In short: Private AI works on company data without handing it to external services you don't control. There are three routes: commercial models with sensitive data filtered out; open models on rented GPU servers in a European data centre; a server bought and installed on your premises. The choice depends on how confidential the data is, how heavily you will use it and which constraints you have to demonstrate.
Private AI is artificial intelligence that works on a company's data without that data ending up in services whose terms the company doesn't control. It isn't a single product: it is a way of organising where the models run and who has access to the data. The choice depends on what you need to protect and how much you will use it.
The three routes
1. Commercial models with filtered data
The commercial models from the big providers are often the most capable. You can use them by first removing sensitive data from requests (names, tax ID numbers, amounts) and putting it back only in the answer. It is the quickest route to get going. It works when documents aren't confidential or can be reliably anonymised: a filter reduces the data sent, but it needs testing on your own documents before you rely on it.
2. Open models on rented GPU servers in Europe
You rent a server with graphics cards (GPUs) in a European data centre and run an open model on it. The data stays in that private environment, used only by you. You don't buy hardware and you don't look after the physical site. You can scale the power up or down as your use changes.
3. A server bought and installed on your premises
The hardware sits in your company and documents stay on the internal network. It is the choice for those with strict constraints or highly sensitive data. The server has to be configured and managed: updates, models, security and backups. We supply, install and manage it; we test the workflows first in our lab on NVIDIA DGX Spark. Outbound connections, remote support and where the backups end up also need checking: the server's location alone is not enough.
How to choose
- Is the data low in confidentiality, or can it be anonymised? Commercial models with filtered data will do.
- Does it contain personal data or trade secrets? A private rented server, or one on your premises, is better.
- Do you need to prove that data doesn't leave the company? A server on your premises is the most straightforward choice.
- Is use irregular, and do you want to start small? Renting lets you begin and change course.
- Will use be heavy and continuous for years? Compare the cost of renting with that of owning a server.
What you need to know about open models
- They have improved a great deal and are good enough for many business tasks: summarising, searching documents, classifying, extracting data, drafting.
- For the hardest tasks, the largest commercial models can still do better. You find out with tests on your own cases.
- The power required depends on the size of the model and on how many people use it at once. That is the first calculation to make.
What you need around the model
- Controlled access with identities, permissions and a log of who asks what.
- Connection to your systems: documents, management software, email, applications. A model in isolation is of little use.
- Updates to models and security.
- Backups of configurations and data.
- Quality tests with real questions, repeated whenever something changes.
A concrete first step
Choose a well-defined use case, for example an assistant for your customer service documents. Work out which data it needs and how confidential that data is. With those two pieces of information, the right route usually becomes clear on its own. How to build the assistant is described in an AI assistant that answers from your documents.
Further reading: AI for business · confidential AI · an AI assistant for your documents · AI Act
Frequently asked questions
What is private AI?
A way of using artificial intelligence in which the models run in a private environment you control: company data doesn't end up in external services on terms you don't set.
Rented server or on your premises: which is better?
Renting in a European data centre lets you start small and adjust the power. A server on your premises is the most straightforward choice for anyone who has to prove that data doesn't leave the company.
Are open models good enough?
For many business tasks, yes: summarising, searching documents, classifying, extracting data and drafting. For the hardest tasks, you find out with tests on real cases.