Klarna's CEO said last month they "went too far" with AI customer service and started rehiring humans. IBM cut 8,000 HR roles into a system called AskHR, then discovered the AI couldn't handle empathy or judgment or any case that didn't fit the script, and started rehiring. Salesforce cut around 4,000 support roles and by December was regretting the call. Forrester says 55% of employers regret laying off workers for AI. Gartner predicts half of all AI-attributed layoffs will reverse by 2027. Microsoft is cancelling Claude Code licenses across one of its largest divisions by June 30 because the token bill blew up at enterprise scale. Uber burned through its entire 2026 AI budget, $3.4 billion, in four months.
The headlines write themselves. AI was oversold. The cost doesn't justify the output. The autonomous-agent future is further away than the vendors promised.
I don't disagree with any of the receipts. I disagree with the conclusion.
The companies walking it back didn't fail at AI. They failed at "replace a labor unit with an autonomous agent." Those are not the same thing.
What I'm actually building, with no technical background
In the last twelve weeks I shipped a real-time call coaching platform for regulated sales teams. It listens to live phone calls, transcribes both sides, runs the rep's words against a compliance script and a Chris Voss negotiation framework, and surfaces specific coaching to the rep in under five seconds. After the call it auto-generates a summary, captures the outcome, and emails the manager a recap of every call their team ran that day. It runs across real customer teams right now. MIT's NANDA report says 95% of enterprise AI pilots fail to deliver measurable ROI. This one delivers it today, on real calls, for real money, in a regulated industry where being wrong has six-figure consequences.
I did not replace anyone with it. I built it to multiply the output of people who were already doing the work. The reps still take the calls. The managers still coach. The compliance officers still own compliance. The product makes each of them roughly five to ten times more productive at the part of their job AI can do well, while leaving them in the seat for the part AI cannot.
That is the design pattern the walkback stories missed.
What actually broke
The autonomy assumption broke. The vendors sold "AI replaces this team," the customer bought "AI replaces this team," and what got delivered was an autonomous system trying to BE the team. Then the autonomous system did what autonomous systems still do in May 2026. It hallucinated. It forgot the context that was load-bearing for the previous decision. It confidently completed a path that should have stopped. It produced output that looked right and was wrong in a way no one caught until a customer complained.
I will give you a concrete example from this week. I was working with my AI assistant on a piece of analysis. In one session, we used a small batch of placeholder data to test the shape of the model. The placeholder was clearly labeled, both of us knew it was placeholder, the work was correct in context. In the next session, hours later, the same assistant pulled in the placeholder data and started reasoning about it as if it were real. I caught it because the numbers didn't smell right. If I had not been paying attention, the next twenty minutes of work would have been built on a fabricated foundation, and the output would have looked confident, internally consistent, and completely wrong.
This is not a bug to be fixed in the next model release. This is the state of AI memory across sessions in 2026. If you hand an autonomous AI agent a workflow and walk away, this is the class of failure you are accepting. It is exactly the failure mode every company that did mass AI layoffs ran into. Not because the AI was bad. Because nobody was watching when it leaked context between turns.
The part the walkback narrative doesn't want you to hear
If your company could afford the headcount before the AI hype cycle, you have already lost the battle if your response was layoffs.
You already had the budget. You already had the people who knew the work. The leverage was never "fire them and save the salary." The leverage was "give them the tools and watch each one of them do the work of five." As a side effect, they spend more of their day managing processes and outputs and less time grinding through repetitive work, which is the work people actually like doing, which is the work that keeps them at your company.
The companies rehiring right now learned that the expensive way. The companies still pretending AI is a labor-cost line item will learn it next quarter.
What "operator as the architecture" actually means
Every project I have built with AI is held together by rules I wrote, not capabilities the AI shipped. There are rules about when to commit code and when not to. Rules about which decisions get a single recommendation and which decisions get a cross-check across three different AI systems before I lock them in. Rules about flushing memory at trigger points so a session does not forget what the previous session learned. Rules about reading the actual error logs before theorizing about the cause. Rules about checking the database before trusting a dashboard. Rules about every doc I touch, every commit I push, every customer-facing surface I ship.
None of this is the AI being smart. This is the operator doing the work of being the rules of the system. The AI executes inside the rules at superhuman speed. Without the rules, it executes the wrong thing at superhuman speed, which is exactly what the walkback stories are describing.
A non-developer can build a real product like this in twelve weeks if the rules are in place. A Fortune 500 with thousands of engineers cannot deliver an autonomous-AI replacement for a team without the rules. The bottleneck is the discipline, not the technology.
If you're reading the rehiring news and concluding AI doesn't work
I would gently suggest the opposite.
AI works. The pattern of using it to autonomously replace humans does not, yet, at any scale that matters. The pattern of using it to multiply the output of the humans you already pay does, today, in production, on real customer calls, for real money, in regulated industries with real liability.
If you laid off your team and tried to replace them, you are right to walk it back. Hire them back. Then give them the tools and watch what they build with them.
~
This article first ran on LinkedIn. Comments are open there. Back to all writing.