Thirty years of building enterprise platforms has taught me one thing: the gap between what executives say in earnings calls and what their operations data shows is where the real story lives.
AI replacement is no different.
The headlines say customer service, driving, and coding are over. The METR data, the Gartner survey, and the Klarna reversal say something different. Here's what's actually happening at Spotify, Uber, IBM, Salesforce, Google and Meta.
If you spent the first half of 2026 reading technology news, you came away with a clear story. AI is replacing customer service teams at 50–80% scale. The Uber CEO is preparing to write off 9 million drivers. Spotify's senior engineers have not written a line of code since December. Coding is over. Driving is over. Customer service is over. Pick your white-collar function and start the clock.
The headlines are mostly real. The narrative around them is mostly wrong.
Having spent thirty years building enterprise platforms — and now building an AI-native one — I have watched a familiar gap open between the press release and the production system. Executives say one thing in earnings calls. Their workforce telemetry says another. And the most interesting story in AI right now is not that the replacements are happening. It is that the replacements being announced are quietly being walked back, redirected, or recategorised once the operational data comes in.
Here is what the evidence actually shows.
Spotify: the senior engineers who stopped typing
On Spotify's Q4 2025 earnings call on 12 February 2026, co-CEO Gustav Söderström made the comment that lit up technology Twitter for a week. The company's most senior engineers had not written a single line of code since December. They generated code and supervised it.
It is true. It is also not what most people heard.
What Söderström described was a shift in what engineering labour looks like at the top of the skill curve, not its disappearance. Architecture, system design, evaluation, specification, code review, and judgment about what to build remain entirely human. The senior engineer is now a director and reviewer of AI-generated output rather than the typist of it. This is a significant change. It is not redundancy. The engineers who stopped typing did not stop engineering. They moved up the stack.
This matters because the same pattern is showing up across the industry, and the framing matters for how organisations plan their workforce.
Google, Meta, Microsoft and Snap: the 75% number
At Google Cloud Next in April 2026, Sundar Pichai disclosed that roughly 75% of new code at Google is now AI-generated, with human engineers reviewing and approving every change. Eighteen months earlier the figure was 25%. The trajectory is steep and deliberate. Meta has set internal targets for select engineering teams to generate more than 75% of their committed code using AI tools by mid-2026. Microsoft sits at around 20–30%, with its CTO publicly predicting 95% by 2030. Snap says 65% of its new code is AI-generated.
These are large, verified, on-the-record numbers from CEOs with little incentive to overstate them in regulated disclosures.
Now look at the headcount data behind them. Google is still hiring engineers. Meta is still hiring engineers. Microsoft is still hiring engineers. Salesforce — which I will come back to — has frozen new engineering hiring but has hired 1,000 graduates and interns and is adding roughly 20% more salespeople. The AI-generated code share is rising fast. The engineering org is not collapsing. It is being restructured around a different unit of output, and the demand for senior judgment is rising, not falling.
The METR randomised controlled trial published in July 2025 is the single most rigorous piece of evidence we have on the productivity question, and it cuts against the hype. Experienced open-source developers using frontier AI tools were 19% slower at completing tasks while believing they were 20% faster — a 39-point perception gap. METR's February 2026 follow-up found that the same developers, after sustained habituation, gained roughly 18%. The gains are real. They require 30 to 100 hours of disciplined practice to unlock. And they accrue most strongly to junior developers (around 40% PR productivity uplift in the Demirer et al. 2026 paper) and least to senior developers (around 7%).
Uber: the number that was off by three orders of magnitude
In a March 2026 appearance on The Diary of a CEO, Dara Khosrowshahi spoke candidly about the future of his platform. The line that travelled was "9 million drivers will not have jobs." The number being passed around in some recaps was 9,000.
The number is 9.5 million. The timeline is 15 to 20 years. Khosrowshahi explicitly said autonomous vehicles will "feather in" alongside human drivers for at least a decade. When asked what happens to those 9.5 million people, his answer was "I don't know."
The more interesting moment in the interview was not the driver figure. It was Khosrowshahi accusing his fellow executives of intellectual dishonesty. He said he had heard private conversations about the magnitude of disruption AI would cause, then watched those same executives go on CNBC and at Davos and tell audiences everything would be fine. He acknowledged the incentive: too much public candour about job displacement spooks investors and complicates fundraising.
That admission — that the AI replacement narrative is in part a capital-markets performance — is the single most useful framing for what follows.
Klarna: the case that ate itself
In 2024, Klarna became the canonical example of AI replacing white-collar work at scale. CEO Sebastian Siemiatkowski announced that the company had eliminated approximately 700 customer service positions and replaced them with an OpenAI-built assistant handling two-thirds to three-quarters of all interactions. It was held up as evidence that the AI labour shift had begun.
By mid-2025, Klarna was quietly rehiring. By spring 2026, the reversal was on the record. Siemiatkowski admitted the company had "gone too far," that the focus on efficiency and cost had eroded customer satisfaction, and that human agents were being brought back under a flexible "Uber-style" remote model targeting students and rural workers. Reports from Business Insider, covered by CX Today, indicate Klarna was pulling staff from marketing, engineering, and legal teams to put them on phones.
This is not a failure narrative. It is a recalibration. But it is also a case study in what happens when an enterprise models the AI business case purely on labour replacement and not on what humans were actually doing.
The implicit knowledge that experienced agents carried — edge case handling, escalation judgment, knowing when a routine question is actually a complex problem — was eliminated with the headcount. Rebuilding it cost more than the original savings projected. Klarna's reversal was not a public announcement in the same way the original replacement was. The unwind happened quietly.
Gartner's October 2025 survey of 321 customer service leaders, published February 2026, shows Klarna was not an outlier — it was the cautionary tale. Only 20% of organisations have actually reduced agent staffing because of AI. 55% report stable staffing while handling higher customer volumes. And Gartner forecasts that by 2027, half of all companies that attributed customer service headcount reductions to AI will rehire staff under different titles. Forrester reaches a similar conclusion and gives the practice a useful name: AI washing — attributing economic layoffs to AI for narrative cover.
The "50–80% customer service headcount reduction" headlines come from vendor case studies and from companies practising AI washing. The Gartner data from 321 leaders shows something different.
IBM: where the press release and the WARN filings disagree
IBM is the cleanest example of how the AI replacement story gets manufactured.
In May 2025, CEO Arvind Krishna told The Wall Street Journal that the company had used AI to take over the work of several hundred HR employees, with its AskHR agent automating 94% of routine HR tasks. He also said total IBM employment had gone up, because the savings funded hiring in programming and sales. The accurate number for AI-driven HR replacement at IBM is around 200 positions. The total layoff figure across IBM in 2024–2026, however, runs into the tens of thousands, with the Q4 2025 round alone affecting somewhere between 2,700 and 8,600 people.
What was actually happening was visible in the job postings. American Bazaar reported that IBM's India job listings jumped from 173 in January 2024 to 3,866 by early 2025, while U.S. postings stayed under 400. Median tech salaries in India sit around $22,000 against $150,000 in the U.S. — roughly an 85% cost reduction per role. Some U.S. employees reported being asked to train their replacements in Bengaluru during their notice period.
Salesforce: the honest version
Marc Benioff has been more forthcoming than most. In interviews through Q1 and Q2 of 2026, he disclosed that Salesforce is not hiring additional engineers in fiscal year 2026, citing roughly 30% productivity gains from AI coding agents across its 15,000-strong engineering organisation. He has also said Salesforce is spending close to $300 million on Anthropic tokens. The replacement framing would treat this as engineering redundancy.
The full picture is different. Salesforce hired 1,000 graduates and interns. It expanded sales headcount by close to 20%. Benioff has been emphatic in subsequent interviews that AI is "hugely augmenting" engineers using Anthropic, OpenAI Codex and Cursor, but not replacing them. His exact line: AI models still need human oversight, and they are not at the point of running themselves.
This is the operationally honest version of the Spotify story. Engineering hiring slows. Engineering doesn't disappear. The savings get reinvested into revenue-generating roles. Total workforce grows. The structural change is real and it is consequential — particularly for anyone planning a career around junior-level coding tasks — but it is not a labour wipeout.
Duolingo: when the memo got ahead of the practice
In April 2025, Duolingo CEO Luis von Ahn published an internal memo on LinkedIn outlining an "AI-first" strategy, including a gradual phase-out of contractors for AI-handleable work, hiring only when teams could prove the work could not be automated, and AI use being factored into performance reviews. The backlash was severe. Users boycotted. Social media erupted.
By August 2025, von Ahn was walking it back in The New York Times. The clarification: "We've never laid off any full-time employees. We don't plan to." The contractor reductions were not new — Duolingo had cut contractors in 2023 and again in late 2024. The memo had simply made the policy more visible than the practice warranted.
The Duolingo episode is useful because it shows the corporate communications layer of the AI replacement narrative quite cleanly. The same set of operational facts was framed first as "AI-first transformation," then as "no full-time layoffs, contractor mix has always varied." Both framings were true. The difference was the audience.
The pattern, named
Across these cases, three patterns repeat and reinforce each other.
The deeper structural observation underneath all of this is that the AI replacement narrative is loudest where the buyer of the story sits furthest from the operation. Boards, investors, and analysts hear "headcount reduction." Operators on the contact centre floor and in the codebase see "volume absorbed with quality risk." When the two perspectives meet — usually 12 to 18 months in, when customer satisfaction scores or DORA stability metrics tell the truth — the rehire begins.
Khosrowshahi at Uber is the only executive of the lot who has said this out loud.
The strategic implications
For enterprise leaders thinking about their own AI workforce strategy, the data points to a small set of operational principles that survive the headline cycle.
The honest position
AI is unambiguously changing what knowledge work looks like. The Spotify senior engineer who has not typed code since December is real. The Google engineer reviewing AI-generated pull requests for three-quarters of new code is real. The Klarna AI agent handling routine customer inquiries is real. The Khosrowshahi prediction that 70–80% of human tasks will eventually be doable by AI is a serious forecast from a serious operator with the data to back the directionality, if not the timeline.
But the version of this story that has dominated the headlines for the last 18 months — wholesale replacement, imminent obsolescence, mass labour displacement — is not what the operational evidence shows. The evidence shows reallocation, augmentation, AI-washed offshoring, and a meaningful subset of high-profile cuts being quietly reversed once quality metrics catch up to the cost model.
The companies that will win the next phase of this transition are not the ones cutting fastest. They are the ones honestly distinguishing what AI can absorb from what human judgment still anchors, and then designing workflows, tooling, and workforce strategies around that distinction. The rest are doing theatre.