Shadow AI is the use of generative AI services or accounts that a company has not approved by employees or contractors for work. A company’s visibility is limited, and what it cannot see must be addressed through policies.

This article organizes how much a company can see from three perspectives, and summarizes tools, implementation steps, contractor management, and legal considerations using diagrams and tables.

Perspective 1Which AI was accessed

Often visible in company network and managed-device logs.

Perspective 2Was the account approved?

Distinguishing company accounts from personal ones requires a way to identify them.

Perspective 3What was the user trying to send?

Checking input content requires tools and makes monitoring more intrusive.

In all cases, a company can only see activity that passes through its network or managed devices, or activity within an AI service the company has contracted for. Verified November 22, 2026.

1. What Is Shadow AI?

Scope of this article

Included
Using an unapproved AI service for work / using an approved AI service with a personal account
Not included
Entering prohibited information into an approved account (treated as a violation of a separate policy)
FigureWhat the figure measuresSource
1 in 5 companiesShare of 600 organizations worldwide that had experienced a breach and reported one involving shadow AI (not the incidence rate across all companies)IBM 2025
$670,000Difference in average breach costs between organizations with high and low levels of shadow AI (not necessarily a cause)IBM 2025
66%People who used AI while believing their company had not approved it (1,250 people in the US, UK, Australia, and Japan; self-reported)PagerDuty 2026
68.9%Japanese companies with a policy to use generative AI (FY2025; 49.7% in the previous fiscal year)Ministry of Internal Affairs and Communications

Sources: IBM, “Cost of a Data Breach Report 2025”, PagerDuty, Ministry of Internal Affairs and Communications, “2026 White Paper on Information and Communications” (in Japanese). Each figure is based on a different population.

Two types of risk (related examples)

Information enteredSamsung (2023)

In the roughly 20 days immediately after ChatGPT was approved, two pieces of source code and one meeting transcript were entered (reported by a South Korean business magazine). An example of allowing use before the rules were in place.

Tool installedDisney (2024)

Malicious software disguised as an AI image generator stole credentials for Disney’s internal Slack (reported; announcement by the US Department of Justice).

2. How Much Can a Company See? — Three Perspectives

The perspective numbers indicate their order. A tool for a higher-numbered perspective does not necessarily cover the lower-numbered ones.

Perspective 1: Which AI was accessed

What can be known
Destination, device used, and volume of traffic (and the user, if the logs can be linked to employees)
What cannot be known
Whether it was actually used for work or how many times the user asked questions
What is needed
Firewall or proxy logs and a device management system

Perspective 2: Whether the user is using an approved account or organization

Use within the contract
Check the usage totals (such as request counts and volume) in that AI service’s admin console
Personal account use
A system that can identify and log the account is needed (availability depends on the product)
Blocking unapproved use
Can be restricted using AI-side features and company proxy settings. A record of a block is not proof that the service was accessible

Perspective 3: What the user was trying to send (text or files entered)

What can be done
Inspect and block content before it is sent (such as pasted confidential information)
What may be recorded
Depends on the product and settings. Some log only the blocking event, while others retain the text as well
What is needed
A proxy that inspects traffic, or browser and device DLP (often requiring an additional contract)

What does invisible use mean?

Company observation points

Company network, managed devices, and managed browsers

+
Records from contracted AI

Only use within that contract

=
Use that leaves no record in either place is invisible

Example: a personal smartphone on a home network using AI the company has not contracted for

3. Tools to Detect and Block Use

SystemPerspectiveWhat it can doExample (verified in official documentation)
Cloud usage discovery (CASB)1List services whose connections have been observed; mark them as allowed or not allowedDefender for Cloud Apps
Traffic-inspecting proxy (SWG / SSE)1, 3Inspect destinations and input content; warn or blockPalo Alto, Cloudflare
Browser management1, 3Log sign-ins to AI services; block pasting and uploadingChrome Enterprise, Edge for Business
Device DLP3Block pasting and copying confidential information on managed PCsPurview Endpoint DLP
Restrict access to approved organizations2Block sign-ins outside the company’s contract (used with the company proxy)ChatGPT Enterprise, Claude Enterprise

Features and required contracts vary by product. Always confirm whether the product can identify accounts and retain full text before adopting it.

Details on product requirements (based on what was verified)
  • Defender for Cloud Apps: Matches connections against a list of more than 30,000 apps. Apps not on the list are not shown by default.
  • Palo Alto AI Access Security: A list of more than 500 generative AI apps. It does not inspect or log AI responses.
  • Cloudflare: Inspecting content requires a setting that decrypts traffic.
  • Chrome Enterprise: Sign-in logging is available with the free Core plan; blocking requires the additional Premium contract.
  • Edge for Business: On Windows devices managed with Intune, all profiles are covered; on unmanaged devices, only work profiles are covered.
  • Microsoft sensitivity labels: Encrypted files are not processed by AI that lacks permission to open them (this does not apply to copied text).

4. What Remains Invisible — Why Technology Cannot Block Everything

ScenarioDestinationAccountContent sent
Company network and managed deviceVisible if logging is configuredIf there is a system to identify itIf there is an inspection tool
Apps on phones and PCs at the same locationVisibleOnly on managed devices with settings deployedMay be invisible due to certificate pinning
New AI serviceMay be missed until it is added to the tool’s listSame as aboveSame as above
Unmanaged device or personal networkInvisibleInvisibleInvisible
Why content may be invisible (vendor documentation)
  • Cloudflare: Content cannot be inspected in apps that use certificate pinning, apps with self-signed or mutual authentication, or traffic using Encrypted Client Hello (ECH), which encrypts the destination name.
  • OpenAI: ChatGPT mobile and desktop apps use certificate pinning, so a company can only block personal accounts on managed devices where it has deployed settings.
  • Microsoft: Some AI services encode content or change the destination before sending it, which can affect blocking systems. There are also limits on the size that can be inspected (4 MB of text; 3 MB of files).

Even on unmanaged devices, use of an AI service contracted for by the company can be checked through that service’s records, but only for use within the contract. Sources: Microsoft, Cloudflare, and OpenAI documentation (linked in the table in section 3).

That is why, rather than trying to block everything with technology alone, it is more practical to provide approved AI and supplement it with policies and training.

5. How to Proceed — Steer People Toward Approved AI Instead of Simply Blocking Use

  1. Understand current usage Use connection logs to identify observed destinations. Supplement what the logs show with interviews about how the services were used. Before collecting logs, decide the purpose and who can view them (section 7).
  2. Prepare AI that people can use Choose a business plan where inputs are not used for training.
  3. Communicate the rules before use begins Set out in policy which AI tools are allowed, what information must not be entered, and how to request approval (How to create guidelines).
  4. Block the highest-risk use first Block services that use inputs for training and tools of unknown origin, and block personal accounts on approved AI services.
  5. Review and update Review the list of connections, number of blocks, and number of self-reports separately, then update the rules.

Items to include in the rules (policy)

ItemExample wording
AI tools that may be usedOnly ChatGPT Enterprise contracted for by the company (personal accounts are not allowed)
Information that must not be enteredCustomers’ personal information, business partners’ confidential information, unpublished figures, and source code
Requesting approval for new AISubmit a request to the IT department; approval is granted after checking how inputs are handled for training and reviewing the terms
Records and who can view themRetain connection records for 90 days; only information security staff may view them
In the event of a violationFirst check with the person and their manager, hear the circumstances, and then decide how to handle it
People coveredExecutives, employees, contract employees, and temporary staff (contractors are covered separately by contract)

Figures in the examples, such as 90 days, are examples of how to word a policy. Set them according to your company’s purpose and policies.

How to get started at different company sizes

SizeWhat to do firstNext step
A few to a few dozen peopleCheck destinations in router logs, choose one business AI plan, and communicate a brief policyIntroduce browser management only for people who handle confidential information
A few dozen to a few hundred peopleUse a CASB to identify destinations, restrict access to approved organizations, and set up an approval processPilot outbound content inspection in departments that handle a lot of confidential information
Several hundred or moreCombine a proxy with device DLP and establish monitoring policiesSet and operate the scope and retention period of records for each department
Blocking method (step 4)Tools neededRecord sensitivity
Block uniformly by destination and accountTools for perspectives 1 and 2Low
Block only submissions containing confidential informationContent inspection toolsMedium (test and verify a setting that does not retain the text)
Retain the text entered in recordsTools that inspect and record contentHigh (complete the procedures in section 7 first)

6. External Contractors — Set Requirements by Contract and Verify Through Disclosures and Deliverables

If a contractor uses AI on their own device and network, your company cannot see it. Instead, divide responsibilities.

WhatHowCaution
Conditions of useSet requirements for the contractor’s company in a contract or memorandumAI tools that may be used, information that must not be entered, and the obligation to disclose use
Whether AI was usedDisclosure when work is deliveredYou can only know what has been reported
QualityInspect against standards agreed in advanceThe deliverable cannot prove whether AI was used
Work that needs to be verified through recordsHave the contractor use your company’s environment and company-contracted AIYou can only see records from that environment

Example disclosure form for delivery

Was AI used?
Used / Not used
AI used
Service name and contract type (business or personal)
Work stages where it was used
Drafting, translation, code, research, and so on
Information entered
Whether materials received from the client were entered, and if so, which ones
What was checked
How figures, sources, and behavior were verified

Three legal points to watch under Japanese law

Disguised subcontractingDo not give direct instructions to workers

Communicate requirements to the contractor’s supervisor and leave day-to-day instructions to the contractor.

Personal dataSupervising contractors is mandatory

Set out in the contract how inputs to external AI services will be handled, and check with legal counsel.

FreelancersSet out the conditions upfront

Specify whether AI may be used and the acceptance criteria when placing the order (there are conditions on the prohibition against refusing delivery; see the details below).

Legal basis and details (to be confirmed with an expert)
  • Disguised subcontracting: Ministry of Health, Labour and Welfare guidelines require the company receiving the work to direct how the work is carried out and evaluate workers. Q&A collection (in Japanese) (First collection, Q2) treats requests to the company to revise work processes as instructions to the company, and direct instructions to workers as disguised subcontracting. This does not directly address conditions on AI use.
  • Personal data: The Act on the Protection of Personal Information requires the party outsourcing work to supervise the contractor (Article 25). It is advisable to track subcontracting through prior reports or approval (guidelines). Agreement may also be documented in a memorandum or written undertaking (Q&A 5-8 (in Japanese)). Whether entering information into external AI constitutes subcontracting is not determined uniformly. Ask legal counsel to check how the provider uses, retains, and deletes inputs, as well as the contractual roles.
  • Freelance Protection Act: When outsourcing to an individual who does not employ workers, the transaction terms must be disclosed (Article 3). When a client that employs workers outsources work for at least 1 month, refusing delivery and other actions are prohibited (Article 5). Unilaterally tightening inspection standards after the fact constitutes refusal to accept delivery (Explanation by the Japan Fair Trade Commission and Ministry of Health, Labour and Welfare (in Japanese)).

7. Legal Considerations When Monitoring

When records are linked to users, even connection records may constitute employee monitoring. The level of intrusiveness depends on the following four factors.

TargetWhose activity?

Everyone’s, or only departments handling confidential information

ScopeWhat is monitored?

Destinations only, or the full text of inputs

PeriodFor how long?

How many days to retain records

AccessWho can view it?

Limit who can access the records

Japan: Four points to consider from the Personal Information Protection Commission (Q&A 5-7 (in Japanese))

1. Purpose
Specify it in advance, set it out in policy, and inform employees
2. Responsible person
Identify the responsible person and define their authority
3. Rules
Establish rules for implementation and ensure that the people operating the system follow them
4. Verification
Check that the rules are being followed

These considerations apply when monitoring employees who handle personal data. Prior notice to labor unions and similar groups, as well as informing employees, is also considered advisable. Consent is not listed among these considerations, but that does not mean it is unnecessary for monitoring in general.

When entering personal data into AI: In 2023, the commission warned that using personal data entered without the person’s consent for training or other purposes could violate the law (warning (in Japanese)). Even with a contract that excludes training, whether information may be entered must be assessed separately.
EUKey points
GDPRMonitoring must be necessary and proportionate, and employees must be informed. Consent is difficult to rely on as a legal basis in the workplace (WP249)
Case lawMonitoring content requires stronger justification than monitoring recipients or volume (European Court of Human Rights, Bărbulescu v. Romania, 2017)
AI ActAI that infers employees’ emotions from biometric data has been prohibited since February 2, 2025 (except for medical or safety purposes)

Points to check with an expert before implementation: where to place the policy in the work rules, whether full input text may be recorded, procedures for using results in disciplinary action, contracts with contractors, and laws in overseas offices.

Summary

What is visible
Only activity at company observation points or within contracted AI. Everything else is invisible
How to proceed
Understand usage → prepare approved AI → communicate rules and start using it → block the highest-risk use first → review and update
Contractors
Set conditions by contract, verify whether AI was used through disclosure, and check quality through inspection
Monitoring
Define the target, scope, period, and access; address the four points to consider; and check with an expert

FAQ

If I use ChatGPT without telling my company, will they find out?

On a company network or managed device, the connection may be logged depending on the logging settings. However, a connection alone does not show whether the service was used for work. If use is prohibited, first check which AI tools are approved.

With a business plan, can administrators see all my conversations?

AIAdmin console analyticsSystem for retrieving message content
ChatGPT Enterprise / EduWorkspace analytics (such as message counts by user; no message content)Compliance Platform (only the workspace owner can be granted access to conversations)
Claude Team / EnterpriseUsage analytics (aggregated data)Enterprise Compliance API (enabled by the Primary Owner)

Verified against each company’s documentation on November 22, 2026. Check the terms of each contract for other AI tools and plans.

Which is better: banning AI entirely or allowing and managing it?

People can still use AI on unmanaged devices even if it is prohibited. The AI Business Operator Guidelines also recommend using AI intended for business. It is more practical to provide approved AI first.

What if a small company cannot afford expensive tools?

Use existing router or firewall logs to check destinations, choose one business AI service, and write a short policy specifying what information must not be entered. Consider content inspection first for people who handle confidential information.

Can I check whether a contractor used AI?

You cannot see a contractor’s devices or network. Set conditions and disclosure requirements in the contract. For work that needs to be verified through records, have the contractor use your company’s devices, virtual desktop, or company-contracted AI.

Update History

  • November 22, 2026Published.

Sources