At Snowflake, the average engineer now works with three or four AI agents daily. One head of engineering in Toronto spends 20 to 30 hours a week interacting with five of them. They review product designs, assist during outages, and carry out coding projects under human supervision. “You don’t have to bother a human for basic questions any more,” he told Al Jazeera.

That’s interesting, but it’s not the story. The story is what happens next.

In some companies, AI systems aren’t just completing tasks. They’re assigning them. Convictional, a Toronto-based AI software firm, has built a platform that translates executive strategy into daily assignments and delivers them to employees through a user-friendly inbox. The work that supervisors used to do (breaking strategy down into tasks, checking progress, providing feedback) is now handled by the platform. Continuously, automatically, without anyone booking a meeting room.

I’ve written before about how AI is unbundling middle management, splitting coordination from leadership and automating the first while making the second more important. That argument still holds. But something has shifted since I made it. The question is no longer whether AI can replace management coordination. It’s whether the humans on the receiving end will let it.

The prediction that missed

The early consensus was that AI would hit entry-level technical jobs hardest. Junior developers, customer service reps, data entry clerks, the roles closest to the work AI already did well. And there’s evidence for that. New graduate unemployment has climbed, and several analyses tie it to AI adoption eating into the bottom rungs of the career ladder.

But the bigger disruption, as the evidence accumulates, is landing squarely on middle management. Middle managers aren’t easier to replace than junior staff. The difference is that their work is disproportionately coordination. And coordination is what AI agents are getting good at fastest.

Middle managers exist for three reasons: to translate executive strategy into operational tasks, to coordinate across teams, and to surface context for decisions. AI agents now do all three. Not perfectly, not in every situation, but consistently enough that the economics have shifted. When Amazon announced 14,000 job cuts last October, executives cited AI’s potential to help the company “operate with fewer layers and greater efficiency.” UPS, Target, and General Motors followed with similar reasoning. January 2026 saw more US layoffs than any January since 2009.

Goldman Sachs estimates 6 to 7% of US workers could lose their jobs to AI adoption, with higher risks for programmers, accountants, legal and administrative assistants. But those headline numbers obscure a structural point: the roles disappearing aren’t the ones doing the visible, measurable work. They’re the ones doing the invisible coordination work that nobody valued until it was gone.

The trust trap

Most companies are getting this wrong, and the reason is psychological, not technical.

Stefano Puntoni, a behavioural scientist at Wharton, has been studying what happens when AI enters the workplace power dynamic. His research shows something counterintuitive: employees are often more willing to delegate tasks to AI than to colleagues. “There’s no social cost,” he says. “You don’t worry about burdening an AI.”

On the surface, that looks good for AI adoption. But the same research reveals a deeper problem. Generative AI threatens employees’ sense of competence, autonomy, and connection. Those aren’t soft metrics. They’re the foundations of why people show up and try. If workers feel threatened by AI, they may want the system to fail. And at scale, as Puntoni puts it, “that guarantees failure.”

This is the trust trap. Companies deploying AI as a management replacement face a contradiction: the technology works best when workers cooperate with it, but the way most companies frame AI deployment (efficiency gains, headcount reduction, doing more with less) actively destroys the cooperation needed for AI to deliver on those promises.

The layoff announcements make this concrete. When Pinterest and HP cite “AI initiatives” as part of the rationale for job cuts, every remaining employee recalibrates their relationship with the AI tools they’re being asked to use. The tools become a threat, not a collaborator. And a workforce convinced the tools exist to replace them will find a thousand small ways to ensure it doesn’t work well enough to prove them right.

The companies that get this

A Harvard Business Review survey of over 1,000 global executives found that 60% of organisations have already reduced headcount in anticipation of AI’s future impact, but only 2% have made layoffs tied to actual AI implementation. The cuts are running ahead of the capability. PwC’s Global CEO Survey backs this up: 56% of CEOs report no revenue or cost benefits from AI yet.

So the job losses are outpacing the actual technology. And the trust implications are serious. You can’t mass-fire people in the name of automation, then ask the survivors to enthusiastically adopt the tools that justified their colleagues’ redundancies. The math doesn’t add up psychologically, even if it looks tidy on a spreadsheet.

Convictional’s founder Roger Kirkness has seen this firsthand. His company builds the AI management tools that could, in theory, justify significant headcount reductions. Instead, Convictional adopted a four-day workweek, framing AI-driven productivity gains as something shared with employees rather than extracted from them.

“Mass layoffs in the name of automation destroy trust,” Kirkness told Al Jazeera. And he’s right. Layoffs aren’t always wrong, but the framing determines whether the remaining workforce cooperates with or resists the AI systems that are supposed to deliver the productivity gains.

The distinction matters enormously. A company that says “AI lets us work four days instead of five” is telling employees that the technology serves them. A company that says “AI lets us cut 20% of the workforce” is telling the surviving 80% that they’re next. Both companies might have identical AI deployments. Only one of them will work.

What management becomes

New Ground Wellness, a Canadian counselling firm, recently turned down a CA$20,000 proposal for an agentic AI intake system that would match therapists with clients. The technology would probably work fine. But their callers told them it would damage trust. Multiple surveys show strong consumer preference for human customer service workers, and in a field built on therapeutic relationships, that preference isn’t a market quirk. It’s the product.

That example sits at one end of a spectrum. At the other end, Snowflake engineers are happily working alongside AI agents because the agents handle the tedious parts (scanning dashboards, chasing data) while the humans focus on higher-level decisions. Nobody at Snowflake is losing their job to the agents. The agents are making the existing jobs less miserable.

The companies getting this right aren’t asking “how many managers can we replace with AI?” They’re asking “what does management look like when coordination is automated?” Those are different questions with very different answers. I’ve sketched one answer in the 50:1 AI org chart: the managers who remain become orchestrators, not supervisors. Part of that shift is making your own expectations legible, which is why I think every manager needs an operating manual.

I wrote about the reskilling myth recently, how the standard “retrain displaced workers” narrative falls apart under scrutiny. The middle management squeeze is the other side of that coin. We’re not just losing jobs at the bottom of the ladder. We’re hollowing out the middle: the layer that translates strategy into execution, that provides context and coaching, that catches problems before they scale.

Some of that work’s genuinely automatable, and automating it makes sense. But some of it’s deeply human: the judgement calls, the relationship management, the ability to read a room and know when a number on a dashboard doesn’t tell the full story. The companies that mistake the second category for the first will discover, expensively, that AI coordination without human stewardship produces fast, confident, wrong decisions at scale.

The question nobody wants to answer

Neither the US nor Canada has clearly defined rules governing AI agents in the workplace. The EU AI Act classifies certain employment-related AI uses as high-risk. The UK’s ICO guidance is clear about rights around solely automated decisions with significant effects. But in North America, regulation is lagging well behind deployment.

That vacuum won’t last. When an employee at a multinational firm wonders whether the online “coach” supporting her development is a human or an AI, and requests anonymity because she’s afraid to ask, something has gone wrong with the social contract. Not with the technology. With how it’s been deployed.

The question isn’t whether AI replaces managers. That’s already happening, unevenly and often badly. The real question is what we want management to be when the coordination layer is automated. If the answer is “nothing, just let the agents run it,” then we’d better be comfortable with the trust deficit that creates. Because every piece of evidence we’ve got says that workers who feel managed by machines they don’t trust will resist, quietly and persistently, until the machines fail.

And the machines will get the blame. But the failure will be human.