I have sat in too many steering committees where a green dashboard was celebrated the week before a rollout quietly collapsed. Login counts were up. Training was "complete." And yet, six weeks later, people had drifted back to spreadsheets. The metrics were not wrong—they were just measuring the wrong thing.
After years running ERP and AI change programs at scale, I have learned one uncomfortable truth: most adoption metrics tell you what already happened, not what is about to happen. If you want to steer, you need leading indicators—signals that predict success while you can still influence the outcome.
The trap of vanity metrics
Vanity metrics feel reassuring because they are easy to collect and almost always go up. Logins, page views, training attendance, number of accounts provisioned. They photograph activity, not value.
A user who logs in once a day to check a number and then leaves is counted exactly the same as a power user reshaping their entire workflow around the new system. That is the flaw. Aggregate activity hides the distribution that actually matters.
If a metric only ever moves in one direction, it is probably not measuring adoption—it is measuring compliance.
The five indicators that actually predict success
Over successive programs, five families of signals have consistently separated the rollouts that stuck from the ones that reverted. None of them are counts of logins.
1. Usage depth
Not "did they open it" but "how far in did they go." Are people using the two features that justified the investment, or only the login screen? I track the ratio of core-workflow actions to total sessions. When depth stalls while activity climbs, that is a rollout living on borrowed enthusiasm.
2. Proficiency and time-to-proficiency
Adoption is not a moment, it is a curve. The question is how quickly a user goes from hesitant to fluent—and how many never make the climb. Time-to-proficiency is my favourite predictive KPI because it is both measurable and coachable. If the median user takes eight weeks to become fluent in a system meant to save them time, you have a design or enablement problem, not a motivation problem.
3. Sentiment and pulse
People vote with their feelings long before they vote with their behaviour. A short, frequent pulse—one or two questions, asked in the flow of work—catches the mood shift weeks before usage data reflects it. This is where I lean on AInspire's real-time pulse approach: instead of a quarterly survey autopsy, sentiment is captured continuously, so a dip in one department becomes a Tuesday-morning conversation, not a post-mortem.
4. Resistance signals
Resistance rarely announces itself. It leaks out as workarounds, shadow spreadsheets, a spike in "how do I export this back to the old format" tickets, or a team that mysteriously never shows up to office hours. I treat these as first-class metrics, not anecdotes:
- Support tickets that ask how to undo the new way of working
- Persistence of legacy tools past the cutover date
- Meetings where the new system is described in the past tense
- Champions going quiet
5. Breadth across the organisation
An average can be poisoned by heroes. One enthusiastic team can carry an adoption number while five others silently opt out. I always look at the distribution: what percentage of teams have crossed the proficiency threshold, not just the headline mean.
Measuring early enough to act
The whole point of leading indicators is intervention. A lagging metric tells you the patient's temperature at the funeral. A leading metric tells you on Tuesday that a specific team is disengaging, so you can call their manager on Wednesday.
This changes the cadence of change management itself. Instead of a quarterly review, you run a weekly signal check:
- Depth flat, activity up? People are complying, not adopting. Revisit the value story.
- Time-to-proficiency lengthening? Fix onboarding or the interface before scaling further.
- Sentiment dipping in one unit? Send a human, not another email.
- Resistance signals rising? Surface the underlying objection—there is always a rational one.
Build the scoreboard before you launch
The mistake I see most often is instrumenting adoption after go-live, when the only data available is whatever the system happened to log. Decide before launch what "adopted" actually means for your organisation, and make it a blend: usage depth, proficiency, sentiment, and the absence of resistance. A single composite adoption score, refreshed weekly, does more to protect an investment than any polished quarterly deck.
Adoption is not a switch you flip at go-live. It is a curve you either watch climb or watch flatten—and the leaders who measure the right signals get to act while the curve is still theirs to shape.
Cédric Bignet is an AI & ERP Change Management expert at Novartis and founder of AInspire. He writes about change management, AI adoption and enterprise transformation.