Frequently asked questions about local AI for businesses

It travels to OpenAI’s servers in the United States. What you write — and the files you attach — may be used to improve their models and is subject to their policies and American law. With Microsoft Copilot the same is partly true: depending on the account and settings, some of the data goes to the cloud. Some features, such as certain Copilot functions on Windows, process part of the work on your own machine, but the conversation and your files still leave home. If you want nothing to leave your computer, local AI is the only option that truly delivers it.

It is artificial intelligence that runs on a computer in your company, not on the servers of OpenAI, Google or Microsoft. It uses open models such as Qwen, DeepSeek, Llama or Mistral, and your data never leaves your network.

For a company's day-to-day work, yes. The best open models of 2026 draft, summarise, analyse documents and write code at the level of the commercial models of just a few months ago, and some accept up to a million tokens of context.

A private ChatGPT-style chat for your team, a search tool that answers from your documents with citations, meeting transcription, agents that automate tasks and models tuned to your line of work. See the uses page for more than a hundred examples.

No. Once installed, local AI works offline. You only need internet access if you want it to search the web or to receive updates.

In Barcelona. We work on site in Barcelona and its metropolitan area, and remotely for the rest of Catalonia, Spain and the European Union.

A chat tool answers when you ask it something, and that is the end of it. An agent is given a goal and works until it achieves it, using tools such as email, the ERP, a spreadsheet or the browser, deciding the steps and checking the result. The chat helps you do the work; the agent does it with you.

It is a program that combines a language model with tools and memory. It can read documents, look up information, fill in forms, write code or prepare reports, and chain several tasks together without you having to ask for each one.

Sort incoming email and draft replies, enter invoices and delivery notes into the ERP, update the CRM after every meeting, compile the weekly report from several sources, or monitor systems and alert you if something fails. Always specific, repetitive tasks.

Yes, if it is designed properly. Our agents have limited permissions, only touch what they are authorised to, log everything and ask for human approval before any important action, such as sending an email to a client or changing data in production.

Yes. The open models of 2026 already reason and use tools well enough to work as agents on your own server, so neither your documents nor your actions pass through anyone else's cloud.

Hermes Agent as the orchestrator, OpenCode and Pi for coding tasks, n8n for simple automations and Cursor for supervision. In our Barcelona lab, dozens of agents work every day with these same tools.

To the cloud. ChatGPT conversations and files travel to OpenAI’s servers in the United States under their policies; Microsoft Copilot follows a similar path, except for certain features processed on the device. If you put client data, medical reports or legal documents in there, you are sending them out of your building. With local AI everything stays on your own server: nothing leaves your network and no intermediary can read it. The “AI at home” guide explains it in detail.

Yes. It is more common than it looks: many people already own a machine at home powerful enough for local AI without knowing it. On Windows, new machines sold as Copilot+ PCs use a neural chip (NPU) for features like Recall, which stores and encrypts your screen snapshots on the computer’s own disk without sending them to Microsoft. But when you ask Copilot something in chat, your question and any attachments still travel to Microsoft’s cloud. On a Mac with an Apple chip (M1 or later) and 16 GB of memory, free apps such as LM Studio or Jan let you download open models and chat with them entirely on the Mac itself, with no accounts or external servers; once a model is downloaded it even works offline. On a Windows PC, the equivalent route is apps like LM Studio or Ollama: they run well with 16 GB of RAM, and better still with an NVIDIA graphics card. The difference is not power, it is where each thing is processed: on your own machine, or in someone else’s cloud.

Not with a personal account. The GDPR requires a data processing agreement, which consumer accounts do not offer. With local AI, the processing happens on your own premises and the problem disappears.

Since February 2025, training for anyone who uses AI in your company (Article 4). From 2 August 2026, telling people when they are talking to an AI or receiving AI-generated content (Article 50). High-risk obligations arrive on 2 December 2027.

Yes, because documents and conversations do not leave your network and are not used to train anyone's models. It is the most prudent option for law firms, advisory firms and medical practices.

Yes. With every project we deliver the record of processing activities, the AI use policy for your team, the transparency notices and an architecture summary for your DPO or your accountant.

It depends on how many people will use it and on the models. For a team of 5 to 20 people, an NVIDIA DGX Spark, an ASUS Ascent GX10 or a Mac Studio is usually enough. We will recommend one in the assessment.

Almost certainly yes, and you may not know it. Machines from the last few years ship with enough memory and graphics to run open models at home: a Mac with an M chip and 16 GB, or a laptop with 32 GB and a modest NVIDIA card, already runs a private ChatGPT, a document search tool or voice transcription. Many people already have the power they need at home without knowing it: before buying anything, it is worth checking what you own.

Yes, it is one of our most common cases in Barcelona. We set it up with models, multi-user chat and document search so that your investment starts working within a few days.

Between 150 and 500 W running all day, similar to a couple of office computers. Current machines are designed to stay on 24 hours a day without making any noise.

Yes. You buy the machine yourself, from whichever supplier you prefer, and it is yours. We help you choose it and, if you already have one, we make the most of it.

Two to three weeks for a ready-to-use local AI. With document search or agents, four to eight weeks, always with a pilot before going into production.

We will tell you. If your data is not sensitive, a cloud solution with EU data residency may be enough and cheaper, and we will help you choose it.

Hug Prat, an IT consultant and developer from Barcelona with more than 20 years of experience, who runs his own local AI lab with two NVIDIA DGX Spark units and a fleet of AI agents.

Yes. We will show you a live demo from our Barcelona lab, with models, document search, voice and agents doing real work.

Yes, and it works well. We use models that write natural Catalan, with vocabulary tuned so they respond quickly in Catalan, which matters for public bodies, schools and companies with Catalan-speaking clients.

Catalan, Spanish and English, both when we deal with you and in the systems we install. The AI can reply in each person's language.