“This is what you do with AI”
Most rollouts start from the same instinct: control the outcome. Pick the tool, write the approved use cases, run the training session, and tell people — this is what you do with AI. It feels responsible. It demos well to leadership. And it produces the same curve every time: a spike of dutiful usage, a couple of weeks of compliance, then quiet.
The people weren’t the problem, and neither was the tool. The prescription was. Nobody’s week matches the approved use cases, because the approved use cases were written by someone who doesn’t live that week. So people did what they were told, found the fit was off, and went back to doing the job the way it actually gets done.
Half the distance, on purpose
At a large global company, in a 300-person org, I ran the opposite play. Everyone got a basic setup: a small CLI tool on their own machine, a personal knowledge base in the Karpathy wiki style — a plain place to put the notes and scraps they were already accumulating. We showed them a few ways to use it. We taught the concepts underneath it, so they understood what the thing was actually doing and could reason about it themselves. Then we stopped.
Deliberately. No mandated use cases. No compliance checklist. Go off and experiment.
That stopping point was the design, not a budget cut. Telling people what to do with AI produces users of one use case. Handing them a working setup and the understanding to bend it produces people who find the use case you would never have written, because it only exists inside their particular week.
What 0.5 means
I think of it as zero to 0.5. You take people half the distance: the setup works, the concepts are in their heads, they’ve watched it do something useful once. The second half, the part where the tool becomes theirs, only happens when they walk it themselves.
And they walk it on their own small information. Not the company’s grand data estate — the inbox, the meeting notes, the decision from March they keep re-finding because it never had a home. That’s where the value showed up: in the small information people already had, suddenly organised and answerable. You can’t prescribe that, because you don’t know what someone’s small information is. They do.
Measured, not hoped
The part that kept this honest: we measured it. Real usage stats, plus feedback from the people actually using the thing — every two weeks. Not a sentiment survey, not hands raised in a town hall. Did usage grow, and what were people doing with it.
Usage tripled every measurement period until it was org-wide. I want to be careful with that sentence, because it reads like a launch-deck line: it was a measured curve, checked on a fixed cadence, and if it had flattened we would have changed course rather than declared victory. That’s most of what I mean by applied AI: adoption is an outcome you check, not a feeling you canvass.
The honest tension
Prescription spreads faster at first. You can mandate a tool into every calendar by Thursday. It also dies, for the reasons above. Experimentation starts slower; the first fortnight looks unimpressive next to a mandate. Then it compounds, because every person who finds their own use becomes a demonstration for the next.
The lever that makes it compound: train the first implementers. The people who got furthest on their own became the ones who taught the next ring, and the thing kept growing without me standing next to it. A rollout you have to keep pushing isn’t adoption — it’s attendance.
Where this lives now
This is the same belief that runs the check stage of how I work: you don’t declare a build adopted, you go back and look at what people actually do with it. And the second brain I run my own work from today is the same shape — a plain knowledge base I walked the second half on myself.