In most companies, the AI programme on the slide and the AI use on the ground are two different things. The slide shows a pilot, a working group and a vendor review. On the ground, a salesperson is pasting a customer email into a personal chat account to draft a reply, a finance analyst is summarising a contract, and an engineer is debugging with whatever tool was fastest that morning.
That gap has a name: shadow AI. It is not a fringe behaviour. The surveys below show it is the normal state in most companies, and the costs show up when something goes wrong. This article explains what shadow AI is, how common it is, why people do it, what the real risks are, and a practical way to bring it into the open without slowing people down.
Key takeaways
- Shadow AI is widespread because AI tools help people do their jobs, and official options lag behind.
- The main risks are data leaving the company, decisions nobody can trace, and costly breaches.
- Bans tend to push use further out of sight; a fast approved path works better.
- Measure the share of AI use that runs through approved, visible channels, and grow it.
What is shadow AI?
Shadow AI is the use of AI tools, models or agents at work without the knowledge or approval of the company's IT, security or leadership teams. The most common form is an employee using a personal chat assistant account for work tasks. Others include browser extensions with AI features, free transcription tools that join meetings, AI features switched on inside software the company already pays for, and small agents built by staff on personal accounts.
It is the AI version of shadow IT, the older pattern where teams signed up for cloud apps without asking. The difference is what flows through the tool. Shadow IT moved files into an unapproved app. Shadow AI sends the content of the work itself, such as customer emails, contracts, code and plans, into a system that may keep it, learn from it or show it to others, with no record on the company's side of what was sent or what came back.
Shadow AI is not the same as misuse. Most people using unapproved AI are trying to do their jobs better and faster. That is the starting point for any sensible response.
How common is shadow AI?
Very common, and it has been for a while.
- Microsoft's 2024 Work Trend Index, based on a survey of 31,000 people across 31 countries, found that "75% of knowledge workers use AI at work today" and "78% of AI users are bringing their own AI tools to work (BYOAI)."
- The same report found that "52% of people who use AI at work are reluctant to admit to using it for their most important tasks," and "only 39% of people globally who use AI at work have gotten AI training from their company."
- MIT's Project NANDA, in its 2025 report The GenAI Divide, found that "only 40% of companies say they purchased an official LLM subscription," while "workers from over 90% of the companies reported regular use of personal AI tools for work," as Virtualization Review reported. The researchers called this a "shadow AI economy."
Read those numbers together. Most people who use AI at work bring their own tools, many would rather not say so, and few have been trained. The company's official AI programme may still be in pilot while the real adoption has already happened, out of sight. That is one reason enterprise AI pilots stall: the pilot competes with tools people already chose.
Why do employees use AI tools you never approved?
The reasons are mostly practical, and they are worth taking seriously because each one points to a fix.
- The tools help. Drafting, summarising and analysing are faster with AI. Someone who has felt that speed will not go back to doing it by hand.
- The approved path is slow or missing. A security review that takes three months loses to a sign-up that takes thirty seconds.
- The approved tool is worse. When the company tool lacks the model, features or context people need, they keep a personal one open in another tab.
- Pressure to do more. In Microsoft's 2024 research, 53% of people who use AI at work worry that using it on important tasks makes them look replaceable, which helps explain why people use it quietly rather than not at all.
- No clear rules. When nobody has said what is allowed, people guess. Most guess that it is fine as long as nobody notices.
- Leadership is behind. The 2024 report found "60% of leaders worry their organization's leadership lacks a plan and vision to implement AI." Employees fill the gap themselves.
What are the real risks of shadow AI?
The risk is not that people use AI. It is what happens to company data and decisions when that use is invisible.
| Risk | What it looks like | Control that helps |
|---|---|---|
| Data leaving the company | Customer emails, contracts or source code pasted into a personal account | Approved tools with data terms, and clear rules on what never leaves |
| Breach exposure | Unapproved tools with weak access control become an entry point | An inventory of AI tools and who can use them |
| Untraceable decisions | A price, a reply or an analysis shaped by AI with no record of the input | AI work that runs through systems that keep a record |
| Wrong answers acted on | A confident draft with an outdated price or a made-up clause goes out | Human approval on anything that leaves the company |
| Compliance gaps | Personal data processed in tools that were never reviewed | A short approved list and a fast path to add to it |
The best-known early case came in 2023. After internal data was accidentally leaked to a public chatbot, Samsung restricted generative AI use on company devices. A spokesperson said, "we are temporarily restricting the use of generative AI through company devices," and staff were asked "not to submit any company-related information or personal data," as TechCrunch reported.
The cost is now measured. IBM's 2025 Cost of a Data Breach analysis, based on research by the Ponemon Institute across 600 organizations, found that a high level of shadow AI "added an extra USD 670,000 to the global average breach cost," against an average of USD 4.44 million. It also found that "63% revealed they have no AI governance policies in place to manage AI or prevent workers from using shadow AI," and that "97% of breached organizations that experienced an AI-related security incident say they lacked proper AI access controls."
Ban, ignore or channel: three responses compared
Companies tend to respond to shadow AI in one of three ways.
| Response | What happens | Typical result |
|---|---|---|
| Ban | Block AI sites and forbid use | Use moves to personal phones and home laptops; visibility falls further |
| Ignore | No policy, no tools, no questions | Use grows unchecked; the first incident sets the policy |
| Channel | Approve good tools fast, set data rules, make use visible | Most use moves into the open, where it can be improved and checked |
A ban can be right for a short time, as a pause while a company decides what to approve. As a long-term policy, it fights the reason people use AI in the first place. Some companies have gone the other way and made AI use an explicit expectation. In April 2025, Shopify's chief executive Tobi Lütke called AI use a "fundamental expectation" for staff, CNBC reported. When use is expected, there is less reason to hide it.
Common mistakes when responding to shadow AI
- Treating it as a discipline problem. Most shadow AI users are trying to work better. Punishing them teaches people to hide use more carefully.
- A policy with no tools. A rule that says "only approved AI" with nothing approved is a ban by another name.
- A review that takes months. If adding a tool to the approved list takes a quarter, people will not wait.
- Rules nobody can remember. A twenty-page policy gets ignored. A short list of what never leaves the company gets followed.
- Only watching tools, not work. Knowing which tools are used matters less than knowing which decisions and messages AI shaped.
- Forgetting agents. Shadow AI is moving from chat to agents built on personal accounts, which can take actions as well as answer. They need owners too.
How to bring shadow AI into the open: a step-by-step plan
- Week 1: ask, without blame.
Run a short anonymous survey: which AI tools people use, for which tasks, and what they wish they had. You will learn more in a week than from months of network logs.
- Week 1: write the never list.
One page: the data that must never go into an unapproved tool, such as customer personal data, contracts, pricing and source code. Keep it short enough to remember.
- Week 2: approve good tools fast.
Pick approved tools for the most common uses found in the survey, with proper data terms. Set a target for reviewing new requests, measured in days.
- Week 3: put AI work where it leaves a record.
For work that matters, such as customer messages, quotes and decisions, use systems that keep the input, the output and who approved it. Start read-only with the sources described in choosing your first sources.
- Week 4: set approval rules.
Decide what AI may draft freely, what needs a person's approval before it leaves, and what AI never does. The setting approval rules playbook covers the format.
- Every quarter: ask again.
Repeat the survey. The share of AI use running through approved, visible channels should rise.
Measure progress with a few numbers: share of AI use through approved channels, time to approve a new tool, incidents involving AI, and share of outbound AI-drafted work with a recorded human approval. For the wider governance picture, see enterprise AI governance, and for agents that act on their own, who is accountable when an AI agent gets it wrong. Agents built quietly by staff should also go through the same tests described in AI agent evals before they touch customers.
How GeniOS approaches shadow AI
Shadow AI grows when the official path does not help people with their real work. GeniOS starts with that work. It reads where a company's work happens, such as email, calendar, chat and the business tools already in use, read-only to start, and forms the company's goals from its own emails, decks and meetings.
It checks every team's pending work against those goals, and what reaches a person is one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands work to a person or an AI agent, and nothing is sent until a human approves.
An item closes only on proof, such as the email that went out, and GeniOS learns from each outcome. AI-assisted work then happens in the open, with a record of what was proposed, who approved it and what followed.
Worked example: shadow AI at a safety-shoe maker
The names, dates and numbers in this example are made up to explain the idea. Vaultex makes safety footwear in Greater Noida. This week, Bata wants 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000, with a vendor review on Friday. Reebok samples are running on Line 2, and Adidas has a vendor audit that needs IS 15298 test reports. Terms are 60 days, and steel toe caps are up 8%.
A short, no-blame survey shows how AI is really used at Vaultex:
- Harsh in sales pastes Bata's emails, including the price and the 60-day terms, into a personal chat account to draft replies. It has saved him hours, and nobody knew.
- Anushka in production uses a free transcription app in supplier calls, including the call where the toe-cap price rise was discussed.
- The quality lead asked a chatbot to summarise IS 15298 requirements for the Adidas audit. The summary missed one test the auditor will ask for.
None of this is bad intent. All of it carries risk: Bata's pricing in a personal account, supplier terms in an unreviewed app, and an audit plan built on an unchecked summary.
Vaultex responds in three steps. It writes a never list: buyer pricing, contract terms and supplier costs stay in approved tools. It approves one assistant with proper data terms for drafting, so Harsh keeps his speed. And it moves the work that matters, the Bata quote, the Line 2 plan and the Adidas pack, into items with an owner, a date and approval before anything goes out. The quality lead's missing test is caught when the audit pack is checked against the auditor's list.
A quarter later, the survey shows most AI use now runs through approved, visible tools. Harsh still drafts with AI. The difference is that the Bata quote, the price it used and the person who approved it are all on record.
Conclusion
Shadow AI is what happens when AI helps people more than the official path does. Banning it pushes it further out of sight, and ignoring it leaves the first incident to set the policy. Ask without blame, write a short never list, approve good tools quickly, and move the work that matters into systems that keep a record and a human approval. Then measure how much AI use runs in the open. To see how GeniOS keeps AI-assisted work visible and approved, see pricing or book a call.
Sources
- Microsoft WorkLab, 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part, 8 May 2024
- Virtualization Review, MIT Report Finds Most AI Business Investments Fail, Reveals GenAI Divide, 19 August 2025
- TechCrunch, Samsung bans use of generative AI tools like ChatGPT after April internal data leak, 2 May 2023
- IBM, 2025 Cost of a Data Breach Report: Navigating the AI rush without sidelining security, July 2025
- CNBC (via NBC San Diego), Shopify CEO says staffers need to prove jobs can't be done by AI before asking for more headcount, 7 April 2025