Sovereign AI
NIS2 in Austrian Hospitals
Introduction
Since 1 October 2026, Austria's NISG 2026 applies, and the health sector is one of its eleven highly critical sectors. That turns every AI tool a hospital uses into a supplier that has to be assessed, documented and defended.
Picture the moment this becomes real. A hospital's CISO sits down to prepare the self-declaration for the Bundesamt für Cybersicherheit. The list of ICT suppliers is long but familiar: the hospital information system, the PACS, the lab software. Then a ward manager mentions that half the team drafts discharge letters with a public chatbot. Nobody approved it. Nobody knows where those prompts are stored.
That chatbot is now part of the hospital's supply chain. This article explains what the NISG 2026 asks of hospitals, where AI already sits in that supply chain, and which questions to put to any AI vendor before the next deadline.
What applies now, and what comes next
The NISG 2026 is Austria's implementation of the EU NIS2 Directive. It was published on 23 December 2025 as BGBl. I Nr. 94/2025 and has applied since 1 October 2026, with no transition period for the substantive obligations.
Hospitals are in scope if two tests are met. The activity falls into one of the 18 sectors, and health is listed in Anlage 1. The organisation is at least medium-sized: 50 or more employees, or annual turnover and balance sheet total each above €10 million (§ 25). Large Anlage 1 organisations count as essential entities (§ 24). Each legal entity runs this check separately, so a hospital group may contain both categories. The WKO NIS FAQ explains the scope rules in more detail.
When | What is due | Legal basis |
|---|---|---|
1 October 2026 | Risk management, incident reporting and management duties in place | § 51 NISG 2026 |
By 31 December 2026 | Registration with the Bundesamt für Cybersicherheit | § 29 NISG 2026 |
By 30 September 2027 (within 12 months) | One-off self-declaration on how the requirements are met | § 33 (1) NISG 2026 |
From 1 October 2028 | Audit report from an independent body, on request | § 33 (2) NISG 2026 |
Timeline overview: SECJUR, NISG 2026. Official updates: Anlaufstelle NISG.
Two points matter most for AI. Management bodies are personally responsible for overseeing the measures (§ 31), so AI use cannot stay an IT side topic. And the obligations explicitly include the security of the supply chain: direct suppliers and service providers have to be assessed and secured by contract.
Where AI already sits in a hospital's supply chain
Most hospitals have more AI suppliers than their supplier list shows. They usually arrive through four doors:
Tools staff use on their own. Public chatbots for discharge letters, summaries, translations or emails to patients. No contract, no assessment, and patient details pasted into a browser tab.
AI features inside existing software. Assistants added to office suites, documentation systems or the hospital information system through an update. The supplier is known, but the AI part often runs with a different provider in a different location.
Specialised clinical tools. Imaging analysis, speech recognition for documentation, coding support. These usually went through procurement, but often before anyone asked where the model runs.
Pilots and projects. A department testing an AI assistant with a start-up, a research collaboration, a proof of concept that quietly became daily use.
Each door creates the same three obligations. The supplier has to be in the risk analysis. The contract has to cover security. And if something goes wrong, the hospital has to report it: an early warning within 24 hours, a full notification within 72 hours, and a final report within a month (§ 34). When patient data is involved, a separate notification to the data protection authority under Article 33 GDPR runs in parallel.
A hospital cannot report on an incident at a supplier it does not know it has. That is why unofficial AI use is no longer only a data protection question. It is a gap in the supply chain picture the NISG requires.
Three questions to ask any AI vendor
The NISG does not prescribe specific technologies. It asks for measures that fit the risk, and for proof that they are in place. For AI tools, three questions cover most of that proof.
1. Where does patient data go when someone types a prompt? Ask for the exact processing location of every prompt, document and output, including subprocessors and the model provider behind the product. "EU hosting" on the website is not an answer if the model itself is called through a provider in another jurisdiction. The answer belongs in the supplier assessment.
2. Who can access it, and under which law? List everyone who could technically see the data: the vendor's staff, the cloud operator, the model provider. Then ask which legal regimes apply to each of them. A foreign authority's access request is a supply chain risk the risk analysis has to name.
3. What is logged, and who holds the logs? When an auditor asks how AI is used, the hospital needs records: which model answered, which department sent the request, when. If those logs sit with the vendor, the hospital depends on the vendor to prove its own compliance.
A vendor that answers all three clearly makes the self-declaration easier. A vendor that cannot answer them is a finding waiting to happen.
What changes when the model runs on the hospital's own servers
When the AI model runs inside the hospital's own infrastructure, the three questions get short answers. Patient data stays on hardware the hospital already controls and already covers in its risk analysis. Access follows the hospital's own access rules. And the usage records stay in-house, where the hospital can show them to an auditor directly.
This does not make AI a supply chain topic of the past. The software that runs the models is still supplied by someone and still belongs in the assessment. But the question shifts from "where does our data go?" to "how is the software on our servers secured?", which is a question hospital IT already knows how to answer.
On-premise used to mean giving up the convenience of cloud AI. That trade-off is smaller now. Open-weight models have become capable enough for everyday clinical writing tasks, and control planes make running them feel like using a cloud service.
This is what Xinity does. Xinity lets hospitals deploy, govern and audit open AI models on their own servers, with a similar to a cloud experience for the staff. Departments get controlled access to the models they need, IT sees which models are running and who uses them, and the records of every request stay inside the building. The platform is developed in Austria and its code is open source on GitHub, so the hospital's security team can review exactly what runs on its servers.
A short checklist before registration
Confirm per legal entity whether the hospital is an essential or important entity, and document the result.
Ask each department which AI tools it actually uses, including unofficial ones.
Add every AI tool and AI feature to the supplier list and the risk analysis.
Put the three vendor questions to each AI supplier and file the answers.
Decide which AI use cases involving patient data should run on the hospital's own servers.
Make sure AI incidents are part of the 24-hour reporting process.
Brief the management body, since oversight is their personal responsibility.
Try it on your own hardware
Xinity offers a 30-day pilot on the hospital's own servers. IT gets a working AI platform with open models, and the team can test real clinical writing tasks without a single prompt leaving the building. Start your pilot or talk to us about your NISG preparation.
This article is general information and not legal advice. For a binding assessment of your organisation's obligations, consult the law text or a legal advisor.
The article was drafted by AI, and reviewed sentence by sentence by the Xinity team.
Sources
Anlaufstelle NISG
WKO: NIS FAQ
Regulation (EU) 2016/679 (GDPR), EUR-Lex