Kaufman Lab · AI architecture
Articles · AI assistants

What data can you put into a public AI chatbot, and what should you keep out?

You can put into a public AI chatbot anything you would not mind showing a stranger: the texts of your own site, general questions, drafts without names. Client contacts, contracts, passwords and keys do not go in, because the service keeps the conversation even with training turned off. You can keep using AI if you first turn off training and remove from the table the name, phone number, email and anything else that identifies a person.

Roman Kaufman, founder of Kaufman Lab · Data as of: October 2026

In August 2026 we ran a five-day pilot of our AI sales manager bot; it takes leads from a website. The test was small, nine conversations. The conversations of such a bot hold client contacts, so we only took a model from a provider that does not train on the data it receives. Whether that was enough depends on where the conversation stays after the chat. The same goes for your leads table: names and phone numbers go to the service along with it.

You can give an AI what you would show a stranger

Check any piece of data with one question: would you show it to someone you do not know? If yes, you can give it to an AI. The table shows the three groups of a site's data, as we divide them in October 2026, and what to do with each before the chat.

GroupExamplesWhat to do
Right awaypage texts, service descriptions, general questions about promotion, drafts without namesgive them as they are
After cleaningreports from Yandex Metrica (Yandex's web analytics tool), the leads table, conversations with a clientremove everything that identifies a person
Neverpasswords and keys, logins to Tilda (a Russian website builder) and hosting, contracts with clientsdo not paste them into the chat

Without cleaning, the AI gets only the first group.

As we see it, the line runs between open text and closed data. Our site rules say that keys, access tokens and server addresses go neither into texts nor into a chat with an AI. The owner enters a key personally.

The leads table is in the second group. Remove the name, phone number and email, keep the request text and the date, and the AI will sort out what people ask about most often. You will check its result by reading it, and what work AI can handle without you is shown in the article on AI employees in business. An export from a CRM (client management software) has more columns, and each one gets the same question: can someone identify a person by it?

A client's name in a foreign chatbot is a legal matter

A client's name and phone number from the leads table fit the definition of personal data in Article 3 of the Russian law "On Personal Data", Federal Law No. 152-FZ as amended on July 26, 2026. In the law's words (our translation), it is "any information relating to a directly or indirectly identified or identifiable individual." Indirectly means the name may be missing if the other fields make it clear who it is about.

152-FZ calls the transfer of such data to the territory of a foreign state, including to a foreign company, cross-border transfer. ChatGPT and Gemini are foreign, and a table with clients in their chat, as we read the definition, looks like such a transfer.

We found no guidance from Roskomnadzor (Russia's data protection regulator) on AI chatbots, so we do not call such input a violation. It is simpler to remove people from the table, and then the question does not arise. How lawyers handle this, with client confidentiality in their documents, is covered in the article what works in AI for lawyers.

The "don't train" switch does not erase the chat

Turning off training settles one question: whether the model learns from your words. The service still keeps the conversation. In ChatGPT, new chats after you turn off training stay in your history. A temporary chat does not go into history, but OpenAI may keep it for up to 30 days for safety. In Gemini with the Keep Activity setting off, new chats are kept for 72 hours.

Google's reviewers read a sample of Gemini chats after disconnecting them from the account. Such chats are kept for up to three years, even if you delete your history. In its Gemini privacy help page, Google itself asks you not to enter anything you would not want a reviewer to see. We checked the terms of both services on October 17, 2026.

We do not call our own leads bot protected, and the reason is the bot itself. Its provider does not train on the data it receives, but the bot's page says plainly: "Personal data protection is postponed until the system scales." In the August pilot there was no separate protection for contacts. If you build your own, ask for it as a separate line in the estimate, like the build and support lines in our AI agent cost breakdown.

Turn off training and remove people from the table

A chat becomes safer not because of the service's name, but because the table has no identifiable people and no keys.

  1. Turn off training right away; it applies to new chats. In ChatGPT it is Settings → Data controls → Improve the model for everyone. The path is described in OpenAI's help article "Data controls in ChatGPT". In Gemini you need the Gemini Apps Activity section and the Keep Activity setting.
  2. Before copying, delete from the table the name, phone number, email and any other fields that make it clear who it is about.
  3. Client data that cannot be cleaned at all goes to a model on your own computer, with nothing sent outside.

For leads from your own site, the first two steps are enough. With a CRM export, you do the second step for every column. An owner with client data that cannot be cleaned needs the third.

Frequently asked questions

Which chatbots are safe for company data?

No chatbot is safe by its name; the risk depends on what you put into it. Even Google asks you not to enter confidential information into Gemini. Turning off training lowers the risk, but the service still keeps the conversation for some time.

How is ChatGPT for business better than the free one?

According to OpenAI, chats on the ChatGPT Business and Enterprise plans are not used to train models by default. On the free plan, you turn off training yourself in the settings. Client names are still personal data there too, so you clean the table the same way.

Can I upload a table without full names?

A table without full names is not ready to upload yet if a person can be identified by other columns. The law counts as personal data anything that identifies a person indirectly, and a phone number or email points to the person even without a surname.

A question about this article?

Message me on Telegram: I answer myself, look at your task and tell you which of this fits it.

Message on Telegram→
More articles
↑ top