Resource
What Is Shadow AI? Examples, Risks, and How to Detect It
Shadow AI is any AI tool employees use for work that hasn't been reviewed or approved by whoever manages the business's IT. Most businesses have more of it than leadership expects, and the risk comes from what goes into those tools: client files, financial records, and internal documents.
Dawn Sizer · CEO and Co-Founder, 3rd Element Consulting · What is shadow AI?
Shadow AI is the use of artificial intelligence tools for work without the knowledge, review, or approval of the people responsible for the business's IT. It is one part of shadow IT, the broader category of unapproved software, cloud services, and devices.
It almost always starts with good intentions. Someone finds a tool that drafts a proposal in two minutes, summarizes a long contract, or writes up meeting notes automatically. It helps, nobody said not to use it, and within a few months half the team is using something similar.
What separates shadow AI from other shadow IT is what people put into it. An unapproved file-sharing app holds the files someone saves there. An AI tool receives whatever someone pastes, uploads, or connects to it. Depending on the tool and the account type, the provider may keep that content, use it to improve its models, or pass it to other services, under terms nobody at the business has read.
Common examples of shadow AI.
Free AI chat tools on personal accounts
ChatGPT, Claude, Gemini, and similar tools used through personal logins rather than a business account. The business has no control over retention or training settings, and no access to the history when the person leaves.
AI features inside software the business already pays for
Writing assistants, summarizers, and AI companions that vendors add to existing platforms, often switched on by default. Nobody adopted a new tool, but how the data is handled may have changed.
AI meeting note-takers
Bots that join calls, record them, and send transcripts and summaries, sometimes to every attendee, including people outside the company. Many connect to a person's calendar and join meetings on their own.
Browser extensions with AI built in
Grammar, summarization, and "chat with this page" extensions that can read whatever is on screen, including email and business web apps.
AI apps connected to work accounts
Tools someone granted access to their mailbox, files, or calendar by signing in with their work account and accepting the permission prompt. That access keeps working until someone revokes it.
AI apps on personal phones
Mobile apps used to photograph documents, transcribe conversations, or draft replies, on devices the business doesn't manage.
Shadow AI vs. shadow IT.
Shadow AI is a category of shadow IT, so the same basics apply: find what's running, decide what's acceptable, and make sure access gets closed when people leave. Three things make AI different enough to treat separately.
The data goes in, not just through. People hand AI tools the actual content of their work, often in full, to get a useful answer.
It shows up without anyone installing anything. Vendors add AI features to existing platforms through routine updates, so the software list can stay the same while the data handling changes.
It spreads faster. AI tools are free or cheap, easy to start, and immediately useful, so adoption outpaces any approval process that wasn't built with them in mind.
For the broader picture, including personal cloud storage, unapproved SaaS, messaging apps, and personal devices, see our guide to shadow IT: examples, risks, and how to manage it.
The risks of shadow AI.
Shadow AI is a security and compliance risk because the business can't protect or account for data it doesn't know is moving. The specific risks tend to fall into a few categories:
Client and company data leaving the business. Content pasted into consumer AI tools can be retained by the provider and, under some account types and terms, used to train future models. Once it's been submitted, there's usually no practical way to pull it back or prove where it went.
Confidentiality and regulatory exposure. Firms with specific obligations around client data carry more risk here. We cover the details for law firms, healthcare practices, and accounting and financial firms separately.
Access that outlasts the employee. Personal AI accounts don't close when someone leaves, and neither does access granted to connected AI apps. Chat histories containing client information stay in accounts the business can't reach, and an app with permission to read a mailbox keeps that permission until someone revokes it.
Sensitive conversations recorded by third parties. AI note-takers can capture HR discussions, legal strategy, and client calls, then store the recordings and transcripts with a vendor nobody vetted.
Unchecked output in client work. AI tools can produce confident, wrong answers. When use is hidden, nobody is reviewing what goes out the door.
Questions you can't answer. Client security questionnaires and cyber insurance applications increasingly ask which AI tools are in use and how data is handled. Without an inventory, there's no accurate answer to give.
How to detect shadow AI.
No single tool finds all of it. These are the places worth checking, starting with what most Microsoft 365 businesses already have:
Connected apps in Entra ID. Review enterprise applications and user consent grants for AI tools people have connected to their work account, such as note-takers, chat tools, and writing assistants. Then tighten user consent settings so any new app that asks for access to mail, files, or calendars needs admin approval.
Sign-in logs. Entra ID sign-in logs show which apps people are signing into with company credentials, including AI tools that offer a "Sign in with Microsoft" option.
Meetings. Look for note-taker bots in Teams meeting participant lists and recordings, and review which third-party apps are allowed in Teams.
AI settings in software you already pay for. Go through the admin settings of the platforms the business already uses, such as CRM, accounting, document, and video meeting platforms. Check which AI features are on and what their data terms say.
Browser extensions. On managed devices, browser policies can inventory extensions and limit them to an approved list. Personal devices that aren't enrolled are the blind spot.
DNS and firewall logs. DNS filtering and firewall logs show which devices are reaching AI services, including ones nobody approved.
Expense reports and card statements. Individual AI subscriptions on company cards or expense reports are one of the most reliable signs of tools nobody reviewed.
Ask people. A short survey asking which AI tools staff use for work usually turns up more than the logs do. Make it clear up front that there is no penalty for anything people report.
Microsoft's dedicated discovery tool, Defender for Cloud Apps, can identify AI apps in use across an organization, but it isn't included in Microsoft 365 Business Premium, which most small and mid-size businesses run, so using it usually means added licensing.
How to manage shadow AI without banning AI.
Banning AI outright tends to push use onto personal phones and accounts, where the business has even less visibility. These steps work better:
Inventory first. Use the detection steps above to build a list of what's running, whose accounts it's on, and what kinds of data are going into each tool. The data matters more than the tool count. An AI tool that drafts internal meeting agendas is a different conversation from one processing client files.
Give people an approved option. Most shadow AI exists because there's no sanctioned alternative. If the business runs Microsoft 365, Microsoft 365 Copilot Chat used with a work account comes with Microsoft's enterprise data protection terms, which is a different arrangement from a consumer chat tool on a personal login. Whatever you choose, it has to be useful enough that people don't go back to the free one.
Put the rules in writing. An AI acceptable use policy and a short approved tools list answer most questions before anyone goes looking for a workaround.
Control app access. Require admin approval before users can connect new apps to their mailbox, files, or calendar, and limit browser extensions on managed devices to an approved list.
Protect the sensitive data itself. Sensitivity labels and data loss prevention policies can flag or block client and regulated data before it goes somewhere it shouldn't. This matters more once AI tools are in the environment.
Make approval fast. A request process that takes weeks guarantees workarounds. Evaluate new tools against a short set of criteria, like the ones in our AI vendor vetting checklist, and give a clear answer quickly.
Include AI in offboarding. When someone leaves, revoke the app consents tied to their account and review any AI tools they used for work, so client information doesn't stay in an account the business can't reach.
Review on a schedule. AI vendors add features and change their data terms often. Compare the approved list against what's actually running at least quarterly.
What a shadow AI policy should include.
A shadow AI policy should be short enough that people actually read it. Most of it fits on one page:
- Which AI tools are approved, and which account to use them with.
- What can never go into an AI tool unless that tool has been approved for it, such as client data, regulated data, financial records, and credentials.
- How to request a new AI tool, who reviews it, and how quickly they respond.
- When AI meeting note-takers are allowed, and how attendees are told a meeting is being recorded.
- Whether personal AI accounts and personal devices can be used for work, and under what conditions.
- Who reviews AI-generated work before it goes to a client.
- What happens at offboarding, including revoking connected apps and reviewing AI tools the person used.
Our Shadow IT and Shadow AI Policy Template covers these points, with a new technology request form and a one-page quick reference for staff. It works alongside a written AI acceptable use policy, which sets the day-to-day rules for approved tools.
Template
Shadow IT and Shadow AI Policy Template
Where to start.
If someone asked you today which AI tools are running across the business, what data has gone into them, and whose accounts they're on, could you answer with confidence?
If not, an IT Environment Review is a good first step. It's free and takes about 30 minutes. We ask a set list of questions about your environment, including how AI tools are being used, answer yours, and send a written summary afterward. If governance work makes sense from there, our AI readiness and governance work covers the inventory, the policy, and the controls.
Common questions
Questions leadership usually asks first.
What is shadow AI?
What is an example of shadow AI?
Is shadow AI the same as shadow IT?
Why is shadow AI a security risk?
How do you detect shadow AI?
Should we ban AI tools to stop shadow AI?
Can you review our current IT provider?
What is an IT Environment Review?
Continue reading
Related work and reading.
What Is Shadow IT? Examples, Risks, and How to Manage It
The broader category shadow AI belongs to, including personal cloud storage, unapproved SaaS, and offboarding.
Read more: What Is Shadow IT? Examples, Risks, and How to Manage ItWhat Should Be in an AI Acceptable Use Policy?
The written rules for which AI tools staff can use and what data can go into them.
Read more: What Should Be in an AI Acceptable Use Policy?AI Readiness & Governance
The inventory, policy, and controls that turn unreviewed AI use into a governed set of tools.
Read more: AI Readiness & GovernanceNext step
Schedule an IT Environment Review.
Talk through which AI tools are in use across your business, what data may be going into them, and what governance would make sense before it becomes a problem.