AI automation without the cloud: the 3 places your AI can run and what each one costs
Dabster Labs · Last updated 20 October 2026
When an automation “reads” a customer email, an invoice or a medical appointment, some AI model sees that text. The question accounting firms and clinics ask us more and more is not “what does the AI do” but “where does it run and who sees my data”. There are three clean answers. Here they are side by side, with what leaves, what stays and what you pay for each, so you can choose without buying anything you do not need.
First, the basics: what actually “leaves” in an automation
An automation has two parts. One is the flow: which email triggers what, where the result is stored, who gets notified. That runs in n8n, the tool we build with, set up separately for each client in their own space. The other part is the AI model that understands text, summarises, classifies or writes. Privacy is decided in that second part, because that is where the content goes. The three options below are about exactly where that model lives.
Option 1: EU cloud with zero retention
The model runs at a European provider, in a separate space per client, configured so the provider keeps nothing of what is sent. We work with Mistral (France, one workspace per client) and with Scaleway (Paris, one project per client with zero retention), the latter mainly for clinics and municipalities. The client never holds a provider key and receives a single invoice, from us.
- What leaves: the text the automation is processing, for the duration of the answer, on servers inside the EU.
- What stays: nothing at the provider after the answer. The flow, the files and the history stay in the client’s space.
- What you pay: a monthly fee per automation plus usage, no hardware. This is our default and the cheapest way to start.
- Where it is not enough: when a contract or an internal rule says explicitly “no transfer to any third party”, even inside the EU.
Option 2: your own server in the EU, operated by us
The AI model runs on a server the client owns or rents exclusively, in a European data centre. No AI provider sits in between: n8n and the model live on the same machine. We already run this for our own internal work, open models on a server in Finland, so we know the limits too: without a graphics card the model has to be small, and small models need testing in Greek before you trust them.
- What leaves: nothing to third parties. Data never leaves the client’s machine.
- What stays: everything, under the client’s control, with backups the client defines.
- What you pay: server rent or purchase, plus maintenance. More expensive than option 1 for a few automations, often cheaper when there are many and they run all day.
- Where it is not enough: when you need large-model quality on difficult Greek text. Then you either add a graphics card or keep a hybrid: local processing, EU cloud only for the final wording.
Option 3: a box in your office
The most talked-about option this year, because there are now small machines, the size of a book, with enough memory to run serious open models locally. The NVIDIA DGX Spark class is the example, with 128 GB of unified memory. n8n talks to the model on the same network, with no cloud call at all. This is the answer for the client who wants to point at a device and say “that is where my AI runs”.
- What leaves: nothing. The AI part can even work with the internet down.
- What stays: everything, in your building. Along with the responsibility for power, backups and physical security.
- What you pay: the hardware once, around 4,000 euros for a machine of this class, plus setup and a monthly maintenance fee. You do not buy it before there is an automation that justifies it.
- Where it is not enough: when the team works from many places and the office has no reliable line, or when nobody wants to be responsible for a machine.
An honest note on quality
Good AI needs good machines. A 4,000 euro box or a server without a graphics card runs mid-sized open models: very good for classification, data extraction, short replies and internal tasks. For top-model quality on difficult Greek text, long compositions or voice, the hardware climbs into the tens of thousands and someone has to maintain it. That is why we almost always recommend the hybrid setup: the data stays on your infrastructure, and only the part that truly needs the large model goes, stripped of anything unnecessary, to EU cloud with zero retention.
The three options at a glance
| Criterion | 1. EU cloud | 2. Server in the EU | 3. Box in the office |
|---|---|---|---|
| Data at a third party | Only during the answer, zero retention | No | No |
| Model quality | Top tier | Good, depends on the machine | Good to very good |
| Starting cost | Low | Medium | High, one-off |
| Monthly cost | Fee plus usage | Rent plus maintenance | Maintenance |
| Who looks after it | Us | Us | Us, on your infrastructure |
| Best for | Most businesses | Many automations, strict contracts | Clinics, accountants with absolute confidentiality |
European models or Anthropic, OpenAI, Google: what you lose and what you gain
The second question always follows: “fine, but is Mistral as good as ChatGPT or Claude?”. The honest answer has two parts. At the top, on the hardest tasks (long reasoning, complex code, fine nuance in Greek), the models from Anthropic, OpenAI and Google are still ahead in 2026. On the tasks that make up 80% of an automation, meaning classification, data extraction, summaries and short answers grounded in your own documents, the gap is small and the client usually cannot see it. There, the European model wins clearly on legal simplicity.
| Provider | Seat and law | EU hosting | Zero retention | Greek | How we use it |
|---|---|---|---|---|---|
| Mistral | France, EU law only | Yes, by default | Yes, per workspace | Very good on everyday tasks | Default for client data |
| Open models on Scaleway | France, EU law only | Yes, Paris | Yes, per project | Good, depends on the model | Clinics, municipalities, sensitive data |
| Anthropic (Claude) | USA, subject to US law | Only through cloud partners in EU regions | By agreement, not by default | Top tier | Only for the hard step, with stripped data |
| OpenAI (GPT) | USA, subject to US law | Yes, EU data residency option | On approval, not by default | Top tier | Voice and complex text, with stripped data |
| Google (Gemini) | USA, subject to US law | Yes, EU regions on Google Cloud | In the business edition | Top tier | Rarely, when the client asks for it |
“Subject to US law” is the heart of it. Even when a US provider hosts the data on servers in Europe, the company can be compelled by US authorities to hand it over. The transfer is lawful under the EU-US data framework and the standard contractual clauses, but it is a transfer, and it has to be written into your privacy policy and your contract with your client. With a European provider that paragraph simply does not exist. So our rule is simple: a European model for anything that touches client data, a US model only for the step that truly needs it, after names and details irrelevant to the task have been removed. Which model handles which step is written in the proposal, so you know before you sign.
How to choose, in three questions
- Is there a written commitment that forbids any transfer? If yes, option 2 or 3. If no, option 1 covers GDPR with fewer moving parts.
- How many automations, how often? Few and occasional: option 1. Many and continuous: option 2 starts to pay off.
- Do you want the machine in your own hands? Only then does option 3 make sense, and only after the automation has proven its value on option 1 or 2.
An example from an accounting firm: sorting incoming emails and pulling data out of invoices starts on option 1, with a Mistral workspace used by that firm alone. If the firm later takes on a client whose contract demands “no transfer”, the same flow moves to an option 2 server without being rewritten. That is the advantage of building in n8n rather than inside a closed tool: the model changes, the work stays. If you are weighing where n8n itself should run, we compared the numbers in n8n self-hosted vs n8n cloud.
What applies to all three
Where the model runs does not solve GDPR on its own. Every option needs a data processing agreement with us, a record of what data enters the automation, minimisation (the model sees only what its step needs) and backups with access control. And in every option the automation must be tested on real Greek text before it takes over a job, because quality in Greek varies a lot from model to model. For where the popular AI tools send your data, see 7 AI employees and where your data goes.
Frequently asked questions
Is EU cloud “private enough” for medical data?
With zero retention, a separate project per clinic and a processing agreement, yes for most uses, such as appointment reminders or message sorting. For patient records or transcriptions we recommend option 2 or 3, and we say so from the start.
Can I start in the cloud and move later?
Yes, and that is the usual path. The n8n flow stays the same, only the model’s address changes. The move is a small task, not a new project.
Do I need a graphics card for option 2?
For classification, data extraction and short replies, not necessarily. For composing long Greek text or transcribing at volume, yes, or you keep the hybrid setup.
Who is responsible if something goes wrong?
In every option Dabster Labs is the processor for the automation, under a written agreement. In option 3 the client additionally carries the physical infrastructure, and we state that clearly in the proposal.
The next step
You do not have to pick an option before you know what you will automate. The process audit maps the tasks worth automating, and for each one we write which of the three options fits and why, with a fixed price before you sign. If you already hold a written “no transfer” commitment from a client or a regulator, tell us in the first conversation, because it changes the design from day one.
