For workplace use

How to Use AI Tools Without Leaking Sensitive Data

Last updated July 27, 2026

An AI assistant is an external system unless your organization has assessed, contracted and configured it otherwise. Before you paste text, upload a document or connect a mailbox, ask two questions: what data is leaving my control, and what can this tool do with my access?

Before entering a prompt

Do not put the following into an unapproved AI service:

passwords, authentication codes, private keys or access tokens;
customer records, health or financial information;
confidential contracts, legal advice or unreleased financial results;
proprietary source code, vulnerability details or incident data;
personal data you do not have permission to process in that service.

Replacing a name is not always enough. A combination of role, location, dates and events can identify a person. If the task can be done with a synthetic example, summary or approved internal tool, use that.

Check the whole data path

The prompt box is only one route into an AI system. Review:

Uploads: PDFs, images and spreadsheets may contain metadata, hidden sheets, comments or revision history.
Chat history and memory: understand whether conversations are retained, used to improve models, or made available to workspace administrators.
Connected services: email, cloud drives, code repositories and calendars can expose far more than the one item you intended to use.
Extensions and plug-ins: every additional service creates another data recipient and another set of permissions.
Outputs and logs: generated content may repeat sensitive input, and monitoring systems may store both.

Consumer, team and API versions of the same product can have different contracts and data controls. Check the terms for the specific plan and configuration you use.

Treat output as untrusted

AI-generated text, code and citations can be wrong even when they sound confident. The person using the output remains responsible for the result.

Verify factual claims against the cited original source.
Review generated code, dependencies and commands before running them.
Do not send AI-generated legal, medical, financial or security advice without qualified review.
Scan output for confidential information before sharing it outside the original audience.
Keep human approval for decisions that affect a person’s access, employment, credit, health or safety.

Connected AI needs smaller permissions

An assistant that can read a document is a privacy concern. An agent that can also send mail, execute code, change records or make purchases is a security boundary.

Grant the minimum data and tools required for the task. Prefer read-only access, narrow scopes, short-lived credentials and a separate test environment. Require confirmation for external communication, deletion, payments, publishing and changes to production systems.

Questions to ask a workplace AI vendor

Is customer data used to train or improve any model, and can that use be disabled?
Where are prompts, files, outputs and logs stored, for how long, and in which jurisdictions?
Who can access the data, including subprocessors and support staff?
Can administrators enforce retention, sharing, connector and model settings?
Are data encrypted in transit and at rest? How are tenant boundaries tested?
What audit logs, export controls and incident notifications are available?
How can all customer data be deleted, and how is deletion verified?
What happens to connected accounts and retained data when a user leaves?

Record the answers. Product settings and terms change; review them on a schedule.

If sensitive information was entered

Stop using the conversation and report it through your organization’s security or privacy process. Do not assume deleting the visible chat deletes every retained copy. Rotate any secret immediately, revoke exposed sessions, preserve the relevant audit details, and ask the provider about retention and deletion under the applicable contract.