Workers are teaching themselves agentic AI faster than employers are training them for it. A new Adobe survey of more than 1,000 full-time U.S. workers who use AI weekly found that 54% learn agentic AI through trial and error, 36% turn to YouTube, and 25% learn from peers. Only a third received any formal training from their employer, and among that group, fewer than half went on to build their own AI agent.
The gap is not just about access to training, it is about what training actually produces. Formal programs are most common in finance and banking (51%) and technology (48%), and least common in education (27%) and retail (29%). But even where formal training exists, it is not translating into practical skill: less than half of trained workers could apply what they learned to build something functional. Adobe’s researchers put it plainly: most workers figure it out as they go rather than rely on formal training.
The original insight here is not that employees are self-directed, that has always been true of new software. It is that the skills gap HR keeps measuring may be a measurement problem, not a capability problem. HRTech has reported that AI skills are becoming a baseline job requirement that most companies cannot measure, and the Adobe data suggests part of the reason: workers are building real fluency outside the systems HR uses to track it. A YouTube tutorial and a Reddit thread do not show up in a learning management system’s completion report, even when they produce a more capable employee than the formal course did.
Source: Adobe
