The AI conversation has split into camps.
The hype camp is loud and simple: every new demo is the end. Engineers are done. SaaS is done. “AI will do everything humans can do.”
The denier camp is also loud, just in the opposite direction: it’s a bubble, it hallucinates, it can’t reason, it’s too expensive, regulation will kill it, it’s just autocomplete with better marketing.
Both camps are reacting to real things. They’re just reacting in extremes.
The camp that matters most is the one that doesn’t have time for ideology: the production camp. The teams using AI inside real workflows — codebases, customer support, finance ops, legal drafts, internal tools. They see the wins. They also see where it breaks.
And once you look from production, the story stops being “everything ends” or “nothing changes.”
It becomes a more practical question.
The question everyone argues about is: “Will AI replace humans?” but the question that matters in real life is: “Where does responsibility land when it makes a bad call?”
That question is the difference between extinction and evolution.
Why the deniers aren’t wrong — they’re just incomplete
Skeptics have real points.
AI can be confidently wrong. It can miss edge cases. It can generate code that looks clean and still causes real security issues. It can write a plausible explanation that doesn’t match reality. And in many companies, it’s still hard to measure what value it’s creating beyond “people feel faster.”
They also notice something else: incentives.
A lot of organizations are “AI-washing” — using AI as a label for decisions that were mostly about budgets, margins, or restructuring. “We cut the team because AI” sounds futuristic. “We cut the team because we overspent” sounds less heroic.
The denier camp also leans on a classic pattern: new tech shows up everywhere, but the big productivity numbers don’t immediately reflect a revolution. That gap has a name economists love to reuse.
They’re not crazy for saying, “Show me the hard proof.”
But denial breaks when it treats AI as binary: either flawless replacement or useless toy.
That’s not how technology actually changes work.
Most tech waves don’t delete jobs overnight. They delete chunks of work. They shift what’s valuable. They raise the baseline. They change what “good” looks like.
That’s what’s happening now.
The hype camp measures the wrong thing
The hype camp has one core mistake: it measures the future in demos.
It sees AI write code and assumes it can run engineering. It sees AI draft a contract and assumes it can carry legal liability. It sees an agent move through a workflow and assumes it can own the outcome.
A job isn’t a demo.
A job is a system: reliability, reviews, exceptions, audits, monitoring, customer fallout, legal exposure, security, and the messy stuff nobody puts in the keynote.
So if you want a clear framework for what changes and what doesn’t, don’t start with “intelligence.”
Start with stakes.
The autonomy ladder
Work isn’t one bucket. It’s a ladder.
At the bottom rung, AI is basically free leverage. Autocomplete. Drafting. Quick transformations. If it’s wrong, you hit backspace.
One rung up, AI becomes a serious tool. Boilerplate code. Test scaffolding. First-pass docs. Ticket summaries. Still manageable, because humans can review and verify.
Then you hit the top rung, where “mostly right” can still ruin your week.
Payments, payroll, credit decisions, pricing changes, access control, regulated compliance workflows, legal review, healthcare steps, hiring filters — anything that can hurt customers, trigger fines, or create liability.
This is where most “AI replaces humans” arguments quietly cheat. They pretend the ladder is one rung.
It’s not.
The gap between “helpful tool” and “autonomous decision-maker” is not closing just because the model got better at writing fluent text.
Because the top rung isn’t mainly about output.
It’s about responsibility.
Even if AI can decide… who is accountable?
This is the part that should be obvious, but rarely gets discussed.
In real life, high-stakes decisions come with a human attached.
A doctor makes a call with their license on the line. A lawyer signs off with malpractice risk on the line. A pilot acts with certification (and lives) on the line. A manager approves a plan and owns the outcome. A finance leader signs a filing and carries legal and reputational exposure.
That pressure isn’t bad. It’s the safety mechanism. It forces care. It forces judgment. It creates a clear chain from decision -> responsibility -> consequence.
Now put an AI agent in that seat.
Even if it makes a decision that looks correct, who takes the blame when it’s wrong? Who gets audited? Who gets sued? Who can explain the reasoning in a way regulators accept?
Not the model.
So the responsibility snaps back to humans anyway. Someone still has to approve. Someone still has to sign. Someone still has to own the outcome.
That’s why “full autonomy” keeps turning into something more boring, more realistic, and more durable:
AI can draft. Humans must own.
You can delegate typing. You can’t delegate blame.
2026 is where the “bubble vs revolution” fight gets real data
This is where the denier camp has been scoring points: a lot of companies can’t show dramatic, measurable impact yet.
A major National Bureau of Economic Research (NBER) working paper based on surveys of nearly 6,000 executives found that most firms reported no meaningful impact on productivity or employment from AI so far, even with broad adoption.
That’s why people keep bringing up the modern version of the “productivity paradox” (often called the Solow paradox): the technology is everywhere — in tools and talk and decks — but the big aggregate productivity gains are not showing up cleanly yet.
At the same time, Gartner put out a prediction that cuts through the hype in a different way: by 2027, half of the companies that reduced headcount and attributed it to AI will rehire people to do similar work (often under new titles).
Put those together and a more realistic picture appears:
- AI is spreading fast.
- Impact is real but uneven, and often modest in the short term.
- Some “AI-driven layoffs” are likely to reverse because replacing a role end-to-end is harder than speeding up parts of it.
That’s not extinction. That’s a messy transition.
Developers are using AI. Trust is still the limiter.
Production data in engineering tells a similar story: lots of adoption, measurable time saved, mixed outcomes.
DX’s AI-assisted engineering report (Q4 2025) describes very high adoption across a large developer sample, with reported time savings measured in hours per week — not days — and quality impact varying by organization.
Stack Overflow’s 2025 developer survey adds the missing layer: trust. More developers actively distrust AI output (46%) than trust it (33%), and only about 3% say they highly trust it. Experienced developers are the most cautious.
That experienced skepticism matters, because it points to the real bottleneck. It’s not whether AI can generate output.
It’s whether teams can rely on it without increasing incidents, risk, and cleanup work.
Same tools, wildly different outcomes: it’s the org, not the model
One of the most important “production truths” is also the least viral:
AI doesn’t deliver the same results everywhere.
DX emphasizes that quality impact varies across companies — and that “average” stories hide extreme differences.
That’s exactly what you’d expect if AI is a force multiplier: it makes strong teams faster, and it exposes weak process faster. A team with clear specs, good tests, disciplined reviews, and strong ops gets leverage. A team without those things gets higher velocity and higher blast radius.
This is why “just roll out AI” often disappoints leadership. AI adoption is not a strategy. It’s a stress test.
If “replacement” were here, agents would finish the whole job
There’s a simple way to cut through hype: test AI on end-to-end work, not toy prompts.
The Remote Labor Index (RLI) was built to measure exactly that: economically valuable remote-work projects evaluated as full tasks. The headline result is sobering: across frontier agent frameworks, the highest-performing system achieved only about a 2.5% automation rate on RLI tasks.
That doesn’t mean AI is useless. It means the “agents replace whole roles next quarter” story is out of sync with measured end-to-end performance.
Again: evolution, not extinction.
Replacement attempts are already revealing the limits
Nothing clarifies this faster than companies that actually tried replacing humans and then had to live with the consequences.
Klarna is a well-known example: after touting AI for customer service at the scale of hundreds of roles, the company later moved back toward human hiring and involvement as service quality issues became clear.
Meanwhile, other big companies are making the opposite bet: they’re investing in humans even while adopting AI.
IBM has said it will triple entry-level hiring for roles many people claim AI can do. Their argument is straightforward: killing the junior pipeline creates a leadership vacuum later. No juniors today means not enough seniors tomorrow.
Dropbox has also talked publicly about expanding early-career programs, explicitly pointing to younger workers being unusually fluent with AI tools.
And Anthropic — the company people like to point to as a “replacement engine” — ended up publishing a much more grounded story when it looked inward.
In Aug 2025, they studied their own team using a survey of 132 engineers/researchers, 53 interviews, and internal Claude Code usage data.What changed first was where AI helps. The most common use wasn’t flashy autonomy, it was very practical: debugging and understanding unfamiliar parts of a codebase. People said Claude showed up in roughly ~59% of their work.
But the most telling stat wasn’t “AI replaced roles.” It was scope expansion. Employees estimated about ~27% of their Claude-assisted work wouldn’t have been done otherwise — the long tail of improvements teams usually postpone: cleanup, internal tools, docs, tests, quality-of-life fixes.
Then comes the autonomy reality check. Most people said they could fully delegate only 0–20% of their work. The default mode was still supervise, verify, revise — especially when the stakes were high. Even as the tools improved (Claude Code completing more actions per run, roughly ~10 → ~20 before needing input), the bottleneck stayed human: review and judgment.
They also surfaced the tradeoffs production teams already feel: concerns about skill atrophy and weaker mentorship/collaboration when “ask Claude” replaces “ask a teammate.”
That’s why the “AI replaces everyone” narrative should be treated carefully. It’s the easiest story to sell to your investors (specially, burning billions in cash). Production teams don’t get to sell stories — they have to ship outcomes. And even inside a frontier AI lab, the steady-state pattern looked like augmentation with guardrails, not “humans out.”
SaaS isn’t dying because agents exist. It’s becoming the system agents rely on.
The “SaaS is dead” story usually goes like this: if agents can do the work, they’ll bypass apps, so the apps disappear.
That’s mixing up “interface” with “infrastructure.”
Agents can skip screens. They can’t skip the things screens sit on top of: the system of record, the permission model, the audit trail, the compliance rules, the retention policies, the integrations, the workflow logic that took years to harden.
If an agent drafts a contract review, where do the contracts live? Who has access? What version is the source of truth? What gets logged for audit? Who can prove what happened when a regulator asks?
That’s SaaS.
Interoperability standards help agents talk to more tools, sure. But they don’t erase workflow gravity. Companies don’t stick with Salesforce, Workday, NetSuite, ServiceNow because they love the UI. They stick because their operating model lives there — approvals, controls, reporting, and history.
So the shift isn’t “SaaS gets replaced.” It’s “SaaS becomes more foundational.” The interface gets thinner, but the governed data layer gets more valuable, because agents need somewhere safe and defensible to act from.
In an agent world, the winners aren’t the apps with the prettiest dashboards. They’re the platforms that can answer the adult questions: who did what, when, why, and who signed off.
That’s evolution, not extinction.
What’s really dying, and what’s growing
Some things are dying.
Work that’s mostly boilerplate. Roles where the main value is producing volume without judgment. Thin SaaS products whose “moat” is basically a dashboard.
But that’s not “software is over.” That’s value moving up the stack.
What’s growing is the part AI can’t carry on its own:
- judgment under uncertainty
- risk tradeoffs
- review discipline
- accountability
- designing guardrails
- knowing when “technically correct” is still the wrong move
That’s why the future doesn’t look like extinction.
It looks like a shift in what counts as valuable.
AI takes more of the routine work. Humans move up the ladder into ownership: deciding what matters, setting constraints, reviewing output, and taking responsibility for outcomes.
Bottom line: evolution beats extinction
Silicon Valley will keep claiming the same thing: “AI will replace humans.” Real companies have to answer a different question: “When it fails, who owns the consequences?”
Because in high-stakes work, responsibility can’t be automated away. Someone has to sign. Someone has to explain. Someone has to carry the risk when “technically correct” still causes real damage.
That doesn’t slow AI down. It just shapes where it lands.
AI won’t erase the workforce overnight. It will compress the boring layers — the boilerplate, the busywork, the first drafts. And it will pull value upward into the parts that can’t be outsourced to a model: judgment, verification, governance, and ownership.
So the winners won’t be the loudest hype merchants or the most comfortable skeptics.
They’ll be the teams in production — using AI hard, shipping faster, and building the guardrails that keep speed from turning into chaos.
The Future of Work Isn’t Extinction. It’s a Shift Up the Ladder. was originally published in Level Up Coding on Medium, where people are continuing the conversation by highlighting and responding to this story.
