On-premise AI: when data cannot leave the building
In healthcare and in professional firms, AI could save hours of work. It almost always stops at the same point: patient data and client documents cannot go to just any external service. That is why we work with language models installed on-premises, inside the clinic or the firm.
The common challenge: data that cannot leave
Two different worlds, the same constraint. In healthcare, health data is treated with particular care under the GDPR (the EU General Data Protection Regulation). In law and tax firms, professional secrecy and the duty of confidentiality cover client documents. The EU AI Act applies to everyone. What these rules require in a specific case depends on the use: it is a matter to assess with your own legal adviser.
What "on-premises" means
The language model (an LLM, the system that reads and writes text) runs on your own machines, or on a dedicated server under your control. Documents stay there. You give up some power compared with the largest cloud services, and we say so upfront; you gain control over the data and over who sees it.
Healthcare: patient data stays in the clinic
A private healthcare provider handles a great deal of text: reports to summarize, clinical records to find, patient requests to route, letters to draft. An on-premises model can handle the first draft and search; the clinical decision always stays with the doctor.
Cloud or on-premises: the differences
| Cloud service | On-premises system | |
|---|---|---|
| Where the data goes | to the provider's servers, outside the clinic | to the clinic's machines or a server under its control |
| Who manages access | the provider, under the contract | the clinic |
| Model power | the largest models available | smaller models: weaker on some tasks |
| Start-up costs | usually low: you pay for use | usually higher: hardware and installation |
| Maintenance | the provider's job | the clinic's or a technical partner's |
| Time to start | usually short | usually longer: installation and tests on real work |
A general comparison, not an assessment of your case. Which option fits, including on the legal side, depends on the case; the legal aspects are a matter to assess with your own legal adviser.
A project in progress
We are working with a medical organization to bring AI into its processes using on-premises language models. We are also looking for professionals to help take it into everyday work: our Network page.
Law and tax firms: client documents stay in the firm
Law firms and in-house legal teams, notaries, accountants, and tax advisers: different professions that all work on other people's documents. Deeds, contracts, minutes, financial statements, tax returns. The issue isn't the medical record; it's the client file, and the first question comes before any tool: where does this document end up when I upload it?
Where on-premise AI helps a firm
- Deadlines and filings: tracking tax returns, VAT, and payroll obligations, where delays and errors cost money.
- Contracts to review: spotting clauses written for another legal system that may not hold in Italy, such as termination, penalties, and unilateral rights.
- Documents in order: procedures, minutes, delegations, and powers of attorney, always traceable.
- Regular checks: reviews repeated over time, not done once.
Examples from our guide Italy for Founders, chapters 03 and 07. The guide
Before you upload a client's document: 8 questions
Questions worth asking of any cloud or on-premises AI service. They are not legal advice: they help you understand where the document goes and who sees it.
- Where are the servers that receive the document physically located?
- Does the provider use uploaded documents to train its models?
- Who at the provider can read the documents, and when?
- How long are documents and conversations kept, and how are they deleted?
- Is there a contract that defines how the provider handles the data?
- Does the document contain data about other people, besides the client?
- Does the client know that AI is used on their document? Whether and when to tell them is a matter to assess with your own legal adviser.
- If an answer is unclear, can the same work be done on-premises?
If your confidential data is something else
Supplier contracts, staff records, projects not yet announced: the method is the same one we use for AI for SMEs.
Frequently asked questions
Can I use a cloud AI service?
It depends on the data and on the contract with the provider. The 8 questions help you find out before, not after.
Do we need a new server?
It depends on the workload and on the model. We assess it after seeing the documents and the work to be done.
Is it as reliable as a cloud service?
On some tasks, yes; on others, less so. We test it on real work before deciding.
Who is responsible if the system gets it wrong?
The person: the doctor for clinical decisions, the professional for what they sign. The system prepares, the person decides.
Does the system comply with the rules?
The rules depend on the specific use: it is a matter to assess with your own legal adviser. We work on the technical and organizational side; we do not give legal advice.