Case Study 02: AI Adoption
From Pilot to Weekly Use
A 35-person team went from 11% to 60% weekly AI use in six months
- 01MeasureWeekly use counted per person before anything else changed. The starting line was 11%.
- 02Train by roleEach role got the three or four tasks AI could take off its week, taught on its own work.
- 03ChampionsPeople already using the tools showed their own team, in their own meetings.
- 04ReinforceThe weekly number went to the sponsor every month, with who had stopped and why.
11% weekly use at the start. 60% six months later.
01: ANALYSIS
The tools were licensed. Use was a rumor.
- The organization
- A 35-person team with AI tools already paid for and a pilot that had run its course. Leadership wanted to know whether the spend was changing anyone’s week.
- What existed
- Licenses, a launch announcement, and a handful of enthusiasts. No count of who used the tools, how often, or for what.
- The learners
- Experienced people with full calendars. They had tried the tools once, got a mediocre answer to a vague question, and gone back to the way they worked.
- The constraint
- No time for a course. Anything that asked for an afternoon away from the work would be skipped, and rightly.
The gap, in one picture
Thirty-one of thirty-five did not use the tools in a typical week.
Diagram of 35 people: 4 used AI tools in a typical week before; the target was a majority, 18 or more, every week.
Before
4 of 35 used AI tools in a typical week
Target
A majority, 18 or more of 35, every week, within six months
02: THE PROBLEM
A pilot without a number cannot succeed or fail.
- The gap
- One person in ten used the tools in a given week. The other nine had no reason to, because nobody had shown them a task of theirs it would shorten.
- Why training alone would not fix it
- A general AI course teaches the tool. The people needed to see their own Tuesday with two hours back in it, and then a reason to keep going after the novelty wore off.
- What success had to look like
- A weekly-use figure the sponsor could read without explanation, moving in the right direction, for long enough to be believed.
Stated as a performance gap
Thirty-one of thirty-five people did not use the AI tools in a typical week. The target was a majority of the team using them every week, measured, within six months.
03: THE DESIGN
Measure first. Then train the work, not the tool.
The model
ADKAR, run in order. Awareness and desire came from the measurement itself: a team that sees its own 11% asks why. Knowledge and ability came from role-based training on real tasks. Reinforcement came from the monthly number and from champions inside the team.
01
Count before you teach
Weekly use was measured per person from week one, before any training. The baseline was the argument for everything that followed.
02
Role-based paths, not a course
Each role got a short path built on its own recurring tasks: the report, the summary, the first draft. Nothing generic, nothing longer than the task it replaced.
03
Champions from inside
The people already using the tools well showed their colleagues in the team’s own meetings. Peer proof travels further than a trainer’s.
04
Reinforcement with a number attached
The weekly-use figure went to the sponsor every month, with the names of people who had stopped and the reason each gave. Lapses were followed up, not counted.
Objectives, stated at a level that can be observed
- Apply
- Each person completes at least one of their own recurring tasks with an AI tool in a given week, and the tool records it.
- Evaluate
- Each person judges whether an AI-produced draft is fit to send, and says what they changed.
- Create
- Champions write a reusable prompt or workflow for a task their colleagues share, and hand it over in a team meeting.
The method, reused
The role-based paths follow a course design I had already published: three starting points, a module per recurring task, and a stated way to measure success. The outline is open on this site.
04: OUTCOME
Four people became twenty-one.
What changed
Weekly use rose from 11% of the team to 60% over six months. Regular users reported an average of two hours a week saved on their own tasks. The sponsor could read the figure without a briefing, which was the point.
Attribution, stated plainly
The measurement, the role-based paths and the champions were the work. The tools and licenses were already there. What the work caused was use; what it did not cause was the tools themselves getting better over the same six months, which they did, and which helped.
05: EVALUATION
Two levels measured, two built in.
| Level | Evidence | Status |
|---|---|---|
| Level 1: Reaction | A one-question check after each role session: would you use this on the task we just did. Recorded in the session, not surveyed later. | Built in |
| Level 2: Learning | Each session ended with the person completing their own task with the tool, observed. Attendance was not the measure; the finished task was. | Built in |
| Level 3: Behavior | Weekly use per person, counted from the tools’ own records: 11% at the start, 60% at six months. | Measured |
| Level 4: Results | Hours saved per week, by each regular user’s own count: an average of two. A time study against the records, rather than self-report, is the next measure. | Measured |
Counted here too
This site counts its own adoption the same way: which engagement each article sends a reader to, by article type, with no cookie and no identity. Measurement is a habit, not a deliverable.
06: WHAT I WOULD DO DIFFERENTLY
Start the time study on day one.
The measure I wish I had
Hours saved came from people’s own estimates. A short time study on three recurring tasks, run at the start and again at six months, would have given the sponsor a figure as hard as the weekly-use one.
The 40% who did not come along
Fourteen people were still not using the tools weekly at six months. Some had no recurring task the tools helped with, which is a fair answer. The next round starts with them and asks that question first.
How I think about it
Three essays on why pilots stall and what holds a rollout, written from this kind of work.
The pilot was never the problem. Nobody had counted.