What is shadow AI, and what should a small team do about it?
Shadow AI is any AI tool used for work without the knowledge or approval of whoever is responsible for the company's tools. On a small team that usually means a personal ChatGPT or Claude account, a free note-taker or a browser extension that someone started using on their own, with client information going in and nobody else able to see it. The fix is rarely a ban. Ask what people use, approve a few tools on work accounts, check their data settings and review the list every three months.
If you're ready to write the rules down, our guide to writing an AI policy for a small team has a one-page template you can copy. This page is about the step before that: finding out what's already happening and deciding what to do about it.
What does shadow AI look like?
| What shadow AI looks like | The risk | What to do |
|---|---|---|
| Client notes pasted into a personal ChatGPT or Claude account | The work and the data sit in an account the company can't see, keep or close when the person leaves | Approve the tool on a work account and move the work there |
| A personal plan with the training setting left on | What people type may be used to improve the provider's models | Check the data settings on every plan people actually use |
| A free AI note-taker that joins client calls | Recordings and transcripts go to a vendor nobody looked at | Pick one note-taker, and tell clients when it's on |
| A browser extension with an AI sidebar | It may be able to read whatever is open in the browser, including email and the CRM | Keep a short list of approved extensions |
| A workflow or agent someone built that can send or edit things | It can act for the business with nobody checking first | Ask for a person to approve anything that goes out |
| AI subscriptions on personal cards and expense reports | Nobody knows the total, or which tools the company is paying for | List every AI subscription and who pays for it |
| Everyone giving AI their own version of the background | The same question gets different answers depending on who asked | Agree on one shared source of company context |
None of these are rare or careless. Most are people trying to get through their week faster.
Shadow AI vs shadow IT: what's the difference?
Shadow AI is a kind of shadow IT, which is any software people use for work without approval. The difference is what happens to the information. A shadow file-sharing app stores what you put in it. An AI tool reads it, may keep it, may learn from it depending on its settings, and writes new work from it that ends up in front of clients.
There's a second difference for small teams. Most guides to shadow AI are written by security companies for businesses with an IT department, so the answer is usually network monitoring or data loss prevention software. A 12-person company doesn't have a security team. The person "responsible for tools" is the founder or an operations lead, and the best way to find shadow AI is to ask.
Take an office manager at a 15-person accounting firm. The firm never bought any AI tools, so on paper it uses none. In practice, two people draft client emails in personal ChatGPT accounts, one uses a free transcription app for meetings, and a new hire installed an AI writing extension. That's shadow AI, and nobody involved thinks they're doing anything wrong.
Why does shadow AI happen?
Shadow AI happens because AI tools are cheap, easy to start and useful, and most companies haven't said which ones to use. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work, and at small and midsize companies it was 80%. The same survey found 52% of people who use AI at work are reluctant to admit using it for their most important tasks.
That second number explains why it stays hidden. People worry that using AI looks like cutting corners, or that it will be taken away. So they keep using it and stop mentioning it.
Small businesses are now building with AI too, not just chatting with it. In Pax8's October 2026 survey of 1,150 businesses with 5 to 499 employees, 35% of those using AI had found tools, workflows or agents running without approval in the last six months. A project manager at a design studio who connects an AI tool to the shared inbox to sort requests is doing something useful. It's still shadow AI if nobody else knows it's there.
What are the risks of shadow AI for a small business?
The main risks are that company information ends up in accounts the company doesn't control, and that nobody can see or check the work AI is doing. For a small team, these five matter most:
- Data in personal accounts. When someone leaves, their chat history, uploaded files and saved projects leave with them. You can't delete what you can't see.
- Training settings on personal plans. OpenAI says data sharing is on by default for personal ChatGPT Free, Plus and Pro accounts, and you can turn off "Improve the model for everyone" in Data controls. It doesn't train on ChatGPT Business or Enterprise data by default. Anthropic asks Claude Free, Pro and Max users to choose whether their chats can be used for training, and says that choice doesn't apply to its Team and Enterprise plans. Either way, the setting belongs to the person, not the company.
- No record of the work. If a proposal was drafted in a personal account, there's no shared copy of the prompt, the background it was given or the earlier versions. When a client asks where a number came from, nobody can check.
- Inconsistent output. Each person gives AI their own version of the company's background, so the same question gets different answers. Our page on why AI gives different answers goes into this.
- Cost spread across cards. Five people each paying for an AI plan on their own card, plus a few usage-based tools on the company card, adds up to a number nobody has seen. In the same Pax8 survey, 49% of small businesses that paid for AI in the past year got at least one bill or usage charge higher than they expected. Our AI cost management guide covers what drives those bills.
The bigger-company numbers point the same way. In IBM's 2025 Cost of a Data Breach report, which studied organizations that had a breach, one in five said the breach involved shadow AI, and those with high levels of shadow AI saw breach costs about $670,000 higher on average than those with little or none. Most of those organizations are much larger than a small team, so read it as a direction, not a forecast for your business.
Picture an account manager at a 10-person agency who leaves for a competitor. For two years she drafted every client recap in her own Claude account. The agency still has the emails she sent, but none of the background, notes or drafts behind them.
Why doesn't banning AI work?
Banning AI usually moves it out of sight rather than stopping it. People who already save hours a week with AI tend to keep using it on their phone or a personal laptop, where you have even less say over the settings. The Work Trend Index finding that half of AI users already hide their use for important work suggests a ban would make that worse.
A ban also sends the wrong message about mistakes. If using AI gets someone in trouble, they won't tell you when they've pasted something they shouldn't have, and that's the moment you most need to know.
There's a real trade-off here. Approving AI tools costs money, since business plans with better data settings cost more per person than the personal plans people already pay for themselves. You also won't approve every tool someone likes, and some people will be annoyed. That's still cheaper than not knowing where client information is, but it's a cost, and it's fine to say so.
A sales lead at a 20-person software company who's told "no AI" will probably keep drafting follow-ups in ChatGPT on her phone. Tell her "use the work account, and keep pricing out of it unless you ask first" and she has a clear path she can follow.
How do you prevent shadow AI on a small team?
You prevent shadow AI by making the approved path easier than the hidden one. For a small team, that's five steps:
- Ask without blame. Post one question in your team chat: "Which AI tools do you use for work, what for, and on which account?" Say up front that nobody is in trouble. Write down every answer, including free tools and browser extensions.
- Approve a few tools on work accounts. Pick the two or three tools people already rely on and set them up on company accounts. Fewer approved tools are easier to keep track of, but they need to cover what people actually do.
- Check the data settings. For every approved tool, check whether what people type can be used for training, how long it's kept, and who in the company can see it. Personal and business plans often have different defaults.
- Give people a better sanctioned path. If the approved tool doesn't have the background people need, they'll go back to pasting it into whatever works. Make the work account the easiest place to get a good answer, with your templates, client background and examples ready.
- Review every three months. AI tools change quickly. Every quarter, ask the same question again, update the list and check any new settings.
When you're ready to turn these answers into rules, use the template in our AI policy for small teams guide. It covers which information is fine, which needs a check and which never goes into an AI tool.
An operations lead at a 25-person property management company could do all of this in an afternoon: one chat message, a short list of three approved tools, a check of each one's data settings, and a calendar reminder for three months from now.
How AgentOS fits
Step 4 is the hard one. People go around approved tools when those tools don't know enough about the business, so they paste in client files and background by hand, often into a personal account.
AgentOS gives your AI tools that context without the pasting. You connect your email, chat, files, calendar and CRM once, and ChatGPT, Claude, Cursor and your own agents connect over MCP and work from the same company knowledge. You choose what each AI tool can see, and anything that goes out, like an email to a client, waits for a person to approve it. You can also describe recurring work in plain language, like a Friday client recap.
It doesn't find shadow AI for you, and it doesn't replace asking your team. It makes the sanctioned path a better one. For the bigger idea, read what a context layer is, or see what AgentOS is.
Questions people ask
What is shadow AI in simple terms?
Shadow AI is AI used for work that the company doesn't know about or hasn't approved, like a personal ChatGPT account used for client emails or a free AI note-taker on sales calls.
What is the difference between shadow AI and shadow IT?
Shadow IT is any unapproved software. Shadow AI is the AI part of it. The difference is that AI tools read what you give them, may keep it or use it for training depending on their settings, and produce new work that goes out to clients.
What are some examples of shadow AI?
Pasting client notes into a personal AI account, using a free AI transcription app on calls, installing an AI browser extension, building an automation that connects AI to the shared inbox, and expensing an AI subscription nobody else knows about.
What are the risks of shadow AI?
Company information ends up in accounts the company doesn't control, personal plans may allow training on what people type, there's no record of how work was produced, answers vary from person to person, and AI costs are spread across cards nobody adds up.
How do you prevent shadow AI on a small team?
Ask what people use without blame, approve a few tools on work accounts, check each tool's data settings, give people a sanctioned path with the background they need, and review the list every three months.
Should a small business ban AI tools?
Usually not. A ban tends to move AI use to personal phones and accounts where you can see even less. Approving a few tools on work accounts, with clear rules about information, works better.
Is all shadow AI bad?
No. Most of it is people trying to do good work faster, and it often shows you which tools are worth approving. The problem is not knowing about it, not the fact that people use AI.
See what your AI tools can do when they all work from the same company knowledge, with you choosing what each one sees.
Try AgentOSSources
- Microsoft and LinkedIn: 2024 Work Trend Index
- Microsoft WorkLab: AI at work is here. Now comes the hard part (2024 Work Trend Index report)
- Pax8: small businesses are building AI faster than they can govern it (October 2026)
- IBM: Cost of a Data Breach Report 2025 (press release, July 2025)
- OpenAI Help Center: Data controls FAQ
- OpenAI: business data privacy
- Anthropic: updates to Consumer Terms and Privacy Policy (August 2025)
AgentOS is built by Devcore. Found something out of date on this page? Tell us at tryagentos.net and we'll fix it.
