An AI hangover at work? What the new disillusionment numbers actually prove

What's going on

Since late October, headlines have been piling up that sound like a hangover after the AI binge. Atlassian reports that far fewer developers are seeing big time savings from AI. Assessment vendor SHL has coined the term "AI Rebounding": nearly one in five US workers replaced by AI were supposedly brought back later. Goldman Sachs is taking AI coding tools away from traders and bankers. Microsoft is offering buyouts to up to 60,000 employees.

So far, these stories have only been reported individually. We put them side by side and checked three things for every number: Who collected it? How was the question asked? What was actually measured, and what was merely estimated? Our thesis: The time saved hasn't disappeared; it's flowing into reviewing and fixing. Hard evidence of a genuine step backward is still missing. And "AI Rebounding" is mostly a marketing term with a kernel of truth.

Atlassian: The plunge from 68 to 26 percent is mostly a methodology problem

According to the State of Developer Experience 2026, only 26 percent of developers say AI saves them ten or more hours a week. Atlassian's own blog says 25 percent; a year earlier, it was 68 percent. At the same time, 49 percent spend at least four hours a week managing AI tools. 65 percent fix AI output that missed the actual goal. Among managers, the figures are 57 and 72 percent.

The methodology is thin:

• Different sample: In 2025, Atlassian and research firm Wakefield Research surveyed 3,500 developers and managers in six countries. In 2026, it was just 1,560 people: 1,155 developers and 405 managers. According to InfoQ, neither edition states a fieldwork period. We couldn't determine from the published materials whether the question was worded identically.
• Self-reports only: Atlassian itself says the results reflect "self-reported experience and perception," not telemetry, meaning automatically collected system data such as cycle times.
• Conflict of interest: Atlassian sells developer tools. A banner above the blog post about the report advertises its own AI assistant, Rovo.

The most striking thing is a contradiction inside the report itself. Despite the collapse in time savings, 87 percent say AI helps them ship faster, and 80 percent say they get more done without working longer hours. Our take: It's more plausible that the 68 percent figure from 2025 was inflated than that productivity collapsed. The report itself offers a clue: 60 percent say AI usage counts as a productivity signal at their company, and 41 percent admit they sometimes use AI more than necessary because it's expected.

SHL and "AI Rebounding": The number doesn't match reality

In June and July 2026, SHL had research firm Propeller Insights survey 2,036 working adults online: 1,029 in the US and 1,007 in the UK. According to Fortune, 19.1 percent of US respondents said they had personally been replaced by AI and then asked to come back. Across both countries, the figure was 16 percent; among 18- to 24-year-olds, 24 percent. Of those brought back, 51 percent wanted significantly higher pay and 42 percent wanted written job guarantees.

The press release gives neither the question wording nor any weighting. It does, however, promote SHL's own "talent mobility" product. Other reports render the question as "laid off or replaced by AI." What exactly was asked remains unclear.

Running the numbers: 19 percent of 1,029 respondents is about 197 people. Extrapolated to the roughly 160 million employed people in the US, that would be about 30 million people replaced and rehired. By contrast, outplacement firm Challenger, Gray & Christmas has counted a total of 193,226 announced job cuts attributed to AI since it started tracking in 2023, including 121,401 in 2026. Even if Challenger only captures announced cuts and misses a lot, the gap is about two orders of magnitude. There are two more biases: only people who currently have jobs were surveyed. And the answers to the follow-up questions rest on just a few hundred people.

The kernel of truth: employer surveys find similar patterns. According to Robert Half, 32 percent of roughly 2,000 US hiring managers say a role eliminated because of AI has been refilled. Reports disagree on whether the 32 percent refers to all respondents or only to companies that made AI-driven cuts. The most common reasons given are missing institutional knowledge (40 percent) and more oversight and quality-control work than expected (38 percent). Careerminds surveyed 600 HR professionals in February and reached similar conclusions. However, all three sources sell HR-related services: assessments, staffing, outplacement.

Goldman and Microsoft: Corporate decisions, not measurements

Goldman Sachs revised an internal policy in October, according to Bloomberg. Claude Code and GitHub tools are now available only to its roughly 12,000 engineers and "strats," specialists who sit between trading, quant analysis and software development. A spokesperson cited "productivity and risk management" as the reason. The in-house AI assistant, which combines models from OpenAI, Anthropic and Google, remains available to everyone. As recently as July, CIO Marco Argenti had praised "vibe coding" by non-programmers, meaning software you describe to the AI in plain language. Gigazine suspects security and cost concerns and points to Simon Willison's "lethal trifecta": access to private data, exposure to untrusted content and the ability to communicate externally. Goldman hasn't confirmed this. Our take: This is a decision about who gets to run agents that execute code autonomously on a network holding client data. It's not proof that AI doesn't work.

Microsoft has, since October 26, been offering a "Voluntary Exit" with severance to up to 60,000 employees up to and including level 70 in the US, Canada, the UK and India. The decision window runs until early December, with departures effective December 31. According to Fortune, Microsoft expects fewer than 5 percent to accept. That would be under 3,000 of roughly 223,000 employees. Microsoft says AI isn't driving the program. The numbers still show a shift: in the fiscal year ending in June, capital expenditures rose from $64.6 billion to $115.9 billion, while headcount shrank by about 5,000. Nadella's "20 million agents" comes, according to Windows Latest, from a June podcast and was framed as a thought experiment. An open letter from "concerned Microsoft employees" now calls the work "joyless and demoralizing." Among other things, it demands that AI usage no longer be tied to pay and performance reviews. That fits with Atlassian's 41 percent who use AI out of pressure to meet expectations.

What independent measurements show

In 2025, METR had experienced open-source developers complete tasks randomly assigned to be done with or without AI. That's a randomized controlled trial, the gold standard for establishing cause and effect. With AI, they took 19 percent longer but believed they were faster. According to METR, self-reports overestimate the effect by more than 40 percentage points. In the follow-up study with 57 developers, the estimates came in at 18 and 4 percent speedups. Both confidence intervals, the ranges in which the true value most likely falls, include zero. METR itself considers the data a "poor proxy": many developers no longer wanted to work without AI and held back tasks. The true effect is probably higher, it says.

Telemetry data from vendor Faros AI shows where the time goes. In Network Perspective's summary, high AI usage raised merged code changes by 98 percent, but code review time by 91 percent. Company-level metrics stayed flat.

The Boston Fed's working paper from September analyzes New York Fed data. Among workers who use AI, only 6 percent strongly agree that they've felt more productive since AI was introduced. 47 percent strongly disagree. Worry about losing their job to AI doubled from about 5 to just over 10 percent.

Stanford's analysis of payroll data from payroll processor ADP finds "no broad, economy-wide displacement." But among 22- to 25-year-olds in highly AI-exposed occupations, employment is 19 percent below the expected level. That's mainly due to fewer hires, not layoffs. This, too, argues against mass rehiring: you can't bring back someone you never hired. In Germany, according to the IAB, only one in four companies used generative AI in 2025, up from 5 percent in 2023.

The bottom line

Our take: these numbers don't prove an "AI hangover." It remains a narrative that sounds good. The most eye-catching figures come from self-reports published by vendors with their own business interests. Atlassian also changed its sample, and SHL's question wording is unknown. What is well documented is a shift: the bottleneck is moving from writing to reviewing. Atlassian's own numbers on fixing output, Faros' longer reviews and Robert Half's "more oversight than expected" all point in the same direction. If you roll out AI without planning for the review step, you're just shifting the work elsewhere.

How big the net effect really is remains an open question. Even METR can't measure it reliably right now. Goldman's restrictions and Microsoft's buyouts are corporate decisions, not measurements. For you as a reader, that means: the next time you see a headline with a percentage, it's worth asking who did the asking, whom they asked, and what they're selling.