Private AI for public bodies in Québec City

Quebec public bodies work under the provincial privacy statute together with the access rules that govern their documents. The modernised obligations add notice and transparency expectations that a tool used on public documents has to accommodate.

The working language of the administration is French. That is a requirement, not a preference, and it applies to any AI system that touches the work.

Québec City holds the ministries and agencies that administer the province, working in French under the provincial privacy statute.

Records obligations here assume French-language documents and French-language answers, which is a real constraint on tooling chosen elsewhere.

Stored data — accounts, conversations, memory and files — sits in Canada, in AWS Canada Central (Montréal). The models that write answers run on NVIDIA GPU hardware located in Canada and operated from Canada. That second part is the one most vendors gloss: storage residency is not the same claim as inference residency, and for a Québec City public body the difference is exactly the one a privacy review will ask about.

If the requirement is that nothing leaves infrastructure you control, Digby can be self-hosted with Docker Compose against any OpenAI-compatible model server. That is a different trade — you run it — and it is documented rather than implied.

1. 01

   Published documents first

   start with material that is already public, so nobody debates whether it was allowed.

2. 02

   One team, one question set

   prove it on the five questions a front counter actually gets asked.

3. 03

   Bring records management in early

   the people who own the schedule should shape what goes into a project.

Québec City firms and organisations often work across provincial lines, and the statute changes at the border. The same public sector material, set up for the next province over: