Shadow AI can be what people use when the approved path is slower than the work. It is easy to treat as a discipline problem. It can also be a demand signal: people found a tool that helps, and the sanctioned route was missing or slower.

Shadow AI is the use of artificial intelligence tools, features or agents for work without the company’s knowledge or approval. It covers a personal chatbot account with work data pasted into it, an AI feature switched on inside approved software, and an automation an employee built with a personal key. The work may still get an ordinary review from a manager or a peer. What is missing is the company’s review of the tool: what data it takes, where that data goes and how far its output can be trusted.

What counts as shadow AI

The test is approval, never the tool. The same chatbot is sanctioned in one company and unsanctioned in the next, and one employee can cross the line by switching from a company account to a personal one.

The forms it takes

  • Personal accounts on work data. A free or personal chatbot subscription, with a contract, a customer list or source code pasted in.
  • AI features inside approved software. The vendor adds an assistant to a product you already bought. Nobody on your side installed anything, and the feature may never pass through an approval.
  • Extensions and note-takers. Browser add-ons that read every page, and meeting bots that record and summarize calls for whoever invited them.
  • Agents and automations. Workflows staff build with their own API keys, acting on company systems under a personal login.

The second form is the one we feared most at Revalize, where we deployed Microsoft 365 Copilot early, in 2024: apps already in use adding AI features that went undetected, because every vendor seemed to want AI in its product details.

How shadow AI differs from shadow IT

Shadow IT is software or hardware used without approval: a file sharing account, a project tracker on someone’s card. Shadow AI shares most of that problem: the company cannot see the data, nobody vetted the tool, and a browser tab can be enough to start.

What the AI version adds, in IBM’s account of the difference[1], is a set of concerns about data management, model outputs and decision-making. A file sharing account stores what you put in it. An AI tool produces something new (a summary, code, a customer email, a recommendation) that enters the company’s work and can shape decisions, from a tool whose accuracy and data handling the company never assessed.

Why people reach for unapproved AI tools

Start with the number. The 2024 Work Trend Index from Microsoft and LinkedIn, which surveyed 31,000 knowledge workers across 31 countries, found that 78% of AI users were bringing their own AI tools to work[2]. Bringing your own tool does not prove the employer refused one. One reaction is still to assume carelessness, or people hiding tools from security.

There is another reading in the same report. In that 2024 data, 52% of the knowledge workers who used AI at work were reluctant to admit using it for their most important tasks, and 53% worried it made them look replaceable[2]. Those are worries about how the work looks to an employer, a different motive from dodging a security control.

Now the company side. IBM’s 2025 Cost of a Data Breach Report, built on breaches at 600 organizations, found that 63% of breached organizations had no AI governance policy or were still developing one[3].

One sample is knowledge workers and the other is breached organizations, so the pair cannot prove what happened inside any single company. As an interpretation, they fit a pattern: people find a tool that makes hard work faster, the employer has no approved version and no clear rules, and the use goes quiet. Treat that as a hypothesis and test it with the questions at the end.

Here is what it looks like up close. A trial of an endpoint monitoring tool flagged one of our employees uploading tax return documents to ChatGPT. The employee was customer facing, so our first read was that client tax returns were leaving the building, a privacy violation and a policy violation in one upload. We had approved only Copilot and Claude at the time, so there was no corporate ChatGPT to use, which meant a personal account, and it later turned out to be a free one. The returns turned out to be the employee’s own, so our response was a stern talk about the danger of putting very private details into a free account.

By the definition above, that falls short of shadow AI on work data: unapproved tool, personal documents. The alert still showed that the approved list lacked the tool this person reached for, and that a personal account was one browser tab away.

What shadow AI costs

The risks of shadow AI are easy to list: data leakage, compliance, bad output. A list is hard to budget against, so here is a bill, with a source on every line that has one.

One caution first: IBM’s figures come from organizations that had already suffered a data breach. They describe that group and are not a rate for all companies.

Breach exposure. Among the 600 breached organizations in IBM’s 2025 report, one in five reported a breach due to shadow AI[3]. In the same report, organizations with high levels of shadow AI saw an average of $670,000 in higher breach costs[3] than those with little or none.

Access control debt. In that 2025 sample of breached organizations, 13% reported breaches of AI models or applications, and 97% of those had no AI access controls in place[3]. The 2026 edition, based on breaches at 602 organizations, reports that more than 20% of organizations reported a breach targeting AI models or applications[4]. The two editions word the measure differently and sample different organizations, so the figures make no trend line.

Sensitive data leaving. An IBM explainer published in October 2024 reports that 38% of employees acknowledge sharing sensitive work information with AI tools without their employer’s permission[1]. With a personal account, that data sits outside the contract terms the company approved and the administrative controls of its own tenant.

Unchecked output. A University of Melbourne and KPMG study surveyed more than 48,000 people across 47 countries between November 2024 and mid-January 2025[5]. In its findings on AI at work, KPMG’s summary reports that 58% of employees intentionally use AI, that 66% rely on AI output without evaluating its accuracy and that 56% are making mistakes in their work because of AI[6]. The summary does not give the exact base for the reliance and mistake figures, so read them as workplace findings, never as shares of everyone surveyed. Undisclosed use can make those errors harder to catch: a reviewer who does not know a model wrote the draft has less reason to check it closely.

Two more lines belong on the bill, and I have no number for either, so take them as argument.

Duplicate spend. Picture five departments expensing five subscriptions for the same job while an enterprise license sits half used.

Learning the company never captures. This line appears on no invoice. Some of the people using unapproved AI have worked out where it helps. When admitting the use means admitting a policy violation, what they learned (the prompts that worked, the ones that failed) tends to stay with them. The company carries the risk and may collect little of the knowledge.

Why banning AI tools does not work

A blanket ban can end your view of the use without ending the use.

Block a chatbot on the corporate network and some of the work can move to a phone, a home laptop or a personal account, where the data still leaves through a channel that is harder to log. The risk of a blanket ban is that it turns a visible problem into a less visible one, and tells capable people that their initiative is a disciplinary matter.

One body of research points the same way. The MIT Center for Information Systems Research worked with more than 70 data and technology executives representing 50 organizations between late 2023 and August 2024[7], and its authors conclude that leaders should set guidelines that let employees experiment, because that reduces risk while keeping the gains.

Granted, some things should be blocked. A tool with no acceptable data terms has no business near customer records, and a regulated workload needs a harder line than a marketing draft. Blocking a specific tool for a stated reason is a control; blocking a category because it is new is a wish.

How to govern shadow AI without banning it

The alternative to shadow tools is a controlled rollout. It has five parts. Each is a recommendation; none is a measured result.

  • Make the approved path faster than the workaround. If the sanctioned tool is weaker, slower or missing, expect people to route around it. Speed of approval works as a security control.
  • Write rules about data, not tools. A tool list needs revising as products launch. A rule such as “No customer records in any account the company does not hold a contract for” does not need rewriting with each launch. Where data lives is the same question it is in data center compliance.
  • Ask before you scan. Try an amnesty inventory (tell us what you use, no penalty) before a block list. The bet is that people answer more honestly without a penalty, and the answers tell you what to buy.
  • Give every request an owner and a deadline. A request with no answer date works like a ban with extra steps.
  • Audit on a schedule. In IBM’s 2025 report, only 34% of organizations with AI governance policies performed regular audits for unsanctioned AI[3]. A policy nobody checks is paper; see My SOC is better than your SOC for the same problem in compliance.

The first item comes with evidence and a caveat. The career record for Revalize, 2021 to 2026, lists 96% weekly Microsoft 365 Copilot adoption, averaging more than 100 interactions per employee a week[8]. An approved tool can become the default. One approved tool was still not enough: Copilot was often the only sanctioned option, and our best users kept pushing past it with their own AI, bypassing policy. The spread of shadow AI stopped only when we went vendor agnostic and adopted central accounts for all the major providers. That is one company’s account; nobody ran a controlled study.

Where I work now, the security, IT and AI teams review each new tool request together: what is the need, does the tool fit the requirements, and does it solve the problem. We never adopt AI for the sake of AI, and tools that are not approved are blocked. That block comes with a review path behind it, which is what separates a control from a ban.

We cannot stop every personal account, so we control what they can reach. We limit MCP connections (the connectors that let an AI assistant read and act on company systems) to an approved list, and lock those to our official company tenants in Claude, Cursor, ChatGPT and the other major vendors. Try to connect to M365 or Azure DevOps through a connector that is not on the list and it fails.

Invite Joe to speak

A test for your own company

You do not need a consultant for this. The NIST AI Risk Management Framework is voluntary[9], and it organizes the work into four functions: Govern, Map, Measure and Manage[10]. The first two are enough for a start: who decides, and what is in use. Four questions cover them.

  1. How long does approval take? Count the days from the last request for an AI tool to a yes or a no. If nobody can find the request, that is your answer.
  2. Would your people tell you what they use? Ask without a penalty attached and compare the list with what you approved.
  3. Which AI features switched on inside software you already own? Check a year of release notes from your ten largest vendors and count the assistants nobody on your side evaluated.
  4. Who audits, and when did they last do it? A name and a date. “Security” and “ongoing” do not count.

Shadow AI can be read as a measure of the gap between what your people need and what you have approved. Narrow the gap and you remove one reason to work around you. Leave it open and the reason stays.

If your best performer found a tool tomorrow that cut a day of work to an hour, how long would it take your company to say yes, and what would that person do in the meantime?

Sources

  1. What is shadow AI?, IBM Think
  2. AI at Work Is Here. Now Comes the Hard Part (2024 Work Trend Index Annual Report), Microsoft and LinkedIn
  3. IBM Report: 13% Of Organizations Reported Breaches Of AI Models Or Applications, 97% Of Which Reported Lacking Proper AI Access Controls, IBM Newsroom
  4. IBM Study: One in Four Malicious Breaches are AI-Enabled, Costing Companies $6 Million on Average, IBM Newsroom
  5. Trust, attitudes and use of artificial intelligence: A global study 2025, University of Melbourne and KPMG
  6. Global study reveals trust of AI remains a critical challenge reflecting tension between benefits and risks, KPMG International
  7. Bring Your Own AI: How to Balance Risks and Innovation, MIT Sloan Management Review
  8. Career Story: 29 Years Leading Technology, Joe Johnson
  9. AI Risk Management Framework, National Institute of Standards and Technology
  10. NIST AI RMF Playbook, National Institute of Standards and Technology

Views are my own and do not represent my employers, past or present.