A private LLM is a language model installed on your servers or a private cloud and connected to your company documents. It works like an internal ChatGPT, but data does not go to external services and each user sees only what they can already access. We know how because we run one: the assistant on our website uses Google’s open-weight Gemma model on one of our servers.
↓ Scroll
According to ISTAT (ICT in enterprises, 2025), among Italian companies that considered AI without adopting it, 43.2% cite privacy and data protection and 45.2% missing or poor-quality data.
Contracts, projects, customer data: material you do not want to paste into a service whose location you do not control.
Colleagues already use AI tools with personal accounts. Without an internal alternative it is hard to stop.
Procedures and manuals exist, but nobody finds them when needed.
With many users and heavy use, licence costs grow linearly.
A private LLM is not always the right answer. Here is how we see it.
Ready immediately, with very capable models. Before choosing, check in the contract terms where data is processed and whether it is used for training.
An open-weight model on a dedicated EU server that we manage. Control over data without hardware on site.
Everything on your servers: maximum control, but you need suitable hardware and maintenance.
We choose an open-weight model suited to the task and the available hardware, or an API with EU processing when that is enough.
We index folders, manuals and procedures. The model answers from them and cites the source document.
Everyone sees only what they can already access: original permissions still apply.
We log questions and answers in line with your internal policy, to improve content and check usage.
A dedicated server we configure, update and monitor.
On site or in your data centre. We handle installation and updates.
Confidential documents on the private model, generic tasks on an EU API.
We quote after the free review. These are the factors that matter.
Larger models answer better but need GPUs with more memory. For many document assistants a mid-sized model is enough.
What counts is how many people ask questions at the same time, not how many have access.
Affects indexing and storage more than the model.
Updates, backups, monitoring and content review are part of the running cost.
The model is not retrained on your documents: it reads them at question time.
If the assistant also answers customers or suppliers, since 2 August 2026 it must disclose that it is AI.
Instructions for users: what to ask, how to check answers, what not to enter.
Processing stays in systems you control. Records, privacy notice and log retention should be set with your DPO.
Point us to a repetitive task: incoming email, documents, data entry. We check whether an AI agent can handle it, with which human checks and with data kept in the EU.
We will email you within one working day to book the 30 minutes. If it is urgent, message us on WhatsApp.
Prefer to talk right away?
Mon–Fri 9:00–12:30 and 14:00–20:00 (CET).
When the assistant must also act in your systems.
The server to run the model on.
Decide before you invest.
How we would have an agent read orders arriving by email and PDF, with mandatory human approval and an exceptions queue. Read the full scenario.