Adoption
You rolled out AI. How do you know it's working?
Six months after the rollout, the board asks whether the AI investment is paying off, and the honest answer in most companies is a shrug dressed as a slide. Licenses purchased and accounts activated are procurement metrics, not adoption metrics. The signals that actually predict value are behavioural, and most teams never look at them.
Vanity metrics and why they mislead
Seat count measures budget. Sign-ups measure the launch email. Even “monthly active users” flatters, because one curious login a month counts the same as daily reliance. The screenshot of an impressive answer, the most common artifact in rollout reviews, proves the tool can be useful once, not that it is useful routinely. None of these numbers move when value moves.
Signal one: repeat usage on real work
The first honest signal is the share of people who return and use AI on consecutive working days, and what they use it for. Habitual usage on real tasks (drafting, analysis, research) is the difference between a tool and a toy. Watch the weekly-returning share of licensed users, and treat a plateau under half as a diagnosis, not a verdict: it usually points at missing enablement or a workflow the tool does not yet cover.
Signal two: task coverage, not message volume
Message counts reward chattiness. The better question is: of the task types this team does weekly, how many routinely involve the workspace? A legal team using AI only to summarise is at one of five; when clause comparison, research with citations, and first-draft review join, that is coverage growing, and coverage, not volume, is what shows up in cycle times. Feature breadth matters here: a team cannot cover research tasks with a tool that has no research mode, or visual tasks with no image generation.
Signal three: where the output lands
The strongest signal is whether AI output leaves the tool and enters the company’s bloodstream: the brief that went to the client, the analysis pasted into the decision doc, the code that shipped. If output consistently dies inside the chat window, the tool is entertainment; if it flows into deliverables, it is infrastructure. Sample it directly: ask teams to flag deliverables that started as an AI draft, and the ROI slide starts writing itself.
Measure inside the governed plane
All three signals share a prerequisite: usage has to happen where you can see it. When work is scattered across personal accounts, adoption is unmeasurable by construction. A single governed workspace with proper audit trails gives you the numerator and the denominator (which teams, which tasks, what share of work) without surveying anyone. Measurement, it turns out, is one more argument for consolidation: you cannot improve, or even defend, what you cannot see.