AI agents

Where does your data stay? Cloud and local AI models

Your database stays in your company. A cloud model gets only the snippets a question needs. A local model runs on your GPU server. Compare the options.

, CTO6 min readAlso available inENAZRU

In this article

Your database stays with you. The ONYX AI agent connects to your 1C, ERP or mail system through a connector on your own server. You choose the model: a leading cloud model or a local model running on your GPU server. With a cloud model, only the part needed for the answer goes to the model. With a local model, the request does not leave your company network.

Snippet: a short piece of data sent to the model to answer a question. For example, one customer's debt line and the number of overdue days. It is not your whole database.

Where does my data stay?

The agent gets access to your system through a connector. The connector is set up on your server and starts in read-only mode. Your database, tables and documents stay where they are.

This does not mean the model never comes into contact with that data. The model needs certain data to answer a question. The difference is that you choose which data goes, by which route and to which model.

In companies where the data is sensitive, this question comes first, so the answer should be open and simple. Below, we explain step by step what each of the two options does.

What happens when you work with a cloud model?

For example, the finance department asks: "Which customer has debt overdue by more than 30 days?"

  1. You write the question.
  2. The agent takes only the rows related to the question from the 1C receivables balances.
  3. Those rows (snippets) are sent to the cloud model you chose.
  4. The model returns the answer, and the agent prepares a reminder draft.

With the cloud option, that part leaves the company. Not the whole customer database, only the rows needed for the answer. Knowing this difference is part of making the choice.

The model choice is also yours. Whichever leading cloud model you pick, such as OpenAI, Anthropic or DeepSeek, it can be connected. The model can change, but the rule stays the same: only the part that is needed is sent.

What is a local model?

A local model runs on your GPU server. The request does not leave your company network. When the data is very sensitive, this is the simplest and clearest choice.

It has a cost too. The company needs a GPU server, and someone with technical knowledge is needed to set it up and maintain it. We define the right configuration together during the AI readiness check.

What is the difference between a cloud and a local model?

TopicCloud modelLocal model
DatabaseStays in the companyStays in the company
What leaves during a requestThe part needed for the answer (snippet)Nothing: the request does not leave the company network
Model choiceOpenAI, Anthropic, DeepSeek and othersThe model on your GPU server
What is neededAccess to a cloud modelA GPU server in the company
When it fitsWhen company rules allow a cloud serviceWhen the data is very sensitive and you want full control

The table is a map for deciding. Once you know which row matters most to you, the choice becomes clear.

Which option should you choose?

Three questions are enough:

  • Is the data in this system customer or financial data, in other words, sensitive?
  • Does your company's internal rule allow the use of a cloud service?
  • Is there a GPU server in the company, or does one need to be built from scratch?

The answers to these questions decide between a cloud and a local option. It is possible to start with a pilot on one request. The pilot is read-only, and the decision is made once the result is visible.

What do we check in the readiness check?

During the 30-minute AI readiness check, we look at four things:

  1. Where your data is. Which system, which table, which server.
  2. What can be connected. 1C version, ERP or mail.
  3. Which data is sensitive. This decides between the cloud and local options.
  4. Which model fits. A cloud model, or a local model on your server.

No access to your systems is needed. The result is delivered within 1 business day.

Nothing runs without approval

The agent reads, calculates and drafts. Sending, writing and ordering happen only after you click "Approve". Every request stays in the log: the request, the source that was read, the draft that was prepared, the person who approved it and the time of execution.

The log shows later what the agent read, who approved what and when. This is a simple but important layer of protection.

For more detail, see How the AI agent works, which covers the approval gate, where the data is and the log.

Check with your own data

AI readiness check: 30 minutes, free. Together we review where your data is, which model fits and what can be connected. No access to your systems is needed. We reply within 1 business day.

For the connection steps, see How do you connect 1C and ERP data to an AI agent?. For the flow from request to quote, see Unanswered quotes: how does a sales agent follow up?.

Frequently asked questions

Does my data go to an outside server?

With the local model option, the request does not leave your company network. With the cloud option, only the part needed for the answer, called a snippet, is sent to the model.

If we choose a cloud model, is my whole database sent?

No. Each request sends only the part needed for it. Your whole database is not sent.

What does a local model need?

A GPU server inside the company. We define the right configuration during the AI readiness check, based on your system and how you plan to use it.

Can the agent change my data?

At the start, the agent only reads. Every write action runs with your approval.

Which data does the agent read?

Only the systems and tables you open. We choose together during the AI readiness check which sources will be opened.

  • AI agent data
  • local AI model
  • cloud AI model
  • company data AI
  • 1C AI agent

Get a quote for your IT project

Tell us about your current IT setup and plans, and we will work out the right approach together.

All articles