Employment and Social Development Canada.
Change management design at the largest department in the federal government. When generative AI was greenlit for government use, I built the educational sites, enterprise comms and safe-use guides that had to make it usable by 35,000 staff of every age and technical background.
Generative AI arrived for 35,000+ people at once.
There are no screenshots on this page. The work sits behind departmental security clearance, so what follows is the problem, the constraints and the decisions, without the artifacts.
Adoption is a design problem before it's a training problem.
ESDC is the largest department in the federal government, a little over 35,000 people. When generative AI was greenlit for government work, that entire population needed to understand it at roughly the same time: what was safe to put into it, what it could reasonably be trusted with, and where the rules sat. The technology was never the hard part. Whether 35,000 people of every age and technical background would use it correctly was, and that was the problem I worked on inside Organizational Change Management, alongside the IT teams who owned the tools.
The output was educational rather than promotional. Sites explaining each tool, enterprise-wide comms written for the department's AI co-champions and pushed to every staff member, infographics, manager toolkits and safe-use prompt guides. I also pitched a set of example prompts for the everyday tasks people could sensibly automate, because abstract guidance about an unfamiliar tool lands far less well than showing someone the exact thing to type.
- A system with no room
- The department publishes inside its own design system, and the tooling around it left very little latitude to build anything that was not already a standard component. I could not design my way out of a dense page. What I could do was work inside it, using imagery, sequencing and structure to carry the weight that layout normally would, so a page about cloud migration read like something a person would finish rather than a policy document.
- The audience was everyone
- There was no baseline I could assume. The department spans every age, every level of technical confidence and every kind of role, and all of them received the same material at the same time. Writing for the people already comfortable with AI would have lost the ones who needed the guidance most; writing only for beginners would have been ignored by everyone else. A lot of the work went into language that condescends in neither direction.
- Everything shipped through review
- Nothing reached 35,000 people without revisions and approvals first. That changes how you work. Decisions had to be defensible to people who were not designers, drafts had to survive several rounds of comment without losing their shape, and anything resting on a subtle visual argument was unlikely to make it through intact.
- Accessibility was the floor
- Everything had to meet Government of Canada accessibility standards before it could publish. That is a hard requirement with real teeth rather than an aspiration in a style guide: contrast, heading structure, alternative text and keyboard order all had to hold up under review. Designing to that standard from the start is faster than retrofitting it, and it is the habit I kept afterwards.
- The brief
- Make generative AI safe to use for a department that did not choose it. Once AI was greenlit for government work, ESDC's 35,000 staff needed to understand it quickly and correctly, and the risk was never that the technology would fail. It was that people would either avoid it altogether or use it in ways that are genuinely unsafe in a government setting. My job inside Organizational Change Management was the layer between what IT had approved and whether anyone would adopt it properly.
- What I decided, and why
- Treat comprehension as the thing being measured, not completeness. The instinct in a large organisation is to document everything, which produces material that is accurate and unread. I pushed the other way: lead with what someone needs in the first thirty seconds, use imagery to break dense technical content into something scannable, and let the exhaustive detail sit lower for the people who want it. The prompt guides are the clearest case. Instead of explaining what makes a good prompt in the abstract, I pitched specific prompts for the everyday tasks people could sensibly hand to AI, so the first thing a reader met was something they could copy and use rather than a principle to interpret.
- Where it landed
- The comms went department-wide through ESDC's AI co-champions, and the guides and toolkits went with them. Closer in, I redesigned the OCM team's own page, which carried the clearest measure on the project: weekly visits, treated as a proxy for whether people spread across different files actually knew what the rest of the team was doing. I also drew the team a logo. That sounds small next to enterprise communications, but it did something the comms could not, which was give a group split across projects one thing to belong to.