CultureAI has found that 81% of workplace AI interactions take place through personal, free or unidentified accounts, leaving most employee AI activity outside enterprise-level controls. The findings come from The Unmanaged Majority: What 1.7 Million AI Interactions Reveal About Enterprise Control, a research report based on nearly 1.7 million AI prompts and interactions across 40 organisations over six months in 2026.
The research also identified 252,480 sensitive data detections across 100 data types, equivalent to approximately one detection for every seven prompts. According to CultureAI, the findings highlight a gap between how organisations believe employees use AI and how these tools are being accessed in everyday work.
The analysis found that approving AI applications and purchasing enterprise licences do not necessarily provide visibility into all employee activity. Workers are using a wide range of AI applications, often through personal accounts, while sharing information that traditional security tools may not be designed to identify.
AI Adoption Is Widespread, but Usage Is Concentrated
CultureAI observed 1,064 distinct AI applications across the organisations studied, with an average of approximately 113 applications per organisation. Nearly half of the applications appeared in only one organisation, while general-purpose chatbots accounted for around 6% of the total.
The findings suggest that organisations relying on industry trends or recommendations from other businesses to create approved AI application lists may overlook tools their own employees are using. CultureAI says application discovery therefore needs to be organisation-specific and continuous.
Only 19% of AI Interactions Used Enterprise Accounts
The research found that only 19% of AI interactions took place through enterprise accounts. The remaining 81% involved non-enterprise access:
- 47%: Personal, free or unauthenticated accounts.
- 34%: Accounts where the licence type could not be identified, mostly on applications that offer only personal accounts.
This means that many AI governance policies and controls may cover only a limited portion of actual workplace AI activity.
The research also found that restricting access to AI applications did not necessarily stop employees from using them. In organisations that blocked AI tools, personal account usage increased by approximately 50% afterwards. CultureAI said this suggests that blocking applications can push activity into accounts that organisations have less visibility and control over.
Sensitive Data Exposure Extends Beyond Credentials
Across the 40 organisations, CultureAI recorded 252,480 sensitive data detections covering 100 data types.
The research identified several categories of sensitive information being shared through AI applications:
- HR records and correspondence: Detected in 90% of organisations.
- Company strategy information: Detected in 85%.
- Health conditions: Detected in 82%.
- Legal case details and financial information: Detected in 72%.
The findings indicate that sensitive information exposure is not limited to a small number of organisations.
Credentials, often considered one of the primary risks associated with data leakage, accounted for less than 1% of detections. Company strategy, financial, HR, legal and health-related disclosures collectively outnumbered credential detections by almost 30 to one.
CultureAI highlighted that information such as employee grievances, medical diagnoses, candidate comparisons and board forecasts can be difficult to identify using conventional pattern-based security tools. Unlike passwords and account numbers, these disclosures do not follow fixed formats, and their sensitivity often depends on context.
AI Data Exposure Raises Compliance and Legal Concerns
The research also points to potential regulatory and legal implications for organisations using AI tools.
Health information may qualify as special category personal data under UK and EU GDPR, requiring an appropriate lawful basis and an Article 9 condition for processing.
From December 2, 2027, high-risk AI systems used in recruitment and employment decisions will be subject to relevant EU AI Act requirements, including human oversight and record-keeping.
Legal teams also face confidentiality and privilege concerns when submitting privileged material to external AI services. The report refers to a case earlier in 2026 in which a solicitor faced a regulatory probe after uploading client documents to ChatGPT.
CultureAI emphasised that having an enterprise agreement with an AI provider does not automatically address every data protection or governance concern. Permission to use an AI application does not necessarily mean that every type of information can be shared through it.
CultureAI Calls for Greater Visibility into Workplace AI Use
Oliver Simonnet, Lead Cybersecurity Researcher at CultureAI, said the research highlights the difference between organisations’ assumptions about AI usage and employees’ actual behaviour.
“AI is now firmly embedded in everyday working life, but our research shows a significant disconnect between how organisations believe AI is being used and what is actually happening. With 81% of AI interactions taking place outside enterprise accounts, it’s clear that traditional approaches to governance are failing to keep pace with employee behaviour.”
Simonnet said organisations need visibility into the AI applications employees access, the accounts they use and the information they share.
“Organisations cannot assume that approving an AI tool or introducing a policy automatically means they have control over how it is used. They need visibility into the applications employees are accessing, the accounts they are using and, crucially, the information being shared. Restricting AI isn’t the answer. Where organisations blocked AI tools, personal account use rose by around half. The better route is to make secure use the easiest option. By continuously monitoring AI activity and applying appropriate safeguards, businesses can protect sensitive information without sacrificing the productivity benefits that make AI so valuable in the first place.”
The full report, including recommendations for security, IT and compliance teams, is available from CultureAI’s website.
Research Methodology
The study used anonymised and aggregated data from 40 organisations using the CultureAI platform. The data was observed over six months, from April to September 2026, before CultureAI controls were applied.
Participating organisations represented legal and professional services, financial services and insurance, technology and cybersecurity, healthcare, housing, charities and education, as well as retail, hospitality and manufacturing.
The research covered:
- 1,668,999 prompts
- 161,599 file uploads
- 1,064 AI applications
- 252,480 sensitive data detections
- 100 sensitive data types, grouped into 18 categories for reporting
CultureAI stated that the research does not identify individual users or participating organisations.
