How to Use AI at Work Without Leaking Confidential Company Data

Why this matters more than people think

In early 2023, engineers at Samsung did something completely reasonable. They used ChatGPT to help fix some code and tidy up meeting notes. The problem: the code was proprietary, the notes were internal, and at the time, anything typed into ChatGPT could be used to train the model. Three separate leaks happened in under 20 days, and Samsung ended up banning the tool on company devices entirely. [1]

Here is the uncomfortable part. Those engineers were not careless. They were doing exactly what most of us do every day: reaching for the fastest tool to get the job done. And that is precisely why this is worth your attention.

The numbers back it up. According to Cyberhaven, around 11% of everything employees paste into ChatGPT is confidential company data, and a meaningful share of workers do it every single week. [2] Meanwhile, the average data breach in 2025 cost organisations $4.44 million globally, and a record $10.22 million in the US. [3] Most of those breaches have nothing to do with AI. But “I pasted the wrong thing into a chatbot” is now a genuine way to land on that list.

This is closely related to what people call shadow AI: staff quietly using tools their employer never approved. If you want the bigger picture on that, we covered it in detail in our guide on shadow AI. For now, let’s keep it practical.

What actually happens to what you type?

Let’s clear up the single biggest source of confusion: not all AI accounts are the same, and the difference is the whole game.

When you use a free or personal ChatGPT account, your conversations can be reviewed and used to improve future models unless you turn that off. That does not mean a human is reading your messages over your shoulder. It means your input can become part of the giant pile of text used to train the next version. Once it’s in there, you cannot pull it back out.

Business accounts work differently. OpenAI states plainly that it does not train on data from ChatGPT Team, Enterprise, Edu, or its API by default. [4] Same story with Microsoft Copilot for work and Claude’s Team plans: your prompts are not used to train the underlying model. The technology is identical. The data promise is not.

The one setting to change today

If you only do one thing after reading this, do this. On a personal ChatGPT account, open Settings, then Data Controls, and turn off “Improve the model for everyone.” That single switch stops your future chats from being used for training.

Even better for one-off sensitive questions: use Temporary Chat (the icon at the top of a new chat). It works like an incognito window. The conversation is not saved to your history and is not used for training. When you’re done, it’s gone.

Claude and Gemini have equivalent controls in their settings. The exact menu names shift every few months, so the rule of thumb is: go into settings, find anything labelled “data,” “training,” or “improve the model,” and switch it off for anything work-related. Takes two minutes. Saves a lot of grief.

What you should never paste into a public AI tool

This is the checklist I give every team I train. Print it, pin it, whatever works. If the information is on this list, it does not go into a free or personal chatbot:

• Anything that identifies a real person: customer names, emails, phone numbers, addresses, health or HR details. This is also a data-protection issue under rules like GDPR, not just a company-secrets one.
• Unreleased or proprietary material: source code, product roadmaps, pricing you haven’t announced, financials, M&A plans.
• Anything covered by an NDA or client confidentiality: if a client would be upset to see it in a stranger’s hands, don’t paste it.
• Login details and security info: passwords, API keys, internal system names. Obvious, but it happens constantly.
• Legal or contract drafts with real party details: fine to ask about a clause in general terms, not fine to paste the whole signed agreement.

Notice the pattern. The issue is almost never the question you’re asking. It’s the raw confidential data you paste in to get the answer. Which leads to the fix.

How to use AI on sensitive work, safely

You do not have to choose between “use AI” and “protect the company.” You just have to be a little smarter about the inputs. Three techniques cover almost everything.

1. Redact and use placeholders. Instead of pasting a real client email, swap the specifics for tags. “Draft a renewal note to [CLIENT], whose contract for [PRODUCT] ends on [DATE], who has raised [ISSUE].” The AI writes a perfect template. You fill in the real details yourself, offline. The model never sees anything sensitive.

2. Use synthetic or sample data. Need help analysing a spreadsheet of real customer numbers? Give the AI a small, made-up version with the same structure, get the method or formula, then apply it to the real file yourself. (We walk through this approach in our guide on using AI for small-business workflows.)

3. Move real work into a business-grade account. If your team regularly needs to put company information into AI, that’s the signal to get ChatGPT Team, Copilot, or Claude for Teams. You get the no-training guarantee, admin controls, and a clear audit trail. For a modest monthly fee per person, it turns a risky habit into a managed one.

The five-line AI policy any manager can roll out

Most teams don’t have an AI policy. They have a vague feeling that they probably should. If you manage people, you can close that gap this afternoon. You don’t need a 20-page document drafted by legal. You need five clear lines that a busy person will actually read:

1. Approved tools: “Use [X] and [Y]. Don’t use random new AI tools without checking first.”
2. The red list: “Never paste customer data, unreleased information, passwords, or anything under NDA into a public AI tool.”
3. The safe default: “For sensitive tasks, use our business account or redact the details first.”
4. The settings rule: “Turn off model training on any personal account you use for work.”
5. The ask-first line: “Not sure if something’s okay to paste? Ask [name] before you do.”

That’s it. The point of a policy isn’t to scare people off AI. It’s to give them enough confidence to use it well. Teams that ban AI outright just push it into the shadows, which is usually worse. We’ve seen this play out again and again, and it’s a big reason so much corporate AI training fails.

Tool-by-tool: where the data settings live

A quick reference for the four tools most non-technical teams actually use:

• ChatGPT (personal): Settings, Data Controls, turn off model improvement. Or use Temporary Chat. Team/Enterprise: no training by default.
• Claude: Settings, Privacy. Free and Pro plans have controls over training; Team and Enterprise plans don’t train on your data.
• Google Gemini: Activity settings let you turn off Gemini Apps Activity. Note that human reviewers may see flagged conversations, so the same redaction rules apply.
• Microsoft Copilot (work version): built for business; your prompts and company data aren’t used to train the foundation models. It’s a natural default for organisations already working inside the Microsoft ecosystem.

One more thing worth saying out loud: AI tools also make mistakes, and a confidential-data leak isn’t the only risk. They’ll invent facts too. If accuracy matters for your task, pair this guide with our anti-hallucination toolkit so you’re protected on both fronts.

What to do this week

Three steps, in order. First, change the setting: open your AI tool today and switch off model training, or start using Temporary Chat for anything sensitive. Five minutes, done.

Second, share the red list with your team. Paste those five “never” items into a Slack message or a team channel. You don’t need permission to make your colleagues a bit safer.

Third, if your team uses AI on real company data more than occasionally, make the case for a business account. The math is easy: a modest per-seat monthly fee versus the cost and embarrassment of explaining a leak to a client.

You don’t have to be a security expert to use AI responsibly. You just have to know what the tools do with your words, and decide, on purpose, what you’re willing to share. That’s a skill any professional can learn, and it’s one worth having.

Related reading

Shadow AI: What It Is, Why It’s Everywhere, and What You Should Actually Do About ItClaude for Small Business: 15 AI Workflows That Plug Into QuickBooks, HubSpot, and the Rest of Your StackWhy Most Corporate AI Training Fails (And What 2,000+ Learners Taught Us About What Actually Works)

Sana Mian — Co-Founder, Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →