This article is part of our AI Agents for Business hub. Want help applying it? Talk to eSEOspace.
Shadow AI is when employees use AI tools without approval or oversight from the company. It becomes more serious when those tools are agents connected to company email, files or accounts.
You may already have it.
This guide is part of our AI agents for business hub. It was checked against primary sources on 9 November 2026. It is general information, not legal advice.
Key Takeaways
- Shadow AI means AI use without formal approval or oversight from your IT function.
- Census data shows about 56% of US workers reported using AI on the job for at least one of 11 tasks.
- Gallup found one in five US employees are unsure whether their employer has integrated AI tools.
- A connected agent can reach live systems, which raises the stakes over a chat window.
- IBM says it might not be feasible to eliminate all shadow AI, so aim to manage it.
What shadow AI is
IBM, which sells AI and security products, describes shadow AI as unsanctioned AI use by employees or end users, without formal approval or oversight. It separates this from shadow IT, which covers unapproved software and hardware generally. IBM says: "Shadow AI zeros in on AI-specific tools, platforms and use cases."
There is a range of severity. Pasting a paragraph into a free chatbot is one thing. Giving an agent ongoing access to your inbox or CRM is another.
How common is AI use at work?
These numbers measure AI use in general. They do not measure unapproved use, and they do not measure agents.
| Source | What it found | Period |
|---|---|---|
| US Census Bureau | "About 56% of U.S. workers said they have used Artificial Intelligence (AI) on the job for at least one of 11 tasks." | March 2026 survey wave |
| Gallup | "Forty-seven percent of U.S. employees now say their organization has integrated AI tools to improve productivity, efficiency or quality, up from 41% in the last quarter." | May 6-20, 2026, 22,573 employed adults |
| Gallup, same study | One in five employees remain unsure whether their organization has integrated AI tools, a share Gallup says is unchanged from prior quarters. | Same |
The two surveys ask different questions, so do not combine them into one figure. We did not find a primary source that measures how much of this use is unapproved.
Still, the pattern is worth noting. Workers report using AI, and one in five cannot say whether the company has adopted it. Neither figure shows how much of that use is unapproved, but it is a good reason to ask.
Why agents raise the stakes
A chat window holds what someone pasted into it. A connected agent can read and act in other systems, using whatever access it was given.
Three things change.
- Standing access. A staff member can approve a connection once, and the agent can keep that access until someone removes it.
- Credentials. OWASP's agentic risk list includes ASI03, Identity and Privilege Abuse. The same roundup says: "Exposed credentials and excessive permissions expanded the damage from individual workloads into connected systems."
- Outside instructions. An agent that reads email or web pages can be misled by what it reads. OWASP's AI Agent Security Cheat Sheet says: "Treat all external data as untrusted (user messages, retrieved documents, API responses, emails)." See our guide on prompt injection for business owners.
That roundup appears on OWASP's GenAI site, and says it is not exhaustive and is not an official OWASP publication.
IBM also lists data breaches, compliance problems and reputational damage. It says: "Using shadow AI can lead to compliance issues, especially regarding data protection and privacy." Sensitive data in a tool you never vetted is the concern here.
How to find out what is in use
Start by asking, and make it safe to answer. If people expect punishment, you will only hear about the tools they are willing to admit to.
- Ask each team what they use and why. Keep it short, and say the goal is to learn.
- Review third-party app connections. Check the admin settings in your email, file storage and CRM for connected apps.
- Check expenses. Look for small subscriptions to AI tools on cards and reimbursements.
- Look at sign-ins and network logs. IBM recommends monitoring through network tools, access controls and audits.
- Write it down. List the tool, the owner, the data it touches and what it can do.
What to do once you know
Rank each tool by what it can touch, then act on the highest risks first.
| What the tool does | Suggested response |
|---|---|
| Drafts text using no company data | Allow, with basic guidance |
| Handles internal or customer data | Check the vendor's data terms, then approve or replace |
| Connects to systems and can take action | Review its permissions now, and remove access if nobody owns it |
| Holds shared passwords or API keys | Rotate those credentials and move to individual access |
Then set up the basics. IBM recommends collaboration across IT, security and business teams, a flexible governance framework, guardrails and regular communication. A practical version for a small business could be short.
- A list of approved tools, and how to request a new one.
- A plain rule about which data must never go into an AI tool.
- A named owner for each connected agent.
- Minimum access for every agent. OWASP says: "Grant agents the minimum tools required for their specific task."
- A simple way to report a mistake with no penalty.
Our acceptable use policy guide is a good starting point, and our staff training playbook covers rollout. Have your attorney review any policy before you adopt it.
Why a ban alone has limits
If AI helps people work, a ban may simply push it onto personal devices, where you see even less. IBM is candid about the limit: "It might not be feasible to eliminate all instances of shadow AI."
A stronger move is to give people an approved option that does the job. That gives them a reason to stop using the unapproved one.
Next step
Run the five-step inventory above this month, starting with connected apps. If you decide to replace unmanaged tools with a purpose-built agent, see our AI agent development service.
Frequently asked questions
Is shadow AI illegal?
We cannot give a blanket answer. IBM says shadow AI "can lead to compliance issues, especially regarding data protection and privacy." The risk depends on the data involved and the rules that apply to your business. Ask your attorney about your situation.
Should I fire someone for using an unapproved AI tool?
That is a decision for you and your HR or legal advisers. Some people use these tools simply to do their jobs better. A safe way to report use will tell you more than punishment will.
How is shadow AI different from shadow IT?
Shadow IT covers any unapproved software or hardware. IBM describes shadow AI as focused on AI tools, with distinct concerns about data, model outputs and decision-making.
Can I block all AI tools at the firewall?
IBM lists firewalls that block unauthorized platforms as one guardrail. Blocking has limits, since staff can use personal phones. Combine it with an approved alternative.
What should I check first?
Start with anything connected to email, files, your CRM or payments. Those connections can act on live data. A tool that only drafts text can wait.
