AI and Russia's Personal Data Law (152-FZ): Staying Legal

In short

Personal data is any information that lets you identify a specific person: full name, phone number, email, address, passport details, a photo, a voice on a call recording. When an employee pastes data like that into ChatGPT or another cloud AI, the company is effectively handing it to a third party — usually to servers outside Russia. That is a direct risk under 152-FZ (Russia's Personal Data Law): since 30 May 2025 fines have multiplied, and a repeat breach carries a turnover-based fine of up to 3% of annual revenue. The practical route for small and mid-sized business: de-identify data before sending it to cloud models, write the rules down in a formal policy, and move bulk processing of raw customer data to a local LLM or a Russian cloud with a data processing agreement. Below is how to do that, without legal jargon and with a self-check checklist.

Important: this article is a practical guide, not legal advice. Bring in a data protection lawyer to audit your documents and processes.

What counts as personal data, in plain words

Personal data is any information relating to a directly or indirectly identified or identifiable person (that is how Article 3 of 152-FZ puts it). The key word is "identifiable": data does not have to contain a name to be personal. A combination of traits that lets you work out who someone is will do.

Almost always personal data:

• full name, phone, email, delivery address, date of birth;
• passport details, SNILS (Russian social insurance number), individual taxpayer number;
• photos, video and voice recordings — including customer calls;
• correspondence with a specific customer and their order history;
• candidate resumes and employee HR records.

Usually not personal data: aggregate statistics ("240 orders this month, average check 3,400 ₽"), information about legal entities with no link to specific people, your own code and copy with no one else's data in it.

A data operator is a company or sole trader that decides for itself what data to process and why. If you have a customer base or even a single employee, you are already an operator — even if you are "just a two-person online store."

Why pasting a customer's name into ChatGPT is a transfer to a third party

A cloud AI is someone else's service, not a program on your computer. The request goes to their servers, is processed there and is stored in their logs. From the standpoint of 152-FZ, three things happen at once.

1. Transfer to a third party. OpenAI, Anthropic or Google are third parties in exactly the way an outsourced call center is. A lawful transfer needs a basis: the person's separate consent or a data processing agreement. Foreign LLM providers work under their own standard terms and do not adapt them to Russian 152-FZ.
2. Cross-border transfer. Most cloud models keep their servers outside Russia, and an operator must notify Roskomnadzor (the Russian data protection regulator) in advance of any cross-border transfer.
3. Loss of control. You cannot delete data from someone else's logs and training sets, and you will not hear about an incident on the provider's side.

The company answers for all of this, not the employee: "the account manager pasted it himself" is not a defense. Until the rules are written down, any such action is a violation that is yours by default. So the first step is not banning AI but a clear AI usage policy for your company.

What you can and cannot send to cloud AI

The working rule: the cloud model should receive the task, not the customer's identity.

Logins and API keys are not personal data, but leaking them hurts more than many fines, which is why they are in the "never" column on purpose. That side of the question gets its own write-up: risks and security in vibe coding.

When de-identification is enough

De-identification means removing or replacing everything that ties data to a specific person: names, contacts, contract numbers, addresses, distinctive details. If the text no longer identifies anyone even indirectly, then as far as the AI is concerned it is not personal data — and that covers most everyday office tasks.

What it looks like in practice:

• Before: "Pyotr Ivanov, +7 916 123-45-67, order No. 4812, demanding a refund." After: "Customer, order for 12,000 ₽, demanding a refund."
• In a call transcript, names become "Manager" and "Customer," and phone numbers and addresses are cut.
• In an analytics export, amounts, dates and categories stay — the contact columns are deleted.

Two traps. First: a combination of indirect traits ("the salon director at 5 Lenina Street") identifies a person just as well as a name — strip out rare details too. Second: manual de-identification only works if the rule is written down and people are trained on it, otherwise the exports will be flying back into chat within a month. How to build those habits in a team is covered in our guide to training employees to work with AI.

When you need a local model, and what it costs

A local LLM is a language model running on your own hardware: the data never leaves the company perimeter, and the question of third-party transfer disappears entirely. It makes sense when raw personal data has to be processed in bulk: transcribing every call, parsing tickets that contain contact details, handling documents with passport data.

Ballpark prices as of July 2026:

• A workstation for open 7–14B models (Qwen, Llama, Mistral): from about 250,000 ₽ — a desktop with an RTX 4090/5090 and 64 GB of RAM. Enough for transcription, ticket classification and draft replies.
• A server for large models (H100 80 GB): 2.5–4 million ₽ from Russian integrators — the "our own ChatGPT" tier for a mid-sized company.
• Renting GPUs in a Russian cloud: roughly 80 ₽/hour for an RTX 4090 and about 340 ₽/hour for an H100, which works out to roughly 60,000–250,000 ₽ a month running around the clock. The hardware is someone else's, but the servers are in Russia and the contract is with a Russian legal entity.

A middle option is Russian cloud models (YandexGPT, GigaChat): servers in Russia, and you can sign a data processing agreement with the provider. That is legally simpler than a foreign cloud, but it does not remove your duties as an operator — consents, notifying Roskomnadzor, a processing policy.

A selection of local models and tools for running them is in the Tools section of the arsenal, and we compared free cloud alternatives in our roundup of free AI models in 2026.

2026 fines and the mistakes businesses keep making

A new schedule of fines (Article 13.11 of the Code of Administrative Offenses) has been in force since 30 May 2025 — this is the main argument to bring to an owner:

• failing to file a personal data processing notification with Roskomnadzor — 100,000–300,000 ₽ for legal entities and sole traders;
• a breach affecting 1,000–10,000 people — 3–5 million ₽; over 100,000 people — up to 15 million ₽;
• failing to report a breach to Roskomnadzor within 24 hours — 1–3 million ₽;
• a repeat breach — a turnover fine of 1–3% of annual revenue, from 20–25 up to 500 million ₽ (overview of the changes — ConsultantPlus).

Since December 2024 there has also been a criminal offense for illegal trade in personal data sets (Article 272.1 of the Criminal Code) — it targets data brokers, but it shows the direction of travel. For a small business the realistic scenario is not a police raid but a complaint to Roskomnadzor from an unhappy customer or a former employee, and an inspection that follows it.

The mistakes we see most often:

1. "We're small, this doesn't apply to us." Nearly every company and sole trader is required to file a processing notification — for free, through the Roskomnadzor website, in one evening. The fine for staying quiet is up to 300,000 ₽.
2. Employees' personal accounts. The company "banned ChatGPT," and the account managers use it from their personal phones. The only thing that works against shadow AI is approved use cases plus training, not a ban.
3. Exporting the CRM "for segmentation." A marketer sends a table of phone numbers to the cloud "so the AI can split them into segments." Segmentation works perfectly well on de-identified data.
4. Believing in blanket consent. A standard customer consent covers processing for your purposes, not transfer to an arbitrary foreign service.
5. AI agents with full access. Automation where an AI agent goes into the CRM and writes to customers on its own needs the same limits: least privilege, logging, de-identification at the input. Practical cards are in the Security section.

Self-check checklist

Those six lines are minimum hygiene. The full AI adoption route, covering processes and people, is in our AI adoption checklist for business.

An honest note about us: qvib is the Quest vibe-coding engine and a free knowledge base, built by one person working with AI. We do not sell implementation services or legal audits; the templates and checklists from these articles are published in the pinned post of the qvib Telegram channel.

FAQ

Can a business use ChatGPT legally at all?

Yes. The law regulates the processing of personal data, not neural networks. Code, copy, marketing, analytics on de-identified data — all fine. The risk starts where data about specific people goes to the cloud without a legal basis.

The customer signed a data processing consent — is that enough?

No. A standard consent covers processing for your purposes: delivery, the contract, mailings. Transfer to a foreign service requires a separate basis, and a cross-border transfer requires notifying Roskomnadzor in advance. In practice, de-identifying is simpler and safer.

Employees have already pasted customer data into AI tools. What now?

Stop the practice, assess the scale, adopt a policy and train the team. A one-off paste of a name into a chat is a breach of processing rules, but not yet a "leak" with million-ruble fines. Systematically exporting the database is a reason to see a lawyer without waiting for complaints.

Is a local model much worse than ChatGPT?

For narrow tasks — transcription, classification, template-based drafts — open 7–14B models are usually enough. Do not expect the breadth of the flagships, so the working setup is this: the local model for data, the cloud for everything else.

Do I need a lawyer, or can we handle it ourselves?

Basic hygiene — the Roskomnadzor notification, the policy, de-identification, training — a small business can handle on its own. A lawyer is needed to audit documents (processing policy, consents, processing agreements) and is essential in any situation involving a real breach.