You rolled out the AI tool. The town hall went well, the "use it, it'll save you time" email went out twice, and a handful of power users folded it into their week. Then you open the usage dashboard on a Wednesday afternoon and it's been flat for two months.
It's tempting to read that as an information problem — the value props didn't land, so send another explainer. But nobody picks up a new habit because they read about it. People use a tool when they have a reason to, know how to, and get a cue to start — and a launch email gives them, at most, the first. Getting AI used is a behaviour-change job, not a comms job: you design for the habit rather than broadcast the access. That's the principle underneath everything in this collection; this article is about applying it to an AI rollout in particular.
AI adoption is a habit, not an announcement
A habit needs three things at the same time: a reason to bother, the ability to do the thing, and a prompt to start. A launch covers none of them for long.
The reason is shakier than it looks. Using AI can feel slower at first — it genuinely is, while you're learning the tool and double-checking its output — and for some people there's a quieter hesitation: why would I automate part of my own job away? The ability is missing because generic tips don't reach the actual work: what am I allowed to use it for, in my role, this week? And nothing in the day says open the agent now — the cue was never there.
Broadcasting louder adds none of the three. Designing for them does.
Start from the blockers — the AI ones are specific
Before you design anything, get concrete about what you want to happen and what's stopping it. That's its own discipline, and designing from the goal and the blockers walks through it.
What's worth knowing here is that AI rollouts have their own recognisable blockers, and they're the thing to surface out loud: I don't know what I'm allowed to use it for. I don't know which of my tasks it's actually good for. It feels slower. What do I do when it gets something wrong? The fastest way to find them — and to find the people already using AI well — is to ask the whole target group in a short evaluation rather than guess from a few volunteers. The examples that come back from your own high performers — how Anna preps a difficult customer call, how Erik summarises a claim — beat any generic productivity list, because colleagues recognise the work as theirs.
Run a focus period, not a webinar
The default is a ninety-minute webinar: slides, a demo, a recording nobody reopens. Afterwards the hard part still hasn't happened — sitting down to work out what to actually do tomorrow. Usually that sitting-down never comes.
Turn it into a short focus period instead — two weeks or so, built as a real experiment rather than a lesson. On Monday, each person picks one behaviour and writes a one-line plan: what they'll try, when, with what input. Midweek, a short prompt lands with a practical tip. At the end, a reflection — you said you'd try this; how did it go, what was hard, what will you try next? Then a second focus period, with something slightly more ambitious. It isn't training before the work. It's the work, with reflection and a second attempt built in.
Protect people through the dip
New AI use gets slower before it gets faster: you're learning, running the old way and the new way to be safe, checking the output. Most people quit in that dip — which is exactly where the manager matters.
The general manager track — loop in, notice, ask a good question — is its own topic, covered in involving managers. What's specific to AI is the job you hand them: protect the time through the slow week instead of reading it as wasted, and name the awkward part out loud. When a manager says the dip is normal, and no, this isn't about automating you out of a job, the hesitation people were carrying silently becomes something they can talk about. Left unnamed, it quietly caps adoption no matter how good the tool is.
Build the sharing in — good usage is discovered, not distributed
Nobody knows on day one what good AI use looks like in your company. It gets discovered in pieces, by the people who try things and notice what worked. Leave that to chance and the knowledge stays in individual heads.
So make sharing the closing move of each attempt. After someone reflects, point them at a shared Threads conversation to post what they tried — I used it for this; here's what worked. Others read it, copy it, and a real library of local use cases builds itself over the rollout. For the stories that don't fit a post, a peer group at the end of a focus period lets a handful of people compare real attempts and pick up the subtler moves. Either way, the sharing is designed, not hoped for.
You bought access. Whether anyone actually works differently is a separate question — and it's the one worth designing for: a habit, protected through the dip, with the good usage spreading on its own.