1. What we mean by private AI
Private AI refers to an artificial intelligence system that operates in an environment dedicated to a single organisation, or in any case under its control, rather than relying on a public service shared with many other customers. The term is not standardised. Some use it for a model installed on their own servers, others for a managed service in a reserved environment governed by a precise contract. When a supplier uses the word, always ask what they mean in practice.
The practical difference concerns three questions: where the data is processed, who can access the systems that handle it, and who decides which model to use and when to update it. A private system is not automatically more secure than a public one. It does, however, give more control over these three points, and that control is why the choice needs to be assessed case by case.
2. How an AI system works inside a business
A generative AI system in a business is made up of several parts. There is a language model, meaning the programme that produces the text. There is an access layer that receives users' requests and verifies their identity. There may be connections to internal documents, and a set of rules that decides what the system can see and what it can do. The model is the most visible part, but it is not always the one that determines the risks or the complexity of management.
When a person asks a question, the request passes from the access layer to the model, which returns the answer. If the system is connected to company documents, the relevant text is first retrieved and then added to the request. Each step may take place in a different location and may leave copies of the data in logs, temporary memory or archives. That is why the useful question is not only where the model runs, but where the data travels and stays at each stage.
3. The options compared
There are three main options, each with different trade-offs. A shared public service can be activated quickly and may be enough for general uses that do not touch confidential information, provided clear rules of use exist. A dedicated environment managed by a supplier offers separate spaces and more precise contractual terms. An installation on your premises, or in your own infrastructure, gives the broadest control but requires skills, maintenance and updates that must be managed over time.
The table is deliberately qualitative. Details vary from supplier to supplier and from contract to contract, so commercial labels are not enough. Ask in writing where requests are processed, which logs are kept and for how long, and who can intervene on the system. A vague answer on any of these points is already useful information about the supplier.
| Option | Where the data sits | Who decides the model | Internal effort |
|---|---|---|---|
| Shared public service | Supplier's infrastructure, under its terms | Low: versions and rules set by the supplier | Low, but the use still needs governing |
| Managed dedicated environment | Space reserved for the organisation, with conditions in the contract | Medium: configuration and models agreed | Medium: a contract owner is needed |
| On-premises or own-infrastructure installation | The organisation's servers or those of a dedicated supplier | High: choice and updating of models | High: ongoing skills and maintenance |
4. When it genuinely makes sense
Private AI makes sense when several conditions occur together. The data processed is confidential or personal, for example customer data, contracts, employee data or projects not yet public. The system must work on internal documents that cannot leave the company's perimeter. Processing personal data requires compliance with the European General Data Protection Regulation (GDPR), and the NIS2 Directive, transposed in Italy by D.Lgs. 138/2024, may apply to some organisations depending on size and sector.
It does not make sense when the use case is generic, for example rewriting an email or summarising a public text, and no sensitive data is involved. Nor does it make sense when nobody in the company is able to govern the system: a private environment without an owner is harder to control than a well-regulated public service. The overall cost depends on scope, number of users, management and level of support, and should be estimated case by case.
5. Data, location and access: three separate questions
Many discussions about sovereignty confuse three questions. The first concerns where the data is physically stored. The second concerns where it is processed. The third concerns who can lawfully access the data, that is, which laws and which authorities can require the supplier to make it available. A supplier may keep data in Europe and still be subject to access requests under the law of the country where it is headquartered. The three questions should therefore be kept separate.
The questions to put to the supplier are specific: where requests are saved and processed, who in the group can access them and under which controls, how logs are managed, what happens to the data at the end of the contract, and how any transfers outside the European Economic Area are handled. Ask that the answers are set out in the contract or an annex, not only in a sales presentation. A legal adviser can help you read them.
6. Risks and controls to plan for
A private AI system reduces some risks and leaves others in place. Misuse by staff remains, as do wrong answers presented as certain, and exposure to hidden instructions in documents or pages that the system reads, an attack known as prompt injection, meaning a text crafted to change the model's behaviour. A closed environment does not eliminate these risks, because they depend on how the system is configured and used. Controls are needed on several levels, not a single barrier.
The basic controls are often the same in every scenario, and they should be chosen according to the type of data and the actions the system can perform. Not all are needed immediately: start with the first five and add the others as the scope grows. Each must have an owner who can say whether it really works. Also record who checks them and when.
- Role-based permissions on the documents the system can consult.
- A log of requests and of the sources used for answers.
- A procedure for human review of outputs that feed into decisions or external communications.
- Tests and checks every time the model or the connected sources change.
- A designated owner, with a technical contact and a business-use contact.
7. Common mistakes in the choice
The first mistake is believing that private automatically means secure. A system installed on your premises, without updates, with shared access and no logs can be more fragile than a service managed carefully by specialists. The second mistake is thinking a contract clause is enough to resolve the data question. The contract is necessary, but it must be verified through technical checks and by an organisation that respects it in practice.
The third mistake is starting from the model. The choice of model matters, but it usually comes after deciding which processes are needed, which data is involved and who is accountable for it. The fourth is trying to do everything at once. Start with a limited use case, measure the result and widen the scope only once the controls are working.
8. The questions the following guides explore
This guide is an overview. Each topic touched on here has a dedicated follow-up guide, and some should be read before others. The most frequent questions concern use by employees, the choice of first use cases, preparing data and the quality of answers. There is no need to read everything in one day: it is best to follow an order that starts with internal rules, because everything else depends on them.
The topics to explore, in order of priority, are these. The first helps set the rules, the second clarifies where unauthorised uses hide, and the others concern data, the choice of use cases and checking answers. Read them in sequence, not all together, and note for each one the decision you need to take in the company. There is no need to do it all at once.
- How to write a policy on employees' use of generative AI.
- Shadow AI: tools used without authorisation and how to identify them.
- Which first use cases to choose, using a method that assesses value and feasibility.
- How to prepare documents and data before connecting them to an AI system.
- Fine-tuning or RAG: which approach your case needs.
- Model hallucinations: how to reduce them and how to verify them.
- Prompt injection and other typical vulnerabilities of language models.
- How to reason about the value of AI without inventing numbers.
9. How to start
The first step this week is simple: list the processes in which someone already uses AI tools, even unofficially, and for each one write down which data goes in and which comes out. This list quickly shows where a rule is needed and where there may be a concrete risk. You do not need a supplier to draw it up, and it does not require special technical skills.
The second step is to take the list to a meeting with the IT manager, the data protection contact and a representative of the department concerned. From there, choose a single low-risk use case, define who governs it and decide which data it may use. Only then does it make sense to ask for a quote, with the questions from section five already in hand.
Frequently Asked Questions
Is private AI always more secure than a public service?
No. A dedicated environment gives more control over where the data is stored and who manages it, but security depends on configuration, updates, access and processes. A system installed on your premises and left without maintenance can be weaker than a service managed with care. Always assess the specific system rather than the label attached to it.
Do I need servers on my premises to have private AI?
Not necessarily. Dedicated environments managed by a supplier, with reserved spaces and defined contractual terms, can suit many businesses. An on-premises installation makes sense when there are particular requirements or solid internal skills. The choice should be made by comparing the data to be processed, the obligations of your sector and the management capacity you genuinely have.
If the supplier says the data stays in Europe, is the problem solved?
Only in part. Where the data is stored is a different question from who can lawfully access it, because a supplier subject to the laws of a non-European country may be required to respond to requests from that country's authorities. Ask for both aspects to be verified and set out in the contract, with the help of a legal adviser.
Which regulations should I consider?
It depends on your sector, your size and the data you process. The GDPR is almost always relevant when personal data is processed. The NIS2 Directive, transposed in Italy by D.Lgs. 138/2024, may apply to certain organisations. The European regulation on AI introduces obligations for those who use artificial intelligence systems in certain cases. An adviser can tell you which of these apply to your situation.
Does private AI eliminate errors in answers?
No. A private system can make mistakes just as a public one can, because the problem of unreliable answers depends on the model and on how it is used. The remedy is a verification process: answers that feed into decisions or external communications must be checked by a competent person, and the sources used must be available for review.
How much does a private AI system cost?
There is no single figure valid for everyone, because cost depends on scope, the number of users, the choice between a managed environment and an internal installation, and the level of support required. Be wary of offers that quote costs before the use case has been defined. Ask for an estimate with its assumptions written out, so you can compare it with other proposals.
Where do I start if nobody in the company looks after AI?
Start with a map of the uses already in place, including informal ones, and a simple internal rule on which data must not be entered into external tools. Then identify one person to coordinate decisions, even part-time. Someone, however, must be accountable for the choices.