An internal ChatGPT: giving staff AI without giving away your data
"We don't need AI" is a phrase that almost always means it is already in use — just without your knowledge. Sales people paste proposals into a chatbot, developers paste chunks of working code, lawyers paste draft contracts. They do it through personal accounts and workarounds, because there is no corporate access: neither OpenAI nor Anthropic serves Russia — subscriptions cannot be paid for legally and sign-in is closed by region. The company ends up with the worst of both worlds: nobody measures the benefit of AI, and the data is already leaking.
Open WebUI is an open web interface for language models that installs on your own server. It looks like ChatGPT: a list of chats on the left, an input box, history, folders, shared prompt libraries. The difference is what happens behind the interface: staff sign in with a corporate login, an administrator manages roles and groups, and every request stays in your log. The interface itself contains no model — it only connects users to one, which is precisely why the choice of model stays yours.
The first path is a model on your own hardware. Open models long ago caught up with commercial ones on everyday work tasks: Qwen, DeepSeek, Mistral, Gemma and Llama are all downloadable, and for Russian there are domestic open models such as T-lite and T-pro from T-Bank. They run via Ollama (simpler) or vLLM (faster under load). As a hardware yardstick: a 7–14 billion parameter model lives on a single 24 GB card and serves a department; 30–70 billion parameter models need 48–160 GB, meaning several cards or a dedicated server. Such a setup is fully self-contained — it works with no internet at all, which for closed network segments is often the only permissible option.
The second path is a Russian provider's API. When buying GPUs makes no sense, Open WebUI connects to any OpenAI-compatible API — that is how YandexGPT in Yandex Cloud or Sber's GigaChat get wired in. You pay for usage in roubles under an ordinary contract, circumvent nothing, and the data stays in Russian jurisdiction. The model is somebody else's, so in practice we often combine the two: sensitive work on your own model inside the perimeter, everything else through an API where it is cheaper and more capable.
A separate kind of value appears once the model is connected to your documents. Open WebUI indexes policies, manuals, technical documentation and contracts, then answers from them with links to the source. "How do I file a business trip under our policy" stops being a trip to HR, and a new hire gets answers without tying up colleagues for a week.
We deliver this end to end: we work through your tasks and pick the model, size the servers or cost out the API, deploy Open WebUI with single sign-on via Keycloak, configure access control, logging and backups, and train both staff and administrators. Write to us via the contact page — we will show you the interface and work out which is better value in your case: your own GPU or an API.