Run Your Company With a Fifth of the Headcount

Here is the number that should keep you up at night: the average company can be run with 20 to 25% of the workforce it has today. A fifth to a quarter. The rest of the org chart is coordination overhead that agents will absorb.

That claim comes out of Peter Diamandis’s “Organizational Singularity” episode, the seed for this piece. What makes it worth taking seriously isn’t the forecast — forecasts are cheap — it’s that the same number reads two ways depending on where you stand. Pessimistically, it’s 75% unemployment. Optimistically, it’s a world with five times as many companies, each smaller, faster, and more inventive than today’s. Both readings live in the same arithmetic. The question is which one the data is trending toward.

So let’s look at the data, because some of it is already on the books.

The cut is already happening at the edge #

Klarna is the cleanest public case. Between the end of 2022 and the end of 2024 its headcount fell from 5,527 to 3,422 — a 38% drop — and CEO Sebastian Siemiatkowski has talked openly about AI’s role in that decline (CNBC, May 2025). Be precise here, because it’s easy to overclaim: the company is explicit that a hiring freeze and natural attrition did much of the work, not AI on its own. The headcount didn’t crater because a model fired 2,000 people; it shrank because the company stopped backfilling roles it no longer needed, while one customer-service assistant launched in February 2024 reportedly covered the workload of 700 full-time agents. AI was the enabler, not the executioner. That distinction matters downstream.

Salesforce is the more controlled experiment. Its customer-support headcount went from 9,000 to 5,000 in 2025 — a 44% reduction — as AI agents took over 50% of all customer interactions and support costs dropped 17% (ODSC, 2025). Marc Benioff has framed this as a structural shift, not a one-off layoff, and the shape of it is the tell: the roles that collapsed were tier-1 routine interactions, while account management, complex escalations, engineering, and product kept hiring. The execution-and-coordination layer thinned; the judgment layer held. That’s not 80% gone. But it is a verified, publicly disclosed 44% cut inside one function at a 73,000-person public company, in a customer-sensitive environment where getting it wrong is expensive. As a data point about where the floor is, it’s hard to wave away.

And the intent is on the record at the top. IBM’s Arvind Krishna said back in 2023 that he could “easily see” 30% of the company’s roughly 26,000 non-customer-facing roles replaced by AI within five years (Computerworld, 2023). Thirty percent of back office, stated as a near-term floor by the CEO of a century-old enterprise, three years ago — not a moonshot, just routine paperwork.

The macro ceiling, and the spring of 2026 #

How far does this go in aggregate? The November 2025 McKinsey Global Institute report puts a number on it: AI agents could already perform tasks accounting for 44% of US work hours, with roughly 40% of all jobs sitting in “highly automatable” roles. Even at mid-point adoption, MGI models $2.9 trillion in annual US economic value by 2030 (Robotics & Automation News, Nov 2025). Read carefully, 44% of hours is not 44% of jobs — and that gap is where the whole argument turns. But it brackets the technical ceiling, and the ceiling is high.

The spring of 2026 is when Big Tech started moving in size. In April, Meta announced a roughly 10% workforce cut (~8,000 jobs) and Microsoft offered buyouts to about 7% of its US staff (~8,750 jobs); year-to-date, the tech sector had shed more than 92,000 jobs (CNBC, April 2026). The attribution deserves a lighter touch than the headlines gave it: Meta only gestured at AI in its efficiency memo, and Microsoft’s buyout wasn’t framed as AI-driven at all. The honest read is that these firms were broadly citing efficiency and AI-investment priorities — reallocating budget toward AI capacity — not announcing that agents had eliminated specific teams. That’s still the thesis playing out. It’s just not the clean causation the phrase “AI layoff” implies.

The optimistic half is also showing up in the numbers #

Here’s what the doom framing misses: the “5x more companies” reading has receipts too.

US business applications hit 1.56 million in the three months to January 2026 — the highest such period on record (CNBC, Mar 2026). And among AI-adopting startups, 79% report hiring more people, not fewer (Startups USA). The AI-native firm doesn’t just need fewer people per dollar of revenue — it makes starting a firm cheap enough that far more of them get started. Investors are already benchmarking that efficiency: an Emergence Capital analysis circulated by Pavilion found AI-native companies generating revenue per employee roughly 7–8x the public-SaaS benchmark of around $300K, with names like Midjourney and Cursor clearing $100M+ ARR on tiny teams (Pavilion / Emergence Capital, 2025). If a handful of people can run a $100M company, the binding constraint on company formation stops being capital or headcount and becomes ambition. That’s the mechanism behind “5x more companies,” and it isn’t theoretical.

The honest counter-case #

Now the part that keeps this from being a hype reel, because the skeptical case is strong and partly correct.

“AI layoff” is frequently a cover story. Of the roughly 1.17 million US job cuts tallied in 2025, only a small fraction were formally attributed to AI in the Challenger, Gray & Christmas tracking data (Development Corporate, 2025). Post-pandemic over-hiring and plain margin pressure account for most of the rest. “We’re restructuring around AI” is a more flattering line for a CEO than “we hired too many people in 2021,” and a meaningful share of so-called AI cuts are the latter wearing the former’s clothes.

Most firms are reinvesting, not cutting — at least so far. EY’s December 2025 AI Pulse survey found only 17% of firms seeing AI productivity gains are using them to reduce headcount, down from the 46% of CEOs who expected to cut back in January 2025. The majority are redirecting the savings into more AI, R&D, and upskilling (EY, Dec 2025). The gap between what executives predicted and what they actually did is the single most important caveat on the whole 20% thesis.

Task automation is not job elimination. The McKinsey number again, read the other way: automating 44% of work hours mostly means changing jobs, not deleting them, because most workers hold several distinct task types and AI tends to take one bucket at a time. The job that survives looks different; it doesn’t necessarily disappear.

History leans the other way. Roughly 60% of US workers today hold occupations that didn’t exist in 1940 — technology has reliably created more roles than it destroyed. Goldman Sachs CEO David Solomon has been reported pushing back on the idea that AI will eliminate 25% of jobs, with Goldman economists modeling closer to 6–7% worker displacement over a decade-long transition (Goldman Sachs). One caveat: the primary Goldman source was unreachable while writing this, so treat that quote and the 6–7% figure as widely reported but unconfirmed at the source. The WEF’s Future of Jobs 2025 points the same way from firmer ground: 170 million jobs created against 92 million displaced by 2030, a net gain of 78 million (WEF, 2025).

And over-automation backfires. Klarna is the cautionary tale on both ends. After its aggressive AI-first hiring freeze, customer complaints about the absence of human support pushed it to start rehiring ahead of its 2025 US IPO, with Siemiatkowski conceding that “cost unfortunately seems to have been a too predominant evaluation factor” (Fortune, Oct 2025). The company that proved you could cut 38% also proved you could cut too far.

What a builder should take from this #

Strip out the forecasting and a usable picture remains. The technical ceiling is real and high — 40% of US jobs sit in highly automatable roles, and 44% of work hours are already addressable by agents. The early movers confirm that 40%+ cuts inside a single function are achievable without the wheels coming off, if you cut the right layer. But the macro picture is reinvestment, not mass unemployment, because the savings are being plowed back in and every collapsed function spawns new firms that need new people.

So the 20% number is probably right about the unit — what one company can run on — and probably wrong as a headline about the economy, because the count of companies is climbing fast enough to absorb much of the slack. The trap isn’t getting automated. It’s sitting in the middle of an org that’s all coordination, in a company that reinvests its AI gains into everything except rethinking its own shape, while a five-person team three time zones away rebuilds your best margin line at a fifth of your cost. Klarna’s real lesson isn’t “cut deep.” It’s that the people who thrive are the ones doing judgment work the agents can’t — and that the cut, when it comes, lands hardest on whoever was only passing information along.