AI at Work

Using AI at Work Without Leaking Company Data

Staff already paste work into AI chatbots. See what ChatGPT, Claude, Gemini and Copilot do with that data, plus a one-page policy and a 30-day rollout plan.

Last updated: November 4, 2026

In most offices, someone has already pasted a client contract, a salary spreadsheet or a chunk of source code into a public AI chatbot to “just tidy it up.” They weren’t being careless on purpose. They were trying to finish work faster with a tool that is genuinely good at it.

The trouble is what happens next. On a personal account, that text may be used to train future models, it sits on the provider’s servers for a while even with training off, and the company can’t see, manage or delete it.

Banning AI tools outright rarely works. People switch to their phones, where you have even less visibility. A better approach is to know where the risks are, give people an approved tool, and set a few rules they can follow. This guide covers what the four biggest providers say about training and retention today, a one-page classification you can copy, and a 30-day rollout plan. It describes general workplace practice, not legal advice.

To keep it concrete, we’ll follow a made-up example: a 35-person marketing agency in Charlotte, North Carolina, where the office manager, Jessica, has been asked to “sort out the AI thing” before a big healthcare client signs.

Where the risk actually comes from

1. Data used for training

On personal plans, chats can be used to improve the provider’s models unless someone changes a setting. Defaults differ and change, so check current help pages. As of November 2026:

  • ChatGPT: on personal Free, Plus and Pro accounts, OpenAI says data sharing is on by default. You can switch off “Improve the model for everyone” under Settings > Data controls and keep your chat history.
  • Claude: since Anthropic’s consumer terms update in 2025, Free, Pro and Max users choose whether their chats and coding sessions can be used to improve Claude. If they allow it, Anthropic may keep that data in de-identified form for up to five years. If they don’t, the 30-day retention period continues.
  • Gemini: for personal Google accounts with Keep Activity on, Google says it uses your activity to improve its services, including training generative AI models, and a subset of chats is read by human reviewers. Google’s own advice is not to enter confidential information you wouldn’t want a reviewer to see.
  • Microsoft Copilot (personal): Microsoft says prompts, responses and file contents in the updated Copilot app, released August 18, 2026, aren’t used to train foundation models. The older version of the app had a training opt-out instead, so which rule applies depends on the version someone is running.

2. Data stored on someone else’s servers

Turning training off does not mean the data disappears. Providers keep conversations for a period to run the service and catch abuse:

  • ChatGPT temporary chats don’t appear in history and aren’t used for training, but OpenAI may keep a copy for up to 30 days for safety purposes.
  • Gemini temporary chats, and chats with Keep Activity off, are kept with your account for 72 hours. Chats that human reviewers have already looked at are kept for up to three years, disconnected from your account, even if you delete your activity.
  • Claude deletes a conversation from its back-end systems within 30 days of you deleting it. Chats flagged for Usage Policy violations can be kept for up to two years.

That may matter for confidentiality agreements, industry rules, or contracts that limit where client data can go.

3. The feedback button

This one surprises people. Clicking thumbs up or thumbs down can send the whole conversation to the provider, even when training is off. OpenAI says the entire conversation tied to that feedback may be used for training. Anthropic stores the related conversation for up to five years. Google says reviewed feedback and the chats around it are kept for up to three years. Tell staff not to rate chats that contain anything sensitive. On Claude Team and Enterprise plans, an owner can turn the rating button off for the whole organization.

4. Accounts you don’t control

When an employee uses a personal account for work, the company can’t manage its settings, review usage, or delete data when that person leaves. If a designer at Jessica’s agency quits with a personal ChatGPT account full of client briefs, the agency can do nothing about it.

5. Connected tools, plugins and agents

AI assistants now connect to email, calendars and cloud drives, and some act on their own. Each connection means more data access. Google, for example, says it does not monitor or secure custom third-party connected apps.

Work tools carry a quieter version of this risk. Microsoft 365 Copilot only shows people content they already have permission to view. If a folder of salary data was shared with “everyone in the company” years ago, Copilot can surface it to anyone who asks the right question. Fix the old sharing settings before you switch it on.

Tools can also change connections without asking, as when Postman auto-installed plugins into developers’ AI coding agents. If your team runs agents that act on their own, our guide to designing safe AI agent workflows covers permissions, checkpoints and logs.

What business plans change

What a business plan mainly buys is control: no training by default, admin-set retention, and accounts you can close when someone leaves.

If your team uses Claude, see our Claude Data admin checklist before the Nov 19 default change.

Provider Business offering What the provider says
OpenAI ChatGPT Business, Enterprise, Edu; API Not used for training by default. Workspace admins control retention; deleted conversations are removed within 30 days unless the law requires otherwise. API data may be kept up to 30 days for abuse monitoring.
Anthropic Claude for Work (Team, Enterprise); API Not used for training by default. API inputs and outputs are deleted within 30 days except in listed cases. Deleted chats leave back-end systems within 30 days.
Google Gemini with a qualifying Google Workspace edition Content not used to train models outside your domain or human-reviewed without permission. Admins set retention: 90 days to indefinite in Workspace apps, up to 36 months in the Gemini app.
Microsoft Microsoft 365 Copilot (work account) Prompts, responses and Microsoft Graph data aren’t used to train foundation models. Admins can set retention for Copilot chats in Microsoft Purview.

In every case, submitted feedback can still be used, and data can be kept longer where the law requires or to investigate abuse. Read the terms for the plan you buy.

A simple traffic-light system

Long policy documents get ignored. A short classification that fits on one page works better.

Level Examples Rule
Green Public information, your own general writing, generic questions, published marketing copy Fine to use with any approved AI tool.
Amber Internal documents, non-sensitive business plans, anonymized data, internal code Only in company-approved AI tools with business accounts and training switched off.
Red Customer personal data, health or financial records, Social Security numbers, passwords and keys, legal matters, anything under a confidentiality agreement Not in AI tools unless a specific, reviewed system has been approved for it.

Most day-to-day use is green or amber, so the policy says yes to most requests, which is what makes people follow it. At Jessica’s agency, drafting a blog post for a client’s public website is green. Summarizing last quarter’s internal campaign results is amber. Anything from the healthcare client’s patient list is red.

Practical rules that help

  1. Provide an approved tool. If you want people off random free accounts, give them a sanctioned option with business terms. When you compare options, check the data terms alongside quality and price, as in step 7 of our guide on how to choose an AI model.
  2. Use company accounts, not personal ones. That way settings are managed centrally and access ends when someone leaves. Add AI accounts to your offboarding checklist next to email and Slack.
  3. Check the data settings yourself. In the admin console, review retention, training, connected apps and feedback. Write down what you chose, why, and when.
  4. Remove details before pasting. Replace names, account numbers and other identifiers with placeholders. “Customer A was charged $49.99 twice in March” works just as well for drafting a reply as the real name and card number. The same habit applies when you use AI to polish everyday messages, as in our guide to writing clearer work emails when English is your second language.
  5. Never paste credentials. Passwords, API keys and access tokens should never go into a chatbot, even an approved one. If someone does paste one, treat it as exposed and rotate it.
  6. Don’t rate sensitive chats. Skip the thumbs up and thumbs down buttons on anything amber or red, or turn feedback off at the admin level where the plan allows it.
  7. Review connected apps. Once a quarter, check which integrations are enabled and what they can access, and remove the ones nobody uses.
  8. Keep a human responsible. Anything AI helped produce is still the responsibility of the person who sends it. That includes checking it for accuracy, as covered in why AI models make things up.

If someone has to use a personal account

If a contractor or new hire must use their own account for now, three habits reduce the exposure:

  • Turn off model training: “Improve the model for everyone” in ChatGPT, the model improvement setting in Claude’s Privacy Settings, or Keep Activity in Gemini Apps Activity.
  • Use temporary or incognito chats for one-off work. ChatGPT, Claude and Gemini say these aren’t used for training, though a copy may be kept briefly (OpenAI says up to 30 days, Google 72 hours).
  • Keep everything green-level only. No client names, no internal numbers.

A 30-day rollout plan

Here is how Jessica could get from “no policy” to “working policy” in a month:

  1. Week 1: Ask the team, without blame, which AI tools they use and for what. You can’t set rules for tools you don’t know about.
  2. Week 2: Pick one approved tool on a business plan. Set training, retention and feedback options, and record the choices.
  3. Week 3: Write the one-page traffic-light list with five or six examples from your own work, then hold a 20-minute walkthrough.
  4. Week 4: Move people off personal accounts, add AI accounts to offboarding, and name one person who answers “is this OK?” questions within a day.

What to tell your team

Keep the message positive. Something like: “We want you to use AI to save time. Use our approved tool, keep red-level data out of it, remove names where you can, and check anything before you send it.” Follow it with a few concrete examples from your own business. Examples stick far better than rules.

If the only options are guessing or waiting a week for an answer, people will guess. That is why the plan names one person to answer quickly.

For small businesses

You don’t need a legal department to get this right. A one-page list of what is fine, what needs care and what is off limits, plus a business account with sensible settings, covers most of the risk. If you’re still deciding which jobs to hand to AI in the first place, start with our AI workflow for small business and apply the traffic-light rules to each job. Revisit it every six months, because the tools and their terms keep changing.

If you handle particularly sensitive data, such as medical, legal or financial records, take professional advice on your obligations before adopting any AI tool for that work. Some business plans offer extra agreements for regulated data; OpenAI, for example, says it can sign Business Associate Agreements for HIPAA purposes on its API. Whether that is enough is a question for your attorney or compliance adviser.

Sources

Kabeer Ahmed

Kabeer Ahmed is the founder and editor of The AI Year, where he explains AI tools, models, agents and AI security in plain English. He works on AI automations for businesses and is the founder of Runnex Logistics and of Aurel Digitals, a web design and automation agency. He writes practical guides for people using AI at work, covers selected news with links to primary sources, and reviews every post before it's published. Reach him at hello@theaiyear.com or on LinkedIn: https://www.linkedin.com/in/kabeerahmed09

Keep reading