How to measure AI adoption in your team (without counting licences)

Almost every AI adoption dashboard answers a question nobody asked. It reports how many seats were bought, how many people logged in this month, and how many finished the training. All three can be at one hundred percent in a team where the work has not changed at all.

The gap is easy to see once you look for it. A person can open the tool every day and use it for the same three low-value tasks they always used it for. A team can complete the training and go back to the workflow they had before, because nobody showed them how the tool fits the job they actually do. Activity is a proxy for adoption in the same way that attendance is a proxy for learning.

What a leader actually needs to know is narrower and harder: for this role, doing this work, has AI changed how the job gets done, and if not, what is stopping it.

It helps to separate what is being measured. Most reporting collapses these into one number and loses the signal.

1. Access: who has the tools

The licences you pay for, mapped to the people who hold them. This is procurement data and you almost certainly have it. It is the denominator for everything else, and on its own it says nothing about adoption.

2. Activity: who opens them

Logins, sessions, prompts sent. Useful for one thing only: finding licences nobody has opened in ninety days, which is a budget conversation rather than an adoption one. Beyond that, activity rewards the person who uses AI constantly for trivial work and misses the person who uses it twice a week on the thing that matters.

3. Capability: how well each role uses them

This is the layer that answers the question, and the one almost nobody measures, because it cannot be read off a usage log. It means watching how a person actually directs a tool on their own work: whether they brief it properly, whether they check what comes back, whether the use is a one-off or part of a repeatable workflow.

4. Outcome: what the week looks like now

Whether the shape of the work changed. Which stages of a process got shorter, which stayed exactly as long, and how many hours moved from one to the other. This is the only layer that is a result rather than an indicator, and it is the one that has to be measured twice to mean anything.

Capability is not a single number. Five things travel together and a team can be strong on some and weak on others, which is precisely why one company-wide score is worse than useless: it averages them into a figure that describes nobody.

• Prompting and direction. Can they brief a tool the way they would brief a capable new colleague, with the goal, the audience and the constraints, rather than a one-line request?
• Tool literacy. Do they know which of the tools they hold is right for this task, and do they have a realistic sense of where its capability ends?
• Workflow integration. Is AI a thing they open occasionally, or is it a repeatable step inside how the job gets done?
• Critical thinking and judgment. Do they check what comes back? The most dangerous user is a confident one who does not.
• Growth and influence. Do they share what works, so one person's improvement becomes the team's?

Of those five, workflow integration and judgment are where most teams are weakest, and they are the two that activity data can never see. Someone can score highly on every usage metric your admin console produces while failing both.

The same capability score means different things in different jobs, and the same weakness costs different amounts.

• In sales, weak verification means AI-generated account research goes to a prospect with a funding round that is two years out of date.
• In marketing, it means a generated claim reaches a campaign with no source behind it, and a customer finds it before you do.
• In operations, it means the automation runs cleanly on the happy path while the exceptions quietly pile up where nobody is looking.

Three people, one score, three completely different exposures and three different fixes. A generic readiness test hides all of it, then a single all-hands workshop gets booked against the average. That is how AI training budget gets spent on the wrong gap.

You do not need a company-wide programme to get a real answer. Pick one team, ideally one where the work is well understood and the tools are already provisioned.

• Map the work first. Write down the stages of a typical week for that role, with rough hours against each. You cannot spot a saving in a process you have not described.
• List the tools that team actually holds, not the ones the company owns. Adoption of a licence nobody was given is not a fair question.
• For each person, look at real output rather than asking them to rate themselves. Self-reported confidence runs ahead of judgment almost everywhere, which is why surveys flatter.
• Score against what that role needs, not against a company-wide bar. Set the bar from your strongest team in that function.
• Record the barriers people name. Time, trust, unclear permission and not knowing where to start are the common four, and they need different responses.
• Separate estimate from result. Hours you think could come back are an opportunity. Hours that came back are a measurement, and you only have one after you run it again.
• Re-measure in a quarter. A standard nobody re-checks stops being one.

Once you have that, four figures carry a leadership conversation, and each one needs a denominator next to it.

• Assessed, out of the team. Coverage first, because every other number is read against it.
• Capability by dimension, by role. Where the team is strong, where it is at risk.
• Tools held versus tools actually used. The clearest budget signal you will get.
• Recoverable hours a week, labelled as an opportunity until a second measurement turns it into a result.

Notice what is missing. There is no single adoption percentage, because there is no honest way to produce one, and a number that cannot be acted on is a number that will eventually be argued with instead.

Be careful with external benchmarks. Most published AI adoption figures measure access or activity, so comparing your capability data against them compares two different things.

Microsoft's Work Trend Index is worth reading for the shape of the argument, and it is where the term Frontier Professionals comes from, meaning the most advanced AI users in their research. It describes a person, not a company. We link it rather than quoting a statistic from it, because a number lifted out of someone else's methodology and dropped into yours is how benchmark claims go wrong.

The more useful benchmark is internal: the level your own strongest team in that function already operates at. It is real, it is current, and nobody can argue that it is unachievable in your company, because somebody in your company is already there.