“I was in a conference in Madrid in May, it was the biggest SAP conference. And I had my keynote and I’m like, ‘Raise your hand if you’re using AI in a meaningful way.’ Now, our survey said 8%. So I expected 8% to raise hands—two people raised hands. Two people,” Adika told Fortune. “That’s the reality.”

That anecdote cuts against the reassuring half of a new survey WalkMe is releasing Tuesday. The company’s third annual AI at Work Pulse Survey, conducted by Propeller Insights among 2,037 working U.S. adults, finds employees are lying less about their AI skills than they were a year ago. The share who admit to pretending to know AI in a meeting fell from 45.2% in 2025 to 28.3% in 2026. The share who admit to passing off AI-generated work as their own dropped from 48.7% to 32.5%.

But the survey—and Adika’s own account of selling AI software into enterprises all day—describes a workforce that hasn’t so much improved as it has simply stopped policing itself. Ninety percent of workers say they feel confident using AI. Only 24.6% say it works on the first try. Half say they’ve spent more time trying to get AI to do a task than the task would have taken manually.

“Workers are not becoming more deceptive; they are becoming more comfortable,” the release states, which is another way of saying the lying didn’t go away so much as it became unnecessary. Nobody needs to fake competence they sincerely believe they already have. Adika explained to Fortune why this growing confidence often blends into overconfidence.

The math doesn’t back up the confidence

Adika’s read is this confidence is mathematically indefensible once you look at a company’s P&L instead of its employee surveys.

“If you have 50,000 employees, so you should save 50,000 hours, let’s call it a week. So you should have saved 200,000 hours a month. Where are the $4 [million] or $5 million in savings?” Adika asked. By and large, he added, “it is not there. It’s not translating to actual P&L savings.” On where the AI adoption story stands, he said: “People feel that it saves time… but they don’t think they can correlate it to actual business results.”

Part of the disconnect, he argued, is a double standard baked into how people judge AI’s mistakes versus their own. “Everybody expects AI to be perfect,” he said. “If AI is wrong 1%, while a human being is wrong 10%, all the focus would be, wow, the AI got it wrong—but he got it wrong 1% while everybody else gets it 10%. There is a big resistance.”

That resistance persists even though the confidence gap runs all the way to the top: 53.6% of employees say they’ve felt a senior leader doesn’t fully understand the AI strategy that leader is publicly championing.

“Everyone, from the newest hire to the executive suite, is learning AI in real time,” said WalkMe’s Global Field CTO KJ Kusch.

The knowledge workers are handing over

Adika argued this overconfidence is especially dangerous because of what companies are actually asking employees to do with AI right now: pour their own expertise into it. He calls the resulting system a “company brain”—a shared memory layer meant to let AI agents operate the way a well-trained employee would. Building it, in his telling, is not a neutral technical upgrade.

“When you’re building the company brain, you’re basically putting a sword on your neck. That’s what you’re doing as an employee,” he said.

The mechanics of that threat, as he described them, are straightforward. A manager who once needed a team of 10 people to execute decisions can, once an AI system is trained on that team’s collective knowledge, get by with two people making calls and one person checking the AI’s work.

“So now you can shrink the team from 10 to three,” Adika said. That leaves the employees being asked to train the system in a bind: The more thoroughly they teach an AI agent to do their job, the less essential they become. “It means that they take all their knowledge, they move it to the AI. Now the AI can do it instead of them. Now they might fire them, right? So it’s a catch,” he said.

That’s the same overconfidence problem in a different light. A workforce that believes it has mastered AI is a workforce likely to hand over its knowledge without fully reckoning with what it’s trading away—and Adika suggested that gap in understanding, more than any deliberate corporate scheme, is what’s currently unsettling employees the most.

“No one has a good answer for that,” he said of where it leaves individual workers.

Why more training won’t fix this

The standard corporate response to shaky AI results has been more training. Adika thinks that response is backwards, because overconfidence is exactly what makes training ineffective—nobody signs up for a class on a skill they’re already certain they’ve mastered.

“You need to train AI like you train a human,” Adika said. “When I hire someone, it takes two to three months to onboard that person… Same goes with the AI. Maybe it’s a bit faster, but people don’t give AI the same attention.”

That tracks with what WalkMe’s own respondents said would actually help: standalone training courses ranked behind better integration between AI and the apps people already use (33.7%) and guidance built directly into those tools (30.2%). Adika’s most concrete explanation for the gap has less to do with psychology than plumbing—AI tools that work perfectly in a sales demo and then hit a wall the moment they touch a company’s actual permissions structure.

“I go, I see a demo. They’re showing me this bot. I can say, hey, how much is [an employee’s] salary, I want to give her a raise, request a spot bonus. The bot is doing that. Everything is great,” he said. “Then my chief security officer says, wait a minute—no one can access salaries that way.”

The promise of one AI assistant collapses back into a maze of disconnected tools requiring separate logins for each task—at which point, he said, employees quietly give up and do it the old way. Multiply that across every department and every vendor’s own AI studio—Microsoft’s Copilot, Salesforce’s Agentforce, SAP’s Joule—and you get what he called “mega chaos.” None of that friction shows up in a confidence survey. It just shows up later, as a missed deadline or a bad decision someone quietly blames on themselves.

The generation with the most to lose

The clearest evidence that confidence and competence are drifting apart shows up in the generational data. Gen Z is the most confident cohort using AI, at 94.1%—and also the most likely to have overstated its skills, with 45% admitting to pretending to be more skilled than they actually are, compared with 13% of baby boomers. That overstatement wasn’t harmless: 31% of Gen Z workers say it caused a real workplace problem—a mistake, a missed deadline, a bad decision, lost trust—versus 7% of baby boomers.

Adika, who runs an AI agent trained on his own work that he says now outperforms him on certain tasks, put the stakes for junior, more automatable roles in the starkest terms of the conversation.

“If I were the developer or if I’m a designer, I would be scared,” he said. “Why do they need me anymore? They can do everything with AI.”

His hope is the bar for human work simply rises—companies building 50 products instead of one—rather than employing fewer people to build the same one.

Asked where this goes next, Adika didn’t reach for hype, which is itself notable coming from an executive whose company sells the fix for the exact gap his own data describes.

“Wow, I wish I knew—I can just guess,” he said. Barring “a mega breakthrough… it would be a little bit better than now, not significant. Just kind of muddling through.”

The workforce, in other words, isn’t lying about AI anymore. It’s just started believing its own press.

For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing.