01 The problem
A national professional services firm with 500+ employees across five Canadian cities had already made the AI investment. Tools were licensed, accounts were provisioned, security had signed off, and the budget line existed. By every measure a procurement process can produce, the firm had adopted AI.
By the only measure that matters, it had not. Active adoption sat at 23 percent.
That number hid more than it revealed. Usage was wildly uneven. A handful of teams had quietly rebuilt their workflows around the new capabilities and would not have given them back at any price. Other teams had not logged in once. Same licenses, same tools, same firm, and two different realities depending on which floor you stood on.
Leadership could see the gap between the teams that had adopted and the teams that had not. What they could not see was a way to close it that did not amount to sending another memo. Memos had been sent. The 23 percent was the result.
This is the most common failure mode in enterprise AI, and it is worth naming plainly: the purchase happens, the behavior change does not. The firm did not need more software. It needed a system for changing how 500+ people work, office by office, role by role, with numbers attached. That is what they hired us to build.
02 The purchase happened. The behavior change did not
It is worth sitting with why this happens, because the firm had done nothing unusual. The standard enterprise rollout goes: buy the licenses, announce the tools, run a kickoff session, wait. Every step of that playbook treats adoption as an event. Adoption is not an event. It is a behavior change, and behavior change has to be engineered, the same way any other system does.
Professional services firms feel this harder than most. The people being asked to change are experts. Their working methods are not habits, they are the product. An auditor's review process or an advisor's drafting routine took years to sharpen, and clients pay for its reliability. Asking an expert to work differently is asking them to be temporarily worse at their job, in front of the people who bill for their time. Unless someone shows them, on their own work, that the new way is better, declining is the rational move. The low adoption number was not laziness. It was professionals protecting their standards from a tool nobody had connected to their actual work.
There is also a quieter cost that finance sees before anyone else. An unused license costs exactly as much as a used one. At 23 percent active adoption, most of that line item was returning nothing.
And the gap compounds in the wrong direction. The teams that had adopted were getting faster every month. The teams that had not were standing still and starting to feel accused. Left alone, that divide hardens into culture. Leadership wanted it closed, and they wanted the closing measured rather than asserted.
03 What had to be true first
Four constraints shaped the program before a single session was scheduled.
Five cities, one program. The firm runs across five offices, and a rollout that radiates from head office arrives everywhere else as a rumor. Whatever we built had to land in every office with the same weight: local training, local champions, one shared standard. No office could be the pilot that the others watched with folded arms.
Role diversity is real. An auditor and a marketer do not need the same skills, and pretending otherwise is how training budgets evaporate. A single generic AI course flatters nobody and wastes everybody's afternoon. The program had to split by role without fragmenting into a different curriculum for every desk.
Compliance anxiety was an adoption blocker, not a footnote. In a firm built on confidential client information, the question that stops usage is not "how do I use this" but "am I allowed to use this here". When the answer is ambiguous, abstaining is professional prudence, and prudent people abstain in large numbers. The program had to make permission explicit: rules built to respect the firm's confidentiality obligations and the privacy law it operates under, written down, and published before anyone was asked to change how they work.
Leadership demanded measurement. Not a sentiment survey at the end of the quarter. Adoption tracked at the individual and team level, visible weekly, so the rollout could be managed like any other operation: with numbers, while there was still time to act on them. If the program worked, the dashboard would prove it. If it stalled, the dashboard would say where.
04 What we built
We ran a firm-wide AI enablement program. Not a lunch-and-learn, a system with six components, each one load-bearing.
- Role-specific training. Different job functions got different programs. An auditor and a marketer do not need the same skills, so they did not get the same course. Each track taught the tools against that role's real tasks, not against demo data.
- Custom prompt libraries. Prompts built from the firm's actual workflows, so the first time someone used the tools, they worked on real work. Nobody had to stare at an empty text box and invent a use case on the spot.
- Governance framework. Clear usage policies that told every person what was allowed outright, what needed review, and what was out of bounds. Compliance stopped being a reason to abstain because the answer stopped being a guess.
- Champions program. Designated champions in each office driving peer-to-peer adoption, because people copy colleagues, not memos. The champions were the program's local presence after the trainers went home.
- Ongoing coaching and office hours. A standing place to bring the workflow that would not cooperate. Most adoption dies in the gap between the course and the desk; office hours existed to close that gap the same week it opened.
- Executive dashboards. Adoption and time saved tracked at the individual and team level, so leadership managed the rollout with numbers instead of anecdotes, and every office could see exactly where it stood.
The components reinforce each other. Training without prompt libraries leaves people staring at a blank page. Libraries without governance leave people afraid to paste anything into them. Governance without champions reads as policy instead of permission. And all of it without dashboards is a story you tell yourself instead of a number you can defend.
05 The tooling behind the program
This engagement was a program, not a pipeline. There is no seven-stage build to draw, no agent doing the work overnight. But programs fail without tooling just as surely as pipelines do, and the parts of this one that people praise, the libraries and the dashboards, only worked because of what sat underneath them.
- Prompt libraries as plain text. Every prompt lives as readable, versioned text the firm owns and edits, organized by role and task. No proprietary app, no vendor lock, nothing to install. When a team improves a prompt, the improvement ships to everyone the same day.
- One usage store. Adoption events land in a single database (Supabase, the same store we use in our system builds), so the dashboards read recorded behavior, not self-reported enthusiasm. Surveys measure how people feel about a tool. The store measures whether they used it.
- Dashboards leadership actually opens. Built as a simple web dashboard on the same stack as everything else we ship, filterable from firm level down to one person. No analyst had to assemble a deck for the rollout to be inspected.
- Delivery inside existing tools. The libraries and the guidance meet people inside the applications they already work in. We did not add a destination app, because every new destination is a place adoption goes to die.
You will notice we have not named the AI tools themselves. That is deliberate. The program is tool-agnostic by design: the firm's licenses were already chosen, and the same structure works whichever vendor sits underneath it. The tools were never the problem. The full reasoning behind our standard stack is on the stack page.
06 The program decisions that made it work
Programs live or die on design decisions that look small from the outside. These are the ones that mattered.
Role-specific beat generic, every time. The training matrix below is the heart of the program: every role shares a foundation and every role ends in governance, but the middle of the path belongs to the job. Auditors drilled on analysis and review tasks, marketers on drafting and repurposing. The fastest way to lose an expert is to teach them something adjacent to their work and call it relevant.
Prompt libraries came from the firm's real work. We built the libraries out of the firm's own documents, templates and recurring tasks, not from a downloaded prompt pack. The first time a person opened the library, they found their Tuesday in it. That moment, recognizing your own work inside the tool, is the moment adoption actually starts.
Governance was published before training began. Most rollouts train first and answer the policy questions later, which means every early user is quietly gambling with their professional standing. We inverted it. The decision tree below went out first, so by the time anyone sat in a session, permission was never ambiguous. Usage cannot outrun policy without casualties, and policy that arrives late arrives as a brake.
Champions were chosen for peer credibility, not seniority. The champion in each office was the person colleagues already asked for help, not the most senior partner willing to hold the title. Authority can mandate attendance; only credibility changes behavior. The champions ran the local office hours, fed workflow problems back to us, and carried improvements from their office to the others.
Dashboards went to the individual level. Team averages hide everything interesting. Individual-level data let coaching land where it was needed instead of being broadcast at everyone, and it let the firm celebrate specific wins instead of vague momentum. It also kept the program honest: an enablement effort that cannot show you who it has not reached yet is not measuring, it is marketing.
07 Before and after
The cleanest way to see the change is one person's week.
Before the program, a knowledge worker at the firm ran their week the way they always had: every draft started blank, every summary was typed from scratch, every review pass was manual end to end. The tools existed somewhere in a browser tab, unused, for all the reasons above.
After the program, the same person runs the same week with the tools wired into their actual tasks, prompts ready for the work they recognize, and clear permission to use them. The output is the same or better. The difference is an average of 8 hours per week returned to each person, measured by the program's own dashboards rather than estimated in a survey.
Multiply that across a firm of 500+ people and the program stops being a training expense and becomes one of the larger capacity decisions the firm has made. The hours did not go to more meetings. They went where the firm's hours always should have gone: client work and judgment.
08 How the firm runs it now
The program was designed to make itself unnecessary, and this is the part most enablement efforts skip. Champions carry the expertise forward in every office. The prompt libraries encode what works, so improvements accumulate instead of evaporating when one enthusiast changes teams. The dashboards show every office what the leading teams are doing, which turns best practice into something visible rather than something rumored.
That structure turned adoption from a mandate into a flywheel. Teams that saved hours talked about it, because people talk about getting their evenings back. Dashboards made the gap between offices concrete instead of deniable. Offices that lagged could see exactly where they stood and exactly who to copy, and copying is the one change program that runs itself.
The firm now owns the whole machine: the libraries, the governance, the champion network, the dashboards. Best practices stopped living in one high-performing team and became the standard everywhere, and the firm built self-sustaining internal AI expertise instead of a permanent dependence on consultants. Including us. We consider that the success condition, not a lost renewal.
09 Results
- Active adoption up from 23 percent to 85 percent across the firm
- An average of 8 hours per week saved per knowledge worker
- Standardized best practices across all five offices
- Self-sustaining internal AI expertise, run by the firm's own champions
- ROI measured at the individual and team level, not asserted
The lesson generalizes: AI tools do not fail on capability, they fail on adoption. Training, governance and visible measurement are what turn a software line item into hours back.
10 What to know before building this
If you are about to run a program like this inside your own firm, three things matter more than the rest.
The program must be designed to make itself unnecessary. If the consultant is load-bearing forever, you bought a dependency, not a capability. Demand a handover plan on day one: who inside your firm owns the libraries, who runs the champion network, who reads the dashboards after the engagement ends. A program that cannot answer those questions is a subscription wearing a strategy costume.
Measure usage weekly, at the individual level. Quarterly adoption surveys arrive too late to manage and too aggregated to act on. Weekly usage data lets you coach the person who stalled this week, fix the prompt that nobody touches, and spot the office that is drifting before the drift becomes identity. Measurement is not the report card at the end. It is the steering wheel.
Laggard offices copy leading offices only when they can see the gap. Nobody changes behavior because an average moved. They change when the dashboard shows their office trailing the office down the highway, doing the same work in fewer hours. Make the gap visible and specific, and the social physics do the rest. Hide it inside a firm-wide number, and the leading teams plateau while the lagging teams never start.
If you are wondering where to begin: publish governance and stand up measurement first, before any training is booked. Permission and proof are the two things every later component depends on, and they are the two cheapest to build.
11 Related work
Adoption is not a separate offering we bolt on at the end. It is the third pillar of every system we build, because a pipeline nobody uses is a cost, not an asset.
The same program structure carries into regulated industries, where governance does the heaviest lifting: see how we approach it for law firms and for private equity. For a build where adoption inside a single firm made the difference between shelfware and a working system, read the legal document automation case.
And if you want to know how we think about making ourselves unnecessary on purpose, that is on the about page. It is not a tactic. It is the operating principle.