Is it safe to put my business information into ChatGPT, or will it train on it?
It depends on the account more than the app. OpenAI says it doesn’t train on data from ChatGPT Business, ChatGPT Enterprise or its API by default, and Claude’s work plans run under Anthropic’s commercial terms. Personal plans can be used for training unless the setting is off. For most small businesses the bigger risk is the personal account on an employee’s phone.
So the question to ask is which account the information goes into, and the answer is usually more than one. The plan you pay for is the easy part to get right. The accounts your team set up on their own are where a lot of it actually goes, and a policy that only says don’t use it fails the moment the habit says to.
What each plan does with your data
These are the providers’ own words, checked in November 2026. The pages change, so check them again before you sign anything.
| Account | Does it train on what you put in? | Where it says so |
|---|---|---|
| ChatGPT Free, Plus and Pro | It can, unless “Improve the model for everyone” is turned off | OpenAI Help Center |
| ChatGPT Business, Enterprise and Edu | Not by default | OpenAI business data |
| The OpenAI API | Not by default, unless you opt in | OpenAI business data |
| Claude Free, Pro and Max | Yes, when “Help improve Claude” is on | Anthropic, August 2025 |
| Claude for Work, Education and the API | Covered by commercial terms, outside that setting | Anthropic, August 2025 |
The business plans also come with things the personal ones don’t. OpenAI lists a Data Processing Addendum, controls over how long your data is kept and, for health information, a Business Associate Agreement on ChatGPT for Healthcare. If a customer or an auditor ever asks how you handle their data, those are the documents you’ll want to hand over.
How to turn training off on a personal account
In ChatGPT, open Settings, then Data controls, and turn off “Improve the model for everyone.” OpenAI’s help page says you can do the same thing with “Do not train on my content” in its Privacy Portal, and either one is enough. There’s a catch on the Free, Plus and Pro plans, because if you give a thumbs up or down on a response OpenAI may use that whole conversation to improve its models even after you’ve opted out. Temporary Chat is the other option, since those chats aren’t used for training.
In Claude, open Settings, then Privacy, and turn off “Help improve Claude.” Anthropic’s August 2025 update says it trains on Free, Pro and Max conversations when that setting is on and keeps that data for up to five years, and that a chat you delete won’t be used for future training.
Ask everyone on your team to check this on their own phone. It takes a minute and most people have never looked.
Does it stay inside my company?
Not entirely, and it’s worth being straight about that. When an AI reads one of your files, the contents go to the provider’s servers so the model can read them. What a business account’s terms promise is that your data isn’t used to train the model, plus controls over how long it’s kept, and that’s the protection you’re actually buying.
In the systems I build, your files stay in your own folders and you choose which AI company reads them, so nothing about the setup ties your knowledge to one vendor. If a firm needs nothing to leave the building at all, a model can run on its own hardware. It’s slower, less capable and a different build, which is why I bring it up for the few businesses that need it and leave it alone for the rest.
The account you didn’t pay for
You can buy the business plan, read the fine print and get every answer above right, and it still only covers the account you paid for. I talk to people every day running their own AI accounts on their own phones to do their jobs, and that’s where a lot of client lists, pricing and proposals end up.
The data points the same way. Harmonic Security, which sells AI governance tools, classified almost two million minutes of workplace AI use and found that when employees use their personal AI accounts, about two-thirds of that time is work. On personal plans those conversations can be used to train the model unless the employee has turned that setting off.
A survey for PagerDuty found 88% of office workers had shared work information with public AI tools. 43% had shared emails, 34% customer data and 31% financial information or confidential documents. Those are people at companies with $500 million or more in revenue, which have IT departments and written policies, so read it as a direction for a smaller shop.
Wall Street already paid for this
The same habit has already been through a regulator once. The SEC says its recordkeeping sweep has charged more than 100 firms since December 2021 and brought more than $2 billion in penalties. The messages at issue were on employees’ personal devices, in iMessage, WhatsApp and Signal, as White & Case summarized it. People did business where it was convenient and the firm couldn’t see it.
Now the SEC is looking at AI. Its 2027 examination priorities say examiners will prioritize firms’ policies and procedures for monitoring AI, which the agency now calls “super intelligence.” Most small businesses will never see an SEC examiner. What the sweep shows is what the habit costs when somebody finally adds it up, and AI on a personal phone is the same habit with one difference, because the conversation can also train someone else’s model on your work.
What to do this week
- Pick one business plan for the whole team and pay for it. One account everyone uses is easier to check than five nobody does.
- Write down in two lines what goes into it and what doesn’t, such as client names and pricing in the business account and nothing about clients anywhere else.
- Have everyone turn off training on their personal accounts, using the steps above.
- Make the company tool faster than the phone. Set it up with the files and instructions people actually need, so reaching for their own account is the slower choice.
A ban by itself fails the moment it’s slower than the habit, and even Harmonic, which sells monitoring, says a blanket ban would interrupt legitimate work without fixing what it can’t see. Your people aren’t trying to hurt you. They’re trying to get their work done faster than the tool you gave them allows, so give them a better tool.
Questions owners ask
Does turning off training delete what I already sent?
No. OpenAI's help page says that once you opt out, new conversations won't be used to train its models, so treat anything you sent before that as sent. Anthropic says a conversation you delete won't be used for future training, which makes deleting the sensitive ones worth the minute it takes.
The setting is also per account. Turning it off on your own phone does nothing for the account your office manager signed up for last spring.
Is Temporary Chat safe for client information?
It's better than a normal chat on a personal account. OpenAI says Temporary Chats don't appear in your history, don't create memories and aren't used to train its models.
It's still a personal account, though, and it depends on someone remembering to open a temporary chat every time. For anything with a client's name on it, a business account is the habit that holds up on a busy day.
Can I put customer data into ChatGPT Business?
OpenAI says it doesn't train on ChatGPT Business data by default, and it offers a Data Processing Addendum for companies that need one. That covers the training question.
What it doesn't cover is what you promised your customers. If a contract or a privacy policy says their information stays with you, read it before anything goes in, because a vendor's terms can't change what you agreed to.
What about health information?
That one needs more than a business plan. OpenAI offers a Business Associate Agreement for ChatGPT for Healthcare and for API healthcare customers, which is the agreement HIPAA work generally calls for, and a regular Business plan isn't the same product.
I'm not a lawyer, and this is where one earns their fee. If your business handles patient information, get the agreement and the advice before you set anything up.
Is a model on my own computer safer?
For keeping data in the building, yes, because nothing gets sent anywhere. A model that runs on your own hardware is usually slower and less capable than the big hosted ones, and it's a different build to set up and keep running.
I haven't measured one against the hosted models on client work yet, so I won't tell you how much you'd give up. For most small businesses a business account with the right terms is the practical answer, and the local model is the option for the firm that can't let anything leave.
Sources
- OpenAI, "Business data privacy, security, and compliance" (checked November 18, 2026)
- OpenAI Help Center, "How your data is used to improve model performance" (checked November 18, 2026)
- Anthropic, "Updates to Consumer Terms and Privacy Policy" (August 28, 2025, checked November 18, 2026)
- Harmonic Security, AI Usage Index 2026 (seven weeks ending April 2026)
- PagerDuty, Shadow AI survey by Wakefield Research (June 11, 2026)
- U.S. Securities and Exchange Commission, press release 2024-186 (2024)
- White & Case, "SEC's Focus on Off-Channel Communications Continues" (October 31, 2023)
- SEC Division of Examinations, Fiscal Year 2027 Examination Priorities (October 23, 2026)
Start with the free assessment
The fastest way to find out whether any of this fits is the free assessment. It takes about 15 minutes, in writing or on a call, and I’ll tell you plainly if the audit isn’t worth it for you.