Alberta has been using artificial intelligence to rebuild the software that runs the provincial government. On Tuesday, Quebec signed a deal to get its hands on how.
The two provinces signed an agreement in Quebec City on July 14. Alberta's Minister of Technology and Innovation, Nate Glubish, signed it with Quebec's Minister of Cybersecurity and Digital Technology, France-Élaine Duranceau, and Quebec's Minister Responsible for Canadian Relations, Jean Boulet.
The deal itself is modest on paper. It runs five years, carries no financial commitment, and either province can walk away with 60 days' notice. A joint steering committee, two people from each province, will draw up a work plan and look for pilot projects to run together. What it commits the two governments to is sharing: AI strategies, governance policies, training materials, and where it makes sense, actual technological assets like source code and documentation.
In plain terms, neither province wants to build from scratch what the other has already built. And right now, Alberta is the one with more built.
What Alberta has actually been doing
The agreement only makes sense against what Alberta has spent the last year and a half building, and that's the real story here.
Alberta's technology ministry maintains the systems for all 27 provincial ministries, from social services to public safety to wildfire response. That's roughly 1,280 applications and 3,400 code repositories, some of it decades old, much of it never properly documented, and most of it never given a systematic security review. The accumulated problems, insecure code, unaddressed bugs, outdated software, run into the billions of dollars to fix. Doing the full modernization the conventional way, the province estimates, would cost about $2 billion and take more than a century.
So instead, Alberta built its own AI tools and pointed them at the problem.
A real example: 466 million lines of code in 20 hours
Anthropic, the company behind the AI models Alberta used, published its own account of the project on July 6, and it puts hard numbers to what the province did.
Alberta's team used Claude Code, Anthropic's coding tool, running on its Opus and Sonnet models, to review the systems the ministry maintains. About 50 AI agents worked in parallel, scanning for security holes, weak points in infrastructure, and gaps in documentation. In roughly 20 hours, they assessed 466 million lines of government code. Alberta estimates the same review done the traditional way would have taken about six and a half years.
The systems involved aren't trivial. They hold tax records, procurement data, and social services case files, some of the most sensitive information the province keeps.
Finding the problems was only half of it. Where the scan turned up a vulnerability, the AI could often write a fix, test it, and build it, with the ministry's own engineers reviewing and approving every patch before it shipped. In some cases the tools rebuilt outdated systems from scratch. One example stands out: a subsidy program portal originally hand-coded in Java about 25 years ago, which took five months to build the first time, was rebuilt in as little as four to five days.
The province also built AI agents that keep watching after the fact. A "red team" agent probes an application the way an attacker would. A "blue team" agent checks its defences against an international security standard and writes up exactly what to fix. Every application gets run against roughly 95 security controls on each pass.
That's where Glubish's "95 per cent less cost" figure comes from. It's a real, specific claim tied to this work, not a vague talking point. As another example, the province says its team built the infrastructure department a new system to manage properties like hospitals and schools in ten months, for $2.5 million.
What the tools are built on
There's a detail worth knowing that Alberta's own release doesn't dwell on: it built this on commercial AI, then gave the method away.
The province built its system using Anthropic's models, the company behind Claude, and open-sourced the entire approach so other governments and companies could copy it. The Velocity White Papers that Glubish keeps referencing, 21 technical papers Alberta published as a blueprint, are that open-sourcing effort. The deal with Quebec is, in part, Alberta finding its first big taker.
Glubish has a bigger ambition attached to all this. He's said he wants Alberta to become "the compute capital of Canada," the place that physically houses the data centres powering AI across the country, run on Alberta natural gas. It's the same push behind Meta's recently announced $13-billion data centre outside Edmonton. The AI-in-government work and the compute ambition are two halves of one strategy: build the tools, then own the infrastructure they run on.
What it means at the counter
Strip away the intergovernmental language and the question for a regular person is simple: does this change anything when you're waiting on the government for something?
The provinces say the goal is faster processing of files, requests and applications, better access to government information, and more responsive service. Quebec's minister framed it as speeding up how the state handles "files, requests, and applications." That's the pitch: quicker, cheaper, less waiting.
Whether it delivers is the open question, and it's worth being clear that most of the evidence so far comes from the governments themselves, and from Anthropic, the vendor whose tools are being used. The cost and speed figures are Alberta's own estimates of its own project. The 466-million-lines and 20-hours numbers come from the company that sold the technology. None of that makes them wrong, and the detail is specific enough to be checkable. But independent verification of a 95 per cent saving on a modernization that isn't finished doesn't exist yet, because the work isn't finished. The claims are plausible and unaudited at the same time.
The training piece
One part of Alberta's push is further along and easier to measure.
The province launched the Alberta AI Academy in September 2025 to train public servants to use AI tools. It says more than 2,000 Alberta public servants have gone through it, and that more than 15,000 people across Canada, from other provinces and the federal government, have used the platform. Alberta has also rolled out Microsoft Copilot and Google Gemini to government staff, with usage and governance policies attached.
That training and workforce material is among the first things the two provinces will pool under the new agreement.
Where this sits
This is one deal in a busy moment. Quebec signed a separate exploratory AI agreement with Canadian company Cohere earlier this year, and has its own multi-year plan to bring AI into the public service. The federal government has a national AI strategy of its own. Governments across the country are moving on this at once, with varying degrees of caution.
Alberta's bet is that it's ahead, and that being ahead is worth something, both the savings it's claiming and the position of being the province others come to rather than the other way around. The Quebec deal is a test of that bet. If Alberta's tools work as well for Quebec as the province says they've worked at home, it's a template others will follow. If the savings prove softer than advertised once the work is actually done, that will show too.
For now, two provinces have agreed to compare notes, and Alberta has a partner willing to try what it built.
Sources:
Government of Alberta, "Alberta and Quebec join forces to put AI to work," July 14, 2026 (alberta.ca)
Government of Quebec, statement from Minister France-Élaine Duranceau, July 14, 2026 (quebec.ca)
Anthropic, "Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities across government systems," case study, July 6, 2026
The Logic, reporting on Alberta's government AI stack and compute strategy (Nate Glubish interview)
Meridian Source, Alberta legacy-systems modernization figures
Statements from Ministers Nate Glubish, France-Élaine Duranceau and Jean Boulet
