The number that gave me pause: 55 percent of employers now say they regret laying off people to make room for AI. That figure — from a Forrester survey — marks the inflection point where early enthusiasm starts running into the actual complexity of work.
What went wrong? The honest answer is a category error that is easy to make and hard to see until it is too late.
Tasks are not jobs
AI agents — software programs that can take a sequence of actions autonomously — are genuinely remarkable at discrete tasks. Give a capable model a well-scoped problem with all the relevant information, and it will outperform most humans for speed and often for accuracy. Research from OpenAI showed this clearly: the same models that look mediocre in real-world deployment approach expert quality when every piece of context is laid out upfront. Same models. Radically different results.
The difference is not the model. The difference is the context.
A task, by definition, comes with its context already provided. A job requires you to bring your own context — the unwritten rules, the relationship history, decisions made three years ago that nobody documented, the things everybody knows but nobody ever wrote down. That is the task-versus-job gap, and it is where the regret lives.
Klarna cut roughly 700 customer service positions to AI, then quietly began rehiring. The CEO described the reversal as offering customers a “VIP experience” — which is a polished way of saying that the AI handled transactional queries fine and fell apart the moment real situations required contextual judgment. The agent knew what to do at each step. It had no idea which steps mattered in which context.
The benchmark that should worry every team
A study across 100 real production codebases found that 75 percent of the best available AI models broke previously working features while performing routine software maintenance. Not because they wrote bad code — because they had no way to know which parts of the system were load-bearing, which changes would cascade, which conventions were unwritten law.
Every individual action was locally correct. The organizational context was missing.
This is not a problem you solve by making the model smarter. The contextual knowledge that makes a job coherent lives in people’s heads, in email threads, in the reasoning behind decisions that got made and then never written down. When that knowledge leaves the organization, the agents are flying blind in exactly the situations where it matters most.
The disaster pattern repeats across domains. Legal agents miss unwritten relationship terms. Marketing agents reopen brand wounds no one thought to document. A particularly stark case: an agent tasked with infrastructure cleanup destroyed 1.9 million rows of student data plus all backups — every action was technically correct, but the agent had no way to distinguish real infrastructure from temporary scaffolding. Twenty-four hours and emergency support later, the data was recovered. The lesson was not that the agent was poorly built. The organizational context simply was not there.
What the human role actually is now
A Harvard study tracking 62 million workers over a decade gives us the clearest signal yet about what is actually happening. At companies adopting AI, junior employment dropped by about 10 percent — driven by slower hiring, not mass layoffs. Senior employment kept rising. The most plausible reading: AI replaces task execution; what survives is the ability to provide context, exercise organizational judgment, and know what “correct” looks like in a specific situation.
That is contextual stewardship — holding institutional knowledge and representing it to machines. It sounds abstract, but it translates into concrete practices.
Document the why, not just the what. When a decision gets made, capture the constraints, the alternatives you rejected, and what would make the decision wrong in hindsight. An agent acting on that decision six months later needs the reasoning, not just the outcome.
Write evaluations. Evaluations — automated checks that encode what “good” looks like for a specific workflow — are the highest-leverage investment most organizations are skipping. And the people who should be writing them are the senior people who actually know what right looks like in your specific situation. This does not require coding. A legal team can check for known relationship terms; a finance team can flag projections that contradict board commitments.
Map where contextual knowledge lives before something breaks. Most organizations have no idea how much invisible context is carrying real weight until a disaster surfaces it.
Where this is heading
Gartner estimates that by 2027, half of the companies that cut staff for AI will rehire for similar functions under different titles. That is not an argument against agents — they are genuinely powerful and improving fast. It is an argument for being precise about what we are giving up when we remove the people who carry context.
The organizations that come out ahead will be the ones that treat contextual knowledge as infrastructure — load-bearing and worth engineering around — rather than as the soft, unteachable stuff that gets squeezed out when the spreadsheet says AI can do the job.
The task, yes. The job is a different story, and the gap between the two is widening faster than most people realize.
Text summarized and optimized using Anthropic’s models and reviewed by a human.