What it is, where it hides, how to find it, and how to govern it without a ban.
Shadow AI is any AI tool, feature or agent used at work without the organization's formal review, approval or control. It can be a chatbot on a personal account, an AI feature switched on inside software the company already uses, a browser extension, or an AI agent acting with an employee's access. The tools are often useful. The risk is that the organization may not know what data went into them, where it is kept, who can see the output, or what the AI was allowed to do.
A survey of office professionals at large companies and one of cybersecurity leaders both find it common. Neither is a census of mid-size organizations, so treat the figures as a signal.
Two-thirds
In a 2026 PagerDuty survey of 1,250 office professionals at companies with at least $500 million in annual revenue, two-thirds of those who had used AI for work said they had used AI tools at work even though they believed company policy did not allow it. In the same survey, 34% said they had entered customer data or information into public AI tools such as ChatGPT, Claude or Gemini.
69%
In a Gartner survey of 302 cybersecurity leaders conducted from March to May 2025, 69% said their organizations suspect or have evidence that employees are using prohibited public generative AI. Gartner predicts that by 2030 more than 40% of enterprises will experience security or compliance incidents linked to unauthorized shadow AI.
The number that matters is the one your own logs and people give you.
Shadow AI is more than a chatbot website, so blocking one is not enough.
Free and personal accounts of ChatGPT, Gemini, Claude and similar tools, on a laptop or a phone. OpenAI and Anthropic each say that, by default, they do not use inputs or outputs from their business products to train their models. Microsoft says prompts and responses in Copilot, as used by organizations, are not used to train foundation models. Google says Workspace customer data is not used to train models without the customer's permission. OpenAI, Google and Anthropic each say their consumer apps can use conversations to improve their models, depending on the user's settings. A personal plan can be a consumer plan even when it is paid (OpenAI's Plus and Pro and Anthropic's Pro and Max are examples).
AI features can be added to email, document, CRM and help desk platforms through routine updates, after the contract was reviewed. Take Microsoft Copilot: Microsoft says it respects your identity model and permissions, so it can surface whatever a user already has permission to open, and loose old sharing settings suddenly matter.
Depending on the permissions they are granted, writing helpers, summarizers and "chat with this page" add-ons can read what is on screen, including the web apps a user is signed in to. InfoSight's blog has covered a Chrome extension caught harvesting AI chats.
AI note-takers join calls, record them and send out transcripts, sometimes because one attendee linked a calendar. Depending on the tool and its settings, everyone on the call may be recorded and the transcript stored with the vendor. Check recording notice and consent, and where transcripts are kept, before approving one. More on the cyber risks of AI note-takers.
An unapproved coding assistant can send proprietary source code and configuration files to an outside service, along with any passwords or API keys left inside them. Check what the assistant sends to its provider, and on what contract terms, before approving one.
Tools that act, not just answer. An agent connected to email, files or a CRM acts through whatever permissions and credentials it was given, sometimes an app permission or API key nobody inventoried. IBM's 2026 breach research notes that AI agents create new non-human identities that often operate with elevated privileges and can be hard to inventory, monitor and govern. It also found that fewer than half of organizations are actively securing non-human identities. InfoSight's advice is to treat AI agents like identities, not employees.
Start with records you already hold. They can give a first picture before any new tool is bought, though gaps will remain.
| Where to look | What it shows | Who usually holds it |
|---|---|---|
| Network, DNS and web proxy logs | Which AI services are being reached, from which devices, and how often | Network or security team |
| Identity provider app consents | AI tools people signed in to with a work account or allowed to read mail, files and calendars (Microsoft Entra ID, Google Workspace) | Identity or Microsoft 365 and Google Workspace admins |
| SaaS admin consoles | Which AI features are switched on inside the platforms you already license, and for whom | The owner of each platform |
| Endpoint and browser inventory | Installed AI apps, browser extensions and coding assistants | Endpoint security admins |
| Expense and card records | Paid AI subscriptions bought on corporate cards or claimed as expenses | Finance or accounts payable |
| Cloud and code accounts | API keys for AI services, and AI tools connected to code repositories | Cloud and development leads |
Logs may not capture personal phones and home networks, and may not show what data went in. A short survey helps close the gap: ask which AI tools people use, for which tasks and with what kind of information, under a no-blame commitment that HR and legal have agreed in advance. The answers become the first draft of your AI inventory and the list of tools to review. Before anyone reads the contents of employee prompts, agree with legal, HR and privacy on notice, purpose and who may look.
The article on how to detect shadow AI on your network takes each of these records in turn: what it can show, what it misses and where to start.
Blocking every AI website feels decisive, but a ban removes the approved route, not the need. The surveys above show the limit of a rule on its own: in Gartner's, 69% of cybersecurity leaders said their organizations suspect or have evidence that employees are using prohibited public generative AI, and in PagerDuty's, people used AI tools they believed policy did not allow. A ban can also push use toward personal phones and accounts, where the organization has less visibility and no business contract terms of its own, and it can lose sight of the use.
Governing AI makes the opposite trade: a safe, approved route for what people already do, and firm lines only where the data or the decision calls for them. Gartner's advice to CIOs points the same way: set clear enterprise-wide policies for AI tool use, audit regularly for shadow AI activity, and build generative AI risk checks into how they assess SaaS tools. In practice:
Artificial intelligence governance, usually shortened to AI governance, is the set of decisions, rules and checks that settle which AI the organization uses, for what, with which data, and who answers for it. IBM's 2025 Cost of a Data Breach Report found that 63% of breached organizations either had no AI governance policy or were still developing one.
Each building block below needs a named owner, not a new department.
Keep a living list of every AI tool, feature and agent in use: what it does, who owns it, what data it touches, which accounts or keys it runs under, and whether it is approved. The NIST AI RMF lists this among its governance outcomes: "Mechanisms are in place to inventory AI systems." Then give each use a tier, so the rules scale with the risk.
| Tier | Typical use | The rule |
|---|---|---|
| Low | Drafting and brainstorming with public or non-sensitive information | Any approved tool, no extra review |
| Moderate | Internal documents and business data that is neither regulated nor confidential | Approved business accounts only, with the data terms reviewed |
| High | Patient, customer financial, personnel or engineering data, trade secrets and other confidential material, or output that feeds decisions about people | A named owner, written review, contract terms, logging and human review of output |
| Not allowed | Uses the policy rules out, such as regulated data in a personal account | Blocked, with a route to request an exception |
Teach the policy with examples from each team's work, and give it a place alongside the security awareness training people already take. Keep the detection sources running, review the inventory on a schedule, and revisit the policy when a major tool or rule changes.
This plan is sized for a small IT team working alongside the day job; a larger or more regulated organization will need longer. The aim by day 90: know what AI is in use, have an approved route for the common cases, and have a policy people have read.
The basics are the same everywhere; the data and the rules change.
The worry is patient information in a prompt, such as a discharge summary pasted in for a plain-language rewrite. Under the HIPAA Privacy Rule, a hospital may let an outside vendor create, receive, maintain or transmit protected health information on its behalf only after it obtains satisfactory assurance, documented in a written contract (a business associate agreement), that the vendor will safeguard the information.
So when a staff member pastes patient information into an AI tool the organization has no business associate agreement with, the rule reaches the tool because of the data, not because the tool is AI. HHS's own guidance on business associates lists a third-party vendor's chatbot on a provider's patient portal, one that handles patients' health information, as an example of a business associate.
The federal banking agencies' Interagency Guidelines Establishing Information Security Standards (the FDIC and the Federal Reserve publish the same guidelines for the banks they supervise) tell each bank to exercise due diligence in selecting its service providers, require them by contract to protect customer information, and monitor them where its risk assessment indicates. NCUA's guidelines direct the same oversight for federally insured credit unions, and NCUA says it has no AI-specific rules but that its existing regulations are technology-neutral and apply to AI. An AI tool that receives customer or member information is a service provider question.
Non-bank lenders and credit unions without federal deposit insurance, where the FTC is their regulator, answer to the FTC's Safeguards Rule, which also requires them to oversee their service providers.
The exposure is designs, process recipes, supplier pricing and engineering data. Because trade secret protection depends in part on reasonable measures to keep information secret, make the rule plain: engineering data goes only into tools approved for it.
Plants add a further concern. AI connected to plant systems needs the same caution as any other change near operations; InfoSight's blog explains why AI in OT is a trust and safety problem first.
Many states now set rules for how their own agencies and employees may use generative AI. Washington, for example, bars agencies from entering non-public data into unlicensed AI tools without a contract.
Public records are the other concern: North Carolina's guidance warns that information entered into a publicly available AI tool is considered released to the public and may be subject to public records requests. City and county IT teams can borrow from both.
An external AI exposure assessment applies the discovery and governance work described above, using an outside team. InfoSight's AI Governance & Exposure Assessment is built for "Any organization handling sensitive data in a regulated environment." InfoSight lists "Where is AI actually being used in our organization?" among the questions it hears in nearly every first conversation. Its published scope:
What you get, as the page lists it: an Executive Summary, a Technical Findings Report, a Mitigator® Risk Dashboard, an AI Governance Roadmap, Quick Win Recommendations and a Readout Meeting. For AI your organization deploys or builds, AI penetration testing tests how an AI system behaves under adversarial pressure: whether it can be persuaded to act against its intended purpose. InfoSight also runs employee security awareness training, which covers phishing, social engineering and AI-driven deepfakes.
It depends on the account and the approval. ChatGPT used through a business account your organization approved and manages, for uses its policy allows, is sanctioned AI. A personal account used for work without the organization's approval, or company data pasted into an AI tool the organization has not formally reviewed or approved, is shadow AI.
Start with records you already hold: network, DNS and proxy logs, app consents in your identity provider, SaaS admin settings, endpoint and browser-extension inventories, and expense records. Then run a short, no-blame survey to catch what logs miss, such as AI used on personal phones.
A blanket ban removes the approved route but not the need, so use may move to personal phones and accounts you cannot see. Approve a small set of tools on business terms, block the services that fail review, and keep firm lines for regulated and confidential data.
An AI acceptable use policy should cover its scope, the approved tools and how to request new ones, the data that may never go into an AI tool, account rules, when a person must review AI output, who may connect AI tools or agents to company systems, disclosure and reporting, and who owns the policy.
No single law makes shadow AI illegal. Whether a particular use breaks a rule depends on the data involved, what the AI is used to decide, and the laws, contracts and policies that apply. When regulated information such as patient records or customer financial information goes into an AI tool, the rules that protect that information still apply. A few states and cities also regulate AI used in decisions about people, such as hiring. This is general information, not legal advice.
AI governance needs one executive who answers for it, such as the CIO, the CISO or the head of risk or compliance, with a working group from IT, security, legal, HR and the business. Each approved tool also needs a named owner.
AI governance is the set of decisions, rules and checks that settle which AI an organization uses, for what, with which data, and who answers for it. The NIST AI Risk Management Framework and ISO/IEC 42001 give that work a structure.
No audit language. No commitment.
Prefer the phone? Call 305.828.1003
Schedule a consultation from one of our expert advisors.
No spam—one security expert follow‑up, that’s it.
Stay informed of the latest cyber trends.