Generative AI in business: separating use, investment and expectations

The question “How many businesses use AI?” sounds straightforward, but the answer depends on what using means. An organization may have allowed some employees to try a tool, be running a limited pilot, have integrated a system into a particular task, or rely on it in routine processes. These are different situations, even if a corporate presentation groups them under one label. A trial can be short-lived and voluntary, while regular use may involve procedures, responsibilities and decisions about how the technology fits into existing work. Treating these stages as interchangeable makes a headline percentage sound more conclusive than its underlying evidence warrants.

When analyzing business adoption, it helps to distinguish at least four stages: exploration, trial, implementation and sustained use. The first may indicate interest; the second, a limited evaluation; the third, an organizational and technical decision; and the fourth, an integration that continues over time. An adoption figure can be interpreted only when it explains which stage it counts and whom it asks. Without those details, it is impossible to tell how many businesses have turned an initiative into a working practice. It is also useful to ask whether the measure records a company’s own assessment, an employee’s experience or evidence that a system is actually part of a process.

What counts as a “business” matters too. A survey may count legal units, business groups or workplaces. The resulting figures are not interchangeable: a group containing several companies may be represented differently depending on the statistical unit used. In Spain, the INE describes its business ICT survey as an operation whose data are collected directly from legal units. That clarification helps define the population being observed, but it should not be mistaken for an individual measure of every team or employee. A legal unit’s response does not, by itself, establish how widely a tool is used within it, whether use is optional, or whether the same practice applies across its locations.

Official statistics are a useful starting point because they document the population studied, the method and the definitions in a more structured way than a commercial announcement usually does. The INE says that its Survey on the Use of ICT and E-commerce in Enterprises aims to obtain harmonized information that is comparable at European level, and that it is prepared in accordance with Eurostat guidance. Its questionnaire covers several technological areas and may include modules on specific topics. This kind of documentation gives readers a basis for understanding what is being measured and how the resulting figures are intended to be used.

That design is valuable for observing the presence of particular technologies across businesses, but it does not automatically turn the survey into a detailed audit of every deployment. A rate of businesses reporting that they use a technology does not, on its own, reveal how many people use it, how often they use it, what tasks it supports, how much is spent on it, or what effect it has on productivity. An adoption statistic answers a question about businesses; it does not replace an assessment of labor or economic impact. A reported use rate can be meaningful while leaving many practical questions unanswered. Those questions require other kinds of evidence, gathered at the appropriate level and with a method suited to the claim being made.

Check the definition before reading the percentage

When reviewing a result, look for the reference period, the population of businesses, any minimum size included, the sector, the counting unit and the exact wording of the question. Also check whether the figure refers to artificial intelligence in general or to generative AI. It is unwise to treat those expressions as equivalent: a broad category may include uses that do not rely on generative models, whereas a survey focused on generative AI defines a more specific technology. If the table does not make the distinction clear, the conclusion should retain that limitation rather than silently assuming a narrower meaning. The source’s definitions are part of the result, not a technical detail that can be set aside after the percentage has been quoted.

Business surveys provide another kind of evidence. They may ask executives about plans, obstacles, budgets or reported uses. They can reveal priorities and perceptions, but their results depend on who responds, how the sample is selected and how the questions are worded. Saying a business “is exploring” a technology is not the same as saying it has deployed it; declaring an intention to invest does not show that the spending has taken place. Even a carefully conducted survey describes the answers collected under its particular method. It should not automatically be read as a direct observation of systems in operation or of the experience of all employees.

Investment announcements and agreements with suppliers describe decisions or commitments communicated publicly. They are relevant to understanding a company’s direction, but they do not, by themselves, establish that a system is operational, reaches the entire workforce or produces the expected benefits. To move from an announcement to evidence of implementation would require later information about deployment, scope, continuity and results, using criteria that make it possible to verify what has actually been put in place. A project can be announced before its final scope is settled, and a commitment can concern future work. Those possibilities do not invalidate the announcement; they simply define what it can demonstrate at that point.

Suppliers may publish aggregate figures for users, queries or activity on their platforms. Such metrics can provide indications of use within the supplier’s own service, but they do not necessarily represent adoption across all businesses or describe the purpose of each interaction. An active account may correspond to an occasional trial; the number of requests alone does not show how many organizations have institutionalized use. The publisher’s commercial interest should also be considered. Before interpreting a figure, look for its definitions, time period, coverage and counting method. A platform activity metric and a business adoption statistic can both be informative, but they measure different things and should be described accordingly.

Two percentages are not comparable merely because both concern AI in business. They may differ by country, year, company size, economic activity, surveyed population and definition of use. They may also measure different moments: the publication date of a report does not always match the period to which its answers refer. Before inferring a trend, it is useful to align these dimensions or explain why the comparison is approximate. If a difference in scope cannot be resolved, the comparison should not be presented as though the numbers came from the same measurement exercise.

A comparison over time is stronger when it retains the same question, methodology and population, or when the source explains changes and makes them possible to interpret. If the definition of the technology changes, an apparent increase may partly result from the new criterion. If the sample changes, the result may reflect a different mix of businesses. A change in the indicator does not automatically prove an equivalent change in business practice. A trend claim therefore needs to distinguish a genuine change in reported behavior from a change in what the survey counts or whom it includes. That distinction is especially important when a technology category is evolving or a survey question is revised.

The survey’s scope matters as well. A national average can conceal differences between company sizes or sectors. A figure collected among organizations that already use advanced digital services should not be presented as representative of the entire business landscape without methodological justification. And the conclusions of a sample of technology managers do not necessarily describe what is happening in workplaces. When reports do not detail these aspects, it is better to present the result as a partial signal, not as a universal measure. A careful account can still report the number, while making clear which population it describes and which questions it leaves open.

A sequence of checks can help assess whether AI is becoming established in a business. This is not an official statistical scale, but a framework for reading evidence without confusing activity with outcomes:

• Exploration: there is interest, introductory training or an evaluation of options.
• Trial: a pilot exists with defined users, tasks and limits.
• Implementation: the tool is integrated into a process and has designated owners and rules for use.
• Continuity: use is maintained over time, and costs, quality, risks and results are reviewed.

Each step calls for different evidence. A press release may substantiate that a company announced a project; it is not enough to prove that use is routine. An internal survey may describe employees’ experiences; assessing results would require defined indicators and an appropriate comparison. An official statistic may report businesses that say they use AI without necessarily detailing each business’s degree of integration. The framework is useful precisely because it keeps these claims separate instead of allowing a single sign of activity to stand in for every later stage.

The most persuasive evidence combines independent, complementary sources: an official measurement with a clear definition, business information about the actual scope of deployment and, when benefits are claimed, an assessment explaining how they were measured. No single source has to answer every question. It is, however, important not to make a limited indicator carry conclusions it cannot support. Consistency across measures, not the volume of announcements, is what makes a more robust interpretation possible. In practice, that means describing what each source establishes, identifying where it stops, and looking for corroboration before moving from evidence of interest to claims of sustained adoption or impact.

The information documented here makes it possible to describe how Spain’s business ICT survey is structured and why its definitions help readers interpret its results. It also supports a methodological distinction between statistics about businesses, statements of intention, activity on a platform and impact assessments. This framework makes it possible to examine future figures more precisely. It is not, however, enough to claim that generative AI adoption is growing in Spain or across Europe. A sound analytical framework should clarify the limits of the available evidence rather than turn those limits into a conclusion about a trend.

To support a claim of growth, comparable results from different periods would need to be assembled, the definition of use would need to remain consistent, and it would be necessary to check that the statistical population had not changed in a way that altered the interpretation. To say that use has become established would additionally require signs of continuity and integration, not just experimentation. And attributing productivity improvements to AI would require distinguishing its contribution from other changes in processes, staffing or investment. Each claim therefore has its own evidentiary burden; evidence sufficient to establish reported use may be insufficient to establish sustained deployment or a measurable effect.

The cautious conclusion is not that adoption does not exist, but that each indicator answers a different question. An official survey can measure reported use within a defined framework; an expectations survey can describe intentions; a company communication can document an announcement; and platform data can offer a partial view of activity. None should automatically be presented as a substitute for the others. Until comparable time series and evidence of sustained deployment are available, the overall trend should remain open. That conclusion preserves what the indicators can show while avoiding stronger claims than the available information supports.