Day one always looks like success
Everyone in the room built something. The demo worked. The screenshots are good. None of that tells you whether the tool survives contact with an ordinary Tuesday three months later, which is the only measure that actually matters to whoever is paying for the rollout.
The launch is the easy part. Ninety days later is the real result.
30 days: did anyone open it again
- Each trained employee used the tool at least once on a real task, not the training exercise, without being reminded.
- The person accountable for each use case can describe, without notes, what it does and why.
- Nothing that broke in week one is still unfixed.
A tool that nobody opened again by day 30 isn't adopted, whatever the training satisfaction score said. This is the checkpoint that catches a rollout that was actually just a demo.
60 days: is it still the human deciding
By now the excitement has worn off, which is exactly when governance either holds or quietly stops mattering. Check that the human in the loop is still checking output, not rubber-stamping it out of habit. An agent that produces a first draft is doing its job. A team that ships every first draft unread has stopped doing theirs.
- Spot-check a sample of the agent's outputs against what a human actually reviewed before it shipped.
- Confirm the escalation path for a wrong or uncertain output is still the one people actually use, not the one written in a slide.
90 days: has the workflow actually changed
The real test isn't whether the tool gets used, it's whether the work itself looks different. A task that used to take four steps and two people now takes two steps and one person's review. That's adoption. A task that still takes the same four steps, with an AI tool bolted onto step three, is not; it's just a more expensive version of the old workflow.
- Name the specific step that used to exist and doesn't anymore.
- Ask the team what they'd lose if the tool were taken away tomorrow. A real answer, not a shrug, is the signal.
- Check whether a second team asked to be trained on the same workflow, unprompted. That's the adoption spreading on its own.
What breaks without checking
A rollout that nobody re-measures after the kickoff quietly reverts. People go back to the old way under deadline pressure, the license renews anyway, and eighteen months later somebody asks why the AI initiative "didn't work," when the honest answer is that nobody checked on it after week one.