At Glance Background
Company documents leaking from a laptop into an unmonitored AI chatbot, illustrating shadow AI

What Is Shadow AI, and How Does It Put Company Data at Risk?

Employees are already using AI tools that IT never approved. Here is how that puts company data at risk, what it costs, why bans fall short, and what an approved option has to include.

Published: | Updated: | Author: Levon Hovsepyan

Shadow AI is using AI tools at work without the knowledge or approval of a company's IT and security teams. The typical case is an employee pasting a contract, a customer email, or a spreadsheet into a public chatbot from a personal account to get a task done faster. The risk is simple: whatever goes into the prompt leaves the company's control, and the company often has no record of it.

This is not a fringe behavior.Microsoft and LinkedIn's 2024 Work Trend Index found that 75% of employees surveyed were using AI at work, and that nearly 80% of AI users bring their own AI tools. That survey is from 2024, so today's numbers may differ, but it shows how quickly everyday use moved ahead of company policy. For leaders in industries that handle confidential data, the practical question is no longer whether employees are using AI. It is which tools, with which data, and under whose terms.

What Counts as Shadow AI

Shadow AI covers more than one behavior, and some of it is easy to miss:

  • Pasting customer emails, contracts, or financial figures into a public chatbot from a personal account
  • Uploading a report or spreadsheet to an AI tool to summarize it
  • Using an AI note-taker in client meetings that the company has not reviewed
  • Installing browser extensions or plugins that read the content of pages and documents
  • Developers pasting proprietary code into a coding assistant
  • Using AI features that were switched on inside already-approved software without a security review

The last item matters because a tool can be approved and still become a shadow AI channel when a new AI feature is added to it.

Shadow AI is related to shadow IT, but it is a different problem. Shadow IT puts company data in a system IT does not know about. Shadow AI also sends that data into a service that may store it, process it under terms nobody at the company has read, and return outputs that employees then use in real work.

What It Costs

IBM'sCost of a Data Breach Report 2025, which analyzed 600 breached organizations across 17 industries, put numbers on this. One in five of the organizations studied (20%) had a breach linked to shadow AI, and those incidents added as much as $670,000 to the average breach cost. They also exposed more sensitive data: customer personal information was compromised in 65% of shadow AI breaches, compared with 53% across all breaches, and intellectual property carried the highest cost per record in the shadow AI cases, at $178.

The same report points to a governance gap behind the numbers. 63% of breached organizations lacked AI governance policies, and only 37% had approval processes or oversight in place. IBM's conclusion is worth repeating: "The real risk isn't AI itself; it's AI without governance."

How Company Data Actually Leaks Through AI Tools

AI data leakage rarely looks like an attack. It usually looks like someone trying to be productive. 5 routes account for most of it:

  • Prompts and uploads. Anything typed or uploaded goes to a third party. Depending on the tool and the plan, that data may be stored, reviewed, or used to improve models, and the terms differ between consumer and business tiers and change over time. That variability is the core of the problem, because nobody at the company has checked the terms for the tool an employee picked on their own. In 2023, Samsung engineers uploaded internal source code to ChatGPT, and Samsung's memo to staff cited concern that data sent to such platforms sits on external servers where it ishard to retrieve and delete.
  • Personal accounts. A personal account has no admin controls, no audit trail, and no way for the company to review or delete what was shared. The chat history lives with the individual, not the organization.
  • Connected apps and extensions. Many AI tools ask for broad permissions to email, calendars, drives, and meetings. A single approval can give a third-party tool continuing access to a large amount of company data.
  • AI features inside approved software. Vendors add AI features to products companies already use, and those features can change where data goes without a new procurement review.
  • Developer tools. Engineering teams are heavy users of AI coding assistants, and source code is often among the most sensitive material a company holds. Pasted snippets can include proprietary logic, configuration, and credentials, and for firms that build software for clients, the code may also be covered by contract. The Samsung case above involved exactly this. A useful policy names which assistants are approved for which repositories and what must never be pasted, such as secrets and client code the contract restricts.

Five routes company data leaks through AI tools.

Why Regulated Industries Are Most Exposed

Any company can leak a meeting transcript. The consequences are different when the data is a customer's financial records, a patient's health information, an insurance claim file, or a government case record. In those settings, a leak can mean regulatory scrutiny, contractual breaches, and lost customer trust on top of the direct cost. Customer personal information is also the data type IBM found most often compromised.

Financial institutions recognized this early. In 2023, JPMorgan restricted employee use of ChatGPT over compliance concerns, and several other large companies issued similar restrictions. Which rules apply to a given company depends on its sector and jurisdiction, so the specifics belong with legal and compliance teams. What holds across all of them is that the channel through which data reaches an AI tool is something the company has to be able to see and control.

Why Banning AI Does Not Solve It

The instinctive response is a ban, and it is worth understanding why bans tend to fall short. Samsung did ban generative AI tools on company devices and networks, but its own memo called the restriction temporary while it worked on a secure environment and its own internal tools. It also could not control personal devices: the company asked employees using such tools on their own devices to keep company information out of them, which is a request, not a control.

When nearly 80% of AI users already bring their own tools, a ban without an approved alternative is likely to move the activity out of sight rather than stop it. The more durable approach combines clear rules with a tool that employees actually want to use.

A Practical Response in 5 Steps

This is a practical sequence, not a formal standard, and the details will vary by company.

  1. Find out what is already in use. Combine an honest, no-blame survey with a review of network and software-as-a-service logs, browser extensions, and third-party app permissions. People disclose more when disclosure is not punished.
  2. Decide what data must never leave. Sort data into a few simple categories (public, internal, confidential, regulated) and state what is allowed for each. A clear line is easier to follow than a long policy.
  3. Approve a short list of tools on business terms. Look for single sign-on, administrator controls, logging, clear retention terms, and contractual commitments about how data is used. Check each provider's current terms instead of assuming, and review AI features switched on inside software you already use.
  4. Write a policy people will actually read. One page, with concrete examples of what is fine, what is not, and who to ask.
  5. Give people a good approved option, and train them. If the approved tool is noticeably worse than the one employees found themselves, they will keep using their own. Pair the tool with short training on what not to paste, and review the setup regularly as tools change. What the approved option has to include is covered in the next section.

Five steps from banning AI to governed, approved AI use

Ownership matters too. IBM's report recommends connecting AI security and AI governance, with data, security, and compliance leaders working together, because treating them separately increases risk and cost.

What an Approved AI Option Takes to Build

Step 5 is where most plans stall, because the approved option has to be as useful as a public chatbot and more trustworthy at the same time. That is an engineering problem as much as a policy one, and a few design requirements decide whether employees adopt it:

  • Identity and permissions. Sign in through the company's existing identity system and, for assistants that answer from internal documents, retrieval that respects the permissions already set on those documents. An assistant that can surface a file the user could not open directly has created a new leak inside the company.

  • A deliberate data boundary. Decide, by data category, where models run: a vendor's cloud under business terms, a private cloud, or fully inside the company's own environment. These carry different risk and cost profiles, covered in more depth in theguide to single-vendor vs. hybrid AI stacks.

  • Logging that someone can review. Who asked what, which sources were used, and how long records are kept, aligned with the company's retention rules. Without it, the approved tool is no more auditable than the shadow one. This is the same principle behind building governance into systems instead of leaving it in a policy document, covered in VOLO's guide tohow a security governance framework works inside an engineering organization.

  • Evaluation and monitoring. Test for accuracy and for permission leaks before launch, and test again when the underlying model changes, since behavior can shift when a provider updates it.

  • Integration where people already work. Much shadow AI is copy and paste between a work system and a chatbot. An approved option connected to the document stores and ticketing tools people already use removes the reason to copy anything out.

  • Fallbacks. A defined path for when the assistant remains unsure, such as routing to a person, so staff trust it enough to keep using it.

Six design requirements for an approved internal AI assistant

Grounding answers in a defined body of content is the principle behind all of this. On one recruitment platform VOLO built AI features for, a retrieval-based assistant answered from the platform's own help content and reduced inbound support inquiries by about 40%. An internal assistant relies on the same principle, with permissions added on top.

Questions Leadership Should Be Able to Answer

Question

Why it matters

Which AI tools are employees using today, approved or not?

You cannot govern what you cannot see

Which categories of data are never allowed in external AI tools?

It gives employees a clear line to follow

Do our approved tools have business terms covering retention and model training?

Consumer and business plans are often treated differently

Who owns AI governance, and do security and compliance work together?

Silos are where gaps appear

Have we reviewed AI features added to software we already use?

This is the least visible channel

What should an employee do after a mistake?

Fast reporting limits the damage

Where to Go From Here

Shadow AI is one part of a broader governance question covered in VOLO'sguide to enterprise AI implementation and itspractical guide to responsible AI and governance for regulated teams. Detecting unusual activity is a security operations task, explained inwhat a security operations center does. The policy side is only half of the answer, since employees need an approved option worth using. VOLO'sAI Platforms & Data team works on the architecture and data boundaries behind that,AI Engineering & Testing covers evaluation and monitoring, and theconsulting practice helps companies decide what to put in place first.

If you are weighing what an approved AI option should look like in your organization, including where it runs, what it can access, and how it is monitored,book a consultation with VOLO to talk it through.

At Glance Background
levon hovsepyan avatar

Levon is an experienced technology consultant leading the strategic direction of VOLO. His work focuses on AI enablement, digital transformation, and how organizations adopt and govern technology at scale.

With a background in engineering and product leadership, he brings a systems-level perspective to technology and business decisions. His writing explores AI adoption, engineering discipline, and leadership in building reliable digital systems in complex, regulated environments.

Levon Hovsepyan Chief Executive Officer

Related Blogs

Cta Background

Subscribe to our Newsletter

Frequently Asked Questions

Still have a question?

Contact us We'll be happy to help you.

Levon HovsepyanNune Darbinyan

Not by itself. Whether it becomes an incident depends on what was entered, the terms of the account used, and the rules that apply to that data under contracts and regulations. Entering confidential or regulated data into an unapproved tool can create a contractual or reportable problem, so involve legal counsel when it happens.

Only where the company has approved the tool and the category of data. Public information is generally low risk. Confidential and regulated data should go only into tools that have been reviewed and are covered by appropriate business terms.

It depends on the type of account, the settings, and the provider's current terms, which differ between consumer and business offerings and can change. Check the provider's documentation and administrator settings directly instead of assuming.

Not inherently. It becomes a legal problem when it breaches confidentiality obligations, contracts, or data protection rules that apply to the information involved.

Common methods include reviewing network and software logs for AI services, auditing browser extensions and connected-app permissions, and running a no-blame survey. No single method catches everything, so companies usually combine them.

Have a product that could benefit from AI?

  • 24 hrs average response time
  • Team of Experts
  • 100% delivery rate