Shadow AI is any AI tool or AI feature used for work that nobody approved or can see.
You can find it with records the business already holds, before buying a detection tool: licence and expense lists, single sign-on and browser logs, admin centre reports and one anonymous question to staff. In Deloitte's May to June 2026 survey of 25,000 UK workers, 31% of generative AI users said they use it without their employer knowing.
This guide is for the COO, head of operations or head of L&D who has heard the term from IT or a board member and now has to explain it, find it and decide what to do first. The pages that rank are written by security vendors for security teams. This one is for the person who owns the staff and the budget, and it needs no security background. Often it is not sabotage. It is someone with a deadline and a free account.
Short version: the ivee app is a free Mac app that sharpens every prompt you send to ChatGPT, Claude, Gemini or Copilot, in one keystroke, so the tools you approve get a better brief every time. Download the ivee app and try it while you read. It is free. 82% of the suggestions the ivee app has made were accepted, from 4 September to 12 November 2026. On Windows? Join the waitlist.
Why does shadow AI matter if you do not work in security?
Shadow AI matters outside the security team because it is a decision about company information that nobody made. A salary spreadsheet goes into a prompt, a client email gets tidied by a chatbot, a draft contract is summarised, and no one who could have said yes or no was asked.
It also leaves almost no trace. In the same Deloitte survey, 46% of generative AI users were using free-to-use tools at work and 17% paid for their own AI tools. A free personal account puts no line on an expense report, and a paid one appears only if someone claims it back.
The aim is simple to state. In ivee's August 2026 AI decision maker survey of 500 UK AI decision makers, one respondent put it as: "I want to wean staff off other AI uses." That is the whole job: find what people use now, then give them something approved that they prefer.
What counts as shadow AI?
Shadow AI covers four things: personal accounts on AI tools, AI features switched on inside software the company already bought, browser extensions with AI built in, and automations someone built on the side. The test is the same for each. Did anyone responsible for the business know it was happening?
Personal accounts. A free ChatGPT, Gemini or Claude login used to rewrite a client email or summarise a document.
AI inside software you already bought. A note-taker added to video calls, or an AI summary button in a CRM. The licence was approved. The AI feature was never reviewed.
Browser extensions. An AI writing assistant installed in Chrome, which can see the text of the page it is helping with.
Side automations. A spreadsheet connected to ChatGPT or Claude through Zapier or n8n by someone who did not ask first.
The last one is where the cost tends to hide, because an automation keeps running long after its builder stops thinking about it. In ivee's research with 500 UK AI decision makers, surveyed in August 2026, one wrote: "Someone has requested access to an automation tool and I haven't allowed it yet, as I do not know the wider risks to the business." That manager is the lucky one, because the request reached them. Shadow AI is the same request from someone who did not make it.
What is not shadow AI?
An approved AI tool used badly is not shadow AI. It is a training problem, and it needs a different fix. A person on a company Copilot seat who writes vague prompts, or who pastes data into an approved tool that your rules say should stay out, is working inside the system. Nobody hid anything and nobody went round an approval. The gap is a skill, or a rule that nobody explained.
The distinction matters because the fixes differ. Shadow AI is solved by making the approved route easier to find, faster or better than the alternative. A skills gap is solved by practice on real work. Mix them up and you either run a discovery exercise on staff who did nothing wrong, or you book a training course while the unapproved tool carries on.
For the skills gap, the quickest fix is practice on real work, which is what the ivee app is built for: it sits inside Copilot, Claude and ChatGPT and coaches people as they go.
How is shadow AI different from shadow IT?
Shadow IT is software or cloud services used without IT's approval, and shadow AI is the same idea applied to AI tools. IBM defines shadow AI as the unsanctioned use of any AI tool or application by employees or end users, without the approval or oversight of the IT department.
Three differences matter to a manager. First, the data moves. Shadow IT usually means a file stored in the wrong place, while shadow AI means text sent into a prompt, and a prompt cannot be unsent. Second, the trail is thin, because free tools leave nothing in the accounts. Third, it often arrives inside tools you already approved, as a feature switched on rather than a product bought. A shadow IT audit will therefore miss part of it.
How do you find shadow AI without buying a detection tool?
Start with five records the business already holds, in this order, and finish with one question to staff.
Expense claims and card statements. Search for AI vendor names such as OpenAI, Anthropic, Perplexity and Midjourney. Paid personal plans appear here when someone claims them back.
Your software licence list. For each tool already bought, open the admin settings and note which AI features are switched on.
Single sign-on logs. In Microsoft Entra, Okta or Google Workspace, look at the third-party apps people have connected with their work email.
Browser records. Where browsers are centrally managed, export the extension list and look for AI assistants.
Cloud app discovery. If you already license Microsoft Defender for Cloud Apps, its app catalogue includes a Generative AI category, and Cloud Discovery can show which of those apps people use.
Then ask one question, anonymously: which AI tools have you used for work in the last month, including on a personal account? Keep it to that one question, because a long survey gets abandoned.
Expect the records to disagree. Each one sees a different slice, so the useful output is a merged list of tool, who uses it and what for. That list is what the first week is built on.
What should you do in the first week?
In the first week, give it one owner, merge what you found into a list, ask staff without blame, decide what happens to each tool, and make the approved route the easy one.
Day one: one owner. Someone in operations or people, with IT as a partner.
Days one to three: run the five checks. Merge the results into the single list.
Day three: ask the question, and promise no blame. Say in writing that nobody is disciplined for answering. A blame-led amnesty gets less disclosure than a no-blame one, for the plain reason that people do not own up to something that can get them in trouble.
Day four: decide each tool. Approve it, replace it with an approved equivalent, or stop it. What data may go in is covered in sorting company data into classes for ChatGPT, and the register that records your answers is covered in how to build an approved AI tools list.
Day five: make the approved route the easy one. People reach for unsanctioned tools when the sanctioned one is slower or worse. The full argument is in why an AI policy alone will not stop shadow AI.
This is where the ivee app earns its place as the better approved route. It sits inside Copilot, Claude and ChatGPT, rewrites a draft prompt when you press ⇧+⌘+Space and suggests the best model for the task, so staff get a better first answer from the tools the company allows. It does not detect shadow AI or report what your team uses, so it belongs to the fix, not the audit.
When do you need a detection tool, and what can nobody see?
You need a detection tool when the estate is too large to read logs by hand, or when a regulator expects you to show what is in use. Some firms do need one, and the records above will tell you which side you are on. Bear in mind what it covers: Microsoft's guidance says monitoring interactions with third-party AI sites in Microsoft Purview requires devices to be onboarded, so a tool sees managed devices and little else.
Nobody can see a personal phone on a home network. A person who pastes a document into an AI tool there leaves no record in anything the business owns. Detection alone therefore cannot be the answer. The rule says what is not allowed, the approved route gives people somewhere better to go, and a named owner for the four AI governance decisions keeps both current.
Get better at using AI while you work
The ivee app is free for Mac. It sits inside ChatGPT, Claude, Gemini and Copilot, sharpens each prompt you write and suggests the right model, so you get better answers with fewer tokens. 82% of the suggestions the ivee app has made were accepted, from 4 September to 12 November 2026.
Download the ivee app for free →
On Windows? Join the Windows waitlist. On a phone, the download page emails you the link to open on your Mac.



