Who needs local AI, and why

A business cloud plan costs less, updates itself and does more. An on-premise system makes sense for one reason only, confidentiality: because it is imposed on you, or because you choose not to entrust certain data to anyone.

Three questions

The first two look like the same question, and they are not. Not being allowed to entrust certain data to an outside provider is one thing; not wanting to, because you do not trust them, is another. The first is an obligation, the second a choice.

1. Are you forbidden? Some data, by law or contract, cannot be processed by an outside provider, not even under a data processing agreement and a promise not to train on it. Technical documentation subject to military or dual-use export rules. Drawings a customer gives you under a non-disclosure agreement that excludes the cloud. Clients who ask you in writing not to use outside services. There is nothing to weigh here: if the prohibition exists, that data has to be processed on the premises, whatever it costs.

The professional secrecy of lawyers, accountants and doctors, on the other hand, does not forbid the cloud by law in any of the countries we examined: what changes is how, from country to country. In Italy, since Law 132 of 2025, professionals must inform clients of the AI systems they use. For many Italian firms the real issue is therefore the second question.¹¹ In Germany a written contract binding the provider to secrecy is required and, outside the EU, comparable protection; the federal bar considers it "not necessary" to send client secrets to models like ChatGPT, and the chamber of tax advisers asks for client consent or anonymised data. In Austria an agreement with the provider excluding training on client data is required. In the United States the American Bar Association requires the client's informed consent. In France the national bar council rules out data covered by secrecy altogether. The Council of Bars and Law Societies of Europe names on-premise systems as the safest option for confidentiality.

2. Do you trust them? If nobody forbids it, the question is whether you trust the providers' promises. The promises are serious: business plans exclude training on your data by contract, they are certified, and so far there is no documented case of business-plan data being used to train a model. But there are concrete reasons not to hand over everything: terms can change, a court can override them, and jurisdiction stays the provider's. Not trusting them, for a certain class of data, is a reasonable choice. But it is a choice, and it has a price: you will find it in the table below. It is worth making for the data that really matters, not for everything. If you are not forbidden and you do trust them, you can stop here: the cloud is enough for you.²² Terms change: since April 2026 GitHub Copilot's individual plans use data for training unless switched off. Courts: in 2025, in the New York Times case against OpenAI, a judge ordered conversations that would have been deleted to be kept. Jurisdiction: Microsoft France's legal director stated under oath that he could not guarantee European customers' data would not be handed to US authorities. We wrote about it in How far to trust ChatGPT and Claude with company data.

3. What does it cost you? The costs of an on-premise system are almost fixed: the machine, the installation, and above all the service of whoever updates it, monitors it and repairs it when it stops, which is the largest item of the bill. Cloud costs grow with the number of seats. Below twenty or thirty people the cloud costs less, always: the sums are in the table below.

The sums

The table compares the three solutions for a company of two, five, ten and twenty-five people. Cloud is a business subscription for each person; local is a machine on the premises, with installation and support; mixed keeps the cloud for everyone and adds a machine for confidential data only. For each case there is the total cost over three years and, next to it, what it comes to per person per month. The gap is largest for the smallest firms: for two people, an on-premise system costs almost four times the cloud. It narrows as the firm grows, but within twenty-five people it does not close.

Where these figures come from is shown by the ten-person case, item by item. The machine is the smallest part of the bill; installation and support weigh more than everything else.

All figures are our estimates, at October 2026 prices, excluding VAT, over three years. They are averages: for a specific company the sums need redoing on its own numbers, and that is the first thing we do in a consultation.

The machine alone costs roughly two years of subscriptions for ten people, but it is a quarter of the bill: the rest is installation and maintenance. And even the yearly running costs alone of an on-premise system exceed those of ten subscriptions. Over three years, break-even sits at around thirty people.

Two trends point the same way: hardware is getting dearer because of the global memory shortage (one of the most popular machines rose 74% in a year), while cloud seats are getting cheaper. And many offices already pay for AI: Copilot Chat is included in Microsoft 365, Gemini in every paid Google Workspace plan.

Quality, and updates

An open model on a desk machine is today at the level of the best cloud models of a few months ago. On questions about one's own documents the difference almost disappears; on general knowledge, expert reasoning and agents working on their own it remains clear. We wrote about it in detail in What runs on a machine in the company, and how far it is from ChatGPT and Claude.

The cloud updates itself: a new model every few months, at no extra cost, sometimes with changes nobody chose. On the premises a model stays as installed until someone updates it, tests it and puts it back into service. The good side is that the same machine, kept updated, runs better and better models: in a year the best models that fit in 128 gigabytes of memory have improved a lot. Web search, deep research and integration with Office or Workspace, on the other hand, remain mainly in the cloud.

The verdict

The answers to the three questions lead to one of three cases.

Nothing forbidden, and you trust the contracts: the cloud. This is the case for most small businesses: design and communication studios, agencies, architects on ordinary projects, trade and service companies, most of an accountant's work, the office work of a manufacturer. What is needed is a subscription in the company's name for everyone who uses it, good enough to give nobody a reason to use their own, and training on what must not go into it.

Part of the data must not leave: the cloud for almost everything, an on-premise machine for that data. By obligation or by choice, there is a part of the archive you do not want to entrust to a provider: a law firm's case files, a medical practice's clinical notes, the drawings a subcontractor receives from customers. Everything else, e-mail, translations, everyday writing, stays in the cloud. It is the most expensive of the three solutions, the Mixed column of the table, and it is chosen for confidentiality, not to save money. To keep costs down, the on-premise machine does one thing only, on one archive.

The data can almost never leave, or there are many of you: everything on-premise. Those working under military or dual-use export rules, those whose clients forbid any cloud by contract, those who must work without a reliable connection. And, on cost alone, companies above thirty people working mainly with documents. Among small businesses they are few.

The written rule

Compliance guides all recommend the same thing: an internal policy on the use of AI, saying which tools may be used, which data may go into them and which may not, and who is responsible. The GDPR asks for adequate organisational measures, the AI Act asks for training of those who use these systems, and a policy is the simplest way to show it has been done. It cannot be missing: it sets the obligation, shows the company has thought about it, and makes it enforceable.

But a rule prevents nothing, and not everyone reads it the same way. Someone who pastes a contract into a chatbot to have it summarised does not feel they are disclosing it: they are using a tool, not sending it to anyone. A person who would never send that document to a stranger may put it into the cloud without a second thought, convinced they have broken nothing. The owner needs to know this: the risk comes not only from those who want to get round the rule, but also from those who follow it in their own way.

Nor does the rule stop a colleague from pasting a contract into the free version of ChatGPT on their own phone, or from photographing the screen. Technical tools, company accounts, network blocks, software that intercepts copy and paste, reduce the risk on company devices, but for a small firm they are expensive, they do not reach personal phones, and monitoring employees has legal limits. Beyond a certain point, the company has to trust its people and hope.

For data that must not leave, then, the strongest defence is not the rule but the way work is done: giving people a convenient tool for confidential work that does not take the data outside. It is one more reason for the mixed case. The on-premise machine is not only a way of not having to trust the provider: it is also a way of not having to rely solely on everyone's good will.

And there is a reason that does not show up in the sums: peace of mind. Knowing where clients' documents are, without having to follow providers' terms or watch over employees, and being able to tell clients so in one sentence. It has a price, the one in the Mixed column, and it works only if someone keeps the machine updated and closed to the outside. But for many owners it is the real reason.

What to do in any case

Whatever the answer, a few things need doing anyway:

• A business plan for everyone, not personal subscriptions. Individual plans use data for training unless switched off, and do not include the data processing agreement the GDPR requires when documents contain personal data. And people use their own subscription when the company's is missing or works worse: giving everyone a good tool is the most effective measure. With a business plan the administrator decides who has access; on the company's computers and network the free versions can be blocked. On an employee's personal phone nobody can check anything, and even on company devices monitoring has legal limits: in Italy article 4 of the Workers' Statute, in Germany the works council has to be involved.
• Inform clients, or ask for consent. In Italy Law 132 of 2025 requires professionals to inform, whatever the tool; elsewhere, as in the United States or for German tax advisers, consent is required.
• Train people. The AI Act asks companies to look after the AI literacy of those who use these systems, and the first obstacle to adoption, according to ISTAT, is precisely the lack of skills.

What we do, in each case

That is why our services are three, and local infrastructure is the last of them.

• Training. For everyone, because the result depends not only on the tool but on how it is used: at which steps of the work, with which instructions, who checks what before the result leaves the office. We teach people to use these tools in their own work, to recognise their mistakes, and to know what may be shared and what may not. It is also the AI literacy the AI Act requires.
• Consulting. For everyone. A cost analysis done on your company's own numbers, not on this page's averages: how many people, which data, which tasks, which tools you already pay for, to see whether the step to an on-premise system is necessary or not. Then choosing the right tools and fitting them into the work: what stays a conversation, what becomes an automated workflow, where a person needs to check. And the data inventory, reading the contract, configuring access and blocks, the notice to clients.
• Local infrastructure. Only for those with data they must not or will not entrust to anyone: a machine on the premises, configured, maintained and updated, alongside the cloud for everything else.

If the cloud is enough for you, we help you use it well, without selling you any machine. The details are on the services page.