There is a German saying I have always liked, partly because it is funny and partly because it captures a very uncomfortable truth about organizations: never trust a statistic you did not fake yourself.
It is usually said with a smile, but there is something brutally relevant in it when we talk about AI adoption. Not because everyone is intentionally manipulating the numbers. That would be too easy, and also too flattering. The more common problem is quieter and more dangerous. Organizations measure what is easy to count, then start believing those numbers describe reality.
This is how leadership ends up being told that AI adoption is strong because licenses were assigned, tools were activated, training sessions were attended, prompts were submitted, or Copilot was technically used. The dashboard looks encouraging. The activity curve is pointing in the right direction. Someone has a slide that says adoption is progressing. Civilization survives another steering committee.
But the actual company may not be changing.
That is the problem.
Most AI adoption metrics are not wrong because the numbers are false. They are misleading because they measure exposure, not transformation. They tell leadership that people have access to AI, or that AI is being touched somewhere in the workflow, but they rarely prove that work is improving, decisions are changing, dependencies are collapsing, or people are becoming more capable. This is why Microsoft’s own research around AI at work is useful but also dangerous if misunderstood: it shows how deeply AI is entering the workday, while still leaving leaders with the harder question of whether that usage changes work in a meaningful way: Microsoft Work Trend Index.
That difference matters enormously.
I have seen this pattern across different transformation waves. The tool arrives. The enablement program launches. Champions are appointed. Usage dashboards are created. Leadership gets reports. Everyone starts using the language of change. But underneath, the operating model often remains untouched. People still wait for the same approvals. Teams still escalate the same decisions. Middle management still protects the same incentives. Experts remain bottlenecks. The organization keeps behaving almost exactly as before, only now it has a shiny technology layer on top of the old friction.
AI makes this worse because it is so easy to appear adopted.
With earlier transformation technologies, adoption usually had a more visible cost. You had to implement systems, migrate data, redesign processes, train users, change operating habits, and fight through real setup friction. That friction was painful, but at least it revealed something. It showed where the organization was resisting change. AI often removes the initial friction, which is why I recently wrote that AI has lowered the cost of starting but raised the standard for finishing. A user can open a tool, ask a question, summarize a meeting, draft an email, generate a slide, or create a transcript. The first step is almost too easy.
And because the first step is easy, organizations can confuse usage with change faster than ever before.
The difference between AI Exposure and AI Adoption Metrics
The core distinction is simple.
AI exposure means people have access to AI and interact with it.
AI adoption means the organization starts behaving differently because of it.
That second part is where the real work begins. It is also where most dashboards become suspiciously quiet, probably because dashboards dislike nuance almost as much as humans dislike accountability.
An organization can have high tool activation and low transformation. It can have thousands of prompts and no meaningful improvement in decision quality. It can have high Copilot usage because people are transcribing meetings, while almost nobody is using AI to rethink how work should be done. It can celebrate enablement while people remain unsure whether they are allowed to use AI in consequential work. It can claim progress while the actual operating system of the company remains untouched.
This is not a theoretical concern for me.
I once complained to people responsible for AI adoption that I had the strong feeling we were moving too slowly as a company. My observation was not based on a dashboard. It was based on behavior. I did not feel that people were actively using AI to improve their work. I did not see enough examples of people changing how they prepared, decided, delivered, learned, or collaborated. I felt strangely alone in transforming my own work with AI.
I have learned, the hard way, to observe my environment carefully when I am too convinced of something. So this bothered me more than it probably should have. I was aware that I might be wrong. Maybe the transformation was happening somewhere I could not see. Maybe teams were experimenting quietly. Maybe my own perception was too narrow. But when I raised the concern, the reply came quickly: I was wrong, because Copilot adoption was high.
The number was thrown back at me as reality.
Later, after leading an internal program to transform back-office departments with AI, I felt proven right. Not because I wanted to win an argument, although let us not pretend that does not feel nice once in a while. I felt proven right because the behavior did not match the metric. High usage did not mean the company had become meaningfully AI-enabled. It did not mean people were experiencing cognitive leverage. It did not mean work had changed.
In some cases, the metric may have been inflated by something as passive as meeting transcription.
And this is exactly where AI adoption metrics become dangerous. A meeting transcript can be useful. I am not dismissing it. But counting transcriptions as proof of AI transformation is like counting people who walked past a gym as evidence of a fitness revolution. Technically, they were near the equipment. Spiritually, we are still eating chips on the couch, Netflixing away.
Passive AI exhaust is not cognitive leverage
This is where companies need to become much more honest with themselves.
AI creates a lot of passive exhaust. Meeting transcripts. Summaries. Auto-generated notes. Suggested replies. Drafts. Search results. Convenience features. These may all have value, but they do not necessarily mean the user has changed their way of working. They may simply mean the tool is present in the background.
True cognitive leverage is different.
Cognitive leverage happens when AI helps someone do work they could not realistically do before, or do it at a level of speed, quality, breadth, or confidence that meaningfully changes the outcome. It happens when a person gets beyond step zero faster, challenges their own assumptions, compresses a research cycle, improves a client conversation, creates a better decision path, or builds something that did not exist before.
That is not the same as having a transcript of a meeting nobody needed in the first place (although, having a notoriously short attention span, I’ve been known to appreciate it and review it frequently).
The distinction matters because AI adoption should not be measured by how often a tool appears in the workflow. It should be measured by whether the workflow itself changes. If the same meetings happen, the same decisions are delayed, the same approvals are required, the same experts are bottlenecks, and the same teams wait for the same answers, then adoption has not happened in any meaningful sense. The organization has added a layer of technical activity without changing the system underneath.
This connects directly to what I have called the Interface Tax: the cost organizations pay when people are forced to work through unnecessary layers between intent and outcome. Many leaders still think AI is another tool added to the workflow. I think that is the wrong mental model: AI is becoming a layer over the workflow itself. It changes how people access knowledge, compose decisions, challenge assumptions, create artifacts, and move from ambiguity to action.
That is why measuring AI like a normal software rollout is so damaging.
If you add a traditional tool, usage can be a reasonable early signal. Did people log in? Did they complete the workflow? Did the new system replace the old process? With AI, usage alone tells you far less. Someone can use AI constantly and produce very little value. Someone else can use AI less frequently but remove a dependency that previously slowed an entire team. One user can generate noise. Another can redesign how work happens.
A metric that treats both as adoption is not neutral.
It is lying to leadership.
The real adoption signal is the “wow” moment
When I look for real AI adoption, I do not start with the dashboard. I look for the moment when people experience a genuine shift in possibility.
I look for “wow” moments.
That may sound soft, and yes, somewhere a KPI purist just spilled coffee on a balanced scorecard. But the “wow” moment is one of the clearest qualitative indicators that AI has moved from enablement to empowerment.
It can show up as relief. Someone realizes that a task that used to feel heavy, repetitive, intimidating, or impossible is suddenly manageable. It can show up as curiosity. Someone sees one use case work and immediately starts imagining five more. It can show up as confidence. A person who felt threatened by AI begins to understand that the tool can extend their ability instead of replacing their value. It can show up as social energy. People become excited to show others what they did, not because they were asked to report usage, but because they found something genuinely useful.
That is adoption.
Not the login. Not the license. Not the transcript. Not the training attendance. The moment when a person sees a better version of their own work become reachable, and wants to keep going.
This matters because transformation is social before it is statistical. People adopt new behaviors when they see value, safety, relevance, and permission. They need to feel that the new behavior helps them succeed in the environment they actually inhabit. They need to know they will not be punished for experimenting. They need examples close enough to their reality that they can imagine themselves doing it too.
This is why enablement alone is often weak. Enablement explains the tool. Empowerment changes the person’s sense of agency.
If a back-office employee uses AI to generate a better process summary, that may be useful. If they then realize they can map recurring pain points, draft a new intake process, challenge an outdated approval flow, and help their department work differently, that is transformation. The value is not in the prompt. It is in the expansion of agency.
Real AI adoption feels like people discovering that they are allowed to build.
AI rewards business builders
This is where the conversation becomes much more interesting than tool usage.
Not necessarily founders, not necessarily executives, and definitely not only the people with the fanciest job titles. I mean people who see friction and instinctively want to remove it. People who understand that work is not just a collection of tasks, but a system of decisions, dependencies, incentives, and outcomes. People who do not stop at “how can I do my task faster?” but ask “how should this work differently now that this capability exists?”
These people are often the first to extract real value from AI because they are not merely consuming the tool. They are recomposing work around it.
This is also why junior employees represent such an interesting opportunity. In many departments, AI adoption is still treated as an extracurricular activity, which is a very corporate way of saying that nobody has fully decided who owns it. So the task gets handed to someone junior, curious, available, or simply unlucky enough to raise their hand.
That could become a waste.
Or it could become one of the most powerful development opportunities available.
There are not many moments where a junior employee can realistically influence how an entire department works. AI creates one. The initial learning curve is often easier for them because they are less invested in existing patterns and more willing to experiment. But the real opportunity is not for them to become the person who knows a tool. The opportunity is for them to become the person who helps the department understand what work can now become.
That requires empowerment from leadership.
A junior person tasked with AI adoption should not be measured only by how many people attended a session or activated a feature. They should be supported to look beyond their own use case and ask: where is the department losing time, quality, confidence, or momentum? Which repetitive knowledge flows could be improved? Which decisions are waiting too long for input? Which documents are recreated again and again? Which experts are constantly interrupted? Which parts of the operating model now deserve to be redesigned?
That is how AI adoption becomes business building.
And it is also why reciprocal trust matters. The people responsible for driving AI adoption should not be incentivized merely to “reach a goal.” They should be trusted and expected to put things in motion that empower people to take the company to new heights. Yes, that sounds grand, but the alternative is counting meeting transcripts and calling it transformation, so forgive me for aiming slightly higher than administrative self-deception.
Performance busyness will destroy the signal
The reason companies fall into bad AI metrics is not stupidity. It is incentive design.
Organizations are full of people who are asked to prove progress. They need something visible. They need something comparable. They need something that fits into a quarterly review. They need a number that can survive being forwarded to leadership without requiring a philosophical essay and three uncomfortable conversations.
So they choose activity metrics.
How many users? How many sessions? How many prompts? How many trainings? How many departments onboarded? How many copilots enabled? How many meeting summaries generated?
These metrics are not useless. They can indicate exposure. They can reveal whether access exists. They can show whether basic enablement has reached the organization. But they become dangerous when they are treated as proof of transformation.
This is performance busyness.
It is the organizational habit of producing visible activity that looks like progress without necessarily changing outcomes. AI is especially vulnerable to this because it produces so much visible activity so quickly. You can create artifacts, summaries, pilots, demos, dashboards, and internal communications at breathtaking speed. The organization can look busier, more modern, and more AI-enabled while still avoiding the deeper work of changing decisions, incentives, and operating models.
Over-reliance on quantitative goal-based incentives makes this worse.
The moment adoption teams are judged primarily by usage numbers, they will naturally optimize toward usage numbers. Again, this is not evil. This is systems behavior. People respond to what the organization rewards. If the goal is activation, you get activation. If the goal is prompt volume, you get prompt volume. If the goal is visible AI activity, you get visible AI activity.
This is why research on responsible and value-oriented AI adoption keeps returning to the same uncomfortable point: value does not appear from deployment alone. For example, BCG’s work on responsible AI connects performance, trust, adoption, and value creation, which is exactly the mix most usage dashboards fail to capture.
But none of that guarantees value.
Worse, it may actively hide the absence of value. A green dashboard can end the conversation too early. It can shut down the person saying, “I do not actually see the work changing.” It can make leadership feel reassured while the organization remains structurally untouched.
That is what happened in my Copilot example. The metric became a shield against observation. Instead of asking why someone close to the work did not feel transformation happening, the organization could point to adoption numbers and move on.
That is exactly backwards.
The contradiction between lived observation and reported adoption should have triggered curiosity. It should have opened a better question: if usage is high, why is behavioral change not visible? Are people using AI passively? Are they using it in low-value contexts? Are they unclear about what is allowed? Are managers failing to create space for experimentation? Are people afraid to show unfinished AI-enabled work? Are we measuring the easiest layer and missing the one that matters?
That is the conversation leadership needs.
Not another dashboard that says the numbers are fine.
The qualitative layer is not optional
The most damaging misunderstanding about AI adoption is the belief that it can be evaluated primarily through quantitative telemetry.
It cannot.
The numbers matter, but they are incomplete. AI adoption changes cognition, behavior, confidence, decision-making, collaboration, and workflow design. These things produce some measurable traces, but the most important signals are often qualitative before they become quantitative.
You need to observe how people talk about their work. You need to see whether meetings change. You need to notice whether people arrive with better starting points. You need to look for reduced dependency on specific experts. You need to ask whether teams are challenging their own assumptions earlier. You need to see whether people feel safe sharing experiments. You need to listen for the “wow” moments.
This is uncomfortable for organizations because qualitative evaluation requires judgment.
And judgment is harder to outsource to a dashboard. And we are accomplices to that: I’m the first person to try to quantify results in Design Thinking experiments to provide short-cuts to decisions (the cool kids call this “data-driven decision making”).
But AI adoption is fundamentally a behavior change problem. That means leadership needs to combine metrics with observation, interviews, examples, stories, workflow reviews, and operational pattern recognition. The goal is not to abandon measurement. The goal is to measure the right thing in the right way.
A company that only tracks usage will learn who touched AI.
A company that studies behavior will learn whether AI changed the business, thus becoming empowered to judge is that change is positive and, if not, make adjustments.
That is the level leadership needs to operate at, especially because AI does not behave like a single-purpose enterprise tool. It is a general-purpose capability that can appear in almost any knowledge workflow. Its value will not always show up neatly in one system of record. Sometimes it shows up as better preparation. Sometimes as fewer handoffs. Sometimes as faster synthesis. Sometimes as stronger client conversations. Sometimes as increased ambition from someone who previously felt blocked by the blank page.
In my own work, that last point is very real.
AI changed what I can attempt. I have limitations in my neurological profile that made starting difficult. Once I got going, I could become a freight train moving at 500 kilometers per hour, which is an image both inspiring and deeply concerning for railway safety. But getting into first gear was always harder than people might assume. AI became the sparring partner that helps me start. It helps me create enough structure to move. It helps me begin, and then my own judgment, experience, and momentum can take over.
That is real adoption.
Not because I used the tool, but because my behavior changed. My ambition changed. My output changed. I am now sharing years of pattern recognition in ways I simply was not doing before. I was never a content machine. AI did not make me one by replacing my thinking. It helped me unlock a part of my work that was previously stuck behind starting friction.
That is the kind of change companies should be looking for.
What leadership should ask instead
If leadership wants to understand AI adoption, it needs better questions.
Not only: how many people have access?
That question matters, but it is the beginning. It is not the diagnosis.
A better question is: where has AI changed the way work gets done?
Where did a dependency disappear? Where did an expert bottleneck collapse? Where did a decision move closer to the team? Where did a cycle time shrink because the first synthesis no longer took two weeks? Where did a junior employee create leverage beyond their role? Where did a manager change expectations because AI made a different operating rhythm possible? Where did customer outcomes improve? Where did employees feel relief, confidence, or excitement? Where did people start sharing what they built because they were proud of the improvement?
Those are adoption questions.
They are harder to answer, but that is precisely why they matter. Easy metrics create easy comfort. Difficult questions create useful truth.
This also requires leadership to accept that AI adoption may not start where the official roadmap expects it to start. Real adoption often emerges from pain. Someone is tired of doing repetitive work. Someone is frustrated by slow coordination. Someone cannot get the attention of an expert. Someone sees that a recurring process is held together by manual effort and goodwill. Someone discovers that AI can remove a piece of friction and suddenly wants to redesign the flow.
Leaders need to find those people.
Then they need to protect them, connect them, and amplify them.
That is how business builders emerge inside organizations. Not through a generic training session, but through permission, visibility, and trust. The goal is not to manufacture AI heroes. Please, no more internal hero narratives. The goal is to create an environment where people who see better ways of working can move those ideas into the operating model.
That is the difference between adoption as reporting and adoption as transformation.
The adoption scorecard I would trust
If I had to design an AI adoption scorecard that I would actually trust, it would include quantitative indicators, but only as the outer layer. Licenses, active users, and feature usage would still be there, but they would not be allowed to pretend they are the whole story.
The real scorecard would include evidence of behavioral change.
It would look at whether teams can name concrete workflows that changed because of AI. It would track whether time-to-first-draft, time-to-analysis, or time-to-decision has improved in meaningful work, not just in toy examples. It would identify which expert bottlenecks have been reduced. It would capture examples of employees using AI to create new capabilities, not just consume convenience features. It would measure whether managers are creating space for experimentation. It would evaluate whether AI-enabled work is improving quality, confidence, and outcome reliability. The evidence should also be connected to the kind of operating-model discipline I wrote about in ethical AI use in companies: attribution, reviewability, accountability, and trust cannot be decorative if leadership wants AI adoption to scale.
It would also collect stories.
I know. Stories. The horror. Someone alert the spreadsheet authorities.
But stories are not the opposite of evidence. Good stories are evidence with context. They show how a capability moved through the organization. They reveal what changed, who changed it, why it mattered, and what conditions allowed it to happen. A story of a junior employee redesigning a departmental workflow with AI may be more valuable than a thousand passive tool interactions. A story of an expert bottleneck collapsing may tell leadership more than a usage dashboard ever could.
This does not mean replacing metrics with anecdotes. It means connecting metrics to lived operational change.
A useful AI adoption scorecard should answer three questions.
- Are people exposed to the capability?
- Are people changing behavior because of it?
- Is the organization learning how to scale the behavior that creates value?
Most companies stop at the first question.
That is why their AI adoption metrics lie.
The leadership trap
There is a final uncomfortable layer here.
Bad AI adoption metrics are attractive because they protect leadership from ambiguity. They create the feeling that transformation can be governed from above through measurable activity. They make change look controllable. They allow leaders to say, “we are progressing,” without asking whether people actually feel empowered, whether managers are changing expectations, or whether the operating model is evolving.
But AI adoption is not a software deployment. It is a shift in how knowledge work is performed.
That makes it messier, more human, and more dependent on trust than many leaders would like. It requires observing the organization, not just instrumenting the tool. It requires listening when someone says the numbers do not match the behavior. It requires distinguishing between passive usage and active leverage. It requires accepting that some of the most important signals will not appear first in a dashboard, but in conversations, examples, experiments, and small moments of genuine excitement. This is also why the World Economic Forum’s AI governance work keeps returning to trust, accountability, and value creation rather than tool deployment alone: World Economic Forum Responsible AI Playbook.
Leadership does not need fewer metrics.
It needs more honest ones.
The goal is not to prove that AI adoption is happening. The goal is to find out whether it is happening deeply enough to matter.
Because if AI adoption is measured badly, the organization will optimize for the wrong thing. It will reward visible usage over meaningful change. It will mistake meeting transcripts for transformation. It will silence pattern recognition with dashboards. It will celebrate activity while the operating model remains stuck.
And then, months later, someone will wonder why all that adoption did not produce the promised value.
The answer will be painfully simple.
The metrics were never measuring adoption.
They were measuring the company’s ability to look adopted.
