Can you put customer data and work documents into ChatGPT

Short answer: personal data about customers and staff should not go into ChatGPT or any other foreign chatbot unless the company has satisfied the cross-border transfer rules in Russian data protection law, and for an employee’s personal account it has not. A contract, a thread with a customer, a CRM export with names and phone numbers all fall in the same bucket. Trade secrets sit outside that law, but the exposure is similar: the text lands on a server you do not control. Banning the tools without replacing them achieves nothing, because people carry on from their phones. What works is a permitted tool inside the company.

We build software and do not give legal advice. What follows is what we see on projects where a company settles this question. The final arrangement is worth agreeing with a lawyer, and if you operate outside Russia the rules that bind you are your own: most jurisdictions regulate this, they simply regulate it differently.

What the law says

Russian data protection law (152-FZ) treats sending data to a server abroad as a cross-border transfer. Since 1 March 2023 the regulator has to be notified before it happens, and extra restrictions apply to countries that are not deemed to provide adequate protection. The United States is not on the adequate list.

On top of that, personal data about Russian citizens has to be collected and stored first in databases located in Russia. An employee copying a customer record into a chatbot breaks both rules at once, and the company answers for it as the data operator.

A first name on its own, an anonymised order number or a contract text with the party details stripped out usually are not personal data. The line runs through whether the text lets you identify a person.

What employees actually send

The policy is usually written about “personal data”, while the leak happens by another route. A sales manager asks for a customer email to be rewritten and pastes the whole thread, signatures included. A lawyer uploads a contract to find the risks in it. An accountant asks for a table to be summarised, and the table has names and amounts in it. Each of them believes they sent “just text”.

Then there is the second half, the commercial one: partner pricing, contract terms, internal procedures. Data protection law does not reach that, but the foreign service’s terms let it keep the conversation, and on free tiers use it for training unless the user has turned that off in the settings.

Why a ban does not hold

ChatGPT is not officially available in Russia, and companies sometimes assume that settles it. In practice people reach it through a VPN, or through Telegram middlemen who forward requests to the same model. The middleman sees the entire text and owes the company nothing.

If the tasks people open a chatbot for are real, a ban simply moves them off the work computer and onto a personal phone. The company stops seeing what goes out and where.

What to replace it with

A Russian cloud model. GigaChat and YandexGPT run on servers inside Russia and are contracted to a legal entity. For drafting, anonymised text and rough copy that is enough. Personal data can also be processed there, under a processing instruction set out in the contract with the provider.

A local model. The model runs on your own server, or on one rented from a Russian cloud, and nothing leaves the perimeter. This is what you need when the prompts contain contracts, medical or financial details, or internal correspondence. What that setup costs and when it pays off is covered in a local LLM for B2B.

Answers from your own documents. People often go to a chatbot for an answer about an internal rule or a contract clause. Search across a corporate knowledge base with a language model on top does that job better, and it stays inside the perimeter. How it is built is covered in RAG on real corporate documents.

Where to start

Ask the heads of departments who already uses chatbots and what for. The answer is almost always wider than expected. Then split the tasks in two: the ones where no person and no secret appears in the text, and the rest. The first group needs a permitted cloud service and a short instruction. The second needs an internal tool, and until that exists those tasks are done the old way.

A ready set of questions for department heads is in the AI rollout questionnaire.

Common questions

Is it fine to send data with the names removed? If the remaining text cannot identify a person, it is no longer personal data. A phone number, an address, or a contract number with a date and an amount often can identify someone, so those come out too.

Does a paid corporate tier of a foreign service solve it? It turns off training on your data, but the transfer is still cross-border, and the legal requirements for that do not go away. Paying for it from Russia is awkward on top.

Who answers if an employee sends a customer’s data? The company does, as the data operator, both to the customer and to the regulator. That is why the chatbot rule belongs in an internal document that staff sign off on.

The rule

Before banning chatbots, find out what people use them for. Hand the tasks with no people and no secrets in them to a permitted service, and move the rest inside the perimeter.