Nearly half of employees use personal AI for work

You must now establish which personal AI services your workforce uses for work, document which data categories are forbidden to them, and provide an approved alternative. Visibility comes before prohibition.

An analysis of 2 October 2026 of personal AI use in the workplace without organisational knowledge argues that nearly half of employees use personal AI accounts for work tasks, often without permission or oversight. The study found that more than a third of this group operate without the organisation's knowledge, and fewer than two in five employees can name which tools are permitted. In our assessment, the practical problem is not that employees adopt AI, but that they do so faster than organisations can establish policy, training and visibility. For a significant portion of your workforce, personal AI has become de facto work infrastructure whilst you cannot demonstrate which data passes through it or under what controls.

Why personal AI use creates a visibility problem

The risk is not adoption itself, but loss of control over data flows. When an employee pastes a confidential document into a public AI service via a personal account, you cannot easily show which data category applied, what consent was obtained, or which human oversight was in place. You also cannot verify the service's training defaults, retention rules or account settings, which vary by provider and subscription tier. Under GDPR accountability requirements, you must be able to demonstrate the lawful basis, processing scope and safeguards for any data your organisation handles. Personal accounts break that chain of evidence.

What failure modes does unmanaged personal AI introduce?

• Data leakage into public services — confidential documents, customer records and financial information entered into tools with unknown retention or training policies.
• Loss of audit trail — no record of which employee used which account, what data was processed, or under what terms of service.
• Uncontrolled training and inference — personal AI accounts may have default settings that use inputs to improve models, creating secondary processing you did not authorise.
• Compliance gap — inability to demonstrate lawful basis, consent or appropriate safeguards for data processed via personal accounts.
• Shadow IT dependency — workflows that depend on unapproved tools become invisible to IT and security teams until a breach surfaces them.

What controls must you be able to demonstrate?

1. Inventory which AI services are permitted for which data categories — document which tools employees may use for public information, internal non-sensitive work, and restricted data, and which are forbidden entirely.
2. Provide an approved alternative for common workflows — if employees need AI for drafting, analysis or coding, offer a service you control or have assessed, with clear terms on data retention and model training.
3. Communicate the rules and test understanding — publish a short, usable policy and measure periodically whether employees know which tools are allowed and why; a one-off announcement is not enough.
4. Establish a low-threshold reporting point for borderline cases — let employees ask before they guess, and track those questions to refine your rules.
5. Audit which services appear in network logs and user behaviour — use endpoint and network visibility to identify personal AI use and feed findings back into policy updates.

How should you approach the transition?

Start with a short, workable set of rules rather than an abstract ban. The study shows that unfamiliarity with rules is the problem; employees who do not know what is permitted have no reason to avoid personal accounts. An approved alternative with clear data handling terms lowers the incentive to use personal services. This is a practical choice, not a security theatre measure. Measure periodically whether your workforce knows the rules, not whether the rules exist on paper. The gap the study exposes is between what employees do and what you can demonstrate. Closing that gap is your responsibility.

Automated tools can help you inventory services and flag risky patterns in network traffic, but the final judgement on acceptable use remains yours. No scanning tool can replace the professional assessment of which workflows genuinely need AI, which data they touch, and what safeguards are proportionate. Use measurement to inform that judgement, not to replace it.