What I’ve learned teaching public servants about AI

Shingai Manjengwa is a PPF Fellow and leads AI education at Mila – Quebec Artificial Intelligence Institute. She served on the federal Working Group on Public Service Productivity. She was previously Director of Professional Development at the Vector Institute and founded the data science education company Fireside Analytics.

In 2023, Anna Jahn, then at Mila, and Sara-Christine Gemson, then at the Public Policy Forum, came to me with an idea: teach public servants and policy leaders how AI actually works. That idea became AI Policy Compass, which Mila and PPF have co-delivered since 2023, and later AI Advantage, a virtual program to teach public servants how to use AI.

I have been calling for AI literacy for years. In June, writing about Canada’s new national AI strategy, I proposed structured productivity training for public sector professionals, because adopting AI raises governance questions: who is accountable, what data is safe to use, and how decisions are made. The strategy, AI for All, includes a National AI Literacy Initiative, launched in September, offering entry-level AI training to all Canadians. That is a win.

But literacy is where the work starts, not where it ends. Many public servants now have approved AI tools, but having them is not the same as using them well. In September 2025, in two of our AI Advantage cohorts built for one federal department, 30 of 33 and 31 of 38 respondents named Microsoft Copilot as a tool they used at work. Yet in a recent cohort, some participants still wanted help using AI to write an email. That is a good place to start AI use. When Excel was new, some only used it to add numbers in columns and some of us graduated to pivot tables and macros quickly. That is where we are with AI today.

Beyond literacy, the next training frontier is adoption: using AI to do the work of government better, improving processes and being able to show it. Literacy and adoption must become part of how the public service operates, not boxes to tick.

Fast forward three years and 18 cohorts. How do we know any of our AI training is working?

My teaching approach to technical education

I began teaching in earnest in 2016, at my own data science education start-up, Fireside Analytics Inc., where I developed courses for IBM Canada and other clients. My very first RFP submission was for the National Energy Board of Canada, now the Canada Energy Regulator, where the goal was to “create a foundation of expertise in data science.” I ran a data science workshop for teams from across the organization, to help them build capacity in data analytics, a culture of experimentation and evidence-based decision making. I still have those slides! I carried that work to the Vector Institute around 2019, where my approach took shape as what I now call the 3 Cs Framework for AI Education: teaching AI through case studies, code and capstones. Each element answers a question learners bring into the room.

Case studies: why does this matter to me? Cases set the stakes with real scenarios from the learner’s own sector, written with practitioners in that field. Anna and Sara-Christine shaped the first policy cases for AI Policy Compass. In October 2025, participants took on AI translation in the federal government. Under the Official Languages Act, service in English and French is a right, not a convenience, and translation requires a large, skilled workforce with a budget to match. The case runs from a policy brief, to how a language model turns a memo into French, to the questions a department would face about law, cost, culture, jobs and trust.

In 2026 we also added a debate session, to teach participants how to disagree about AI. We present a case, ask participants to pick a position and vote, debate it, then vote again to see who has shifted. In one session the case was AI in child welfare, and most participants’ gut response was an immediate no. After working through the details and the data, some changed their position. The case itself is not the point. The muscle we are exercising is assessing the facts and the data, and articulating a position on AI respectfully, while still fulfilling the mandate to serve.

Code: how does it work? Learners who may never have programmed a computer, and possibly never will, must still see AI demystified as code and mathematics. The point is not to make them programmers. AI tools invite people to treat them as minds, and today’s tools are particularly prone to sycophancy: telling users what they want to hear. Knowing what AI really is matters more as headlines describe AI as if it were a person, a habit known as anthropomorphization. Recently, the discussion has turned to AI sentience, AI rights, and AI agents showing human-like traits such as collusion, deception and coordination. Seeing what a model actually is behind the clean user interface changes how learners judge its outputs, spot biases and notice how the tools affect the people using them. Even our instinct to say please and thank you to a chatbot is a telling signal, one that may shape how we make AI policy, or do policy work with AI.

In one exercise, learners open a coding notebook: a web page that holds Python code anyone can run, with nothing to install and no programming experience needed. The code looks at a video of a parking lot, checks each space and counts the ones that are free. Participants can change the one number that decides whether a space is free and watch the count change. Systems like this are common at airports and shopping malls. Those who are more technically advanced go further and see the same system fail at dusk, at night and in snow. Some are happy to watch the exercise during the lesson, but everyone is encouraged to interact with the code. Everyone ends on policy questions: who authorized the camera, does the resolution pick up darker skin tones, who decides how sensitive it should be, and what happens if it misses a child in a parking space?

Capstones: can I use it? Every course is designed to end with a capstone built on the participant’s own work: a process map, a proposed use case, or an implementation plan for a real task. In AI Advantage, a first assignment builds toward the capstone and the second is the full capstone, both with personalized feedback. In AI Policy Compass, the capstone is now the Start AI exercise, described below. I insist on homework; my team insists on making it optional. We find a healthy balance, and most people submit. One participant wrote: “I really liked the practical part, which ‘forced’ me to take the time to work on a real case” (translated from French).

Most of our cohorts are open enrolment, so anyone can sign up. A manager who has never written a line of code can sit beside an analyst who programs every day. I treat that as a feature: exercises have a low floor and a high ceiling, so no one is lost and no one is idle. It has a cost, though. One participant wrote: “Sometimes I didn’t want to ask questions because I knew they were way too basic.” That is why office hours sit alongside the three Cs. Technical topics are challenging, and people need time with an instructor to ask the question they would not ask in front of the group. Participants appreciate the office hours.

The three elements are the core, case studies, code, and capstones. Almost everything else can change to fit the audience, and in our programs it does:

• Format: two consecutive days in person for AI Policy Compass; short live virtual modules with assignments in between for AI Advantage, which suits public servants across the country.
• Language: cohorts in English and in French.
• Level: a pre-course survey tells us where a cohort is starting, and the introduction shifts between AI 101 and AI 201 to match.
• Cases: policy dilemmas for policy leaders; workplace processes and change management for public servants. Examples closer to home have the greatest impact on learning.
• Capstones: done alone or in a team, marked with personalized feedback or discussed in plenary, and learners are invited to check their own work with an AI tool.
• Who teaches: a core team, joined by guest experts and co-developers who bring specialist depth. Technical expertise and depth in AI are what build trust in the learning materials, even when a course is aimed at entry-level AI users.

What I’ve learned in three years of teaching AI

I have taught public servants in other courses and programs too, but these lessons come from AI Policy Compass and AI Advantage: 18 cohorts since 2023, 16 of them with survey data.

Satisfaction scores don’t tell you what to change. Across five cohorts of AI Policy Compass, 95 percent of respondents rated the program 4 or 5 out of 5. That was good news, but it did not tell us what to fix. In a mixed-level cohort, a 2 out of 5 cannot tell you whether the course was too technical or not technical enough. Our survey looked like most course evaluations: about 17 questions, mostly 1-to-5 ratings of overall satisfaction, each session, and self-assessment statements, plus whether you would recommend the program. We were distilling machine learning into a few hours, and the specific challenges showed up in the written comments, not in the ratings.

Traditional course design also assumes that a curriculum is set well before a course runs and revised between offerings. That rhythm has served education well. With the speed of AI though, it can mean holding to a plan while the data from the current cohort is already telling us to change.

Two comments told us what the numbers could not. After the first session of the first cohort, which I taught, one participant called the coding exercise “basically just box clicking.” Another wrote that “pressing play for every code didn’t exactly translate into understanding what was happening.” They were right: running code is not the same as understanding it. Learners have to change something, understand what happened, and see why it belongs in a course about AI.

Three open questions add what ratings can’t. We now ask three open-ended questions at the end of each course, alongside the usual ratings, not instead of them:

1. What concepts stood out to you?
2. What did you like most about the course?
3. What would you change?

Ratings track a program over time; the open answers tell you what to do next. Reading hundreds of them used to be too slow, which is why most evaluations stopped at the numbers. Now AI does a first pass and a person reviews it: we teach people about AI, and we use it to learn from our own teaching. The answers are specific. In AI Advantage, process mapping stood out most, in about two of every five answers about concepts. Two of the most common requests were for more hands-on practice and for examples built on the tools people are allowed to use: “most government employees can only use Copilot.” And participants told us the work was travelling: “I took the use cases back to my team to discuss and brainstorm.”

Ask those three questions and, whether participants love the course or hate it, you will know immediately. You will know specifically what is working and what is not. We take that feedback seriously. I have been doing this work long enough to know that the people you train have a multiplier effect: they move on to other departments and teams, and they bring others with them. Word of mouth is our single greatest source of participants. In eight of the eleven cohorts where we asked, “from a colleague” was the most common way people heard about the program.

The audience moves faster than the curriculum. ChatGPT was released to the public in November 2022. Before that, I had been teaching with older, more primitive language models such as BERT. In our early cohorts, many participants had never touched a large language model. In 2023, one wrote: “We got AI-101 and was expecting AI-201.” By 2025, most participants in our departmental cohorts were already using Copilot at work. In our open cohorts, Copilot use at work rose from 55 percent of respondents in May 2025 to 97 percent in October 2026. In 2026, one suggested our coding challenge “could be shifted to vibe coding.”

So we must change quickly too. After the first cohort, our log recorded that the introduction was not pitched at the audience’s level. We added more nuanced pre-course survey questions to gauge how familiar each cohort was with AI, and by March 2024 the introduction had indeed become “AI 201.” Participants asked for French and for two full days in person, and both followed in 2024. Ratings of the coding exercise rose from about 3.2 and 3.6 to between 4.0 and 4.3.

Still, every cohort is different, and there are always AI 101 learners just starting out. In March 2026, we added the Start AI exercise partway through an AI Policy Compass course because the group needed it. A significant share of the course submissions described their policy challenge as needing to “start AI.” A number of learners had been tapped on the shoulder to become AI champions in their departments, and they were looking for help, for community and for concrete first steps.

The Start AI exercise gives them those steps. Participants list up to five priority tasks, pick one and define the problem it solves, restate it as something to maximize, minimize or optimize, map the process, mark where AI could help, choose one step, and set a baseline to measure the impact. It can be done alone or in a team, with pen and paper. We knew it had been the right pivot when one participant said they would “definitely reflect on [it] when I’m back at my desk on Monday,” and others said they would run the exercise with their teams back in the office.

Keep a technical practitioner on the course team. That pivot, and the period around it, highlighted two needs. The first is a technical practitioner on staff, someone who makes sure changes in industry reach the classroom while they are still current. That agility does not happen by accident. Our course team always includes a data scientist, instructional designers, and experts who know how adults, and perhaps specifically policymakers, learn.

The second is technical rigour. In June 2026, U.S. export controls switched off Anthropic’s Claude Fable 5, one of the most capable AI models available, for every user in the world, Canadians included, for nearly three weeks. In July, Hugging Face, a platform at the centre of open-source AI, disclosed a breach that OpenAI later attributed to its own AI agents. Sovereignty and security are no longer abstract. Public servants need to learn the real challenges of AI and still have the confidence and trust to adopt it, within the governance their team or department sets. That makes technical rigour in what we teach a must-have, even when it comes through guest speakers and co-developers. The only constant is change, in AI and in the learners we teach. A course that stands still is out of date by the next cohort.

Be realistic about what a course can show. A capstone completed during a course shows that someone can apply what they learned to their own work. It does not show that they did, or what it was worth. That takes a follow-up after the course, and a baseline set before it. Promising more than that sets training up to disappoint at the leadership level.

We can do more than show that people took a course. We must. But we cannot leap from “this many people took the course” to “productivity rose by that much” unless the course was designed to build a discipline that goes beyond AI tools to processes and workflows. For AI learning and use cases to take hold, organizations must consider their operational standards and become more data-driven in how they operate.

If public servants have AI tools, Canadians should be able to see what those tools cost and what they return. I would like to know how much we are paying for AI licences across government, how much of that goes to a handful of frontier AI companies, and how learning outcomes are being measured. That is a lot to chew on. The more immediate focus can be what we can control: grounding training in the data and processes public servants work with every day, and building ways to make return on investment and productivity measurable.

I served on the federal Working Group on Public Service Productivity. Our report said it plainly: “Without reliable data, it is difficult to assess the effectiveness and efficiency of government services.” The national AI strategy aims for a 3 percent increase in GDP from labour productivity. If productivity is the goal, it has to be measurable, starting with the training we give the people expected to deliver it.

That is why the Start AI exercise ends with measurement. Before anyone touches a tool, participants write down the task, the process, the one step where AI could help, and the baseline they will measure against: time saved, errors avoided, turnaround, quality. As the exercise puts it: there is no productivity gain without measurement.

Treat assignments as inputs to operations. A capstone is more than homework. It is an opportunity to get feedback on an AI issue participants are actually grappling with. It is a process map, a proposed use case or an implementation plan built on a real task, and it should go back to the participant’s team. If people have taken the time to train, what they produce in course assignments should serve the organization’s larger AI transformation.

When several people from one organization take a course together, their capstones add up to a picture of that organization: how its staff think about AI, what they can already do, where data and rules get in the way, and which opportunities are worth pursuing. I have delivered that kind of cohort report to employers in confidential engagements. I dream of doing the same in government, where the people who can remove blockers and approve change could see for themselves what their teams have in mind. These assignments are a rich dataset that employers funding training should consider exploring, with the right disclosures and terms agreed between learners and employers.

The three greatest barriers to AI adoption in the public service

I must now also be a good steward to the learners who have passed through our programs, and share perspectives that have come up time and time again. From our data and my conversations with learners, these are the three greatest barriers to AI adoption I have seen:

1. Leaders who cannot receive the change. Many learners feel that even their best work, a pilot pitch, a use case, a process improvement, will go unheard if their leaders have not done similar training, cannot receive the ideas, or are paralyzed by fear and have zero appetite for risk. Before one course even began, a participant hoped to find “a way to defuse my superiors’ fears about the constraints of using AI in my workplace” (translated from French). And yes, there can be managed risks in AI adoption, as in aviation and health care. In one 2024 cohort, participants asked for sessions tailored to the senior managers who shape the policy space.
2. In the news. Public servants are held to a higher standard. The fear of ending up on the front page of the Globe and Mail works like a subroutine that never stops running in the background (forgive the computer programming analogy). You can’t see it and it never switches off, but it shapes behaviour, including whether or not people adopt AI.
3. Personal worries. Then there is a bigger bucket of individual concerns. Will AI take my job? Will this harm my kids? We don’t have the right tools in government. What is the impact on the environment? Will this make me less intelligent? Is my data safe? Will it kill us all? Some of these show up clearly in our surveys: participants asked for examples built on the only tools they are allowed to use, and one wrote that the course “did not address at all the impacts of AI on the environment. As public servants, this is very much a concern that is central to our use of AI.” Others worried about their own thinking: one hoped to use AI “without succumbing to cognitive offloading,” and another described AI as “a crutch that risks reducing human intelligence” (translated from French). We don’t have answers to all of these concerns, but AI Policy Compass now has a longer chapter on AI and sustainability, and our cyber security case study is about what to consider if your AI chatbot gets hacked. One participant who described themselves as non-technical called it “very informative and eye opening.” We wrestle AI issues to the ground and arrive at informed positions, with a focus on AI safety and public trust.

I feel a responsibility to share these sentiments plainly, as I have heard them, in case those leaders read this.

Making AI training count

The next AI course you procure or teach must come with a data requirement: a baseline before it starts, open questions, and a suitable follow-up to assess impact and return on investment.

If you fund or commission AI training

1. Treat literacy and adoption as an operating model, not a checkbox. Build learning into how teams plan, buy and use AI, not into a one-time course. AI education is a process, not an event. 
2. Ask for evidence of use, not just completion or enjoyment. Not “did they complete it or like it?” but “what did they take back to work?” Expect every course to end in work your teams can use, and ask to see it.
3. Write the follow-up into the contract. Three months after a course, ask whether the work is in use and what changed against its baseline. I would like to build this into our public service programming, but it may need to be written in at the procurement stage. Course requesters don’t readily see the world this way, and demand is pushing us to scale rather than run fewer, possibly more expensive programs, even when we believe in the complete model.
4. Report spending and results together. What the licences and the training cost, and what they returned. Agreed definitions of success, thoughtful spending on access and measurable outcomes breathe accountability and trust into AI programming. In the interest of disclosure: I benefit when leaders see it working.

If you design or deliver AI training

1. Train on the work people do and the tools they have. Use case studies from their context and exercises with the tools they are permitted to use. This one is tricky: we don’t know what we don’t know, and exposure to other tools can help. But government has real constraints on which tools can be used and how. Training must reflect that.
2. Teach what AI is, not just which buttons to press. Show learners how it works, where it fails, and why it agrees with us so readily. The buttons keep changing, and keeping up with the interface alone is real work. We have seen people at the same level, in similar departments, with different versions of Copilot: some have the agents button and some do not, and the browser version differs from the desktop one. Teach the fundamentals and enough of the tools to get going, so people can adapt to each software update.
3. Early in the adoption cycle, end every course with a capstone and a baseline measure. One task, one AI-assisted step, one measure, set before the tool is used widely. We know this is simplistic; the goal is to start, and build from experience. Measurement is a field of study in its own right. Token usage, for example, is a fraught measure of adoption: one high-quality image can use as many tokens as several pages of text. More tokens do not mean more value.
4. Add open questions to your evaluation. Keep the ratings, and ask what concepts stood out, what participants liked, and what they would change. Use AI to read the answers, and have a person review them. Please.
5. Build course teams that can adapt. Pair technical instructors with people trained in instructional design and education, so courses can change as fast as the tools. AI education is interdisciplinary, and it is a team sport.

Canada has started to invest in AI literacy. The next test is whether we can show what that investment returns. With the right data, we can.

The author wrote this article from her own experience, program records, and survey results. She used Claude (Anthropic) to help analyze the survey data, check facts, draft and edit text.