AI at Work

Using AI at Work Without Leaking Company Data: A Practical Checklist

By Lena Ortiz / November 8, 2026

Quick answer: Treat a personal chatbot account like a postcard: fine for public or made-up text, risky for anything your employer or customers would not want read. Consumer accounts can store chats, use them to improve models and, in some cases, let human reviewers read samples. Company-managed accounts usually add contract promises against training, plus admin controls. When unsure, use the approved tool, swap real details for placeholders, or ask first.

Key takeaways

Which account you are signed into matters more than which chatbot you use.
Vendor policies differ by plan and change, so read the page for your plan and check its date.
Customer data, source code, credentials, contracts, unreleased financials and health data stay out of unapproved tools.
Placeholders and smaller excerpts reduce risk but do not make sensitive material safe.
If your company has no AI policy, asking for one is a reasonable request.

What happens to what you paste

Personal (consumer) accounts

Personal plans tend to have looser defaults. Google’s Gemini Apps Privacy Hub (updated October 2026) says that with Keep Activity on, chats can be used to improve Google services, including training generative AI models, and a subset is read by human reviewers. Reviewed chats are disconnected from your account but kept up to three years, even if you delete your activity. The page advises against entering confidential information you would not want a reviewer to see.

Anthropic’s Privacy Center says consumer chats can be used for training if you allow it in the Model Improvement setting, or if a conversation is flagged for safety review. Incognito chats are never used. With the setting on, de-identified data may be kept up to five years, and thumbs-up or thumbs-down feedback stores the whole conversation for up to five years. Deleted conversations leave back-end storage within 30 days.

Company-managed accounts

Work accounts come with contracts. Microsoft’s enterprise data protection page says Copilot and Copilot Chat used by organizations fall under its Data Protection Addendum, with Microsoft acting as data processor, and that prompts and responses are not used to train foundation models. Your organization’s permissions, sensitivity labels and retention policies apply, and interactions can be audited. Web searches run through Bing follow separate handling.

Google’s Workspace privacy hub says customer content, including prompts, is not used to train generative AI models outside your domain without permission, and admins set how long Gemini conversations are kept. Anthropic’s Commercial Terms bar it from training models on customer content from its services. OpenAI’s API documentation says API data is not used for training unless you opt in, with abuse-monitoring logs kept up to 30 days by default. We could not open OpenAI’s ChatGPT plan pages, so check yours directly.

These are promises about the vendor. They do not make every document fair game: your employer’s rules still apply, and your admin may log and retain what you type.

Policies move

These pages get revised. The FTC has warned that quietly loosening data-use terms after collecting data can be unfair or deceptive, but do not count on enforcement to protect you. Re-check the page for your plan and note its date.

What counts as sensitive, and what has gone wrong

What counts as sensitive

Customer data: names, emails, account details, support tickets.
Source code and credentials: private repositories, architecture notes, keys, tokens, passwords.
Contracts and legal drafts: terms, pricing, negotiation positions.
Unreleased financials: forecasts, deal and board material.
Health data and other regulated records: Microsoft’s page says its HIPAA support covers properly configured setups, not web search queries.
Employee and HR data: reviews, pay, complaints.

In the National Cybersecurity Alliance’s 2025 survey, the types of workplace data people reported sharing with AI included internal company documents (50%), client data (44%) and financial data (42%).

Documented incidents and surveys

Samsung, 2023. Bloomberg reported on a company memo saying engineers had accidentally leaked internal source code by uploading it to ChatGPT. Samsung restricted generative AI on company devices and networks, citing worries that data sent to outside servers is hard to retrieve or delete and could be disclosed to other users. An internal survey found 65% of respondents saw these services as a security risk.

Shared chats, 2025. ChatGPT briefly offered a “make this chat discoverable” box when creating a share link, which let search engines index the conversation. Engadget reported that Fast Company found thousands of such chats in Google results, and OpenAI removed the option. Lesson: a share link is publishing.

Surveys. The National Cybersecurity Alliance’s 2026 report (5,000 adults in five countries) found 32% of workplace AI users had shared sensitive company information with an AI tool without their employer knowing, and 43% said their employer had not clearly explained what is acceptable. Microsoft’s 2024 Work Trend Index (31,000 people) found 78% of AI users bring their own tools to work. Security vendor Cyberhaven’s 2026 report, based on its own platform data, put sensitive material at 39.7% of data movements into AI tools. That is vendor-reported, so treat it as a signal, not a census.

Safe to paste, ask first, never paste

This table assumes a personal or unapproved tool. An approved company tool may allow more, so your policy wins.

Safe to paste	Ask first	Never paste
Published or public text	Internal drafts and meeting notes	Passwords, API keys, tokens
General questions (“how do I write this spreadsheet formula?”)	Private code snippets, even small ones	Customer, patient or employee personal data
Your own draft with names and numbers removed	Contracts with names redacted	Signed contracts, NDAs, legal advice
Made-up examples using placeholders	Internal process docs or wiki pages	Unreleased financials, deal or board material

Not sure which column something belongs in? Treat it as “ask first.” Yes, everyone has wanted to paste the whole messy document and type “make this sound smarter.” That urge is exactly why a checklist helps.

A 10-point checklist and how to redact

The checklist

Check which account you are in, personal or company.
Use the company-approved tool for work content.
Read the vendor’s data page for your plan, and note its date.
Opt out of model improvement where you can, but do not treat that as permission.
Remove names, emails, IDs, account numbers and file metadata.
Replace specifics with placeholders.
Paste the smallest excerpt that gets the job done.
Never paste credentials, even to debug.
Do not publish share links; check sharing settings.
Review the output before it leaves your desk, and report slips quickly.

Redacting with placeholders

Example: “Draft a reply to Dana Ortiz at Acme about the $48,000 renewal due March 3” becomes “Draft a polite reply to [Customer] about a renewal due soon, with a price increase of [X]%.” Keep a private note of what each placeholder means, then swap the real details back in yourself.

For code, replace hostnames, keys, internal URLs and company names. Placeholders are not magic: a rare project detail can still identify you, so keep excerpts short.

Getting a policy: how to ask, and what a good one contains

How to ask your manager or IT

Keep it short and practical: “I’d like to use AI for drafting and summarizing. Which tools are approved? Which account should I use? What can I put in them: customer data, code, contracts? Is there a written policy, and who do I tell if I make a mistake?”

What a good company AI policy contains

A list of approved tools and the accounts to use.
A data classification that says what can go where.
Prohibited data, in plain examples.
Human review of AI output before it is used.
What gets logged, retained and who can see it.
Vendor review criteria: training use, retention, admin controls.
Incident reporting with no-blame wording.
Training, an exceptions process and a review date.

NIST publishes a voluntary AI Risk Management Framework and a generative AI profile that policy writers can borrow from.

Common mistakes

Assuming opting out of training makes any paste safe.
Assuming a company account is private from your employer.
Pasting “just a little” code or a customer list.
Using a personal account because it is faster.
Forgetting that screenshots and uploads count as pasting.
Trusting AI output without checking it.

Frequently asked questions

If I turn off “improve the model,” can I paste work data?

Not automatically. That setting covers training, not your employer’s rules. Google’s page says it still uses chats to respond to you and to protect Google, its users and the public even with Keep Activity off, and keeps them 72 hours.

Does a company account mean my chats are private?

Private from vendor training, usually. Not necessarily from your employer. Microsoft’s page notes audit support and your organization’s retention policies, and Google lets admins set retention. Assume your employer can look.

What if I already pasted something I should not have?

Tell your manager or security team promptly, and write down the tool, time and content. Deleting may not erase every copy: Google keeps reviewed chats up to three years, and Anthropic removes deleted chats within 30 days.

Are AI features inside tools I already use safer?

Often, if they run under your company’s work account and contract. Microsoft’s page says to check the terms of any agents you add, because their data handling can differ.

Before you decide

Before you hit paste, take ten seconds. Which account is this? What is in the text? Would I be fine seeing it in a screenshot in the company chat? If any answer is shaky, use the approved tool, add placeholders, or ask.