
What is shadow AI? Learn why employees use unapproved AI tools, the real risks for your data, and a practical way to manage AI at work without banning it.
On this page
- What is shadow AI?
- What’s the difference between shadow IT and shadow AI?
- How common is shadow AI? Key statistics
- Why do employees use shadow AI?
- Common examples of shadow AI at work
- What are the risks of shadow AI?
- Why banning AI tools doesn’t work
- How to prevent shadow AI in six steps
- 1. Find out which AI tools people already use
- 2. Offer approved tools people will choose
- 3. Write a short AI acceptable use policy
- 4. Set up admin controls before rollout
- 5. Train employees with real examples
- 6. Review every quarter and fast-track new requests
- Meeting AI: The shadow AI category hiding in your calendar
- How Fireflies helps teams bring meeting AI under control
- FAQs
- Should companies ban ChatGPT and other AI tools?
- How do you detect shadow AI?
- What should an AI acceptable use policy include?
- Are AI notetakers a security risk?
- Conclusion
Eight in 10 employees use AI tools their employer hasn't approved, according to UpGuard's 2025 State of Shadow AI report. That unapproved use, commonly called shadow AI, reaches well past chatbots into meeting bots and the AI features built into software the company already licenses.
Most of it starts with convenience. Employees reach for tools they already know when the approved route is missing or slow, and those tools take in client notes and meeting recordings before IT has checked where that data goes.
In this article, we'll walk through what counts as shadow AI and the six steps IT can take to move employees onto approved tools without slowing teams down.
What is shadow AI?
Shadow AI is any use of AI tools or features for work that a company's IT or security team hasn't approved or doesn't know about. That approval covers the account as well as the tool.
An AI tool can be approved, but if an employee signs in with a personal account and pastes client notes into it, that use still counts as shadow AI, because the personal account sits outside the company's admin settings and its contract with the vendor. Microsoft and LinkedIn's 2024 Work Trend Index calls the habit of bringing personal AI tools to work BYOAI (bring your own AI). However, once IT approves that tool and moves the employee onto a company account, the same use is no longer shadow AI.
What’s the difference between shadow IT and shadow AI?
Shadow IT is any software or device employees use for work without IT approval, and shadow AI is the part of shadow IT made up of AI tools. The two differ most in what happens to company data. A cloud storage app keeps the files it's given, while an AI tool sends them to a model, and an AI agent can act on them inside connected apps.
Shadow IT and shadow AI compare on five points.
| Shadow IT | Shadow AI | |
|---|---|---|
| What it is | Software or devices used for work without IT approval | AI apps, features and agents used for work without IT approval |
| Typical examples | Personal cloud storage for work files, a team's project management app | A personal chatbot account, an AI notetaker on a work calendar, an AI agent with inbox access |
| What happens to data | Stored in an account IT doesn't manage | Sent to an AI model, and with agents, acted on inside connected apps |
| Main risk | Company data IT can't secure or remove | Company data kept or used for training under terms nobody reviewed, plus agents acting with an employee's permissions |
| How it spreads | A team signs up with a work email or pays with a company card | An employee signs in, grants access to a calendar or inbox, or turns on an AI feature inside an approved app |
How common is shadow AI? Key statistics
Shadow AI reaches most of the workforce, and the security incidents tied to it are rising. Four recent reports measure it, and each counts something different (what employees admit to, AI app traffic, what security leaders suspect and confirmed breaches).
- Employee use. In UpGuard's 2025 State of Shadow AI report, 8 in 10 employees said they used AI tools their employer hadn't approved, and 70% knew of sensitive data being shared with AI tools at work.
- Personal accounts. The Netskope Cloud and Threat Report 2026 found that 47% of generative AI users reached AI apps through personal accounts, down from 78% a year earlier. The share using accounts their organizations manage climbed from 25% to 62%.
- Security leaders' view. In a 2025 Gartner survey of cybersecurity leaders, 69% of organizations suspected or had evidence that employees were using prohibited public generative AI. Gartner forecasts that by 2030, more than 40% of enterprises will have security or compliance incidents linked to shadow AI.
- Incidents. IBM's 2026 Cost of a Data Breach Report found that 43% of the security incidents studied involved shadow AI, up from 20% a year earlier. Those incidents cost an average of $5.39 million, against a $4.99 million average for all breaches.
Why do employees use shadow AI?
Employees use shadow AI because unapproved AI tools are often the quickest way to get work done. The reasons usually fall into six groups.
- Habit. Employees reach for the AI tools they already use at home, because they already know how those tools work.
- No approved option. When IT hasn't approved any AI tool for a task, such as summarizing a call or drafting a reply, employees use whatever they can find.
- Approved tools that fall short. When the company version lacks a feature people rely on at home, they go back to the version that does the job.
- Slow approval. In a 2026 Merge survey, 28% of employees said IT's approval process was so slow that they didn't bother asking. A tool nobody requests never reaches IT's review queue.
- Unclear rules. Employees who don't know the company's AI policy have to guess which tools and data are allowed.
- New AI features in approved apps. An app IT already approved can get a new AI feature in an update, such as a summary button in a help desk tool. The app passed review before that feature existed, so nobody has checked where the feature sends data, and employees can start using it the same day.
Common examples of shadow AI at work
Shadow AI at work usually looks like an ordinary shortcut, taken with a personal account or a quick permission grant. Common examples include the following.
- Personal chatbot accounts. A sales rep pastes a customer contract into a personal chatbot account to pull out the renewal terms. The contract then sits under the chatbot's consumer terms, with no company agreement behind it.
- Personal AI notetakers. An employee signs up for a notetaker with a personal email and connects a work calendar, and the bot starts joining meetings IT doesn't know about.
- AI browser extensions. Before installing some extensions, Chrome warns that they can "Read and change all your data on all websites." An AI extension with that reach can see webmail, CRM records and internal tools open in the same browser.
- AI summaries switched on by a team admin. A video meeting app the company already licenses adds AI summaries, and a team admin turns them on for every call before IT has looked at them.
- AI agents on personal logins. An operations lead lets an agent draft replies from a work inbox and update deal records in the CRM. In BlackFog's November 2025 survey, 51% of employees said they had connected AI tools to other work systems without IT approval.
- Personal coding assistants. A coding assistant reads the files open in a developer's editor, and those files can hold proprietary code and API keys. In the Netskope report, source code made up 15% of data policy violations involving personal apps.
What are the risks of shadow AI?
Shadow AI puts company data, compliance and money at risk, because IT can't see or control the tools involved. The risks fall into seven groups.
- Data leaks. Prompts, files and recordings leave the company and, depending on the provider's terms, can be kept or used to train future models. In IBM's 2026 Cost of a Data Breach Report, 49% of security incidents involving shadow AI led to data loss or compromise.
- Compliance violations. GDPR requires a contract or similar legal act with any vendor that processes personal data on a company's behalf (Article 28), and HIPAA generally requires a business associate agreement before health data goes to a vendor (45 CFR 164.502(e)). A personal AI account has neither. In about one in five shadow AI incidents in IBM's report, the organization paid a fine. Regulated teams that record client calls also have meeting compliance rules to follow.
- Recording consent gaps. Federal law allows recording when one party consents (18 U.S.C. 2511(2)(d)), but Florida requires prior consent from every party (Fla. Stat. 934.03(2)(d)). An AI notetaker set to join every meeting can record someone in Florida whose consent nobody asked for, and calls with people in Europe raise GDPR call recording questions too.
- No audit trail. Personal accounts sit outside the company's admin console. IT can't pull logs after an incident, put the data on legal hold or delete it when the employee leaves.
- Wrong answers in decisions. In Mata v. Avianca, a federal judge in New York ordered two lawyers and their firm to pay a $5,000 penalty in June 2023 for filing a brief that cited court opinions a generative AI chatbot had invented. Nobody had checked the citations.
- Agents with too much access. OWASP calls this risk "excessive agency," meaning an AI system holds more permissions than its task needs. Its example is an email assistant that only needs to read mail but can also send it, so a malicious incoming message can trick it into forwarding sensitive data.
- Paying twice. Teams that expense their own AI subscriptions can end up paying per seat for tools that overlap with software the company already licenses. Finance may see the tool before IT does, as a line on an expense report.
Why banning AI tools doesn’t work
Banning AI tools tends to push shadow AI out of IT's sight without ending it. Most companies already block something. The Netskope report found that 90% of organizations block at least one generative AI app.
Employees route around those blocks. UpGuard found that 41% of employees find a way around company attempts to block applications, and a tool blocked on a work laptop is still a tap away on a personal phone. A ban also takes away the time those tools save, which gives employees a reason to keep looking for a way in.
The workable alternative is to approve business versions of the tools people already use and keep blocks for tools that can't meet the company's data rules.
How to prevent shadow AI in six steps
The most reliable way to prevent shadow AI is to make the approved route the easiest one for employees and keep it that way as new tools appear. These six steps form a cycle IT can repeat every quarter, starting with an honest count of what's already in use:
1. Find out which AI tools people already use
Shadow AI discovery starts with records IT can already pull. In the Google Workspace Admin console, the Accessed apps list (under Security, then Access and data control, then API controls) shows third-party apps employees have authorized with their work accounts and how many users each one has. Microsoft Entra ID lists the apps employees have granted access to under Enterprise apps, and its user consent settings decide whether employees can grant that access on their own.
Expense reports and corporate card statements catch the AI subscriptions people pay for themselves. A short anonymous survey catches what no log shows, such as a personal account on a personal phone, and people answer it more honestly when the company says up front that answers won’t lead to discipline.
2. Offer approved tools people will choose
Employees move to approved tools when those tools do the job at least as well as the ones they found themselves. The AI tool inventory is the shortlist, and the tools with the most users are the ones to replace first. Often the fix is a business version of the same product, with admin controls, SSO (which lets employees log in with their work account) and a contract that covers how company data is stored and used.
Ask the employees who found the original tool to test the replacement before rollout. They know what the approved version has to match. For a starting list, a roundup of the best AI tools for business groups options by business function.
3. Write a short AI acceptable use policy
An AI acceptable use policy tells employees which AI tools are approved and which kinds of data can go into each one. The simplest version sorts data into tiers. Public material can go into any approved tool, internal documents only into tools under a company contract, and client, health or financial data only into tools approved for that data. Anything covered by an NDA or a legal hold stays out of AI tools entirely.
The policy also says when employees must tell clients or colleagues that AI drafted or recorded something. For meetings, that means announcing any AI notetaker at the start of the call and following the strictest consent rule that applies to anyone on it. Companies that want a fuller AI governance policy can build from NIST's AI Risk Management Framework, released in January 2023, and its Generative AI Profile from July 2024. Both are free and voluntary.
Use this checklist as the starting point for the company's policy.
- Approved AI tools, and the rule that work happens in company accounts only
- Data tiers, with the tools each tier is allowed to go into
- Banned uses, such as pasting passwords, API keys or anything under NDA
- How to request a new tool, and how long a decision takes
- When a person has to check AI output before it reaches a client, a court or a regulator
- When AI use must be disclosed, and to whom
- Recording rules for AI notetakers, including how participants are told at the start of a call
- Who owns the policy and how often it gets reviewed
4. Set up admin controls before rollout
Admin controls are what make an approved tool safer than the personal version of the same product. Check which plan includes each control before buying, then turn them on before anyone gets access.
- SSO. Employees sign in with their work identity. When IT disables a work login, access to the tool goes with it.
- SCIM. An open standard (IETF RFC 7644, September 2015) that lets the identity provider add and remove users in other apps automatically, which closes accounts on an employee’s last day without a manual ticket.
- Data retention. Admins set how long prompts, files and recordings are kept and when they’re deleted.
- Permissions. Roles decide who can view, share or export content. For AI agents, grant read access only where the task needs no more.
- Audit logs. A record of who accessed, exported or shared what, which is the trail IT needs after an incident and the one personal accounts never produce.
5. Train employees with real examples
Shadow AI training works best when it uses the company's own examples and names the approved tool for each one. Only 39% of AI users in Microsoft's 2024 Work Trend Index had received AI training from their company.
Build each session from what the inventory turned up, with names removed. A sales team learns which notetaker is approved for client calls. Engineers learn which coding assistant the company contract covers and which files it can read. For everyday habits, such as writing prompts and checking output, point employees to the guide on how to use AI at work.
6. Review every quarter and fast-track new requests
Shadow AI returns whenever new tools appear faster than IT can approve them. Rerun the AI tool inventory each quarter and compare it with the approved list. The same check is the time to review what each AI agent can access and to cancel seats nobody uses.
A fast path for requests matters as much as the review. Publish how long a decision takes and meet that deadline. Each request should end with IT approving, restricting or blocking the tool, and each of those decisions fits a different kind of tool.
| Decision | When it fits | Example |
|---|---|---|
| Approve | A business version exists with SSO, admin controls and a contract that covers storage and model training | A chatbot plan under the company's contract, open to every team |
| Restrict | The tool is useful but only safe for certain data or teams | A writing extension allowed for public marketing copy and kept away from client records |
| Block | No business version exists, the terms allow training on inputs with no way to opt out, or the tool wants access its task can't justify | An extension that asks to read every website for a job that needs one |
Meeting AI: The shadow AI category hiding in your calendar
Meeting AI is an easy kind of shadow AI to miss, because it enters through a calendar permission and then joins calls on its own. An AI notetaker becomes shadow AI when an employee signs up and connects it to a work calendar before IT has reviewed the tool. The employee can sign up with a personal email or a work email. A work address alone doesn't put the account under company admin settings or a company contract.
Connecting the calendar takes a single permission screen, and from then on the notetaker can see every event with a meeting link. Depending on its join setting, it can enter client calls, job interviews and board meetings as well as internal standups. Participants see the bot in the call, but the recordings and transcripts land in an account IT doesn't manage.
Any AI notetaker can end up in this position. The risk comes from the unmanaged account. The same product inside a company workspace, under admin settings and a company contract, is an approved tool.
Before approving or blocking a meeting AI tool, IT can check five things.
- Who can connect a work calendar. In Google Workspace, setting the Calendar service to Restricted means only apps an admin has marked Trusted, or approved for specific data, can read calendar data. In Microsoft Entra ID, user consent settings decide whether employees can approve apps on their own.
- Which meetings the bot joins. Look for an admin setting that applies to everyone, such as joining only internal meetings or only meetings someone invites it to.
- Where recordings and transcripts live. Ask who in the company can reach them and how long they're kept.
- Whether the vendor trains AI models on meeting content. The answer belongs in the contract, and it should hold on every plan, free tiers included.
- How participants are told. Check what notice goes out before and during the call, and whether anyone on the call can stop the recording or remove the bot.
How Fireflies helps teams bring meeting AI under control
As an AI Teammate, Fireflies gives companies an approved version of the meeting AI employees already use. Its team settings line up with the five checks IT runs on any meeting AI tool.
Admins decide who gets into the workspace and what each person can connect. An Advanced Settings option, off by default, automatically adds users who sign up with the company's email domain, which pulls stray accounts under the same admin settings. On Business and Enterprise, Integration Permissions set who can connect each integration (all members, selected members and user groups, admins only or no one). With SSO and SCIM on Enterprise, access starts and ends with the company's identity provider.
Admins also decide which meetings Fireflies joins. In Team settings, under Recording & Privacy, the Auto-record meetings options range from every calendar event with a meeting link to only meetings someone invites Fireflies to. Admins can apply one choice to the whole team or let teammates pick their own.
On Enterprise, admins also control where meeting data lives and how long it stays. Custom data retention sets when meetings are deleted, and Private Storage keeps meeting data in a storage location the company chooses. Audit logs through the API record activity, and a super admin role can open any meeting in the workspace.
The rule on AI training is the same for every customer, on every plan. Fireflies.ai does not use customer data to train any AI models. Your personal data is never used to train AI models. Users own their data. Fireflies is SOC 2 Type II, GDPR, and HIPAA compliant, and ISO/IEC 42001 certified. HIPAA is available on Enterprise with a signed BAA and Private Storage. SOC 2 Type II and GDPR compliance also apply on every plan, including Free.
Fireflies also lets participants know when it's in a meeting. Team settings can turn on an email that tells attendees an hour before the meeting that Fireflies will record it, with a link to decline. During the call, the AI Notetaker joins as a visible participant and pins a notice in the meeting chat. Participants can remove it from the call, and on Zoom, Fireflies won't join if anyone declines the recording consent prompt.
Take Notes on the Fireflies Desktop App captures a call without a bot. Participants get an audio message instead, plus a camera watermark when the user selects Fireflies as the camera and a chat message on Mac where the meeting platform supports it. All of these notices are on by default, and a user who turns one off is responsible for telling participants.
For the security detail behind these controls, see the Fireflies security page, along with is Fireflies safe and how Fireflies approaches security.
FAQs
Should companies ban ChatGPT and other AI tools?
No, companies shouldn't ban ChatGPT and other AI tools outright. A blanket ban tends to push that use out of sight, because employees can switch to a personal phone or another tool that isn't blocked. A safer route is approving business versions of the tools employees already use and blocking only the ones that fail the company's data rules.
How do you detect shadow AI?
To detect shadow AI, check the records IT already holds and ask employees directly. Google Workspace's Accessed apps list and Microsoft Entra ID's Enterprise apps list show which tools employees have authorized with work accounts. Expense reports reveal paid AI subscriptions, and an anonymous survey with a promise of no discipline surfaces personal accounts on personal devices.
What should an AI acceptable use policy include?
An AI acceptable use policy should name the approved AI tools and which kinds of data can go into each one. It should also cover banned uses, how to request a new tool, when AI use must be disclosed, when a person checks AI output and how people on a call are told that AI is recording it.
Are AI notetakers a security risk?
AI notetakers can be a security risk when employees connect them to a work calendar before IT reviews the tool, even with a work email, because recordings and transcripts then sit outside company admin settings and contracts. Any notetaker can end up this way. An approved tool with admin controls, a contract that rules out training on meeting content and clear notice to participants lowers that risk.
Conclusion
Shadow AI tells IT what employees need. The tools on a shadow AI inventory show which jobs employees needed done and where the approved options fell behind.
Approved tools and plain rules do the rest. Once employees know which tools are approved and which data can go into each one, asking IT becomes faster than working around it, and the next unapproved tool shows up as a request.
For meeting AI, the approved tool has to clear two bars. Employees drift back to their own tools when the approved one falls short, and IT can't approve a tool that fails the five checks for meeting AI. Fireflies, an AI Teammate, clears both. Employees get transcripts, summaries and answers from AskFred across their meetings, while admins in a company workspace decide who connects it, which meetings it joins and where recordings live. Its data terms hold on every plan, Free included.
Give your team the meeting AI they already want, with the controls IT needs.

Lynn Wang is a content strategist and editor with 8+ years covering technology, AI, and fintech. She has run newsroom operations and published 500+ articles, turning complex, technical topics into clear, accurate writing for mainstream readers.


