“They quoted $54 million. Alberta said no.”

Fri, 15 May 2026 19:04:25 +1000

TL;DR — Alberta’s Provincial Government replaced two failing legacy IT systems in 10 months for an estimated $2.64 million total. Vendors had quoted $54 million for one system, almost certainly heading to $100 million. AI tools — particularly Gemini’s vision capabilities — compressed months of requirements work into minutes. The bottleneck wasn’t technology. It was leadership willing to ask whether there was another way. Almost everyone wins when governments do this. Except the large IT vendors and consultants who’ve been pricing uncertainty for decades.

The quote came in at $54 million. For one system. Over four years. With no guarantee it wouldn’t balloon to $100 million by delivery day - which is roughly what history says you should expect from large government IT projects.

As Nate Glubish, Alberta’s Minister of Technology and Innovation, puts it from direct experience:

“A project like this would more realistically land north of $100 million by the time it’s actually delivered. And by then, the technology has evolved so much that the finished product is already outdated before you even turn it on.”

Alberta’s Ministry of Infrastructure - part of the Provincial Government of Alberta, Canada - needed to replace two critical systems: one tracking 4,000 government-owned buildings and $12 billion in assets, the other managing capital construction across 500-plus active projects. Both were failing. Staff were literally spending their working days copying data from one broken system and pasting it into another. Full-time human bridges between software that couldn’t talk to itself.

Three separate replacement attempts between 2016 and 2024. All three failed.

Glubish has written a detailed first-hand account of what the government did next. It is worth reading in full. The short version goes like this.

After a complete review into the failures, something unusual happened. The Deputy Ministers looked at a $54 million quote - for half the solution - and killed the procurement.

Not because the vendors were incompetent. The bids reflected exactly how large government IT projects get scoped and priced. Vague requirements create expensive proposals, because vendors price in all the uncertainty they’re inheriting. That’s a system problem. The system has been running that way for decades.

Instead, they asked a different question: what if a small team of public servants, equipped with modern AI tools, just built it themselves?

What they built, and what it cost

The PRISM Initiative launched in June 2025. A small internal team - domain experts who understood the business processes, not just developers who could write code - started building working software and putting it in front of real users every two weeks.

Ten months in, 643 government employees are using both systems in production every day. The team has spent $858,000. At the current monthly burn of around $118,000, the total cost to fully deliver both systems sits at an estimated $2.64 million.

That’s a 95% cost reduction. Both systems instead of one. Already live, not theoretical.

What AI actually did here

AI did not build these systems. The team did. But AI made them dramatically faster in ways that go well beyond autocomplete-for-code.

Infrastructure had over 50 hours of recorded walkthroughs of their legacy systems - staff filming themselves navigating every screen, every workflow, every data entry process. In a traditional project, you pay consultants to watch those recordings, interview users and produce a requirements document. It takes months and costs millions.

The PRISM team fed those recordings into Google Gemini’s vision capabilities. The AI processed the video frames, analysed every screen and workflow, and transformed 50-plus hours of recordings into structured requirements, user flows and data models. At roughly one cent per image processed, running in parallel, the conversion took minutes.

Not days. Minutes.

Beyond requirements capture, AI compressed prototyping to the point where developers were building working interfaces live on video calls with users. Staff could see their feedback take shape in real time. And once they saw that, something shifted - a handful of non-technical Infrastructure employees started using rapid prototyping tools themselves to sketch interface ideas and hand them directly to the development team.

Government employees who are not software engineers, communicating their ideas visually in real time to the people building their systems. That’s a culture shift you can’t mandate. It emerges when people see AI delivering real value and realise they can participate.

The thing that made it possible

Glubish is explicit that AI amplifies skilled people - it does not replace the need for expertise, judgement or deep understanding of what users actually need.

Cohen McLeod, the director who led the project, put it plainly: “AI makes development a lot faster, but not on everything. We’ll be working faster than anticipated, then hit something that we just have to do by hand.”

What AI does is compress the grunt work - requirements capture, prototyping, iteration, routine code generation - from months into days. That compression changes the economics of everything downstream.

Why it keeps going wrong everywhere else

The problem with the traditional model is structural, not accidental.

That ambiguity doesn’t disappear when a vendor wins the bid. It gets priced into the contract - by the vendor, in their favour. Every gap in the requirements specification becomes a scope negotiation later. Every policy decision that lands after signing - and governments change policy constantly - hits a contract that was never designed to absorb it.

In Australia, the Victorian Auditor-General has documented the pattern in detail. In a recent review of a major government technology contract, VAGO found the department had signed an agreement with an overly optimistic delivery schedule and unresolved intellectual property risks. Months were lost to standstill. When policy changed mid-contract - as it always does - sections of the system paused while commercial implications were worked through. The auditor-general’s finding on value for money was a single word: “unclear.”

That is not a description of one bad project. It is a description of what the procurement model reliably produces. Vendors are rewarded for managing scope, not for delivering what users actually need. By the time the system finally arrives - late, over budget, already ageing - the requirements that scoped it have been overtaken by the world it was meant to serve.

The systems nobody writes about

Every government runs hundreds of internal systems that never make headlines. Grant management platforms. Asset registries. Permit tracking tools. Workforce scheduling systems. Compliance reporting databases. Most of them were built 10 to 20 years ago. Many are approaching end-of-life on ageing infrastructure, running on databases the vendor no longer supports, maintained by staff who can’t take leave without putting the system at risk.

That is exactly where Alberta’s Ministry of Infrastructure was two years ago.

Each of those systems will eventually hit a procurement list. A requirements brief will be drafted. Vendors will be invited to bid. Proposals will come back with four-year timelines and seven-figure price tags. Scope will drift. Costs will grow. The finished product will be outdated before it goes live.

That cycle is not inevitable. It is a default - one that persists because nobody has yet had sufficient reason to question it.

Alberta gives you the reason. AI tools have compressed the development economics to the point where a small team of capable people - with genuine domain expertise and leadership backing - can deliver production software for a fraction of the vendor quote, faster than the procurement process would have concluded.

What would actually need to change

The shift is not technical. Every government that can issue a procurement can also access the AI tools Alberta used. The technology is not the bottleneck.

What Alberta had - and what is genuinely rare - was leadership willing to internalise the risk that has always been outsourced to vendors. That means a Deputy Minister who can look at a $54 million quote and say no. It means Ministers who will back internal teams to attempt something the public service hasn’t tried in decades. It means a procurement culture that treats building your own software as a legitimate option, not an eccentric one.

The cultural change required is specifically about confidence - the confidence of senior public servants that their own organisations are capable of something this significant, and the confidence of their ministers to let them try.

Glubish frames the ambition plainly:

“If a small team can save taxpayers over $50 million on a single IT project, imagine what becomes possible when thousands of people are equipped with the same tools and training.”

Almost everyone is a winner

Better government technology - built faster, aligned to actual user needs, responsive to policy change - produces better public services. Citizens get faster processing, fewer errors and interfaces that reflect how work actually happens. Taxpayers get systems at a fraction of the cost, without the nine-figure blowouts and the paying for the same failure three times.

Public servants get something rarer still: the experience of building something that matters, seeing it used, and improving it in response to real feedback. That is more meaningful than maintaining a vendor relationship for a decade.

There is one group that does less well under this model. The large IT vendors and management consulting firms that have built profitable businesses around the complexity of government procurement - the ones who price uncertainty, manage scope and bill by the day for requirements that a Gemini API call now processes in minutes.

Their margins get squeezed.

Good.

Sources:

Rambling Rows