What Customer Data Can Restaurants Share With ChatGPT and Claude?

Tuesday night, the marketing director of a seven-location burger group pastes last month's reservation export into ChatGPT. Names, phone numbers, allergy notes typed in by the host stand: everything goes in. Was that allowed? Partly, and only on the right plan. When restaurants share customer data with ChatGPT or Claude, the plan matters more than the tool.

It was the question we heard most during our September 30, 2026 webinar. Can we put guest data into ChatGPT or Claude? A close second followed. When a guest leaves a review, can we let AI write and post the reply for us?

This article builds on our guide to using ChatGPT and Claude in a restaurant group. It starts from the files your teams handle every week, from the reservation book to the schedule, and sorts what can go into AI from what must stay out.

Key takeaways

• The plan decides: on ChatGPT Business or Enterprise and Claude Team or Enterprise, your conversations do not train the models by default.
• Menu allergens are fine: menus, recipe cards and dish descriptions can go into AI with no special precaution.
• Guest allergies never are: an allergy noted next to a guest's name is health information, treated as sensitive data under California law.
• Reservations, loyalty, schedules: strip names and contact details, or work from totals by shift.
• No GDPR, but rules still apply: the CCPA, about 20 state privacy laws and the FTC all reach restaurant groups.
• Sensitive reviews: food poisoning, allergic reactions, injuries or a named employee get an AI draft, never an AI-published reply.
• Hallucinations: an invented holiday hour or a dish wrongly labeled gluten-free gets checked before it goes live.

Summarise this article with:

C ChatGPT C Claude P Perplexity G Grok

Do ChatGPT and Claude train on your data?

On personal plans, they can. On business plans, not by default. That distinction matters far more than picking one assistant over the other.

At OpenAI, consumer conversations help train the models as long as the "Improve the model for everyone" setting stays on, according to its data controls FAQ. For Business and Enterprise, OpenAI does not train on your data by default, per its enterprise privacy page.

Anthropic took a different route. Since its August 28, 2025 consumer terms update, every Free, Pro or Max user chooses whether their chats train the models. If they opt in, chats are kept for five years, versus 30 days otherwise. Claude Team, Enterprise and the API sit outside that choice.

Personal vs. business plans, side by side

In most groups, both coexist. Headquarters pays for a team workspace, while a chef or general manager uses a personal account on their phone. The table compares the four situations.

The takeaway is simple. A personal account is fine for testing the tool on a menu or public reviews. As soon as an internal file is involved, move to a business plan: only that tier brings the contract, access management and no training by default.

Your restaurant files, from green to red

A restaurant group produces the same files every week. One rule sorts them: the less a file identifies a person, the more freely you can hand it to AI.

The classic trap is allergies. Your dishes' allergens are not personal data, and AI can help you check them across the whole menu. A guest's allergy is a different story. Typed into a reservation note, it becomes health information about a named person.

For the reservation book, export totals rather than lists. Covers per time slot, no-shows by day of the week or average party size are enough to plan a holiday rush, without a single name.

Reviews, finally, are safe once reviewer names are removed. Our method to analyze restaurant reviews with AI walks through that preparation, location by location.

No GDPR, but US rules still apply

The US still has no federal privacy law like Europe's GDPR. That does not leave restaurant groups in the clear. Three layers apply, and a multi-unit group often meets more than one.

The first is California's CCPA. It covers businesses with more than $26.6 million in annual revenue, the threshold set by the California Privacy Protection Agency for 2025, among other criteria. Health information counts as sensitive personal information under the law.

The second is the states. About 20 now have a comprehensive privacy law in force, according to this state law tracker. Washington's My Health My Data Act goes further, with no size threshold and a private right of action. A guest's allergy could plausibly fall under it.

The third is the FTC, which polices unfair and deceptive practices, including around reviews. In practice, four habits cover most of the risk.

• Sign the contract: OpenAI offers a data processing agreement for Business and Enterprise, and Anthropic builds one into its commercial terms.
• Share only what you need: remove names and phone numbers before uploading an export, and turn off training where the plan allows it.
• Write a one-page AI policy: who uses AI at headquarters and in each location, which files are off-limits, who reviews before anything goes live.
• Loop in counsel: state laws differ, and in Canada (PIPEDA) and the UK (UK GDPR) similar principles apply.

The policy matters as much as the contract. One general manager pasting their team's schedule into a personal account undoes everything else. Present it at your next GM meeting, then open business seats for those who need them.

The same care applies to connectors, which open your inbox or Drive to AI. Here are the admin settings to put in place before connecting ChatGPT or Claude to your restaurant tools.

The reviews AI should never answer alone

AI handles everyday reviews very well, a lukewarm burger or a long wait at the bar. Some reviews, though, should never get a reply without a human check. On Google, a reply is public, signed by your location, and the reviewer is notified. Google's own reply tips say to never share a reviewer's private information.

During our webinar, we gave Claude a review reporting food poisoning, with no special instruction. The generated reply implicitly admitted fault. Posted as is, it would have been read by your future guests, and by your insurer.

Five kinds of reviews should be flagged and put on hold.

• Health: food poisoning, feeling sick, allergic reaction or intolerance.
• Safety: accident, injury, a fall, a foreign object in a dish.
• People: a server named or accused, an argument in the dining room.
• Serious accusations: discrimination, hygiene, fraud.
• Threats: a lawsuit, an attorney, a report to the health department.

For these, AI can suggest a draft. The general manager approves it before it goes live, without repeating the guest's name or health condition. Write this rule into your AI project in your brand voice.

One more line not to cross. The FTC's rule on consumer reviews, in force since October 21, 2024, bans fake reviews, AI-generated ones included. Knowing violations can cost up to $53,088 each, under the current penalty table. AI writes your replies, never your reviews.

For everyday reviews, our 30 Google review response examples give you a safe base. Across a group, sorting has to become automatic: the Malou AI agent flags sensitive topics in reviews across every location.

Hallucinations, check before you post

A hallucination is a mistake stated with confidence. An invented Thanksgiving hour, a dish listed as gluten-free when it is not, a sales figure that exists in no file. The risk drops when AI works from your data, but it never disappears.

Three habits are enough. Ask for the source of every number, have each location check hours, prices and allergens, and never copy an answer as is. Our restaurant review management guide shows how to keep replies consistent once they scale.

Want to see where your group stands before rolling AI out? Run a free visibility audit across all your locations.

Frequently asked questions

Does ChatGPT use my data for training?

On Free, Go, Plus and Pro plans, yes, as long as the "Improve the model for everyone" setting stays on. You can turn it off in Data Controls, and Temporary Chats are not used for training. On ChatGPT Business, Enterprise and Edu, OpenAI does not train its models on your data by default.

Is ChatGPT CCPA compliant?

Compliance depends on how you use it, not only on the tool. On Business or Enterprise, OpenAI offers a data processing agreement and does not train on your data by default. You still need to limit what you share, keep sensitive data such as guest allergies out, and set an internal AI policy for your teams.

Can I upload my OpenTable or Resy guest list to ChatGPT or Claude?

Not as is. A guest list holds names, phone numbers and sometimes allergy notes, which count as health information. Export totals instead, such as covers per time slot or no-shows by day, and use a business plan. Those figures are enough to prepare a busy service or adjust a schedule.

Can AI reply to my restaurant reviews on its own?

For everyday reviews, yes, if it knows your brand voice and each location's context. For reviews that mention food poisoning, an allergy, an injury, an employee or legal action, no. An unguarded reply can admit liability, so these reviews should be put on hold and approved by a manager.

Is it legal to use AI to write restaurant reviews?

No. The FTC's rule on consumer reviews, in force since October 21, 2024, bans fake reviews, including AI-generated ones, and knowing violations can cost up to $53,088 each. Using AI to draft your replies to real reviews is a different matter: it is allowed, as long as a human checks sensitive cases.

At Malou, we help restaurant groups manage their local visibility — listings, reviews, content — across every location, and reply to reviews at scale without losing control of sensitive cases. Test your group's visibility with our diagnostic tool.