Published on September 17, 2026
The Swiss Parliament has chosen its AI. Here's what it's still forbidden to do with it.
On September 4, 2026, the Federal Assembly's Administrative Delegation approved the deployment of Pia, the Swiss Parliament's AI assistant. A national operator, data stored in Switzerland, open-weighted models: everything typically described as a sovereign AI. And yet, Parliament forbids it from reading its own most sensitive documents. This prohibition is what's interesting.
Pia, in one minute
Pia — short for "Parlamenti Intelligentia Artificialis" — is the AI assistant made available to members of parliament, their personal staff, group employees, and staff of the Parliamentary Services. Swisscom is responsible for its operation and data storage, exclusively in Switzerland. Users choose from several open-weighted models, which are regularly updated, and the entered data does not train the models. The pilot program will run for approximately one year, starting with the autumn 2026 session, with a maximum budget of CHF 150,000 drawn from the existing IT budget of the Parliamentary Services.
Everything one needs to be sovereign. And it's still not enough.
Let's review the list of criteria. Hosting in Switzerland: yes. Operator subject to Swiss law, outside the scope of the US CLOUD Act: yes. Open-weighted models, therefore auditable and replaceable: yes. No training on entered data: yes. On paper, Pia ticks all the boxes usually used to describe an AI as sovereign.
And despite that, according to ICTjournal, Its use is currently limited to classified documents up to the "INTERNAL" level. Not above. It also cannot perform in-depth internet searches or access current data.
In other words: Parliament has built the most sovereign infrastructure within its reach, and yet it has still set a ceiling. If Swiss accommodation were sufficient to resolve the issue, this cap would have no reason to exist.
INTERNAL, CONFIDENTIAL, SECRET: what this ceiling means
The Ordinance on the Protection of Information of the Confederation (OPrI, RS 510.411) defines three levels of classification:
- INTERNAL (art. 7) — information which, if disclosed by unauthorized persons, may harm the interests of the country.
- CONFIDENTIAL (Art. 6) — information whose disclosure may have prejudice to the interests of the country.
- SECRET (Art. 5) — information the disclosure of which may have an serious harm to the interests of the country.
Pia stops at the first rung. The two levels above remain out of reach of a system that nevertheless runs at Swisscom, on Swiss soil. The reason has nothing to do with the country of hosting: The more sensitive a document is, the less willing we are to let it leave the system where it originated. — even to go to a data center located thirty kilometers away.
While we were debating, the minutes were already being sent out.
The other half of the story is less glorious. According to watson.ch, Parliamentary staff ran entire committee meeting minutes through consumer AI tools—ChatGPT, DeepL—to have them summarized, before disseminating these summaries to interest groups. Parliamentary services themselves reportedly transmitted classified documents via Microsoft Outlook, despite a 2023 directive explicitly prohibiting the use of cloud services for classified information.
National Councillor Mauro Tuena: "There is no justification for downloading the minutes." His colleague Gerhard Andrey: "It's a matter of Swiss security."«
These two facts do not contradict each other; they complement each other. When the official tool is too limited for the actual work, people find another one. Shadow AI is not a discipline problem, it is a symptom of insufficient coverage.
What this means for your business
You don't have any documents classified as SECRET. You have something else, and the hierarchy still exists, even if no one has ever written it down: a sales brochure is not a contract under negotiation, a contract under negotiation is not a labor court case, and none of this is a medical certificate in an HR file.
The relevant question is therefore not "where is my data". It is: How far is my AI allowed to go, and what specifically prevents it from doing so? Answering this question requires three things: knowing what you have, deciding what has the right to leave, and having a system where what should not leave actually does not leave.
When the document cannot be output, the template comes to it.
The Swiss cloud provides the first layer of protection: Swiss jurisdiction, no subcontractors subject to the CLOUD Act, and compliance with nLPD. For the majority of an SME's processes—supplier invoices, orders, reminders, expense reports—this is the right level, and it's what we deploy most often.
Above that, there remains a category of documents for which even the best hosting in the world is insufficient, because the constraint is no longer legal but physical: the file must not leave the premises. This is precisely why WivenLLM It runs locally, on your own machines. The document doesn't travel; the template is already there.
Parliament made the same argument with a budget of 150,000 francs and Swisscom as the competitor. It's not a question of resources, it's a question of scope.
The right question to ask yourself
PIA is good news: it's a state choosing a national operator, open models, and storage within its territory rather than subscribing to a foreign platform. But the "INTERNAL" cap is the instructive part, and it's the part no one is talking about.
Before asking a supplier if they are sovereign, ask them where they store her ceiling — and what it offers you for what is above it.
Frequently asked questions
What is Pia, the AI assistant of the Swiss Parliament?
Pia («Parlamenti Intelligentia Artificialis») is the AI assistant approved on September 4, 2026, by the Administrative Delegation of the Federal Assembly. It is operated by Swisscom, with data stored exclusively in Switzerland, and relies on open-weight models chosen by the user. The pilot program will run for approximately one year, starting in the autumn session of 2026, with a maximum budget of 150,000 Swiss francs.
Why can't Pia handle confidential documents?
Its use is limited to documents classified up to the "INTERNAL" level as defined by the OPrI. The CONFIDENTIAL and SECRET levels remain excluded: the sensitivity of a document depends not only on the country where it is hosted, but also on the number of systems and intermediaries through which it passes.
Is hosting in Switzerland enough to make an AI sovereign?
No, and the Pia case demonstrates this. Swisscom, Swiss soil, and open weighting models were not enough to grant access to confidential documents. Hosting is a necessary condition, not a sufficient one: beyond a certain level of sensitivity, only local execution guarantees that the document does not leave the system.
How can an SME apply the same reasoning?
By classifying its documents before choosing its AI. What can be shared goes through a Swiss cloud compliant with nLPD; what must not leave the premises is processed locally, on the company's machines. This is the principle of WivenLLM.
To learn more: what "sovereign" really means in Switzerland, and the functioning of WivenLLM, our model that runs on your machines.
Your documents don't all have the same level of sensitivity. Your AI should know that.