What Is AI Adoption and How to Measure It Properly

AI adoption is the degree to which artificial intelligence has entered a company's daily work and processes, in a stable and repeated way. It is not the number of licenses purchased, and it is not the number of people who tried a tool once. It measures how much AI changes the way work gets done.

The distinction matters because budgets are decided on this reading. If adoption is overstated, you renew licenses nobody uses. If it is understated, you lose sight of where AI is already producing value.

Licenses, usage and real adoption: three different things

Companies often confuse three levels that should be kept apart.

• Licenses. They measure what was purchased and assigned. This is spend and distribution data, not behavior. An assigned seat can sit idle for months.
• Usage. It measures who opens the tool and how often. This is activity data: it says someone is there, not what they get out of it.
• Real adoption. It measures whether AI is part of a work process: whether a function uses it regularly for a specific use case, with results the department recognizes as useful.

A hypothetical example: a company assigns 1,000 seats of an AI assistant. In the first month, 700 people log in. After three months, weekly logins drop to 250. Licenses are still 1,000, usage has fallen sharply, and real adoption probably covers only a subset of departments. Yet the number that reaches the executive committee slide is often the first one.

Why the company-wide average misleads

An average usage rate hides the differences you need in order to decide. One department may use AI every day on three recurring tasks while another has never built it into its work. The average merges them into a value that describes neither.

For this reason, adoption should be read along three axes, always in aggregate form.

By department

The department is the natural unit for seeing where AI has taken hold and where it has not. It lets you compare similar functions, spot who is ahead and understand which experience can be replicated elsewhere.

By role

Within the same department, roles have different needs. People who write contracts, analyze data or manage clients use AI in ways that are not comparable. Reading adoption by role shows where the problem is training, where it is the tool, and where there is simply no sensible use case.

By use case

The use case is the most useful axis for linking adoption to value. Knowing that "legal uses Claude" says little. Knowing that the legal department uses it for the first read of standard contracts, at a steady frequency, lets you assess whether that process has improved. It is also the starting point for measuring impact, covered in how to measure AI ROI.

How to measure adoption: data sources

There is no single reliable source. You need at least two that correct each other.

1. Vendor usage data. Vendor admin consoles expose, to varying degrees, per-user usage data: logins, frequency, volumes. It is objective, but it covers only official tools and, when there are several vendors, must be normalized to be comparable.
2. Organization map. Per-user data becomes useful only when tied to department and role. This requires alignment with company HR records, handled so that only aggregated results are returned.
3. Surveys. They surface what logs cannot see: what AI is used for, how satisfied people are, what obstacles they meet and which unofficial tools are in use. The phenomenon is explored in shadow AI, risks and how to manage it.
4. Use case register. A census of use cases declared by departments provides context: without it, a usage figure does not say which process it belongs to. It is the basis of any reading by use case.

Which indicators to use

A few precisely defined indicators are worth more than a crowded dashboard. A reasonable baseline is the following.

• Active user rate, weekly and monthly, by department and role.
• Frequency of use: how many days per week an active user turns to the tool.
• Depth of use: whether use stays on simple tasks or extends to more structured work.
• Use case coverage: how many of the recorded use cases have recurring users.
• Retention: how many users active in a given month are still active three months later.
• Share of use outside official tools, estimated through a survey.

Retention is often the most revealing figure. An initial peak followed by a decline shows that launch enthusiasm did not turn into habit.

The most common mistakes

Counting seats as users. An assigned seat is not a person using the tool. The ratio between the two is exactly what you want to know.

Measuring a single tool. When several assistants coexist in the company, looking only at the main one gives a partial picture. Adoption is about AI in work, not about one product.

Ignoring unofficial use. Part of the use may happen on personal accounts and never appears in company vendor data. Without a survey, the estimate is systematically too low.

Assessing individuals. Data on people should always be presented in aggregate by role or department, never as a judgment on one person. Besides being common sense in relations with employees, in Italy it interacts with the Workers' Statute (Law no. 300 of 20 May 1970), covered in monitoring AI use and the Italian Workers' Statute.

Mistaking adoption for impact. High adoption does not prove that work has improved. It is a necessary condition, not a sufficient one. Impact requires a separate measurement on well-defined processes.

How to start: an operational sequence

1. Define what counts as "use" (for example, at least one session per week) and apply the same definition to every tool.
2. Collect usage data from each vendor and normalize it.
3. Map users to department and role, reporting only aggregated results.
4. Run a short survey to estimate use outside official tools and the real use cases.
5. Record the main use cases by department.
6. Build a reading by department, role and use case, and repeat it every quarter.
7. Use the result to decide where to train, where to consolidate tools and where to measure impact.

Step 7 is the one that is often missing. Measurement exists to support decisions, so it helps to link it to spend: reading cost and usage per seat together is the method described in how to reduce AI license spend.

Limits to declare

Even a well-built measurement has boundaries. Usage data says how much and when, not how well. Surveys depend on the honesty and memory of respondents. Intensive use does not imply effective use, and occasional use can be valuable if it concerns a rare and critical task. Adoption is best presented as an indicator to interpret together with departments, not as a score.

Where Delphis fits

Delphis is a vendor-independent platform that connects spend, adoption, impact and regulatory obligations in a single model. On adoption, it integrates usage data from multiple vendors, ties it to department, role and use case, and sets it alongside surveys. Data on people stays aggregated. For more, see the adoption and ROI page.