Someone on your team is about to paste a customer list into ChatGPT to clean it up. Or a support rep is dropping a full email thread into a chatbot to draft a reply. Or your bookkeeper is asking Gemini to summarize a spreadsheet that still has account numbers in column D.
This happens in every small business now, usually with good intentions and no policy behind it. The work gets done faster. The risk is invisible until it isn’t. And the part most people get wrong is assuming that a paid subscription means their data is private. For the consumer versions of these tools, that assumption is simply false.
Here’s what’s actually true as of 2026, and a checklist you can hand to your team this week.
A paid subscription doesn’t make it private
By default, the consumer versions of the three big assistants use your conversations to improve their models. ChatGPT on Free, Plus, and Pro. Claude on Free, Pro, and Max (Anthropic switched to a train-by-default posture in August 2025). Gemini, where Google also has human reviewers read a sample of chats. In all three cases, the default setting on a normal paid account is not the private one.
“Training on your data” sounds abstract, so be concrete about what it means. The text you paste can be retained, can be reviewed by humans in some cases, and can influence future versions of the model. It won’t pop out verbatim to another user next Tuesday. But you’ve handed a copy of that information to a third party, under terms most people never read, and you can’t pull it back. If that text was a customer’s name, address, medical detail, or payment information, you’ve just made a data-sharing decision on their behalf without telling them.
It helps to remember what these systems actually are under the hood: a pattern-matcher trained on an enormous pile of text, and your inputs can become part of that pile.
Business and enterprise tiers are a different animal
This is the good news, and it’s the single most useful thing to know. The business and enterprise tiers generally do not train on your inputs by default:
- ChatGPT Business, Enterprise, and Education
- Claude for Work / Enterprise, and the Anthropic API
- Google Workspace with Gemini (the business and enterprise editions)
If your company runs on any of these, your exposure is dramatically lower, because the vendor has contractually agreed not to use your prompts to train the model. The trap is the mismatch most small teams don’t notice: the company pays for a business plan, but individual employees are logged into their own personal free accounts on their own laptops. The org has protection. The actual work is happening outside it. Closing that gap is often the highest-value privacy move a small business can make, and it costs nothing but a policy and an afternoon of account cleanup.
The checklist
Hand this to anyone on your team who touches AI tools.
- Classify before you paste. The one habit that prevents most incidents. Before pasting anything, ask: would I be comfortable emailing this to a stranger? Personally identifiable information (names tied to addresses, account numbers, health details, government IDs; see the FTC’s guidance on protecting personal data and the general definition of personally identifiable information) does not go into a consumer AI tool. Ever. General, non-identifying text is fine.
- Strip identifiers instead of pasting raw. You rarely need the real data to get the AI’s help. Replace “Maria Gonzalez, 4412 E Palm Ln, overdue $3,200” with “Customer A, overdue balance.” The model drafts your collection email just as well against a placeholder. You get the output without exposing the person.
- Know which account you’re actually in. Have every employee confirm they’re using the company’s business/enterprise account, not a personal free login, for any work involving customer or company data. This one check neutralizes most of the default-training risk.
- Turn off training on any consumer account that touches work. If someone must use a consumer tier, flip the setting. These menus move around and get renamed, so treat the paths as approximate rather than exact: look in settings for a data-controls or privacy section, then a toggle worded roughly like “improve the model for everyone” (ChatGPT), a model-improvement switch (Claude), or your activity setting (Gemini). If you can’t find it, search the tool’s own help for “training” or “data controls.” Know the limit: opting out stops future training, but it is not a full erasure of what you’ve already sent, and retention windows differ by tool.
- Watch the file uploads and integrations, not just the chat box. Pasting text is the obvious risk. The quieter one is uploading a whole spreadsheet or connecting the tool to your Google Drive or inbox. That’s a much larger surface. Treat any “connect your data” feature as a decision that needs the same classification check as a paste, times a hundred.
- Prefer a governed system for anything involving real records. If your goal is to make the AI genuinely useful against your own documents and customer data, build it into a proper workflow instead of pasting records into a public chatbot. Whether that’s a retrieval setup like the one I described in this piece on RAG for business data or a purpose-built AI-workflow tool, the rule is the same: check how it stores, retains, and trains on your data the way you’d vet any vendor, so your records aren’t quietly feeding someone else’s training pipeline.
- Write it down. A half-page policy beats a great intention. Name which tool and tier is approved, state that customer PII never goes into consumer accounts, and say what to do when someone’s unsure. People follow a rule they can point to. They don’t follow a vibe.
The one-sentence version
If you remember nothing else: the consumer versions of these tools can learn from what you type, the business versions generally don’t, and the fix is mostly about which account your team is logged into and what they strip out before they paste. None of this requires a compliance department. It requires a checklist and ten minutes of everyone’s attention before the next time they hit paste.
This is a general privacy overview, not legal advice. If you handle regulated data (health, financial, or anything under a specific privacy law), confirm your obligations with a qualified professional.
