Is AI Actually Replacing Jobs? What the 2026 Evidence Shows

How many layoffs are actually caused by AI?

Few, on the employers' own statements. Challenger, Gray & Christmas counted 1,206,374 announced US job cuts in 2025, and AI was cited in 54,836 of them, about 4.5 percent. Yale Budget Lab found no discernible labor-market disruption in the 33 months after ChatGPT's release.

The layoff ledger is the closest thing to a direct count. Challenger, Gray & Christmas's year-end 2025 report recorded 1,206,374 announced job cuts, up 58 percent from 761,358 in 2024, with AI cited in 54,836 of them. The leading stated causes were DOGE actions at 293,753, market and economic conditions at 253,206, closings at 191,480, and restructuring at 133,611. These are employer announcements, so the AI attribution is whatever the employer chose to say, and both overstatement and understatement are possible. The same report counted 507,647 announced hiring plans, the lowest year-to-date total since 2010, and the hiring channel is where the real signal in this post shows up.

The academic reads agree on the aggregate. Yale Budget Lab's October 2025 analysis, academic work that is not peer-reviewed, concluded that “the broader labor market has not experienced a discernible disruption since ChatGPT's release 33 months ago,” measuring how fast the occupational mix shifts against the computer and internet eras, as reported by ZME Science. Daron Acemoglu's peer-reviewed macro estimate, published in Economic Policy, caps expected gains at “no more than a 0.66% increase in total factor productivity (TFP) over 10 years,” a ceiling far too low to carry a replace-the-workforce story.

Did Klarna and Duolingo really replace workers with AI?

Both companies claimed it, then partially reversed. Klarna said its assistant did the work of 700 agents in February 2024; fifteen months later its CEO was recruiting humans again and admitting quality had dropped. Duolingo announced an AI-first contractor policy, then its CEO said AI was not replacing employees.

Klarna wrote the canonical claim itself. The February 2024 press release said the company's OpenAI-built assistant handled 2.3 million conversations in its first month, was “doing the equivalent work of 700 full-time agents,” and was “estimated to drive a $40 million USD in profit improvement” in 2024: an estimate, never an audited result. Fifteen months later, CEO Sebastian Siemiatkowski told Bloomberg, as reported by CX Dive: “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.” Klarna began recruiting human agents again while AI kept handling about two-thirds of inquiries.

Duolingo followed the same arc: Luis von Ahn's April 2025 memo said it would “gradually stop using contractors to do work that AI can handle,” and his follow-up weeks later, quoted by Entrepreneur, said “I do not see AI as replacing what our employees do (we are, in fact, continuing to hire at the same speed as before).” The cases still standing carry the same grade problem. IBM CEO Arvind Krishna told The Wall Street Journal that AI replaced several hundred HR roles, yet “Our total employment has actually gone up, because what [AI] does is it gives you more investment to put into other areas.” And Marc Benioff said on The Logan Bartlett Show: “I've reduced it from 9,000 heads to about 5,000, because I need less heads.” Salesforce sells the agents it credits, and no independent audit exists.

Where is AI measurably replacing workers?

At the entry level, through hiring. Stanford's Digital Economy Lab measured employment of workers aged 22 to 25 in the most AI-exposed occupations at 19 percent below the trend of less-exposed peers, on ADP payroll data through June 2026. The gap operates through reduced hiring of young workers, not through firings.

The Stanford Digital Economy Lab's “Canaries in the Coal Mine?” working paper, by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, revised August 12, 2026 and not yet peer-reviewed, tracks high-frequency ADP payroll records covering millions of US workers. Employment of young workers in the most AI-exposed occupations, software development and customer support among them, “now stands 19% below where it would be had it kept pace with that of their less-exposed peers.” The mechanism finding matters most: the decline “operates primarily through reduced hiring of young workers rather than increased separations.”

The gap also lands exactly where substitution logic says it should: “Declines are concentrated in occupations where AI usage primarily substitutes for human tasks; where usage primarily complements workers, employment is flat or rising, especially for experienced workers.” Nobody in the Stanford data was fired by a model. The measurable event is quieter: the junior requisition that never opens. A company can report record headcount and still be substituting, because substitution shows up in the hires not made, which is precisely where announced-layoff counts cannot see it.

Is AI automating jobs or augmenting workers?

Both, split by channel. Anthropic's Economic Index classified 52 percent of Claude.ai conversations as augmentation and 45 percent as automation in November 2025, while automation dominates API traffic at roughly three-quarters. The strongest causal study of AI at work found productivity up 14 percent, with retention rising.

Anthropic's Economic Index report of January 2026, first-party usage telemetry rather than an employment measure, classified Claude.ai conversations in November 2025 as 52 percent augmentation against 45 percent automation, with fully delegated “directive” conversations down seven points to 32 percent. The API side inverts: “automated use remains dominant in 1P API traffic, reflecting its programmatic nature,” at roughly three-quarters of interactions. Programmatic use is where agents live, so the split reads as a leading indicator: people using AI directly lean toward augmentation, and the automation share arrives through the API.

The strongest causal evidence points to augmentation for workers who already hold the job. Brynjolfsson, Li, and Raymond's staggered rollout study of a generative AI assistant across 5,179 customer support agents, an NBER working paper, found productivity up “14% on average,” a 34 percent improvement for novice and low-skilled workers, and that “AI assistance improves customer sentiment, increases employee retention, and may lead to worker learning.” That is the same customer support job family Klarna and Salesforce cite for substitution. Where the measurement is rigorous, the workers got better and stayed longer.

Will AI replace jobs eventually?

Unknown; the measured trend is recomposition, not disappearance. Indeed found software postings up almost 15 percent in the year after Claude Code launched, with 71 percent of the increase in senior roles. My own expectation is that substitution keeps concentrating on the junior requisition rather than the existing employee.

Demand is recomposing rather than vanishing. Indeed Hiring Lab measured that “US software development job postings have grown by almost 15% since the launch of Claude Code in late February, 2025, while overall job postings fell by 7% over the same period,” with the stated caveat that “correlation does not imply causation.” Composition carries the message: 71 percent of the increase was senior roles, and software postings remain about 27.5 percent below their pre-pandemic level. Demand came back. The junior end of it largely did not.

Policy is moving ahead of measurement. Tobi Lütke's April 2025 Shopify memo, reported by TechCrunch, made not-hiring the default: “Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.” My position, a founder's prediction rather than a measured fact, is that this gate becomes the norm as agent capability improves: substitution keeps operating through hiring, teams of 45 to 60 people keep holding milestones that once took hundreds, and the management flattening on the record at Meta, Amazon, Google, and UPS continues. The rest of the agentic organization series argues that end-state with the counter-evidence attached, as this post does.

What should a lean team plan for if hiring stays flat?

Coordination load. A team that holds headcount flat while agents absorb execution ends up supervising more parallel workstreams per person, whether the team is a seed-stage startup or a pod inside a large company. The evidence in this post says the jobs mostly remain; the default new hire does not.

A lean team on the flat-headcount path ends up with each person supervising several parallel streams of agent and human work at once. Middle management existed to buy coordination, and the teams in question either never hired that layer or are watching their company remove it. So the binding constraint stops being execution capacity and becomes the shared plan: keeping one true picture of the work while agents produce output faster than a weekly meeting can absorb.

That constraint is the one Nazr is built against, for a five-person startup and for a pod inside a 200,000-person incumbent alike. Nazr is a system of record for teams shipping with agents: one canonical plan, with calls, commits, and agent reports arriving as signals from the tools a team already runs, changes landing as accept-or-dismiss proposals with provenance, and approved work routed to agents with commits linked back as proof. The Jira comparison sets out where that sits relative to an issue tracker: one layer up, where the plan lives.

The verified record as of August 2026: AI was cited in about 4.5 percent of announced 2025 US layoffs, the two marquee replacement claims were partially reversed within about fifteen months, and the one substitution signal that survives scrutiny is a 19 percent entry-level employment gap that operates through hiring, not firing.