How to Measure AI Adoption: The Five Numbers That Matter More Than Your Licence Count
AI adoption is not how many people have access to an AI tool. It is how many people use it every week on real work, and what changed because they did.
Five numbers cover it: weekly active users against seats, use cases running with a named owner, time or cost returned against a before number, which tools and connectors are actually used, and how many teams have at least one use case of their own. All five can be collected without a data project, and together they answer the only question a board cares about — is this changing how the company works?
Why licence counts lie
The most common AI metric in a board pack is the number of licences bought. It is also the least informative. A licence measures a purchase. It says nothing about whether anyone opened the tool after the first week, and nothing about whether the work that matters changed.
The pattern behind it is familiar. A company buys seats for everyone, announces it, and usage follows the same curve almost everywhere: a burst of curiosity, then a collapse to a small group of heavy users. Six months later the licence count is unchanged and looks like progress. The actual adoption is a dozen people.
Asking people whether they like the tool is not much better. Satisfaction scores after a training session measure the session. What you want to know is what people did on an ordinary Tuesday three weeks later.
The five numbers
| Number | What it tells you | How to get it |
|---|---|---|
| Weekly active users ÷ seats | Breadth. How much of what you bought is used, week to week. The trend matters more than any single week. | The admin analytics in your AI tool. |
| Use cases running, with an owner | Depth. Recurring tasks that now run with AI, each with a named person responsible. Ad-hoc questions do not count. | A short list, kept by whoever owns AI internally. |
| Time or cost returned | Value. For each use case, how long it took before and how long it takes now, in hours, kronor or quality. | Written down when the use case starts, checked a month later. |
| Tools and connectors actually used | Where the work is happening. Which connectors, projects and skills are called, and which were set up and forgotten. | Admin analytics, or the logs of whatever connects your AI to your systems. |
| Teams with a use case of their own | Spread. Whether AI lives in one enthusiastic team or has reached sales, finance, operations and the rest. | The use case list, grouped by team. |
Track the second. A use case with an owner and a before number is the unit that turns into every other metric: it produces weekly usage, it returns measurable time, and it shows which tools matter. A company with twenty of those has adopted AI, whatever its licence count says.
Leading and lagging: which to watch when
The five numbers do not move at the same speed, and reading them at the wrong time leads to the wrong conclusion.
Leading indicators move in days: weekly active users, and which connectors and projects are used. They tell you whether the rollout is taking hold. In the first month they are the ones to watch, and a drop in week three is the early warning that people have gone back to how they worked before.
Lagging indicators move in months: time returned, and the spread of owned use cases across teams. They tell you whether it paid off. Judging a rollout on them after four weeks is too early; not having a before number to compare against at week twelve is too late.
How to collect them without a data project
None of this needs a dashboard project or a data team. Three habits cover it.
- Use the analytics you already have. Business plans for the main AI assistants include admin views of who is active and how much. For Claude, we have written about the admin analytics and cost controls, and for individuals, Claude Reflect shows how they actually use it.
- Write the before number on day one. When a team moves a task to AI, one line: what the task is, who owns it, how long it takes today. That line is the whole measurement system for value, and it is the one almost everyone forgets.
- Keep one list. Every use case, its owner, its team, its before and after. Reviewed once a month by whoever owns AI internally. A spreadsheet is enough.
As AI use spreads across several assistants and many connected systems, the fourth number gets harder to see from any single tool’s analytics. That is one of the reasons companies route those connections through one place — we explain the idea in what an MCP gateway is.
What good looks like at 30 and 90 days
Not benchmarks to hit, but the shape of a rollout that is working, and the one sign at each stage that it is not.
| By | Working | Warning sign |
|---|---|---|
| Day 30 | Most seats are used every week. Each team has at least one use case with an owner and a before number written down. | Usage peaked in week one and has fallen every week since. |
| Day 90 | Several use cases per team, the first after numbers are in, and at least one team has found a use case nobody planned. | All the use cases still sit in the team that asked for AI in the first place. |
Four ways to measure the wrong thing
- Counting prompts. A thousand quick questions can mean less than one weekly report that now builds itself. Volume measures activity, not value.
- Surveying satisfaction instead of behaviour. People can like a tool they never use. Ask what they did with it last week.
- Measuring the tool instead of the work. The question is not whether Claude is being used, but whether the monthly close, the proposal or the client follow-up got faster or better.
- Starting without a before. Without the before number there is no after, and the next budget conversation becomes a matter of belief.
The companies that get this right report AI adoption the way they report pipeline: monthly, owned by a named person, with the same few numbers every time. It is a leadership metric, not an IT one, because the value only appears when the way people work changes. For the rollout that produces those numbers in the first place, see our thirty-day plan for rolling Claude out to a team.
Setting up the use case list, the owners and the before numbers, and reviewing them every month until adoption runs on its own, is the core of Deployed Partner. It starts with a Deployed Kickstart, where each team leaves with its first use case and its first before number.
Frequently asked questions
How do you measure AI adoption?
Measure what people do, not what was bought. Five numbers cover it: weekly active users compared with seats, the number of use cases running with a named owner, time or cost returned against a before number recorded when each use case started, which tools and connectors are actually used, and how many teams have at least one use case of their own. All five can be collected from the admin analytics in your AI tool and a single list of use cases reviewed monthly.
What are the most important AI adoption metrics?
If you track only one, track the number of use cases running with a named owner and a recorded before number. A use case is a recurring task that now runs with AI, not an ad-hoc question. It is the unit that produces every other metric: it creates weekly usage, returns measurable time and shows which tools matter. Weekly active users against seats is the best early indicator of whether a rollout is taking hold.
Why is the number of AI licences a bad measure of adoption?
A licence measures a purchase, not use. Usage after a company-wide rollout typically rises briefly and then falls to a small group of heavy users, while the licence count stays the same and looks like progress. Licence counts say nothing about whether people use the tool every week or whether any work became faster or better.
How do you measure the ROI of AI?
Per use case, against a before number. When a team moves a recurring task to AI, write down what the task is, who owns it and how long it takes today, in hours, cost or quality. Check the same measure a month later. The sum across use cases is the return. Without the before number recorded at the start there is nothing to compare against, which is the most common reason companies cannot show AI ROI.
How often should AI adoption be measured?
Watch the leading indicators, such as weekly active users and which connectors are used, weekly during the first month of a rollout, because a drop in week three is the early warning that people have gone back to their old way of working. Review the lagging indicators, time returned and the spread of use cases across teams, monthly. Judging a rollout on lagging indicators after only a few weeks is too early.
Who should own AI adoption metrics?
A named person in the business, reporting to leadership, rather than IT alone. The value of AI appears when the way people work changes, so adoption is a leadership metric. The companies that do this well report it monthly, the way they report pipeline, with the same few numbers every time.
Found this useful? Send it to someone who needs it.