Is It Safe to Use AI With Customer Data? A Small Business Guide
Is it safe to use AI with customer data? Yes, on the right plan with clear rules. What AI vendors keep, what never to paste, and a 10-step policy for your team.
On this page
Yes, it can be safe to use AI with customer data, if three things are true: you use a business plan that doesn't train on your data by default, your team follows clear rules about what never gets pasted, and you meet any rules your industry already has. The biggest risk for most small businesses isn't the AI company. It's an employee pasting customer details into a personal account.
This guide is general information, not legal advice. If you're in a regulated industry, run your policy past your attorney or compliance person.
We deal with this daily. East Coast Mortgage handles borrower files under federal privacy rules, and Wander Luxe Lodgings handles guest details for seven homes. Both run on software we built with AI inside it.
What happens to the data I paste into AI?
Three things matter: whether it's used for training, how long it's kept, and who at the vendor can see it.
Training
This is the question most people mean when they ask "is it safe."
- Business plans: OpenAI says it does not train on ChatGPT Business or Enterprise data by default (OpenAI enterprise privacy). Anthropic says it will not use inputs or outputs from Claude's commercial products, including Team, Enterprise and the API, to train models, unless you send feedback or opt in (Anthropic privacy center).
- Personal plans: On ChatGPT Free and Plus, training is on by default and you can turn it off in Data Controls (OpenAI help). On Claude Free, Pro and Max, it depends on the model improvement setting you choose (Anthropic).
Retention
Not training on your data isn't the same as not keeping it. For example, Anthropic says deleted Claude chats leave your history right away and are removed from back end storage within 30 days. If you allow training on a personal plan, de-identified chats can be kept up to 5 years, and chats flagged for policy violations can be kept longer (Anthropic retention). Every vendor has a page like this. Read it before you sign up.
Feedback buttons
A small detail that catches people: on Claude, clicking thumbs up or down can send that whole conversation to Anthropic for training, even on a business plan. Team and Enterprise admins can turn the rating buttons off for the organization (Anthropic). Check whether the tool you use works the same way.
Which AI plans are safer for customer data?
| Plan type | Trains on your data by default? | Admin controls | Use with customer data? |
|---|---|---|---|
| Free personal accounts | Often yes, or depends on settings | None | No |
| Paid personal accounts (Plus, Pro, Max) | Depends on settings | None | Avoid |
| Business team plans (ChatGPT Business, Claude Team) | No | Yes | Yes, with a policy |
| Enterprise plans | No | More: audit logs, SSO, retention controls | Yes |
| API inside your own software | No, for both vendors above | You control it | Yes, built carefully |
The last row is how we do it at our companies. When AI is built into your own app, you decide exactly which fields it ever sees.
What customer data should never go into AI?
Put this list in your policy word for word.
- Passwords, PINs and security answers.
- Full credit or debit card numbers.
- Bank account and routing numbers.
- Social Security numbers and full dates of birth together with names.
- Driver's license and passport numbers.
- Health information, unless you have a signed business associate agreement (more below).
There's almost never a business reason to paste these. "Draft a reply to a customer about a late delivery" works just as well without their card number.
Do industry rules apply to AI?
Yes. The rules you already follow don't go away because a tool is new.
Financial businesses
The FTC's Safeguards Rule covers many nonbank financial businesses, including mortgage brokers, tax preparers and some auto dealers. It requires a written security program, and covered businesses must notify the FTC within 30 days of discovering a breach involving 500 or more consumers' unencrypted information (FTC Safeguards Rule guidance, FTC on notification). An AI tool that sees customer financial data belongs in that program as a service provider.
Health information
HHS guidance says a covered entity or business associate can use a cloud service for protected health information only after signing a business associate agreement with the provider (HHS on cloud computing). AI vendors offer BAAs only on certain products. Anthropic, for example, offers one for Claude Enterprise once an owner turns on HIPAA settings, and for its API, and lists features that aren't covered (Anthropic BAA page). OpenAI also offers BAAs for eligible business customers; check its current list of covered features.
State privacy laws
Many states, including Virginia, have consumer privacy laws that cover how businesses collect and share personal data. Thresholds and duties vary by state, so ask your attorney whether one applies to you.
How do we set up AI safely? A 10-step policy
This fits on one page. Here's what we'd put on it.
1. Pick one approved tool and plan
Name the tool, the plan, and who pays for it. Everything else is unapproved for work.
2. Ban personal accounts for customer work
This is the single most important line. It closes the biggest gap.
3. Turn on the admin settings
Single sign-on if available, two-step login, and turn off any setting that shares data for training. If your plan has rating buttons that send chats to the vendor, decide whether to keep them.
4. List what can never be pasted
Use the six items above.
5. Share the minimum
Teach the habit: first name and the facts of the request, not the whole customer record.
6. Limit connectors
Connecting AI to email, drive and calendar is powerful. Connect only the accounts that job needs, and review them every quarter.
7. Give people only the access they need
At Wander Luxe, interns see the operations app through a read-only role in the database that can't change anything. At East Coast Mortgage, the portal has per-person permissions that widen or narrow what each user can do. AI tools deserve the same thinking.
8. Keep a person in the loop
AI drafts. A person approves anything that goes to a customer. Our nine-agent operations team works this way: agents brief, sort and draft, and a person approves anything that goes out.
9. Ground AI in your own content
When AI talks to customers, limit it to your approved information. Our guest assistant answers from the site's own content and never guesses at prices.
10. Review it twice a year
Vendors change their terms. Put a reminder on the calendar to re-read the data pages and check who still has access.
Is it safer to build AI into our own software?
Often, yes, because you decide what the AI sees. When we built the Closing Disclosure reader for East Coast Mortgage, the AI reads one document and fills in the deal for a person to review. It doesn't browse the whole file system. The same thinking applies to pulling data from invoices, which we cover in How to Automate Data Entry From PDFs and Invoices.
Custom isn't always the answer, though. For most small teams, a business plan plus a written policy covers the risk. If you're choosing between tools, start with ChatGPT vs Claude for Business, and see How to Use Claude for Your Small Business for setup.
The short version
- AI can be safe with customer data on a business plan that doesn't train on it by default.
- The biggest risk is staff using personal AI accounts for customer work. Ban it in writing.
- Never paste passwords, card or bank numbers, Social Security numbers, or health data without a BAA.
- Your industry rules, like the FTC Safeguards Rule or HIPAA, still apply when AI is involved.
- A person approves anything a customer sees, and you review vendor terms twice a year.
Questions people ask
Q: Is it safe to put customer data into ChatGPT or Claude?
A: It can be, on a business plan that does not train on your data by default, with clear rules on what staff may paste. Personal accounts used for work are the bigger risk, because their training settings may be on.
Q: Does AI keep the information I type in?
A: Usually yes, for some period. Anthropic, for example, says deleted Claude chats are removed from its back end within 30 days, and longer retention applies if you allow training or a chat is flagged. Read each vendor's retention page.
Q: Can I use AI with health information?
A: Only if you have a signed business associate agreement with the AI vendor and you use features it covers. HHS guidance says a cloud provider handling protected health information needs a BAA first.
Q: What customer data should never go into AI?
A: Passwords, full card numbers, bank account numbers and Social Security numbers. There is almost never a business reason to paste them, and leaving them out removes most of the risk.
Q: Do I need an AI policy for a small business?
A: Yes. One page is enough: which tools are approved, which plan, what can and cannot be pasted, and that a person reviews anything a customer sees.