Companies have spent the past few years buying access to artificial intelligence. A new study suggests many have not made the matching investment in how people should use it.
Research released by The Conference Board on July 28 found that 55.1 percent of surveyed workers use generative AI or AI agents daily or weekly. Only 33.3 percent had participated in employer-provided AI training during the previous six months. More than a quarter said their organization offered no AI training at all.
The findings come from interviews with 35 enterprise leaders and a global survey of nearly 1,300 workers. They do not prove that every workplace has the same gap. The public summary does not provide enough demographic detail to treat the sample as a universal measure. It does, however, identify a practical problem that businesses can test inside their own operations.
AI adoption can spread through individual experimentation. Reliable business use cannot.
Many workplace programs still concentrate on AI literacy and basic prompting. Those skills help employees understand what a generative model can do and how to ask it for a useful response. They are a reasonable starting point, but they do not prepare a team to redesign a process around the technology.
Applied capability is more demanding. A worker using an AI agent, which is software that can carry out a sequence of actions toward a goal, needs to decide what the system may access, where a person must approve its work and how failures will be detected. Someone integrating AI into customer support must understand escalation, privacy, record keeping and quality checks, not only how to phrase a request.
The Conference Board found that fewer organizations were developing skills such as managing agents, integrating AI into workflows and applying it to strategic business problems. Independent coverage from HR Dive and Learning News highlighted the same distinction between introductory training and preparation for changing roles.
This matters because a license can be counted immediately while capability is slower and harder to measure. A company can report that thousands of employees have access to an assistant even if most use it for occasional summaries. That is different from changing how a sales proposal is prepared, how maintenance is scheduled or how an insurance claim is reviewed.
Federal Reserve research offers useful context. Its review of several US surveys found that estimates of workplace AI adoption vary widely depending on who is asked and how the question is framed. At the end of 2025, one survey estimated that 18 percent of firms had adopted AI, while an individual survey put work-related generative AI use at about 41 percent of the labor force. The difference is a warning against equating employee experimentation with material operational adoption.
Only 48 percent of workers in The Conference Board survey said their employer provided enough work time to build AI skills. A similar share said they had sufficient tools, access and resources.
That finding exposes a common accounting error. Training is often treated as a course to purchase rather than work that requires capacity. If employees are expected to learn through experimentation after completing their normal workload, the organization is shifting the cost onto them and making progress dependent on personal time and confidence.
The study’s enterprise leaders emphasized hands-on experience and managerial support. That points toward a different model. Teams need protected time to test a real task, compare the result with the existing process, document failure cases and decide whether the change is worth keeping.
The most useful unit is not a generic AI workshop. It is a supervised workflow experiment. A finance team might test invoice classification on historical data. A marketing team might compare research briefs produced with and without an assistant. A service team might measure whether suggested replies reduce handling time without lowering accuracy or customer satisfaction.
Each experiment should have an owner, a narrow business outcome and a review point. It should also record what employees had to learn beyond the tool itself. Those adjacent skills, including evaluation, data handling and exception management, are often where the real training requirement appears.
The study argues that most employers are improving skills for current roles while doing less to prepare people for roles that will change substantially. That does not mean companies should predict exactly which jobs will disappear or emerge. Such forecasts remain uncertain.
It does mean workforce planning should begin before a tool is deployed at scale. If an agent takes over the first draft of a task, the human role may move toward verification, judgment and handling unusual cases. Those responsibilities need to be designed, taught and rewarded. Otherwise, a company may automate the visible step while leaving workers responsible for an undocumented layer of correction.
The missing angle in many discussions of AI training is that the goal is not tool fluency. It is organizational reliability. Businesses should measure whether employees can recognize weak output, intervene safely and improve a process over time.
The practical consequence is straightforward. Before buying another block of AI licenses, leaders should examine a few high-volume workflows and ask whether staff have the time, access and authority to learn how those workflows should change. If the answer is no, the constraint is not the model. It is the operating system around the people using it.