35% of Generative AI Users Experience Near-Misses, "AI Overconfidence" Leading to Copy-Pasting and Confidential Information Input as Main Causes

Global security manufacturer Cyber Security Cloud Inc. (Headquarters: Shinagawa-ku, Tokyo; President and CEO: Toshihiro Koike; hereinafter "the Company") conducted a survey on the actual state of generative AI business use targeting 300 office workers who use PCs daily for work.

Survey Summary

- Over 60% of office workers use generative AI for business, with approximately 15% in a "shadow AI" state*. - Over half of users input internal documents into generative AI—including financial and customer information. - Approximately 35% of generative AI users have experienced near-misses, and about 1 in 7 (14%) have actually faced problems.

*Shadow AI: Employees using AI tools (such as ChatGPT, Gemini) not approved or managed by the company for business purposes based on their own judgment.

1. Over 60% of office workers use generative AI for business, 15% are in a shadow AI state

When 300 office workers who use PCs almost daily for work were asked about their generative AI business usage, it was revealed that 67% (201 people) use it in some form.

Looking at the breakdown of users, while 52% "use tools officially approved by the company," approximately 13% "use it without clear company rules," and about 2% "use tools not permitted by company rules," indicating that approximately 15% are in what is known as a shadow AI state.

Regarding the tools used, Microsoft Copilot was the most common at about 49%, followed by Google Gemini (about 36%), ChatGPT personal accounts (about 35%), and ChatGPT corporate contracts (about 25%). While 90% of usage environment is company-issued PCs, a certain number of users also access it from personal PCs (9%) and personal smartphones (8%).

*For ChatGPT, data handling policies differ between personal accounts (free/paid) and corporate contracts, so usage patterns (personal/corporate) are differentiated for the purpose of shadow AI considerations.

2. Over half of users input internal documents into generative AI—including financial and customer information

Approximately 20% have experienced inputting financial and sales data, about 16% customer names and business partner information, about 13% contracts and legal documents, and 10% personal information (names, contact details). It was revealed that approximately 60% of users have experienced inputting some form of business information into generative AI.

On the other hand, only about 52% "always consider" the possibility of information being shared externally, while about 23% "hardly or not at all consider" it, highlighting a disparity in risk awareness.

3. Approximately 35% of generative AI users experience near-misses, about 1 in 7 actually face problems

Among the 201 people who use generative AI for business, approximately 35% answered that they "have had a near-miss experience." Of those, about 14% said it "actually became a problem," revealing that risks are not merely potential.

Lack of verification of output content is cited as a background factor. Approximately 37% of users answered that they "often or sometimes use generative AI output as is for business without confirmation or modification," suggesting that overconfidence in generative AI is a cause of near-misses.

Furthermore, approximately 43% of users "mostly or intentionally do not share" the fact that they used AI for work, revealing a common practice of low transparency. Additionally, approximately 45% answered that company AI usage rules "do not exist" or "they don't know," indicating a latent risk of operations being left to individual discretion.

Summary

This survey revealed that while the business use of generative AI is rapidly expanding, corporate governance frameworks are not keeping pace. Approximately 15% of users are in a shadow AI state, and over half input highly confidential information such as internal documents, financial data, and customer information into generative AI.

Furthermore, approximately 35% have experienced near-misses, some of which have developed into actual problems, indicating that risks are already apparent. Additionally, about 45% answered that AI usage rules "do not exist" or "they don't know," highlighting a situation where usage is expanding while being left to individual judgment.

Moving forward, to safely and effectively utilize generative AI for business, it is essential to promote a three-pronged approach: "visualization of usage," "establishment of rules," and "employee education," beyond just tool implementation.

Survey Overview

- Survey period: April 21-22, 2026 - Survey method: Internet survey - Survey body: Cyber Security Cloud Inc. - Research commissioned to: Rakuten Insight, Inc. - Valid responses: 300 people - Survey target: Office workers who use PCs daily for work

Cyber Security Cloud Inc. (https://www.cscloud.co.jp)