A business analyst at a mid-size insurance company had her phone propped against a coffee mug, playing a YouTube tutorial at one and a half speed while she worked. Some guy in a hoodie was explaining how to write prompts that produce usable output from the assistant her company had licensed four months earlier. She had a spreadsheet open on the other monitor and kept pausing the video every forty seconds to try what he had just said. I asked who had trained her on the tool. She looked at me like I had misread the room, and said nobody had. There had been an email with a link in it.

That is close to how much of the American workforce is picking up the technology their employers keep describing as a remaking of their jobs. Jobs for the Future surveyed 3,020 people in late 2025 and found that 9% get their information about AI from their employer1, well behind the 31% who get it from social media1. Among people who said they wanted to grow more comfortable with the tools, 62% planned to experiment on their own and 53% planned to watch YouTube tutorials or informal courses1. JFF is a workforce nonprofit that argues for heavier employer investment in training, so the result flatters a position it already holds, and you should know that before you carry the number into an executive meeting. The sample is large and weighted to the census, and the pattern holds across the groups they broke out.

Nobody was in the room when the decision got made

The finding that should stop a senior leader is not really about training. 56% of workers said their employer has never consulted them about how AI tools should be used in their work1. Among people with less than a year on the job, that reaches 70%1. Over the same stretch, the share of workers who said they adopted AI at their employer's direction climbed from 11% to 18%1. The instructing went up. The asking stayed where it was.

I spent years coaching teams through Agile transformations, and I watched a lot of organizations relearn the same lesson at considerable expense. We put the people who do the work in the room for a practical reason. The knowledge about which tasks are worth handing to a machine, and which ones only look repetitive because the judgment buried inside them is invisible from three levels up, lives with the person doing them and nowhere else. You cannot procure that knowledge, and it does not show up in a job description. An executive can name the function. Only the analyst can tell you that the twelve minutes she spends reconciling the file are the twelve minutes where she catches the errors.

An adoption dashboard tells you the tool was opened. It says nothing about whether anyone who understands the work ever decided it belonged there.

What most companies ran over the last two years was a waterfall project. Central decision, procurement cycle, launch date, a dashboard counting seat usage, requirements written by people who would never touch the thing. Agile did not invent the alternative. It put a name to something people had already noticed worked better, on the kind of work where the old project management habit of treating people as resources to be assigned kept producing worse results than treating them as people whose judgment was worth having. Plenty of the organizations that spent a decade and a small fortune on Agile never actually adopted it. They ran the ceremonies and kept the old instincts underneath, and the AI rollout is simply where they stopped pretending. What they had been all along is process-focused project management, working from the belief that when people will not follow a simple process, the fault lies in how the process was designed, and never in whether the people doing the work thought it was worth following.

What happens when nobody asks

Use went up modestly, from 35% of workers to 38%1. Readiness went the other way. The share saying they feel not at all prepared to use AI in their job rose from 10% to 17%1. Sentiment moved too. In 2024, more respondents said AI was doing more good than harm to people's ability to find work and build wealth, by 45% to 41%. A year later that had turned over, with 44% saying more harm and 38% saying more good1.

Read that from the top of the house and the temptation is to call it resistance, then treat it with more mandate. My read is duller and I think more likely. People were handed a general-purpose tool with no account of how it applied to their particular work, told to figure it out, and then measured on whether they had. They went to YouTube, learned a version of the tool that nobody at their company had vetted, and formed their opinion of it out of that experience. Dave Birss of The Gen AI Academy, quoted in LinkedIn's 2026 talent report, put it about as plainly as it can be put: most leaders "have simply bought some licenses and are crossing their fingers that productivity will magically appear."2 That same report, produced by a company that sells corporate learning and therefore has an obvious stake in the conclusion, puts 86% of companies below the bar it sets for building and moving skills at the pace the work now demands2.

The cheapest thing on the table

Asking costs almost nothing. An hour with eight people who do the work, spent before the license is signed rather than after the adoption numbers disappoint, would tell an executive more about where the tool will actually earn its money than any vendor deck ever has. It would also convert a group of skeptics into a group with some authorship in the outcome, which is most of what decides whether a rollout survives its first difficult quarter.

There is a version of this where the analyst with the coffee mug ends up teaching the tool to her whole department, because somebody thought to ask her what she had worked out on her own. That version costs about the same as the email did.

If you rolled out an AI tool in the last eighteen months, who did you ask first, and what did they tell you that changed the plan?