No. 12 · Policy & People

The AI Academy: Investing in People

A graduated, open-source program that trains public servants to work with artificial intelligence, from prompting through building enterprise-grade applications.

The Alberta AI Academy began with what we learned in the field. When Technology and Innovation launched the AI Maximalist program in early 2025, we selected people through a call of interest, independent of their technical skills and independent of their AI experience, on the strength of their desire to take part. Around ten percent of the workforce volunteered, and for operational reasons we were able to take roughly five percent into the first cohort. Those first sixty-five people became a foundational core, learning to apply artificial intelligence across four domains: project management, application delivery, communications, and financial analysis. What we observed watching them work became the basis for everything the Academy teaches today.


## §01 From the Maximalist program

Those first four teams gave us a deep look at which technologies were effective in a government context, and at the range of responses to being placed in a learning environment with something as new as artificial intelligence. Because these people self-selected into the program, they came with an above-average willingness to engage, and they were a positive, collaborative, and open group, so gathering feedback on what was working was straightforward. The team told us plainly where they felt they needed help.

What stood out was that the group bifurcated, between those who were immediately comfortable with artificial intelligence and those who struggled with it. That told us an effective adoption strategy would have to serve both the early adopters and those who were more skeptical or who needed more assistance along the way. A non-technical background was no barrier to success with these tools. People who did have a technical background, though, found concepts like building a harness or thinking in structured data more familiar, which gave them an early advantage on the foundational ideas.

The founding cohort 65 / ~5%. About sixty-five volunteers, roughly five percent of the workforce, formed the founding cohort of the AI Maximalist program. Close to ten percent had volunteered.

The clearest signal was the demand for a formal training program. The feedback made it plain that expanding the Maximalist program would require formalized training, and that the training could not be limited to prompting or a tour of the tools, because many people were already well past that on their own time. We needed a graduated program.


## §02 Human-centered AI

The shape of the Academy was also informed by a conversation with a colleague, JP Lalonde, then the Head of AI at the Impact Assessment Agency of Canada. We were reflecting on how negative the public conversation around artificial intelligence had become, and how worried many of our friends, family, and peers in government were about its potential impact. Much of that worry was amplified by social and popular media, by predictions of displacement and job loss from voices with a fairly negative read on the technology. As two people who had been using these tools heavily, four to six hours a day for several years, JP and I felt the concerns were overblown, and that they misrepresented the opportunity AI creates.

We saw an abundant opportunity to transform government in a positive way. Drawing on his long experience in technology delivery and digital transformation, JP coined a term we have since adopted in Alberta: human-centered AI. When we sat down to design the Academy and its levels, we built them around that principle, and around a commitment that no public servant would be left behind. Everyone would get a fair chance to see and assess these tools for themselves. With so much fear in the conversation, we also wanted to show people the positive side: how AI can enable and empower staff, and how it democratizes access to technology, especially for those who have not had the chance to build a strong technical background.

These tools let people do far more than they could before, and we see generative AI as a real benefit for workers. Artificial intelligence will change the nature of work, and the most valuable thing we can give our peers is a solid foundation to be ready for it. Through all of these conversations I have been consistent with our teams: this is about investing in people. We carry decades, in some cases centuries, of technical debt. We have case backlogs that stretch back months and years. And we have a government set on reducing red tape, so that businesses can get on with business and Albertans can get on with their lives. AI is a way to take on that accumulated work and to reduce those systemic barriers, without weakening the policies they serve. This became the rallying idea behind the Alberta AI Academy: to empower staff to use these tools well, and to ease the anxieties and fears around them.


## §03 A graduated program

We envisioned a tiered program covering the fundamentals of prompting, the fundamentals of agents, which were becoming popular in early 2025 but had not yet taken off, and the building of applications, which remains one of our most expensive and laborious activities. In June 2025 we pitched the idea of an Alberta AI Academy to our minister, Nate Glubish, who endorsed proceeding wholeheartedly. We had proposed a September launch, which left a narrow window to develop the material. Our primary source of inspiration was the Maximalist team, who through surveys and through their own experience told us which lessons landed and where they had faced the most uncertainty and frustration.

On September 9, 2025, we launched the Academy with Level 1 and an initial site of curated articles, learning modules, and tools, an integrated chat agent, and cohorts of students learning together and sharing their progress. We split the material into ten days, building from foundational concepts on day one through homework to a capstone project by day ten. We ran it on a cadence of a couple of weeks on and a couple of weeks off, alternating Level 1 and Level 2 while we developed the more complex Level 3 material.

Level 3 was the hardest course to create, because even those leading the Maximalist teams did not yet know how to explain its concepts, and we were still building the tooling and the structures that let people learn effective patterns and avoid bad habits. There is no single predetermined way to apply these tools well; it is up to the individual. Most people, though, do not flourish in pure uncertainty. They want some structure and consistency, a path they can follow. Like much of IT, there is no one way to do a thing, only more and less effective ways, and plenty of opinionated ones. To build an academy we had to make some bets, finding concepts sticky enough to remember, accessible enough for a layperson, and principled enough to drive repetition. You can learn to play an instrument or to paint any number of ways, and yet the beginner still has a few principles to master before creativity has somewhere firm to stand.


## §04 Level 1: prompting and trust

Level 1 focuses on prompting: how to interact with a chat-based agent, and how to build patterns of interaction that move our use of technology away from a transactional system and toward a human conversation, where each side adds something and the exchange has real give and take. We teach a prompting framework called RICECO, a six-step formula for a well-formed prompt that runs from the role you give the AI through to the output format you ask for, with instruction, context, examples, and constraints in between. We pair it with a second framework called TRUST: verify what the AI tells you before you rely on it. We sum it up in three words: verify, then trust.

We also teach a range of enterprise tools: Copilot in Microsoft 365, Gemini Enterprise, and our own platform, Albert, which offers tools, agents, prompts, and frameworks, and lets people build and deploy their own agents for unclassified workloads. Different people gravitate to different tools by temperament and experience. Some reach for the more complex agents early, while others are content to paste a prompt into Copilot to structure and coordinate their day within a wider team.

A good part of Level 1 is teaching people what to avoid. Used like a Google search, the tools disappoint. Trusted blindly, they mislead. Fed sensitive data on a platform that is not classified for it, they create risk. Expected to be perfect on the first try, or to understand a process they have never seen, they fall short. None of this is obvious, and none of it is advertised by the AI companies, so we teach it directly, alongside a mindset of careful use and plenty of room for people to bring their own use cases for coaching.

Level 1 satisfaction ≈ 9 / 10. Average satisfaction reported across Level 1 cohorts. Level 1 has since been compressed to a single week to lower the time commitment.

Feedback on Level 1 is consistently high, averaging around nine out of ten. Since the start of the Academy we have compressed it to a single week to make it more accessible and to ask less of people's time, while reminding them that even a deceptively simple chat interface still takes time to master. In our experience it is somewhere between two and three months before most people use these tools well on a consistent basis. Level 1 lays a foundation, and graduates are encouraged to keep using the tools as they return to their work.


## §05 Level 2: reusable agents

Level 2 builds on that foundation and introduces reusable agents. Here we use Gemini Gems, GitHub Copilot Agents, and our own AgentBuilder Console, which gives people a structured way to describe their processes and to assemble agents and tools that run repeatable workflows, in a linear or a branching pattern, ending in a range of finished documents.

For anyone working in a repeatable system, this is valuable, because Level 2 teaches the rule of three: if you are likely to do something more than three times in a week, take the time to build the agent and to document the workflow, so you can pass that work to the agent and share it with your peers. In a unit of people doing similar work, a shareable workflow means one person's investment of time raises the performance of the whole unit and inspires the people around them.

The agentic material is well subscribed, though we see our first fall-off here, since for many people simple prompting remains enough. People often take a break between Level 1 and Level 2 rather than running them back to back, both for operational reasons and to ease the learning curve, and they can repeat any level as often as they like, or work through the material self-serve on the Academy site whenever they want a refresher. We also bring people into a broader community of practice, where we coach specific use cases and the right tool for each. The more advanced needs surface in that community: a particular team's process or form, a dataset locked inside an IT system the AI cannot reach, questions about exporting and uploading. These are the cases where a custom solution becomes worthwhile, and they lay the groundwork for Level 3.


## §06 Level 3: building enterprise applications

In Level 3 the cohort learns to work with sophisticated coding agents and with the harness: the skills, standards, guides, hooks, and templates needed to build highly effective, enterprise-grade applications. For most people this is a significant step change in complexity and technical knowledge, and even for technical participants it is demanding. In our most recent cohort we saw a meaningful drop-off, with many participants not finishing or graduating, because of the technical nature of the work.

There is a misapprehension that AI will replace the coder, the architect, or the client engagement team. In practice it amplifies them, letting them produce greater output at greater velocity. The topics are genuinely complex, so we break them down and make them as accessible as we can. We lead students through the three stages of the factory covered in Pronghorn, Nexus, and Velocity; we teach the well-built harness; and we walk through the four approaches to government modernization. We coach even non-coders to understand what goes into an enterprise application, and the hundreds of controls and decisions that sit behind even the basic template they start from.

We then teach participants to build their own harness, tailored to the work they do in the organization, and we challenge them across a series of projects to lead and to demonstrate to the class how they are using the tools. The Velocity leaderboard comes in here, where people take on challenges and share their progress. Producing a quality output takes real rigor. It always has, and AI does not change that. Anyone looking for a shortcut, who skips the thinking that goes into a solution, soon finds the AI failing to deliver the outcome they were after, while those who do the work excel.

"Persistence beats cleverness. Treat the AI as a servant on a platter and it will fail you; treat it as something you have to verify, and you will catch what it quietly skipped. The skill is knowing how to check." · AI Academy Level 3, Welcome Package

Level 3 is challenging for most people, and teaching it has been just as instructive for us. I have taught Level 3 for more than twenty-five days over the past year, as the lead instructor across the levels and primarily on Level 3 myself. That time in the room shows us where people are understanding the material and where we need to invest more in the fundamentals and in building comfort.

DM Janak Alford introduces the Level 3 Academy cohort. Video: https://youtu.be/tH09fiayhdU


## §07 Warrior, wizard, craftsperson

The Level 3 room held no hierarchy beyond facilitator and participant. The class included a deputy minister, three assistant deputy ministers, executive directors, directors, managers, employees, students, and interns. Everyone followed the same lessons and built solutions as part of the learning journey, with no organizational rank in play. Instead of titles and job descriptions, we asked people to choose a persona: warrior, wizard, or craftsperson.

The framing is a deliberate, lateral nudge away from job titles. We ask people to set aside 'I am a full-stack developer' or 'I am a solution architect' for two weeks, and to consider instead which kind of work they actually want to do with these tools. The ambiguity of the framing is the point, and it doubles as a filter. A game needs two willing players, and the people who decline the invitation tend to end up watching from the sidelines.

Each persona aligns with a different way of working in the AI era. The warrior ships, where speed, delivery, and a green leaderboard are the measures that matter, the cadence of the AI Garage and the AI Factory. The wizard strategizes, taking on legacy modernization, architecture reviews, and the refactoring that changes a team's trajectory, the work that points toward Government 3.0 and the higher approaches to modernization. The craftsperson builds infrastructure for others: harness files, shared templates, and reusable evaluation scripts, and the tools like Pronghorn, Nexus, and Velocity that enable everyone else. The personas are also where the cohort's code lives. For Level 3 we took a different approach to the learning itself, building a dedicated site to tie the lessons and concepts together, and producing a large body of code, much of which is shared within this collection of white papers.


## §08 The anti-slop mindset

As these papers have shown throughout, some bad habits are emerging with the use of artificial intelligence. One is the production of long-form, unvalidated content that takes minutes to generate and days for a subject-matter expert to read and vet. Another is jumping past the architecture, the standards, the requirements, and the tools straight into a vibe-coded application that offers immediate gratification and little long-term reliability, while loading a different kind of technical debt onto the organization. The Academy works to steer people away from both.

Both habits come from a reasonable human pull toward outcomes, which is healthy in itself, and yet pursuing outcomes without rigor leads to failure in any domain. An athlete who skips the stretching, the sleep, the food, and the rest breaks down, gets injured, and ends up further back than where they began. A public servant who does not take the time to learn how to use and trust AI will produce unreliable content, damage the organization's reputation, and erode the public's trust in the institution, and will likewise end up further back than where they started.

Earlier papers noted that AI agents can code roughly a hundred times faster, while we aim for a twentyfold improvement overall. The gap is deliberate. The goal reaches past producing the app, and past producing the artifacts, to engagement, conversation, comfort, change management, and awareness, all the things that come with a team of people learning to work together. So we treat AI adoption as a mindset to instill rather than a target to hit. When a measure becomes a target it stops being a good measure, and some organizations will take the time to instill the mindset properly. Some people will take to it and some will not, and the commitment is to keep growing along the path without expecting perfection.


## §09 Investing in people

A great deal is made of how much power and how many resources these tools consume. I think we are judging a technology at a single point in time. The efficiency of these systems is expected to improve roughly a hundredfold over the coming years, which means close to a hundred times less power for an equivalent amount of intelligence. The tools we use today are prototypes for a future where intelligence is available on demand. My responsibility to staff in the Government of Alberta is to make sure they are ready for that future, with relevant and meaningful skills, so they meet it as the architects and builders of a world where scalable intelligence sits at everyone's fingertips.

To be clear, the genie is not going back in the bottle. Whether or not we build data centres, intelligence on demand has arrived, and it falls to each of us to learn to be effective and to offer meaningful value in that space. Our approach stays human-centered. We invest in people, and we make sure everyone has the opportunity to take part.

That investment comes with a compact. The organization does everything it can for its people, with weeks and months of training, tens of millions of dollars in tools, and the content to learn from. In exchange, staff meet us with an open mind and a commitment to learn these tools and to use them well. Being effective at work will come to require a working knowledge of artificial intelligence, in the same way it already requires the use of technology today. We ask for open-mindedness, and in exchange we offer opportunity. The result is an empowered workforce, able to reimagine how government works and how services can be delivered to Albertans in the most meaningful way possible.

"We ask for open-mindedness. In exchange, we give opportunity." · Janak Alford, Deputy Minister, Ministry of Technology and Innovation

Alberta's AI Academy is available at albertaaiacademy.com, and we share all of our materials openly. Level 3 is being refactored now and will be released shortly. The lesson, in the end, is a simple one: working through the frustration to become proficient at something you could not do before is the skill that takes people through life. That confidence is what graduates leave with, and it is what they carry back into the day job.

Tags: ai-academy, training, change-management, personas, harness, prompting, agents, open-source

Open the interactive version