Almost every organisation now has staff using AI assistants, and almost none have a policy. The usage is not the problem. The problem is that client contracts, customer records, incident details, and source code are being pasted into services nobody evaluated, under terms nobody read, from accounts the business does not control.

Why bans fail

An outright prohibition does not stop usage; it moves it to personal accounts on personal devices, where you have no visibility, no data controls, and no record. The organisations handling this well provide an approved tool and are clear about what may go into it.

What the actual risks are

  • Confidentiality: client data shared with a third party you have no agreement with, potentially breaching your own contracts.
  • Training data: consumer tiers frequently use inputs to improve models. Business and enterprise tiers usually do not — but you have to check, and the setting is sometimes off by default.
  • Retention: conversations may be stored and reviewable, including by the provider's staff for abuse monitoring.
  • Accuracy: confidently wrong output acted on without verification, which matters enormously in regulated or safety-relevant work.
  • Regulatory: under GDPR and similar regimes, pasting personal data into an unassessed processor is a compliance problem regardless of outcome.

A policy that people will actually follow

Keep it to one page. Name the approved tool and the account type. List what must never be pasted in: client personal data, credentials, unreleased financials, anything under NDA, and case or incident detail identifying an individual. State that output must be verified before use in client work. Then explain why, briefly — policies people understand get followed, and policies that only prohibit get ignored.

Configure the tooling properly

Use a business tier with data-training disabled and administrative control over accounts. Enable single sign-on so access is removed at offboarding. Review retention settings. Where you use AI features inside existing products, check what those features send externally — the answer is frequently not what people assume.

Consider what you build with, too

If you are integrating a model into your own product, the same questions apply and the exposure is larger: what leaves your environment, what is retained, what happens if the provider has an incident, and what you tell your own customers about it.