Security & Privacy
Using AI at Work Without Exposing Confidential Data
Choose suitable tasks, minimize what you share and check the account's data controls before putting workplace information into an AI service.
Start with the task and the information
AI can help draft an agenda, explain a public document or suggest a spreadsheet formula. The risk changes when the prompt includes customer details, unreleased plans, access credentials or an internal investigation. Before choosing a model, decide whether the material is appropriate to send to that particular service under your organization's rules.
Separate a useful task from its sensitive inputs. To ask for an email structure, you usually do not need to paste the real customer's message. To learn a formula, you can use made-up column names and sample numbers. Keeping the task while replacing the data often preserves much of the benefit.
Use an approved account, not a familiar logo
A personal chatbot account and an organization-managed service can have different terms and controls even when the product name looks similar. Confirm which account you are signed into and whether your employer has approved that use. A paid badge alone does not prove a service meets a confidentiality requirement.
Provider policies differ. Google's Gemini API terms describe content handling for unpaid services and regional exceptions. Groq documents retention exceptions and organization controls. Read the terms for the actual service, endpoint and region rather than transferring an assumption from a different product. If nobody can confirm the approved arrangement, keep the confidential material out.
Reduce and replace before you paste
Use a practical three-step check: remove information not needed for the task, replace identifiers with synthetic values, then read the prompt again as if it would be shared outside the team. Names are only one kind of identifier. Contract amounts, unusual dates and a detailed incident description can identify an organization or person even after names are removed.
Try a safer version of the same task. Instead of pasting a named customer's overdue invoice, ask: 'Draft a polite payment reminder using placeholders for customer, invoice number, amount and due date.' Fill the placeholders locally after reviewing the draft. Do not assume that replacing names alone makes a detailed real record anonymous.
Be careful with files, connectors and generated actions
An uploaded spreadsheet or document may contain hidden sheets, comments or metadata you did not intend to share. A connector can expose much more information than one prompt. Review its scope and account permissions before enabling it. For a work task, ask whether the service needs access to the entire drive or just a small approved collection.
Treat output as a draft, not a command from a trusted colleague. Check factual statements and calculations, inspect code before running it and confirm consequential actions in the actual business system. The NCSC's guidance on language-model risks is a useful starting point for understanding why fluent output should not be mistaken for verified advice.
Know what to do if something was shared
If you accidentally submit confidential information, stop further sharing and use your organization's reporting process. Record the service, account, approximate time and type of information without copying the sensitive content into more channels. An administrator may need to revoke access, review retention options or contact the provider.
Deleting a conversation is not evidence that every log or downstream copy has disappeared. Keep the response proportionate to the data and seek the responsible security or privacy team's guidance. The goal is not to ban every helpful AI task; it is to make the approved path clear enough that people can use it without guessing about confidentiality.
Primary references
Sources checked 1 November 2026. Follow the current documentation for changes to features, limits and pricing.