AI used for work outside your organisation’s approved systems or processes. It can include personal accounts or unreviewed AI features in existing software. Visibility starts with understanding what staff actually use.

A prompt, uploaded file or connected tool can share information with a service you have not assessed. Client emails, pricing, contracts and staff information are examples to consider. We help distinguish approved sharing from unresolved exposure.

That depends on the service and its supply chain. A review should identify the app provider, model or hosting providers and relevant subcontractors, then examine processing purposes, storage locations, retention, access and deletion terms. The supplier’s actual arrangements matter more than a broad “secure AI” label.

No. Training use, storage, retention and service-improvement arrangements differ by product, account, contract and settings. “Not used for training” also does not mean “never processed or stored”. Check the exact service and any connected tools rather than assume all AI handles data in the same way.

A business account is a starting point, not a complete access review. Microsoft describes protections for its work or school Copilot Chat, including no training of foundation models on prompts and responses. Still check which product, account, integrations and permissions staff actually use. Other tools require their own assessment.

Only where the available evidence supports that conclusion. We would agree sources, time period, coverage and measurement first. App-use counts or traffic volume alone do not prove which records or sensitive fields were shared. Historical gaps and unsupported devices must stay visible; an estimate must be labelled as one.

It may, if its permissions or connected systems allow it. Reviewing a pasted prompt is different from reviewing an integration with access to mail, shared files or a CRM. We would examine granted access, intended purpose, ownership and how that access can be reduced or revoked.

A policy needs a practical route for staff to follow: approved tools, clear data-handling rules, accountable owners, an exception process and reviews when use changes. Our approach joins those decisions to an evidence-based inventory.

The NCSC recommends understanding why staff use unapproved tools and providing suitable alternatives. A useful first step is to identify the work being done, agree which tools and information are appropriate, and make the approval route usable.

That is not our default proposal. We would agree a proportionate scope, permissions, transparency and handling rules before collecting evidence. Start with the least intrusive sources that answer the agreed question. Any review of content or personal information needs specific consideration and authorisation.

Yes, that is the proposed delivery model: agree who owns systems, evidence, approvals and remediation before work starts. We would coordinate findings and actions with your nominated team. Access, responsibilities and any implementation work would be agreed in the engagement scope.

An agreed-scope AI inventory, a map of known data flows and permissions, a prioritised action log, and recommendations for approved uses and ongoing review. Findings would distinguish confirmed evidence, assumptions and unknowns. Coverage, delivery, pricing and review cadence must be agreed before work begins.