A new global study from The Conference Board finds that as generative AI and AI agents become a daily or weekly tool for more than half of the workforce, most employer training programs are still built for the job employees have today, not the one AI is about to hand them. The gap is not a lack of training budget. It is training aimed at the wrong target.
The training that exists is the wrong shape
The research, based on interviews with 35 enterprise leaders and a global survey of nearly 1,300 workers, found that 55% of workers now use generative AI or AI agents daily or weekly. Only a third, 33%, say they have taken part in employer-provided AI training in the past six months. Nearly three in ten, 28%, say their employer has offered no AI training at all. Published as “Skilling for AI: Critical Factors for Navigating AI Disruption” on June 8 and released to press on July 28, the study frames this as a preparation gap rather than an adoption gap: employees are already using the technology at scale, and the organizations behind them have not caught up.
Literacy, not fluency
The more specific finding is about what the training that does exist actually teaches. Conference Board researchers Matt Rosenbaum, Marion Devine, and Diana Scott found that most programs concentrate on foundational AI literacy and basic prompting, the entry-level skill of asking a chatbot a good question. Far fewer build the more advanced capability of directing and managing AI agents, despite evidence cited in the research that workers who direct agents toward a goal outperform those who simply prompt for an answer. In effect, employers are teaching workers to talk to AI at the exact moment the more valuable skill is learning to delegate to it.
What the enterprise leaders say actually works
The 35 enterprise leaders interviewed for the study converged on a different model than the one most training catalogs still reflect. Rather than treating a course as the unit of learning, the leaders described combining formal instruction with social learning, peers teaching peers, and experiential learning, workers practicing on live, low-stakes tasks, and stressed that hands-on experimentation and practical application were the parts that actually produced capability. A prompting course delivered as a one-time module, in other words, is close to the least effective format available, even though it remains the most common one companies deploy.
Time and resources are the other half of the gap
Training content is only part of the shortfall. Fewer than half of workers, 48%, agree their organization gives them sufficient work-hour time to build AI skills, and a nearly identical share, 48%, say they have adequate tools, access, and resources to do so. That means a majority of the workforce is being asked to develop AI fluency without the time or the systems to do it during the workday, which pushes the learning into employees’ own hours instead.
HRTech has already reported what happens next: workers fill that gap themselves. A recent Adobe survey found that most employees are teaching themselves agentic AI through YouTube and trial and error rather than waiting on formal programs, and that even where formal training exists, fewer than half of trained workers can apply it to build something functional. The Conference Board data explains the mechanism behind that pattern: when employer training does not cover the skill workers actually need, and does not give them protected time to build it, self-teaching is not a workaround, it is the default.
What it means for the HR leader
The practical risk here is not that workers are unprepared. Adoption is already at 55%. The risk is that HR has no visibility into where real capability sits. HRTech has previously covered how AI skills are becoming a baseline job requirement that most companies still cannot measure, and this study points to the same blind spot from a different angle: even the training programs meant to build measurable, trackable skill are calibrated to yesterday’s use case. A completion certificate for a prompting course tells HR very little about whether an employee can now direct an AI agent through a multi-step workflow, which is the skill the research says actually moves outcomes.
That mismatch has a compounding cost. As more of the workforce reaches for AI agents rather than single-prompt tools, the organizations that trained only for prompting will find their skills inventory increasingly describes a workforce that no longer exists. Workforce planning, succession pipelines, and even job architecture built on outdated skill taxonomies will start producing decisions based on the wrong baseline.
What to do now
Three moves close the largest part of the gap without waiting for a full curriculum rebuild. First, audit existing AI training content specifically for whether it teaches directing and managing an AI agent toward a goal, not just prompting for a single answer, and rebuild the modules that stop at literacy. Second, treat time as a budget line, not an assumption: block protected work hours for AI skill-building explicitly, since fewer than half of workers currently report having any. Third, build a formal channel to capture the self-directed learning already happening on YouTube, in trial and error, and among peers, so that real capability shows up in HR’s skills data instead of staying invisible until a project surfaces it. The workforce is not waiting for HR to catch up. The measurement systems should not wait either.
Source: The Conference Board
