What Alberta’s AI numbers mean for IT pricing

The next three-year, eight-figure technology estimate to land on a Canadian CIO’s desk might have to answer a different question.

What part of it still depends on people doing the work the old way?

Alberta has now put some public numbers behind that question. At the Velocity Symposium in Edmonton last week, Janak Alford, Alberta’s deputy minister of technology and innovation, walked attendees through how the province is using AI to map, repair, and reduce the application and code behind government services.

The event followed Alberta’s release of the Velocity White Papers on July 6, a set of 22 open-source documents that show how the province is doing this work and hand the playbook to anyone else who wants it.

Nate Glubish, Alberta’s minister of technology and innovation, said one project had drawn a bid of $54 million over three years. He said the province is developing a solution for about $2.5 million, with a first version working within 10 months.

The province also used AI to scan more than 466 million lines of code in 20 hours, work Alford discussed at Upper Bound in May.

According to a case study Anthropic published alongside the white papers, the compute cost was about $2,000. 

That figure doesn’t include the people, systems, and preparation required to build the scanning process, but it shows what can happen once the machinery is in place.

Alberta’s figures give buyers something concrete to challenge. A vendor quoting three years and eight figures now has to show which parts of the work still require that much time and labour.

Buyers with maps negotiate differently

Most organizations have a better inventory of their office chairs than their software.

Alberta knew it had an aging and complicated set of systems, and the AI scan gave the province a much clearer view of what was inside.

Alford said the scan covered more than 4,000 applications and mapped their code to almost 2,200 business capabilities. The AI assessment recommended maintaining about 1,500 repositories, repairing 1,375, consolidating 739, and rebuilding roughly 400. 

The province is still validating those classifications manually, he added.

Alford said the current process leaves the government guessing what vendors can do, and vendors guessing about undocumented systems. 

Once the province had a clearer map of what it was running, it had more to test a bid against.

“The vision is to use AI not to vibe code 4,000 more applications. It’s to get rid of 2,000 applications along the way,” said Alford.

Alberta’s assessment identified 739 repositories that could be consolidated, and Alford said the broader goal is to reduce the number of applications the province maintains.

Fewer systems can reduce maintenance, licensing costs, integration work, and the number of applications security teams have to defend.

What AI removes and what it adds

Once the scan flagged code that needed updating, Alford’s team ran it through what they call the AI garage. 

Five AI agents get the same problem and each writes a plan for how to fix it. The agents review each other’s plans, a lead agent chooses one, and a separate agent does the work.

For repositories that can be updated, the AI work takes about an hour on average. It checks dependencies, looks for vulnerabilities and exposed secrets, attempts the fix, and prepares a pull request for review.

This is the work buyers should expect to see reflected in a project estimate.

Dependency updates, code documentation, vulnerability triage, repetitive migrations, and other well-defined tasks can absorb thousands of hours when people handle them one repository at a time.

A July 2026 study followed 802 developers and 196,212 pull requests at a mid-sized, AI-focused company. By April 2026, merged pull requests per developer had reached 2.09 times the pre-mandate baseline.

The review load per reviewer roughly doubled over the same period. Automated reviews overtook human reviews.

Adoption wasn’t randomized, so the researchers can’t say AI caused all of it. But at this company, higher output came with more review work. A good estimate would account for both sides.

On the event’s closing panel, Alford was joined by Cory Janssen of Edmonton-based AltaML, John Cousens of Google, Jeremy Crossman of Microsoft, Natasha Clarke, former deputy minister of cyber security and digital solutions in Nova Scotia and now an executive advisor at Kainos, Cole Cioran of Info-Tech Research Group, and Magne Vange of Cisco.

The discussion covered project pricing, AI-generated code, employee adoption, process design, organizational change, and accountability.

Janssen said AltaML works with law firms and is looking for clients ready to pay a fixed price for the work rather than by the hour. He sees the same opening in engineering and accounting. 

He used a hypothetical to describe what fixed-fee pricing could look like on a project of that scale.

“That $20 million project, I’m going to do it for two,” he said. “And if I screw up, that’s on me.”

Janssen was putting a price on work he says can now be done differently. Google offers one clue about how much the production side has changed.

At Google, AI is already generating most new code. Cousens put the figure at roughly 70%. Google’s published number is 75%, with engineers approving the output. The missing number is what all that approval costs, and whether it reduced the total project bill.

Vange focused on the rest of the workforce. Cisco is rolling out personalized AI assistants to about 90,000 employees, alongside company-wide training and internal knowledge sharing.

Back to Janssen’s hypothetical, buyers still need to ask what’s included in the $2 million and what was included in the $20 million.

A lower estimate needs to account for integration, testing, training, security reviews, data cleanup, and the people required when the AI gets stuck.

Alford said the province has used the process on hundreds of applications, and estimated AI gets about 70% of the way through much of the work before engineers step in. He said the province is still validating the changes manually. What’s left may need more planning, or involve systems the agents can’t modernize.

What still costs money

AI can read and repair the code. The decisions, compromises, and workarounds that produced the system don’t show up in the code itself.

That context may live in a ticketing platform, a forgotten SharePoint site, an integration diagram, or the head of the employee who has spent 15 years keeping the application alive. Alberta is now working to connect those sources so its AI systems can see more than the code itself.

Then people have to use the new system.

Employees need training, processes need to change, and privacy, security, legal, and audit teams still have questions to work through. Someone has to approve the result, and someone has to answer when it fails.

Clarke saw this dynamic across her time in Nova Scotia’s government

“If we don’t start to look at those incentives, if we don’t start to look at how we fund things, how we hire people, procurement, that machinery has a way of grinding down transformation,” she said.

Code may arrive faster while training, approvals, and organizational work keep their old timetable.

What to ask before the next bid

The next bid meeting is where this becomes useful.

Ask the vendor to show which tasks AI will perform, which tasks still require engineers, and how those efficiencies changed the estimate.

Ask how many existing applications the project will remove. If the new system leaves the old one running, the buyer pays for both.

Ask how the output will be tested, who approves it, and what happens when the automated checks fail.

Ask what information the vendor needs to understand the existing environment. A vague request allows uncertainty to become expensive.

Glubish said Alberta’s new understanding of what it runs has changed what it expects from technology companies.

“We are a more sophisticated customer today than we were two years ago,” said Glubish.

The next time a three-year, hefty-price-tag quote lands on a CIO’s desk, Alberta’s figures give the buyer another public comparison to put beside it.

Final shots

• AI efficiency should show up in a vendor’s price. So should the cost of review, integration, and accountability.
• A buyer that can’t map its own systems is likely to pay more for uncertainty.
• An estimate is easier to defend when it separates automated tasks from work that still needs experienced people.