On-premise AI · itSilesia

On-premise AI for industry: BellBox: an AI expert in a box, inside your network

BellBox is a physical computer installed inside your internal network, where itSilesia configures MCP servers that connect AI to your company's data, so employees can talk to your company's knowledge through a regular browser, while no document or query ever leaves the building.

Private AI on your own data, with no cloud and no per-user subscription, deployed on-site by a software house from Gliwice with 15 years of industrial projects.

Book a call about BellBox See what you get
Client's internal network ● online
Employee's browser
▼
BellBox
AI model + MCP servers, no outbound connection
▼
Technical documentation
Production data
Service procedures
Next data source (MCP)
outbound traffic: 0 kB
0
documents and queries leaving the company
8–12 mo.
return on investment at 200 users
5–15 days
to connect the next data source, instead of 4–8 weeks
1
contract covering hardware, model, integrations and deployment
On-premise AI: what it is

Private AI on your own data: nothing sent to the cloud

An on-premise AI deployment means the language model runs on hardware sitting in your own server room, not on a third-party vendor's servers. An employee asks a question in the browser, BellBox searches your technical documentation, service procedures and production data for the answer, and the entire query stays inside your internal network.

We connect data sources through MCP servers, an open standard where every system is a separate, reusable component. This lets your local AI assistant grow together with the project's scope: from a single process to further plants, without rewriting integrations from scratch.

local AI model AI without the cloud MCP servers RAG on technical documentation one-time license

Who we build on-premise AI for

→ Production and maintenance: technicians ask about service procedures and machine documentation instead of digging through folders.
→ Energy, mining, heavy industry: production data under a security policy that can't leave the building.
→ Design and engineering offices: knowledge from hundreds of projects available in one place, with no risk of leaking intellectual property.
→ Companies under audit and NIS2: IT gets a solution that passes security review instead of another exception to approve.
Value proposition

Five reasons an AI project doesn't die at stage two

Each point starts with a problem we see at industrial companies, and ends with what the client actually gets out of it.

01

Your data never leaves the building: the box sits in your own server room.

Problem

Many industrial companies block AI projects not because they don't see the value, but because IT won't approve sending technical documentation and production data to a third-party vendor's servers.

What we do

That's why itSilesia installs a box with the AI model physically inside the client's network, with no outbound connection.

Result

As a result, the project passes security review on the first attempt, instead of getting stuck for months in a security-policy exception process that usually ends in a rejection anyway.

02

You pay for the box once, not per employee, every month.

Problem

Many companies start their AI journey with a 30-person pilot that looks cheap, then discover that covering 300 employees across three plants means hundreds of thousands of PLN a year under a per-user model, and the project dies exactly when it starts working.

What we do

That's why itSilesia sells the box as a one-time cost, where the number of users doesn't affect the price.

Result

That means the investment pays back in 8–12 months at 200 users, and the production director can extend access to further plants without going back to the CFO for a bigger subscription every time.

03

An expert in one process, not a general-purpose company chatbot.

Problem

Companies that deploy a general-purpose AI assistant usually find, six months in, that only a handful of the workforce actually uses it, because a technician who once got the wrong service procedure won't ask a second time.

What we do

That's why itSilesia configures the box as an expert on one specific process, and only signs off the deployment once it clears an accuracy threshold on a question set built together with the client's own expert.

Result

That means the client deploys a tool people actually use, and has a hard number to show the board instead of a vendor's promise.

04

We connect the next data source in days, not quarters, thanks to MCP.

Problem

Companies starting an AI-on-their-own-data project usually begin with the ambition to connect every system at once, and a year later have two sources connected and a budget burned on integrations instead of value.

What we do

That's why itSilesia builds the box on the MCP standard, where every source is a separate, reusable component.

Result

That means we connect the next system in a few days to about two weeks instead of 4–8, and the client expands scope in stages instead of paying upfront for integrations whose value isn't proven yet.

05

One partner: hardware, model, integrations and deployment in your network.

Problem

An on-premise AI deployment today typically needs three vendors: a hardware integrator, a model provider and a software house for integrations. And when the AI expert gives a wrong answer, the client's IT director is left alone to referee a dispute they're not equipped to judge.

What we do

That's why itSilesia takes the whole scope under one contract: hardware, model, MCP integrations, interface and deployment in the client's network.

Result

That means the client has a single point of accountability for a working result, and isn't paying dozens of hours a month of their own IT's time coordinating three companies that all point fingers at each other.

What the rollout looks like

One process, one sign-off, a measurable result

01

Choosing the process

Together we pick one specific process for the AI expert to work in.

02

Installation in your server room

The box goes into your internal network, with no outbound connection.

03

Connecting sources via MCP

Every data source is a separate, reusable component, added in stages.

04

Sign-off on an accuracy threshold

We sign off the deployment only once it clears an accuracy threshold on a question set built with your own expert.

Frequently asked questions

What IT departments ask before deploying on-premise AI

How is on-premise AI different from a company ChatGPT account?

In a cloud model, your query and any attached document go to the vendor's servers. In an on-premise model, the model runs on a box in your own network, so technical documentation and production data never leave the company.

Does BellBox need an internet connection?

No. The box runs inside your internal network with no outbound connection, which is why the project passes security review on the first attempt instead of waiting on a security-policy exception.

How much does the deployment cost, and what does the price depend on?

BellBox is a one-time cost, where the number of users doesn't affect the price. At 200 users the investment pays back in 8–12 months, and extending access to another plant doesn't raise any subscription.

What systems and data can be connected?

Technical documentation, service procedures, production data and company systems. Every source is a separate MCP component, so we connect the next system in a few days to about two weeks instead of 4–8.

How long does the deployment take, and how does sign-off work?

We start with one process. We sign off the deployment only once it clears an agreed accuracy threshold on a question set built with your own expert: you get a hard number, not a vendor's promise.

Who's responsible for the hardware, the model and the integrations?

itSilesia takes the whole scope under one contract: hardware, model, MCP integrations, interface and deployment in your network. One point of accountability instead of three vendors pointing at each other.

Let's talk about one process in your company

A short call is enough to work out whether BellBox makes sense for you, and which process to start with.

+48 781 66 53 35 lukasz.lipka@itsilesia.com
Contact
Łukasz Lipka
itSilesia
ul. Wincentego Pola 16
44-100 Gliwice, Poland