Everyone Can Talk About Agents. Almost No One Can Ship One.
Three years ago, if someone dropped "vector database" or "retrieval-augmented generation" into a conversation, you assumed they'd actually touched one. Today those words show up in job postings, LinkedIn bios, and interview answers from people who learned them the same afternoon, from the same tool they're claiming mastery of. AI didn't just make code easier to write. It made expertise easier to fake, because the model can hand anyone a fluent one-liner on a topic they've never shipped in production.
The scale of the gap between claimed and real usage backs this up. Adoption numbers at the organizational level look almost universal now, north of 90% of companies say they use AI somewhere. But daily individual usage sits closer to one in five workers. That's not a rounding error. That's a company telling its board, its customers, and its recruiters one story while the actual day-to-day tells a different one. Interviews are the same story shrunk down to one person.
The output of confident talkers has a name now
Here's the part that should worry any manager who's been impressed by an interview answer: researchers have started tracking what happens when people who sound competent produce actual deliverables. They call it "workslop" - AI-generated work that looks finished on the surface but falls apart the moment someone tries to build on it. Recent surveys put the share of workers who've received it in the last month at around 40%, and each incident costs the receiver close to two hours to untangle. Half the people who get handed workslop start quietly rating the sender as less capable and less trustworthy than before.
Read that against the hiring problem and it clicks into place. The candidate who can talk fluently about agents but has never shipped one isn't a hypothetical risk. And unlike a bad hire from five years ago, who might at least produce something obviously broken, this one produces something that looks done. That's worse, because it takes longer to notice the rot.
Even the real ones are cracking
Here's the twist nobody wants to hear: verifying that someone actually knows how to use these tools doesn't fully solve the problem, because the people who genuinely do know how to use them are running into a different wall. An eight-month study inside a 200 person tech company found that when employees got legitimately good at AI, they didn't end up with more free time. Their to-do lists just expanded to absorb whatever time got freed up, and work started bleeding into lunch and evenings. One engineer in the study put it plainly: you'd think you could work less, but you end up working the same or more.
So the hiring bar can't just be "show me you can build it." It has to include "show me you can build it and still go home at six." A team of skilled people quietly grinding themselves into burnout is not a stable asset, no matter how good their demo looked in the interview.
What to actually do about it
Stop asking candidates to explain concepts and start asking them to touch your actual problems. Pull a real, bounded ticket from your backlog - not a leetcode puzzle, not a take-home that could've been generated in ten minutes - and watch what happens when they hit the parts that don't have a clean answer in any tutorial. Give them your actual tools, your actual data quirks, your actual legacy mess. The person who's actually done this work will ask different questions than the person who's memorized the vocabulary. They'll ask about rate limits, about what happens when the model changes its behavior mid-quarter, about who monitors the thing after it ships. Those questions don't come from a prompt. They come from having been burned once already.
Track outcomes over token counts internally, too. If someone on your team claims a workflow saved the business real time, ask them to show the before-and-after, not the screenshot of seventeen agents running. If a workflow only works when it's hand-fed exact context every time, that's not automation, that's a very elaborate macro with extra steps.
And build in room for the good ones to actually slow down. If your best AI-literate engineer is the one working sixteen-hour days because the tools made everything feel possible, you don't have a productivity story. You have a burnout story wearing a productivity costume, and it'll show up on your attrition numbers before it shows up anywhere else.
The teams that win this next stretch won't be the ones with the most confident AI talkers in the room. They'll be the ones who figured out, quietly and without a LinkedIn post about it, how to tell the difference between someone who can describe the work and someone who's actually done it.
