In our previous article, we introduced 12 KPIs for measuring AI adoption and connecting it to business value, workflow by workflow. The question we hear most since is a practical one: where does the data actually come from?
Most companies already have much of what they need. AI usage is recorded in platform logs. Operational performance sits in systems such as Jira, CRM and ERP. Costs are captured in finance. Risk events appear in security and compliance systems. And some of the most important information comes from the people doing the work.
The challenge is bringing these sources together to understand whether AI is changing how work gets done, and whether that change matters to the business.
AI adoption measurement starts with the workflow
The first mistake in measuring AI adoption is treating it as an organisation-wide technology metric.
An enterprise dashboard can show how many employees have access to AI, how many use it and how frequently they interact with it. These numbers are useful, but they don’t tell you whether AI is improving the work that matters.
A better starting point is the workflow.
Consider a customer service team introducing AI to help resolve requests. Platform logs can show how many agents use AI. The ticketing system can show whether requests are resolved faster, how many cases each agent handles and how often tickets are reopened. Interviews can reveal whether agents spend less time drafting responses, more time handling complex cases or simply more time checking AI-generated answers.
Together, these sources provide a much more complete picture than any AI adoption dashboard alone.
This is why AI adoption should be measured through a combination of usage, workflow performance, quality and economic impact. Each answers a different question.
The data is already there. It just lives in different systems.
Most AI adoption metrics draw on three sources: system data, interviews and analysis.
System data is the starting point. AI administration consoles record usage. Ticketing, CRM and ERP systems capture the movement of work. Accounting systems record licence, usage and implementation costs. Security systems provide evidence of incidents and policy violations.
The advantage of system data is consistency. It can be collected repeatedly without asking people to estimate what happened.
But systems rarely capture the whole story.
Interviews with employees and managers are particularly important when measuring what happens to the capacity AI releases. A system may show that a task takes less time, but it cannot tell you whether the team used those hours to handle more work, take on new responsibilities or simply absorb the workload more comfortably.
Analysis connects the two. It turns operational improvements into financial estimates, compares actual results with business-case expectations and identifies how AI adoption is progressing across priority workflows.
The combination matters. System data without context can produce misleading conclusions. Interviews without operational evidence can produce optimistic estimates. Analysis without either is little more than a business case.
Measuring whether AI is actually changing work
The most important AI adoption metrics are not necessarily the ones found in AI platform dashboards. They are the measures that show whether the workflow itself has improved.
Take three basic indicators.
End-to-end lead time measures how long work takes from the moment a request enters the process until it is delivered. The relevant timestamps usually already exist in ticketing, CRM or ERP systems. The important distinction is between measuring the AI-assisted task and measuring the entire workflow. Drafting a response faster does not improve customer service if the response still spends two days waiting for approval.
Throughput per person measures how much work gets completed relative to the capacity available. It can reveal whether AI allows a team to handle more work without increasing headcount. But the units need to be comparable. Producing twice as many reports is not a productivity improvement if the reports are substantially smaller or require more corrections.
Rework rate provides an important counterweight. Ticketing systems often record reopened cases, corrections and work returned for revision. Interviews can help identify the corrections that happen informally and never appear in the system. Without this measure, faster output can look like improved productivity even when downstream teams are spending more time fixing it.
These metrics need to be read together. Faster lead time and higher throughput are useful only if quality remains acceptable.
The hidden cost of AI adoption
One of the most frequently overlooked AI adoption metrics is review burden.
AI can accelerate the production of an initial output while increasing the time required to check it. That review effort is often missing from the original business case.
Where workflows have formal review stages, ticketing and project systems can provide timestamps. Where they don’t, reviewers need to estimate the time they spend checking AI output, using a consistent question and measurement period.
The same issue applies to AI costs. Licence and usage charges are relatively easy to find in accounting systems. Allocating shared licences across workflows is harder, but necessary if leaders want to understand the cost of producing one unit of work.
Cost per unit of output is therefore more useful than cost per user. It connects AI expenditure to the work being delivered.
And then there is the question of capacity.
AI may release hours without creating any immediate financial saving. Managers need to establish what happened to that capacity: was it used to produce more output, support new activities, avoid recruitment or reduce overtime?
This distinction is critical. Capacity released is not automatically capacity redeployed, and capacity redeployed is not automatically economic value.
Realised net value requires a further step. Operational improvements need to be translated into financial terms using conversion assumptions agreed with finance before results are known. Additional output, redeployed capacity and avoided costs should be valued separately to prevent double counting.
The result is not necessarily a precise ROI figure. It is a defensible estimate of the value actually generated, based on evidence rather than theoretical time savings.
Measuring AI adoption across the portfolio
Workflow-level measurement answers whether AI is improving individual processes. Leadership also needs to know whether the organisation is adopting AI in the right places and whether the portfolio is delivering what was expected.
This requires a different set of measures.
Priority workflow coverage asks how much of the work that matters most to business performance has actually changed. It starts with a fixed list of priority workflows agreed with business leaders. Adoption means AI has become part of how the workflow operates, not that a pilot has been launched or a tool made available.
Time to adoption measures how quickly a workflow moves from launch to sustained use. Platform logs can track weekly active users in the target group, but reaching an adoption threshold once is not enough. Usage needs to persist.
Expected versus actual gains compares the original business case with measured results. The baseline and expected improvements should be recorded before AI is introduced. Otherwise, organisations end up reconstructing expectations after the results are known.
Risk needs to be considered alongside performance. Security and IT logs can reveal confirmed AI incidents, while network monitoring, expense records and interviews can help identify unapproved AI usage. But a low incident count is not necessarily evidence of low risk. It may simply indicate that monitoring is limited. Incident severity and monitoring coverage matter as much as the count itself.
Taken together, these measures allow leadership to distinguish between a portfolio that is generating AI activity and one that is changing how the organisation operates.
The measurement discipline matters more than the data platform
There are four conditions that make AI adoption measurement credible.
First, establish the baseline before AI is introduced. Without knowing how the workflow performs today, it is difficult to establish whether anything has improved.
Second, agree on the definitions before collecting results. What counts as an adopter? What constitutes a completed unit of work? Which costs are included? How will capacity be valued? These decisions should not change after the results arrive.
Third, assign measurement ownership to the people who understand the work. IT can report usage, but workflow owners need to validate operational improvements. Managers need to confirm how capacity is used. Finance needs to agree how those improvements translate into economic value.
Finally, separate the measurement cadence from the decision cadence. Operational metrics can generally be refreshed monthly. Portfolio reviews and decisions about whether to scale, fix or stop a workflow are better made quarterly, when there is enough evidence to assess sustained performance.
From AI adoption metrics to business decisions
The data required to measure AI adoption is rarely the biggest obstacle. Most companies already have the systems, records and people needed to produce a credible picture of what is happening.
What they often lack is a consistent way to connect those sources to the work AI is supposed to improve.
That is the difference between reporting AI adoption and measuring it.
An active-user dashboard can tell leadership whether employees are using AI. A workflow scorecard can tell them whether the work is getting faster, better or cheaper. A portfolio view can tell them whether those improvements are happening where they matter and whether the organisation is getting the returns it expected.
Ultimately, AI adoption is not a technology metric. It is a measure of how effectively an organisation changes the way it works.
And the purpose of measuring it is not to prove that AI is being used. It is to give leadership the evidence to decide where AI deserves more investment, where the operating model needs to change and where the business should stop.




