IT operations AI that runs its own model in the EU
Incident data is full of names, hostnames and the shape of your network. Some IT operations tools send it to a large AI provider; others run their own model in the EU. Here is why the difference matters and how to check what a vendor actually does.
By Abhishek Singh, Founder of Alexus AI, 12+ years building enterprise IT platforms. First published 10 November 2026.
Is there IT operations AI that runs its own model in the EU?
Yes. Some IT operations tools run a self-hosted open-weight model (a model whose weights are published, so anyone can run it on their own servers) in the EU, instead of sending prompts to a large AI provider's API. Alexus AI is one: by default its model runs in the EU on infrastructure it operates, calls to outside AI providers are off unless a customer switches them on, and personal data is filtered out before anything is sent.
Many AI features in IT operations, monitoring and service management products work differently. They pass incident text, logs and tickets to a hosted model from an AI provider, sometimes with an EU data region, sometimes not. Neither approach is wrong, but they put your data in different places, under different contracts, so it is worth knowing which one a tool uses before you connect it to production.
Why does it matter where the model runs?
IT operations data is more sensitive than it looks. A ticket or an alert can carry names, email addresses and usernames, IP addresses, hostnames, error messages that expose versions, and sometimes a secret pasted into a log. A dependency map adds the shape of your estate: what talks to what, and what breaks when it fails. Whoever runs the model sees the prompts built from that data.
- GDPR: an AI provider that processes prompts for a vendor is a sub-processor (Article 28), so it belongs in the vendor's data processing agreement. If data goes outside the EEA, Chapter V needs a legal basis, such as an adequacy decision (for the US, the EU-US Data Privacy Framework covers certified companies) or standard contractual clauses.
- NIS2: an AI service inside your incident tooling is part of your supply chain under Article 21(2)(d), so its security and its availability during an incident become your concern.
- DORA: for a financial entity, an AI service that supports ICT incident handling is an ICT service provided by a third party, and Article 28(3) requires every such contractual arrangement to be in the register of information, including subcontractors behind critical or important functions.
- Fewer parties: every extra provider that sees your incident data is one more contract, one more breach notification chain and one more place your network map lives.
- Laws outside the EU: some public-sector and regulated buyers also ask whether a provider is subject to non-EU laws that can require access to data, such as the US CLOUD Act. Where the provider is established matters for that question, not only where its servers are.
What are the three ways an IT operations tool can use AI?
- It calls an AI provider's hosted model. The tool sends a prompt to the provider's API and gets an answer back. This gives access to the largest models with no infrastructure to run. Your data goes to the provider under its terms, and an EU region, if offered, changes where it is processed, not who processes it.
- It runs a provider's model in the vendor's own cloud account. Some clouds let a vendor deploy a provider's model in a chosen region inside its own account. Data stays in that region and that account, but the cloud provider and its terms still apply.
- It runs an open-weight model itself. The vendor downloads a published model and runs it on servers it operates, in a location it chooses. Prompts never reach an AI provider at all. The trade-off is that the vendor carries the cost and the operations work, and the largest proprietary models are not available this way.
What should 'runs its own model in the EU' mean?
The phrase is easy to say and hard to check. A claim worth trusting is specific on each of these points:
- Where inference runs: the country and cloud region of the servers that run the model, and who operates them.
- Which model: the model family and its licence, so you know it is open-weight and self-hosted rather than a call to someone else's API.
- Fallbacks: whether the tool can ever send a prompt to an outside provider, for example when the local model is busy, and whether that is off by default and controlled by you.
- Filtering: what is removed from a prompt before it leaves the vendor's environment, if anything ever does.
- Logging and retention: whether prompts and answers are stored, where, for how long, and who can read them.
- Training: a written commitment that your data is not used to train or fine-tune models, including the vendor's own.
- Sub-processors: the list in the data processing agreement, which should match the answers above.
What should you ask a vendor about its AI?
- When your AI answers a question about our incident, which servers process the prompt, in which country, and who operates them?
- Is any prompt ever sent to an outside AI provider? If so, is that off by default, and can we see and change the setting?
- What do you remove from a prompt before it goes anywhere outside your environment?
- Do you store prompts and answers? Where, for how long, and can we delete them?
- Is our data used to train or tune any model? Where is that commitment written?
- Which sub-processors handle AI workloads, and where are they established?
- Does the AI act on our systems, or only advise? Who approves an action?
Is a self-hosted model good enough for IT operations?
For most IT operations work, the hard part is not language, it is facts: which service runs on which database, what changed at 13:55 and who approved it, which incident looked like this one last month. A model of any size gives poor answers if it has to guess those facts, and good answers if they are retrieved from records and put in front of it.
That is why grounding matters more than model size here. A tool that answers from a current graph of your systems, your change records and your incident history, and shows the records it used, can get strong results from an open-weight model. A tool that relies on the model's general knowledge will not, whichever model it uses.
How Alexus AI does it
Alexus AI is a Dublin-built operational intelligence platform. Its default model is a self-hosted open-weight model run in the EU, and the platform is hosted and processed in the EU. It reaches models through an abstraction layer, so you are not tied to one vendor. Sending anything to an outside AI provider is off by default: it is a setting you control, and personal data is filtered out before anything leaves. Your data is not used to train models.
Answers are grounded in a live graph of your services and dependencies, built read-only from the tools you already run (status on the connectors page), with the change and incident record around it. AI-proposed causes are stored apart from confirmed ones, and a person approves any action. See the security page for how data is isolated per customer, and the platform for what the assistant does. Alexus AI is early-stage and working with design partners. This guide is general guidance, not legal advice.
Frequently asked questions
- Does an EU data region make an AI provider's model GDPR-safe?
- It helps with where data is processed, but it does not settle the question. The provider is still a sub-processor that must be in the data processing agreement, and a transfer can still arise if staff or sub-processors outside the EEA can access the data. Check the agreement, the sub-processor list and the transfer mechanism, such as the EU-US Data Privacy Framework or standard contractual clauses.
- What is an open-weight model?
- A model whose trained weights are published, usually under a licence that allows others to run it on their own infrastructure. That lets a vendor run the model on servers it controls, in a location it chooses, so prompts are never sent to the company that trained the model. Licences differ, so check the terms for commercial use.
- Is self-hosted the same as on-premises?
- Not necessarily. Self-hosted usually means the vendor runs the model on infrastructure it operates rather than calling an AI provider. On-premises means it runs inside your own environment. Both keep prompts away from AI providers; they differ in who operates the servers and where they sit.
- Does Alexus AI send our data to OpenAI or another AI provider?
- Not by default. The default model is self-hosted and runs in the EU, and sending prompts to an outside AI provider is off unless you switch it on. If you do, personal data is filtered out first. Your data is not used to train models.
- Which matters more for IT operations AI, the model or the data?
- The data. Most IT operations questions are about facts in your estate: what depends on what, what changed and who approved it, what happened last time. A model answers those well only when the facts are retrieved from current records and shown with it, so a grounded open-weight model can outperform a larger model that has to guess.