "nearly 40% of workers say they prefer external AI solutions for their better features." MIT went further, finding that shadow AI "often delivers better ROI than formal initiatives."
Read together, those findings say shadow AI is a demand signal. Employees have found AI that helps them, and the approved tools often haven't kept up. A ban with no better option doesn't remove that demand. It just moves it further out of sight, and the Microsoft data suggests much of it is already there.
The goal isn't to stamp out every AI tool outside IT. It's to make the governed path the easiest and most useful one.
What should AI governance actually do?
Think of it as six jobs. Each ties to a gap in the data above.
- See what's in use. You can't govern tools you don't know about. IBM lists regular audits for unsanctioned AI among the governance controls it tracks.
- Approve quickly. An approval process people will actually use, with a clear answer in days, not quarters.
- Control access. Apply role-based access and multifactor authentication to AI models and data, the controls that 92% of organizations with an AI-related breach lacked in IBM's study.
- Keep sensitive data in bounds. Decide what may leave your environment and enforce it before data reaches an outside model.
- Keep a record. Gartner lists "compliance tracking and audit trails, and security controls for data and model access" among the critical tools for responsible AI. (Gartner, 2026)
- Join governance and security. Policy owners and security teams need one picture. Only 19% of organizations in IBM's study said they coordinate the two.
The first five are controls. The sixth is about people. All six fail if the sanctioned tool is worse than the one employees found on their own.
Where does a context layer fit?
A context layer is designed for the "useful sanctioned option" half of the problem, with governance built in. In Nvitis, the Governance pillar sits across the whole record:
- Your team sets what information AI can use, by whom, and where, in one place instead of in every tool.
- Sensitive data is designed to be redacted before it reaches an external AI model.
- Answers come from your own record with their sources attached, and each one leaves an audit trail.
- Access follows roles, so people and AI agents see what they're allowed to see.
Because the same layer also connects your systems and keeps the reasons behind decisions, the governed path is designed to answer questions a personal chatbot can't, like why a policy changed or what was decided last spring. That's the point: make the approved tool the better tool. (We covered why that context matters for AI results in Why Do Enterprise AI Pilots Stall?)
Be clear about what it doesn't replace. IBM's leading root causes, such as compromised APIs and cloud misconfigurations, need identity, application, and cloud security work. A context layer is one part of an AI governance program, not the whole program. For buyers who ask, Nvitis's SOC 2 and ISO reports are available on request.
A checklist for CISOs, CIOs, and general counsel
- Inventory AI use. Include personal accounts, browser plug-ins, and AI features inside tools you already pay for.
- Ask employees why. Learn which tasks drive shadow use. That's your list for the sanctioned option.
- Publish a short policy. What's allowed, what data may never go into an outside tool, and who to ask.
- Set a fast approval lane. Name an owner and a turnaround time.
- Lock down access. Role-based access and multifactor authentication on every AI model, agent, and data source.
- Require sources and logs. Any approved AI should show where an answer came from and keep a reviewable record.
- Put governance and security in one room. One shared inventory, one incident playbook.
- Recheck every quarter. New tools appear faster than annual reviews.
Frequently asked questions
What is shadow AI? Shadow AI is employees using AI tools their organization hasn't approved, such as personal chatbot accounts or unvetted plug-ins. IBM defines it as cases "where workers use unapproved AI."
How much does shadow AI cost? In IBM's 2026 study of 602 breached organizations, incidents involving shadow AI cost an average of USD 5.39 million, up from USD 4.63 million the year before. The global average for all breaches in the study was USD 4.99 million.
How common is shadow AI? Microsoft and LinkedIn's 2024 Work Trend Index found that 78% of AI users bring their own AI tools to work. MIT NANDA found workers at over 90% of surveyed companies used personal AI tools for work, while only 40% of companies had bought an official LLM subscription.
Should we ban public AI tools at work? You can restrict them, but a ban alone rarely meets the need that drives their use. Pair limits with a sanctioned tool that is useful, governed, and easy to get.
What are AI access controls? They limit who and what can reach an AI model, agent, or data source. IBM names role-based access and multifactor authentication, and found 92% of organizations with an AI-related breach lacked proper AI access controls.
How do you govern employee AI use? Inventory what's in use, publish a short policy, approve tools quickly, control access, keep audit trails, and give people a sanctioned option good enough to choose.
Next step
If you want employees on a governed path that's actually worth using, book a demo at nvitis.com/contact. We'll show you how access rules, sources, and audit trails work on your own systems.
Shadow AI tells you what your people need. Governance decides whether they get it safely.
Sources
- IBM and Ponemon Institute. Cost of a Data Breach Report 2026. July 2026. Report PDF: https://assets.ibm.com/adobe/assets/urn%3Aaaid%3Aaem%3A21111142-1251-4369-86fb-57b82f5bb108/original/as/Cost%20of%20a%20Data%20Breach%20Report%202026.pdf ; press release: https://newsroom.ibm.com/2026-07-29-ibm-study-one-in-four-malicious-breaches-are-ai-enabled,-costing-companies-6-million-on-average
- IBM. "Is rising AI adoption across the US workforce creating shadow AI risks?" Nov 3, 2025 (IBM-sponsored Censuswide survey). https://www.ibm.com/think/insights/rising-ai-adoption-creating-shadow-risks
- Microsoft and LinkedIn. "AI at Work Is Here. Now Comes the Hard Part." 2024 Work Trend Index, May 8, 2024. https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part
- Challapally, A., et al. The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025. Copy used: https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
- Chandrasekaran, A. "Why 50% of GenAI Projects Fail — And How to Beat the Odds." Gartner, Jan 26, 2026. https://www.gartner.com/en/articles/genai-project-failure