Contents
- 01The full picture first— Actions, costs and how it runs
- 02Three drivers of the shift— Why now
- 03Six organizational trends— Common threads across the sample
- 04Deep company cases— Conclusions and numbers first; timelines expand on demand
- 05Side-by-side comparison— Paths, rhetoric, costs and evidence strength
- 06The real practice— How AI-era companies actually run
- 07Where did middle managers go— Five transition paths
- 08Research & industry views— What analysts and academics say
- 09Conclusions & outlook— Five takeaways and a manager self-check
- 10Video sources— 19 video evidence items, expandable
- 11References— The full link list, expandable
The Full Picture First: Actions, Costs and How It Runs
Three pillars answer three questions: what happened, what it cost, and how the organization keeps running.
The first two items (Google) are the sample of "precisely cutting management roles"; the other five show this shift does not stop at management — the layoffs at Amazon, X, Block and Salesforce genuinely reached engineers and frontline roles. Nvidia's ~60 direct reports is "long-standing flatness", not "change".
Measured de-layering
Management-role measures
Workforce contraction
Percentages and headcounts use different scales
Three Drivers of the Shift
Why now? Why management?
- The traditional value of the middle layer is "information routing" — and routing is precisely the first job AI is good at. That is the core of the argument by Jack Dorsey (Block co-founder and Block Head) and Roelof Botha (Sequoia Capital partner, Block's lead independent director), and the starting point for understanding this shift.
- Multiplied individual output shrinks headcount: Salesforce froze engineering hiring, Amazon signaled a leaner corporate workforce; with fewer people, the management layer contracts with them.
- The competitive clock now ticks in months: Google's stated motive for cutting management roles was "fewer layers, faster decisions"; Coinbase wrote "no more than 5 layers below CEO/COO" into policy; Xbox compressed as many as 14 layers to 5 or fewer — and Nvidia's extreme co-design moves communication into collective settings rather than reducing it.
Coordination cost collapse
For decades, the core value of the middle layer has been "information routing" — summarizing upward, breaking work down downward, aligning sideways. Jack Dorsey (Block co-founder and Block Head) and Roelof Botha (Sequoia Capital partner, Block's lead independent director), in "From Hierarchy to Intelligence", co-published on March 31, 2026 (also carried on Sequoia's site), put it directly: middle managers are essentially information routers, and AI can do the routing better. The essay traces bureaucracy back two thousand years to the Roman legion — the eight-man contubernium as the smallest unit — then through the Prussian general staff and the railroad companies of the 1850s, to today's org chart.
Per-person output multiplies
AI coding assistants, agent-based support and automated workflows multiply what a single engineer or operator can produce. Salesforce, citing Agentforce, announced it would add no new software engineers in 2025; Amazon's June 2025 memo forecast "efficiency gains" from AI and a leaner corporate workforce. With fewer people, the management layer naturally contracts.
The competitive clock speeds up
Model capability now jumps month by month. Google reorganized twice within 2024 alone; Jensen Huang says Nvidia must "move at the speed of light", and his ~60 direct reports exist precisely to delete the middle of the decision chain — he calls the method "extreme co-design": not less communication, but all communication in collective settings. When the strategic window is measured in weeks, layers tend to mean delay.
Six Organizational Trends
Common threads distilled from the public moves of 11 companies; not every company shows every one of them.
- The Great Flattening is the shared move across companies — but from the second half of 2025 it evolved from "cutting only management roles" into layoffs that reach the front line.
- Merging research and product teams is not an end state: Meta's MSL split just 7 weeks after it was formed, showing organizations must keep adjusting to match the model release cadence.
- The need for coordination does not disappear: spans of control widen (Gallup average 12.1, Coinbase allowing 15+) while Chief of Staff postings more than double in a year — a signal that coordination work may be redistributing, not yet proof of cause and effect.
The six trends resolve into three simultaneous changes: structures get thinner, team boundaries get redrawn, and authority and coordination get reorganized. See the relationships first, then check the company evidence item by item.
From hierarchical routing to system-level coordination
AI takes over part of the relaying, summarizing and status tracking; people move closer to the problem, the judgment call and the responsibility.
Thinner layers, wider spans
Team boundaries keep getting redrawn
Power moves to the front line; coordination is redistributed
The Great Flattening
Google, Amazon, Microsoft and Meta each cut into their management layers between 2023 and 2026: Google cut 10% of management roles cumulatively (the precise 2025-08 measure: 35% fewer managers with fewer than 3 direct reports); Amazon raised its IC ratio by 15%+, then added roughly 30,000 corporate layoffs in 2025.10–2026.01; Microsoft laid off more than 15,000 people in 2025 with the explicit goal of increasing span of control; Meta's 2023 "Year of Efficiency" turned many managers back into individual contributors, followed by three rounds of cuts in 2026. Business Insider called it a "crushing" of middle managers.
AI research and product merge — then split again
Google folded the Gemini App team into Google DeepMind; Meta merged model and product teams into Meta Superintelligence Labs in 2025.06 — then split it into four teams just 7 weeks later in 2025.08. The lesson is that the merger itself is not an end state: putting model people and product people in one org does not mean the org shape will hold still.
Engineer-led culture returns
At X under Musk, "the people who write code" replaced the management layer as the center of power, and the organization grew even more engineer-centric after CEO Yaccarino's departure; Nvidia has long run an "engineer CEO + extremely flat" structure; on Anthropic's Claude Code team "everyone writes code" — product managers, engineering managers, designers, finance and data scientists all commit code directly (per team lead Boris Cherny). "Founder Mode" lives on as a Silicon Valley catchphrase.
Human-AI hybrid teams become the headcount unit
Inside Salesforce, roughly 30–50% of work is already done by AI agents; some Coinbase teams are experimenting with "one person + N agents" pods. The new org-design question is no longer "how many people report to one manager" but "how many humans plus how many agents report to one owner".
Wider spans of control: the verifiable numbers
Amazon's internal guidance raised the minimum direct reports per manager from 6 to 8 (AWS internal guidance revealed by BI 2025.01, relayed by Fortune 2025.09); after its reorg Coinbase allows leaders to carry 15+ direct reports (TechCrunch / Fortune 2026.05); Gallup finds the average US manager's direct reports rose from 10.9 (2024) to 12.1 (2025); Google defined the same class of metric precisely with "35% fewer managers with fewer than 3 direct reports". Morgan Stanley, working from Jassy's 15% target, estimated Amazon could trim roughly 13,800 manager positions and free $2.1–3.6 billion a year. Wider spans mean managers cannot micromanage — status tracking and report-rollups shift to AI.
Coordination roles reorganized: rising Chief of Staff demand
With the middle layer cut away, executives' information bandwidth tightens — postings for "Chief of Staff" (a corporate aide-de-camp role, close to running the office of the CEO) more than doubled over the prior year (Bloomberg 2026.07 citing Revelio Labs; the report discloses no base or methodology), which organization-design commentators have called "the bill for the Great Flattening". Block's alternative is the "Directly Responsible Individual": named ownership with the ability to pull in resources ad hoc, but no standing management layer; the player-coach model has drawn named criticism (a Fortune 2026.05 column: it "misunderstands the value of good managers").
Deep Company Cases
This chapter walks through the 11 companies roughly by how prominent their org-change moves have been: Google, Amazon, X, Nvidia, Meta, Microsoft, Salesforce, Block, Coinbase, OpenAI and Anthropic. Each case includes a timeline, the key numbers and primary sources.
- The 11 companies' moves fall roughly into three types: structural reorganization (Google, Meta, OpenAI), layoff-driven (Amazon, Microsoft, X, Block, Coinbase), and long-standing flatness or deliberate leanness (Nvidia, Anthropic); Salesforce is the case of agents directly replacing headcount.
- There is often tension between official rhetoric and actual moves: Amazon denies its layoffs were AI- or cost-driven, Block laid people off first and published the argument afterward — read each case's timeline together with its "tension" note.
- No new structure has proven to be a stable end state: OpenAI's dual-track setup lasted about 14 months, Meta's MSL 7 weeks, and Google rewired its reporting lines again in 2026.08.
The moves are accelerating — and shifting gears: 2023–2024 was mostly structural reorganization, layoffs intensified from H2 2025, and by 2026 the practices are being written into policy and made public.
All events (table)
| Company | Date | Type | Event |
|---|---|---|---|
| 2023-04 | Reorg / reporting-line changes | Brain + DeepMind merge | |
| 2024-04 | Reorg / reporting-line changes | Parts of Research join GDM | |
| 2024-10 | Reorg / reporting-line changes | Gemini App joins GDM | |
| 2024-12 | Layoffs / management cuts | 10% of management roles cut | |
| 2025-08 | Layoffs / management cuts | Sub-3-report managers −35% | |
| 2026-08 | Reorg / reporting-line changes | Hassabis steps down as CEO; lines rewired | |
| Amazon | 2024-09 | Policy and practice | Jassy memo: IC ratio +15% |
| Amazon | 2025-10 | Layoffs / management cuts | 14,000 laid off |
| Amazon | 2026-01 | Layoffs / management cuts | 16,000 more laid off |
| Amazon | 2026-07 | Layoffs / management cuts | AGI-org layoffs |
| X / xAI | 2025-03 | Reorg / reporting-line changes | xAI acquires X |
| X / xAI | 2025-07 | Reorg / reporting-line changes | CEO Yaccarino departs |
| X / xAI | 2026-02 | Reorg / reporting-line changes | xAI reorg; into SpaceX |
| Nvidia | 2023-12 | Policy and practice | HBR: ~60 reports, no routine 1:1s |
| Nvidia | 2026-03 | Policy and practice | Lex Fridman: extreme co-design |
| Meta | 2023-03 | Layoffs / management cuts | Year of Efficiency 10,000 (with 2022.11, ~21,000 total) |
| Meta | 2025-06 | Reorg / reporting-line changes | MSL formed |
| Meta | 2025-08 | Reorg / reporting-line changes | MSL split into four |
| Meta | 2026-01 | Layoffs / management cuts | Reality Labs layoffs |
| Meta | 2026-03 | Layoffs / management cuts | Hundreds more cut |
| Meta | 2026-08 | Reorg / reporting-line changes | Project OT stumbles (Reuters) |
| Microsoft | 2025-05 | Layoffs / management cuts | ~6,000 laid off |
| Microsoft | 2025-07 | Layoffs / management cuts | ~9,000 laid off |
| Microsoft | 2026-07 | Layoffs / management cuts | Xbox 3,200; layers 14 → ≤5 |
| Salesforce | 2025-06 | Policy and practice | AI does 30–50% of work |
| Salesforce | 2025-09 | Layoffs / management cuts | Support team cut 4,000 |
| Block | 2026-02 | Layoffs / management cuts | ~40% laid off |
| Block | 2026-03 | Policy and practice | DRI / IC / player-coach roles |
| Coinbase | 2025-08 | Policy and practice | AI coding-assistant mandate |
| Coinbase | 2026-05 | Reorg / reporting-line changes | ≤5 layers; pure manager abolished |
| OpenAI | 2025-05 | Reorg / reporting-line changes | CEO of Applications created |
| OpenAI | 2026-07 | Reorg / reporting-line changes | Simo to part-time advisor |
| OpenAI | 2026-08 | Reorg / reporting-line changes | 13 execs depart; infra line to Katti |
| Anthropic | 2026-06 | Policy and practice | CEO has 1 direct report (Fortune) |
| Anthropic | 2026-08 | Publishing their own playbook | Startup practice guide |
Nvidia is the "always been this way" template; Anthropic is "deliberately lean" (its reporting structure is single-source information). Learning from the current state of a company that never had to transform is a different exercise from learning from the process of one mid-transformation.
| Structural reorg | Layoff-driven | AI substitution | Long-standing flat / deliberately lean | |
|---|---|---|---|---|
| ● | ● | · | · | |
| Amazon | · | ● | · | · |
| X / xAI | ● | ● | · | · |
| Nvidia | · | · | · | ● |
| Meta | ● | ● | · | · |
| Microsoft | · | ● | · | · |
| Salesforce | · | ● | ● | · |
| Block | · | ● | · | · |
| Coinbase | ● | ● | · | · |
| OpenAI | ● | · | · | · |
| Anthropic | · | · | · | ● |
● = a main practice verifiable in the company's 2023–2026 public moves; one company can occupy two types. The four-way split is our synthesis, not official company language. Anthropic's "long-standing flat / deliberately lean" placement rests on a single 2026.06 snapshot; whether it holds cannot be confirmed.
Each company starts with its conclusion and key numbers; expand for the full timeline and sources.
Google / Alphabet
Pichai: repeated reorgs + leaner management + the 2026.08 leadership reshuffleGoogle is the big company that has reorganized most often in the AI era: two major restructures within 2024 alone, the core logic being to stitch the model team (DeepMind) directly to the product team (Gemini App); in August 2026 it completed its biggest rewiring of reporting lines in years.
−10%manager / director / VP roles (cumulative, 2022.09–2024.12)−35%managers with fewer than 3 direct reports (2024.08–2025.08; many stayed on as ICs)2major reorgs within 2024
Google / Alphabet
Pichai: repeated reorgs + leaner management + the 2026.08 leadership reshuffle- 2023.04Google Brain and DeepMind merged into Google DeepMind, reporting to Demis Hassabis — ending the era of two parallel, competing AI labs.Why: competing with OpenAI/Microsoft required concentrating compute and talent, and avoiding internal wheel-reinventing.
- 2024.04–10Two reorgs within one year: in April, parts of Research moved into Google DeepMind and a Platforms & Devices division was created; in October, the Gemini App team moved into Google DeepMind, the Assistant team into Platforms & Devices, with Search and Ads leadership adjusted in step.Why: put application teams and model teams at zero distance, shortening the path from model capability to product feature.
- 2024.12.18Pichai confirmed at an all-hands: 10% of manager / director / VP roles had been cut, as part of the efficiency drive launched in September 2022.Why: fewer layers, faster decisions — a response to the OpenAI threat.
- 2025.08.27VP of People Analytics and Performance Brian Welle gave a more precise measure at an all-hands: over the past year the number of managers with fewer than 3 direct reports fell 35%, many of whom stayed on as ICs; ten product areas offered voluntary exit programs at the same time.This supersedes the vaguer "a third of all managers cut" version that had circulated — the cuts targeted managers of teams below a certain size, not management roles across the board.
- 2026.05.27On The Verge's Decoder podcast, Pichai described Google's post-ChatGPT response as more than layoffs — a redesign of "who owns how technology flows": AI infrastructure was consolidated under a new SVP position, and a "Chief AI Architect" role was created — held since 2025.06 by Koray Kavukcuoglu (SVP, reporting directly to Pichai), of a piece with his expanded remit in 2026.08.
- 2026.08.05Google completed its most consequential rewiring of reporting lines in years: Demis Hassabis stepped down as Google DeepMind CEO to become GDM Chair and Alphabet Chief Scientist, continuing to lead Isomorphic Labs; Jeff Dean, the "godfather-tier" chief scientist with 27 years at Google, left to found Discovery Loop (with Google as investor and cloud partner); DeepMind CTO Koray Kavukcuoglu was promoted to Senior Vice President, reporting directly to Pichai, jointly overseeing Gemini model development, frontier research, the Gemini App and developer teams with Josh Woodward.This round further tightened the reporting chain from AI models, applications and developer teams up to the CEO.
- 2026.08.06–10Follow-up reporting sketched the backdrop: Time said DeepMind's communications, legal and marketing functions were folded into their Google counterparts; Fortune reported months of low morale at DeepMind before the reorg, several senior researchers leaving over July–August, Gemini 3.5 Pro repeatedly missing internal timelines, and the center of gravity shifting from London to Mountain View.These accounts are each single-outlet reporting, unconfirmed by Google; recorded here as background only.
The oft-quoted Google "20% more efficient" line traces, verifiably, to September 2022 — Pichai's remarks at the Code Conference — not to the "2023 all-hands" some reports cite. This report uses the 2022.09 dating.— Sourcing note
Amazon / AWS
Jassy: anti-bureaucracy + IC ratio → 30,000 laid off across two 2025.10–2026.01 roundsAmazon was the company that made "flattening" the most engineered and most quantifiable: a hard numeric target for every S-Team organization, plus a "bureaucracy mailbox" and a culture reset. But from H2 2025, its own actions broke the "flattening without layoffs" narrative — two corporate layoff rounds totaling roughly 30,000 people, in direct conflict with the earlier line.
+15%IC : manager ratio-increase target (by Q1 2025)1,500 / 455bureaucracy-mailbox complaints / process changes shipped~30,000combined corporate layoffs, 2025.10–2026.011,800+engineers among them (per WARN filings in four states, 2025.11)
Amazon / AWS
Jassy: anti-bureaucracy + IC ratio → 30,000 laid off across two 2025.10–2026.01 rounds- 2024.09Jassy's internal letter: every S-Team organization must raise its individual-contributor-to-manager ratio by at least 15% by Q1 2025, alongside a return to five days in the office.His words: "Fewer managers will remove layers and flatten organizations more than they are today."
- 2024Created the "bureaucracy mailbox": employees can report pointless process and meetings directly. By September 2025 it had received about 1,500 complaints and driven 455 process changes.Why: Jassy said he wants Amazon operating like "the world's largest startup".
- 2025.01Leaked documents showed Amazon hit the 15% target not through layoffs but by giving managers more reports (per Entrepreneur).That claim held in early 2025, but was broken by the two layoff rounds from 2025.10 — "flattening without layoffs" is no longer the whole story at this company.
- 2025.06–09Published a culture memo tying performance reviews directly to the 16 Leadership Principles; in his June memo Jassy forecast AI-driven efficiency gains and a leaner corporate workforce.
- 2025.10.28 – 2026.01.28Two rounds totaling roughly 30,000 corporate layoffs — among the largest in company history: about 14,000 roles on October 28; per CNBC's November 2025 analysis of WARN filings in four states (covering only about 4,700 people), more than 1,800 engineers were cut in that round, far beyond "middle managers only"; another ~16,000 followed on January 28, 2026. Notably, pressed by analysts on the October 30 Q3 earnings call, Jassy was explicit that the layoffs were "not really financially driven, and it is not really AI driven right now... it is culture" — attributing them to the organization being "too big, too many layers", not to "AI replacing people".
- 2026.07AGI-org layoffs: on July 22 CNBC reported cuts to the Nova model team (headcount undisclosed); Business Insider then reported the AGI Lab (the Adept team absorbed in 2024) was shut down, the Nova flagship model moved to maintenance mode, and Pieter Abbeel now leads a new frontier-model team. The AGI org had already been folded into Peter DeSantis's SVP organization in 2025.12, with former head Rohit Prasad departing.Beyond "compressing layers", Amazon's AI organization itself is being rapidly reshuffled.
- AWS sideIn 2024.06 Matt Garman succeeded Adam Selipsky as AWS CEO, followed by sweeping leadership changes across global sales, channels and the generative-AI organization.
X (formerly Twitter)
Musk: an extreme-flattening testbed for "engineer sovereignty"X is the most radical sample of this cycle: first physically delete the organization (roughly 80% of employees gone), then rebuild it as an extreme-flat, engineer-driven structure; merged into xAI in 2025 and, with xAI, into SpaceX in 2026 — fully AI-native.
~80%cumulative headcount reduction (2022.10–2023.04, Musk's own BBC figure)7,500 → ~1,500headcount change (cross-verified by multiple outlets)
X (formerly Twitter)
Musk: an extreme-flattening testbed for "engineer sovereignty"- 2022.10–11After the takeover Musk cut about 50% of staff, then issued the famous "extremely hardcore" ultimatum: accept "long hours at high intensity" or leave — hundreds left, and headcount fell from about 7,500 to about 1,500 (roughly 80% cumulative by 2023.04, Musk's own figure in a BBC interview).The result: the management layer and middle processes were removed essentially wholesale.
- 2022–2023Rebuilt as "Twitter 2.0 / X": multiple observers noted the new organization is run by "the people who write code", with engineers reporting directly to Musk and product decisions made in the open over Slack/email. Synthesizing The Verge's and Fortune's reporting, X in this period no longer had an org chart in the traditional sense — engineers, product teams and Musk himself shared one information channel, and whoever owned something simply claimed the task directly.Why: Musk regards the management layer itself as information loss.
- 2025.03.28X was acquired by xAI in an all-stock deal (valuing it around $33 billion), folding the social platform and the AI lab into one.
- 2025.07.09CEO Linda Yaccarino resigned — she had been the only executive buffer since the takeover; Musk then took direct control, pulling sales, product and engineering together. The announcement came the day after the controversy over the Grok chatbot's inflammatory outputs, but per people familiar speaking to NBC News, the decision "had been brewing for over a week" and was not directly caused by the Grok incident.The consensus across outlets: X had cut roughly 80% of staff in the Musk era, and Yaccarino's exit marked the end of the "two-headed" structure.
- 2026.02–08On February 2 SpaceX acquired xAI in an all-stock deal (X included); on February 11 Musk announced an xAI reorg, saying the first build "wasn't built right" and would be "rebuilt from the foundation" — 6 of the 12 co-founders had already left; in June the merged SpaceX completed its IPO; on August 5 X product head Nikita Bier departed with no successor named.X is no longer an independent-company sample but a product line inside the SpaceX–xAI complex.
Nvidia
Jensen Huang: ~60 direct reports, almost no one-on-ones — "extreme co-design"Nvidia is the most valuable company of the AI era and the most thorough sample of flat structure — an organization that did not "transform" into this shape but has always run this way, now studied industry-wide as the template.
~60direct reports to the CEO (his own figure; media accounts 36–60)≈0routine one-on-ones (HBR's version: "rarely")Top 5employees' ad-hoc "top 5 things" emails straight to the CEONo long-term planrolling forecasts replace the annual plan
Nvidia
Jensen Huang: ~60 direct reports, almost no one-on-ones — "extreme co-design"- Current structureHuang says he has about 60 direct reports and does almost no one-on-ones, preferring group discussions with everyone present and "top 5 things" emails. The figures are not consistent: Fortune (2024.11) used "60"; he himself said "60, maybe more" on the Lex Fridman podcast in 2026.03 and "as many as 55" in an earlier Stanford talk; a Business Insider report in 2025.10, citing an internal list, showed 36 direct reports — the number drifts with reorgs and counting method, and this report uses his own ~60.His explanation: "The more direct reports a CEO has, the fewer layers are in the company." Information flows transparently within the group instead of being filtered layer by layer.
- 2026.03.23On Lex Fridman podcast #494, Huang defined his method as "extreme co-design": "No conversation has just one person in it. So I don't do one-on-ones — we pose a problem and everyone solves it together." Earlier, in a February 2025 Fortune interview, his version was: "They never hear anything I say to only one of them."
- Management philosophy"No long-term plans, no fixed job duties" — he asks everyone to "move at the speed of light", with the organization designed around information flow rather than rank.
- Industry influenceAcross 2024–2026, Huang's "no 1:1s, many direct reports" model became the reference point for AI-era CEOs — at Stripe Sessions in 2025.05, host John Collison unprompted compared Zuckerberg's "25–30 person core team" to the Huang model, while Zuckerberg stressed the two are not the same.
Meta
MSL merged then split + three rounds of layoffs, ~10,000 people, in 2026Meta was the first mover of this Great Flattening — and the best evidence that merging organizations is never one-and-done: the top-level AI team it created in 2025 split again after just 7 weeks, followed by three rounds of contraction in H1 2026.
7 weeksMSL from formation to split (2025.06.30 → 2025.08)+220% / +36%code changes vs user-visible new features during Project OT (year over year)+40%major technical / security incidents over the same period (year over year)~21,000cumulative layoffs 2022.11–2023.05 (10,000 in the Year of Efficiency + 11,000 before)
Meta
MSL merged then split + three rounds of layoffs, ~10,000 people, in 2026- 2023.03Zuckerberg's open letter: "Flatter is faster" — the Year of Efficiency would "remove multiple layers of management" and ask many managers to become individual contributors (ICs); another 10,000 laid off and about 5,000 open roles closed.
- 2025.06.30Zuckerberg's internal memo announced Meta Superintelligence Labs (MSL), consolidating model development, product teams and fundamental research (FAIR); former Scale AI CEO Alexandr Wang (28) became Meta's first Chief AI Officer, and former OpenAI researcher and ChatGPT co-creator Shengjia Zhao was named MSL Chief Scientist on July 25 of the same year.
- 2025.08.19Just 7 weeks later, MSL split again into four teams: TBD Lab (frontier models), FAIR (fundamental research), Products and Applied Research, and MSL Infra; two months after that, on October 22, Meta cut about 600 more AI roles across FAIR, product AI and infrastructure teams."Merge, then split" shows that merging by itself cannot guarantee collaborative efficiency — the split, equally, was about giving each team focus.
- 2026.01–05Three consecutive rounds of contraction at Reality Labs: in January, about 10% of the division cut (multiple VR studios closed, budgets trimmed by up to 30%); in March, "hundreds" more across at least five units including Reality Labs, Facebook, recruiting and sales; on May 20, a company-wide round of about 8,000 people (roughly 10% of global staff) — explicitly not aimed at low performers, but at reorganizing teams company-wide into AI-led "pods" (internal codename Project OT), with Chief People Officer Janelle Gale saying 7,000+ people were moved into newly created AI teams. A ~1,000-person Reality Labs developer-tools team piloted dropping traditional functional titles for three roles — AI Builder / AI Pod Lead / AI Org Lead — with AI systems assisting promotion evaluation.
- 2026.03.06Created a new Applied AI Engineering organization reporting to CTO Andrew Bosworth, carving data and engineering functions out of MSL; the NYT reported friction between Alexandr Wang and Bosworth and Chris Cox, and Yann LeCun had left in 2025.11.The third boundary redraw since MSL's merge-then-split.
- 2026.01.28On the Q4 2025 earnings call, Zuckerberg told investors plainly: "Projects that used to take a big team can now be done by one very talented person," so the company is "elevating individual contributors and flattening teams".
- 2025.05In conversation with Stripe president John Collison at Stripe Sessions, Zuckerberg said he directly runs a "core team" of about 25–30 people (not a company-wide 60-person arrangement) and holds no regularly scheduled one-on-ones — adding "it is not that there are no one-on-ones, there are just no fixed-cadence ones; I am talking with these people all the time". The host, unprompted, compared this with the Jensen Huang model.
- 2026.08.26A Reuters special report laid out Project OT (Organization Transformation) in full: drafted at a January 2026 leadership offsite, the plan envisioned AI taking over much of the work of thousands of employees, with smaller human teams supervising "virtual workers"; internal documents had assessed cutting some teams by up to 60%. Two waves were planned for May and November; Zuckerberg canceled the second late on May 19, so only the first round of about 8,000 was executed. Internal data showed AI-platform code changes up 220% year over year — but user-facing new features up only 36%, and major technical and security incidents up 40%.This is currently the most important piece of counter-evidence on "AI-native reorganization": a surge in output volume did not convert proportionally into delivered value. The Reuters original is paywalled; cited here via Engadget / CTV and other secondary accounts.
"Flatter is faster. In our Year of Efficiency, we will make our organization flatter by removing multiple layers of management... ask many managers to become individual contributors."— Mark Zuckerberg, open letter, 2023.03 (official)
Microsoft
Nadella: 15,000+ roles cut in the name of "fewer layers, fewer managers"Microsoft's two big 2025 rounds (about 6,000 in May, about 9,000 in July — over 15,000 for the year) wrote org-theory language into the explanation: CFO Amy Hood's official phrasing was "fewer layers, fewer managers", while the internal target Business Insider cited was widening the span of control.
~15,3002025 full-year layoffs (of ~228,000 employees, per the 2024.06 annual report)3,200Xbox layoffs, 2026.07 (~1,600 immediate)14 → ≤5Xbox management-layer compression target (GeekWire)
Microsoft
Nadella: 15,000+ roles cut in the name of "fewer layers, fewer managers"- 2025.05.13About 6,000 laid off (under 3% of staff). BI, citing internal information, said the goal was precisely to widen spans of control and cut management layers; ironically, per the same report, the lists included several managers with only 1–2 reports — and later tallies showed a sizable share were frontline engineers.
- 2025.07.02Another ~9,000 cut, taking the year past 15,000 (of roughly 228,000 employees). Xbox CEO Phil Spencer wrote in an internal memo that teams would "follow Microsoft's lead in removing layers of management" — the official line remained fewer layers and faster decisions, while acknowledging many frontline engineers were affected.
- 2025.10.01Judson Althoff rose from EVP and Chief Commercial Officer to "CEO of Microsoft's commercial business". Nadella wrote that this would let "our engineering leaders and me" focus on data-center buildout, systems architecture, AI science and product innovation. The promotion and the mid-year layoffs came from different documents and should be read as concurrent but independent adjustments, not cause and effect.
- 2026.02–07On February 20 Xbox chief Phil Spencer retired and Asha Sharma became CEO of Microsoft Gaming; April brought the company's first-ever voluntary exit program (open to about 7% of US employees); on July 6 Xbox announced about 3,200 cuts within FY27 (about 1,600 that day) and the divestment of four studios, with Sharma saying "we spread ourselves too thin".
Salesforce
Benioff: the extreme case of an "AI agent workforce"Salesforce is the bluntest case of "AI directly changing headcount": the CEO personally quantified the share of work AI performs and cut the support team; engineering hiring had been frozen even earlier (2024.12).
30–50%share of internal work done by AI agents (Benioff's figure)9,000→~5,000customer-support workforce (Benioff's own account)4,000roles confirmed cut in 2025.09
Salesforce
Benioff: the extreme case of an "AI agent workforce"- 2025.02Benioff announced no new software engineers in 2025 — because Agentforce and related tools had lifted engineering productivity, with "engineering velocity incredible".
- 2025.06–07Bloomberg/Fortune interviews: AI agents already perform 30%–50% of the work inside Salesforce; sales roles, by contrast, are growing.
- 2025.08.29On The Logan Bartlett Show, Benioff first said "I need less heads": "I've reduced it from 9,000 heads to about 5,000, because I need less heads" — cutting the customer-support workforce from about 9,000 to about 5,000.
- 2025.09.02CNBC confirmed 4,000 roles cut, quoting the podcast line above; support tickets had fallen sharply thanks to agents, and support engineers were not backfilled.Note: this widely circulated quote originates from the 08.29 podcast interview; CNBC's version is secondhand, not the primary source.
- 2026.02–08Three smaller rounds: under 1,000 in February (marketing, product, analytics and the Agentforce team), then dozens-to-hundreds each in June and August (Agentforce, MuleSoft, Marketing Cloud); the filings did not cite AI as a reason. Agentforce annualized revenue passed $1 billion in May.
Block (Jack Dorsey)
Behind the theory paper, a near-halving layoffWhere the giants have been "doing", Jack Dorsey (Block co-founder and Block Head) and Roelof Botha (Sequoia Capital partner, Block's lead independent director) tried to turn it into theory — but the theoretical essay itself is best read as an after-the-fact justification of a near-halving reorganization, not a pure thought experiment.
~40%share of staff cut in one round on 2026.02.26 (~4,000 people)~18%next-day stock move after the announcementDRIthe DRI mechanism: named ownership, no standing management layer
Block (Jack Dorsey)
Behind the theory paper, a near-halving layoff- 2026.02.26About a month before the essay, Block had already answered with action: a single round of roughly 4,000 layoffs, about 40% of all staff, taking headcount from over 10,000 to under 6,000. Dorsey wrote to shareholders: "A much smaller team, using the tools we are building, can do more and do it better." The stock jumped over 20% after hours on announcement day and closed up about 18% the next day.
- 2026.03.31Dorsey and Roelof Botha published their co-authored essay "From Hierarchy to Intelligence", carried simultaneously on Block's site and Sequoia's site. It traces bureaucracy back two thousand years to the Roman legion — the eight-man contubernium as the smallest unit — then through the Prussian general staff and the railroad companies of the 1850s to today's org chart; the core claim is that the middle layer's core function is information routing, AI routes better, and therefore middle management in the traditional sense can be abolished.
- Block practiceThe organization was rebuilt on the "Directly Responsible Individual (DRI)" model: three roles — IC / DRI / player-coach — where a DRI takes named ownership of cross-team problems and can pull in resources ad hoc, with no standing management layer; still rolling out as of 2026.08 (Q2 earnings call: "the reorganization is on track"). Some organization-design commentators worry the risk is "responsibility without authority" — a criticism with no named primary source aimed at Block specifically, recorded here as observation only.
- 2026.04–08On April 2, on Sequoia's podcast, Dorsey said the ideal would be "no layers at all — all 6,000 people reporting to me", with a plan to compress management from 5 layers to 2–3 within a year; on August 5 the Q2 earnings call said the "intelligence-centric reorganization is proceeding to plan", per-engineer code changes were up 150% since the start of the year, Square shipped 130 features in H1 (more than 3x the same period last year), and full-year guidance was raised.The mirror image of Meta's Project OT: what Block has published so far is positive data — but the measures differ (code-change volume vs features shipped) and cannot be compared directly.
- Ripple effectsForbes / Fortune read the essay alongside Amazon's and Meta's de-layering moves; but set against the 02.26 layoff timeline, the more accurate reading is that the essay is an after-the-fact justification of a large reorganization that had already happened, not a purely theoretical manifesto.
Coinbase
Armstrong: at most 5 layers below the CEO, "pure manager" role abolishedCoinbase is the most aggressive layer-slimming case in the sample: its May 2026 reorg compressed the company to at most 5 layers below the CEO and abolished the "pure manager" job category outright.
≤5 layersmaximum layers below CEO/COO (from 2026.05, written into policy)~14%share laid off in 2026.05 (~700 people)15+direct reports a leader may carry (permitted, not required)40%share of day-to-day code that is AI-generated (Armstrong's own figure, 2025.09)
Coinbase
Armstrong: at most 5 layers below the CEO, "pure manager" role abolished- 2025.08Armstrong gave engineers one week to adopt AI coding assistants (GitHub Copilot / Cursor); those who did not were let go.
- 2025.09.03Armstrong said on X that AI-generated code makes up about 40% of Coinbase's day-to-day code, targeting over 50% by October 2025.
- 2026.05Announced layoffs of about 14% alongside a restructuring: the "pure manager" position was abolished in favor of the "player-coach" (managing a team while remaining an active IC); the hierarchy was compressed to at most 5 layers below CEO/COO, with leaders allowed 15+ direct reports (permitted, not required); some teams are piloting "AI-native pods" — one person directing multiple AI agents to do work that used to take an engineer, a designer and a product manager.
OpenAI
A research/applications dual track that lasted only ~14 monthsOpenAI answered scale with a new CEO title while keeping a "dual-head" structure on the research side — but that architecture itself is iterating fast, with visible adjustments by 2026.
~14 monthshow long the "CEO of Applications" structure lasted13executives departed in 2026 (TechCrunch roundup)
OpenAI
A research/applications dual track that lasted only ~14 months- 2025.03The research side formalized "dual heads": Mark Chen became Chief Research Officer, co-leading with Chief Scientist Jakub Pachocki.
- 2025.05.07Former Instacart CEO Fidji Simo took the newly created role of CEO of Applications, reporting directly to Sam Altman; COO Brad Lightcap, CFO Sarah Friar and CPO Kevin Weil, who had reported to Altman, now report to Simo. Altman used the change to concentrate on research, compute and safety, writing on X that the new setup would let him "increase focus on research, compute and safety".
- 2025.09The ~14-person Model Behavior team (responsible since GPT-4 for shaping model "personality" and reducing sycophancy) was folded into the larger Post Training team; its lead, Joanne Jang, left to found OAI Labs, a new team focused on human-AI interfaces.
- 2025.09–2026.04Kevin Weil moved from CPO to VP of AI for Science (announced 2025.09), driving the creation of the OpenAI for Science team; on 2026.04.17 Weil departed and the team was dissolved into the research groups.
- 2026.07.09Fidji Simo, managing a chronic condition (POTS), stepped back from full-time "Applications CEO" to part-time advisor — the "research/applications dual-CEO" structure lasted about 14 months. OpenAI named no single successor, splitting the duties among Greg Brockman (product strategy), CFO Sarah Friar and Jason Kwon; COO Brad Lightcap had already moved to special projects in 2026.04 and left to found a company on 08.11.
Anthropic
No official headcount figure — a deliberately small "research-meets-engineering" teamAnthropic publishes no official headcount and third-party estimates diverge widely; but the consensus across reports is that, relative to its revenue and model capability, the team is kept lean deliberately — not under cost pressure.
1direct reports to the CEO (Fortune, single source)2,300–5,600third-party headcount estimate range (no official figure)
Anthropic
No official headcount figure — a deliberately small "research-meets-engineering" team- HeadcountSources disagree: Revelio Labs puts it between about 3,173 (2025) and 4,020 (2026.03); Fortune reported about 2,300 at end-2025; Tracxn gave about 5,600 in 2026.07. Anthropic publishes no official number, so this report presents the range "2,300–5,600".
- 2026.02.19Claude Code lead Boris Cherny (who started Claude Code as a side project on the Anthropic Labs team in 2024.09) told Lenny's Podcast that Claude Code had grown to about 4% of public GitHub commits, with daily active users doubling in the month before the interview; the team deliberately thins functional boundaries — in his words: "On the Claude Code team, everyone writes code. Product managers write code, engineering managers write code, designers write code, folks in finance write code, data scientists write code."
- 2025.05Judging that it could not sufficiently rule out elevated risk, Anthropic formally activated ASL-3 protections.
- 2025.10.02Rahul Patil was appointed CTO, with co-founder Sam McCandlish becoming Chief Architect; product engineering was folded into the infrastructure and inference teams.
- 2026.02.24Published the fully rewritten v3.0 of its Responsible Scaling Policy (RSP), adding two mechanisms — a "frontier safety roadmap" and "risk reports" — approved by the Responsible Scaling Officer (RSO) and CEO, overseen by the board and the Long-Term Benefit Trust; the dedicated team breakdown (Frontier Red Team, Alignment Science, Trust & Safety, RSP Team) is in the 2024.10 v2.0 announcement.A rare case of an AI company adding process to itself — "safety" built out as a concrete division of teams, not a slogan.
- 2026.06.18Fortune reported Anthropic's unconventional reporting structure: CEO Dario Amodei has exactly 1 direct report (a Chief of Staff), with every other executive reporting to President Daniela Amodei.The opposite end of the spectrum from Nvidia's "60 direct reports" — showing there is no single answer to "flat" even inside AI companies. Fortune, single source.
Side-by-Side Comparison
All called "flattening" — but the path, the rhetoric and the cost differ company to company.
- Under the same "flattening" label, the companies' quantified targets are mutually incomparable: Google counts "managers with fewer than 3 direct reports", Amazon counts the IC ratio, Coinbase counts the number of layers.
- The "typical rhetoric" column shows the different explanations each company gives for the same move — efficiency, culture, speed, AI — the rhetoric itself is information that needs decoding.
- The evidence-strength column at the end is a filter to apply before quoting: official/firsthand statements can be checked directly; single reports and analytical inference are directional at best.
The four companies count four different things — the numbers cannot be added or ranked; Microsoft has only the "fewer layers" language and no published management-cut figure, so the chart shows the four metrics the other three have published. Together they still point the same way: flattening is turning from slogan into verifiable metric.
Dozens of direct reports to the CEO has become a template, and minimum direct-report counts are being written into policy; as spans widen, coordination costs shift onto new roles. Note Gallup's 12.1 is a mean pulled up by a few very large teams — the median is still 5–6.
Table view
| Group | Item | Value |
|---|---|---|
| Direct reports to the CEO | Nvidia · Jensen Huang | 36–60 (his own figure ~60) |
| Direct reports to the CEO | Meta · Zuckerberg's core team | 25–30 |
| Direct reports to the CEO | Anthropic · Dario Amodei | 1 |
| Managers' direct reports: policy floors, permitted ceilings and the industry average | Amazon · minimum reports (internal guidance) | 6 → 8 |
| Managers' direct reports: policy floors, permitted ceilings and the industry average | Coinbase · reports allowed (permitted, not required) | 15+ |
| Managers' direct reports: policy floors, permitted ceilings and the industry average | Gallup · US manager average | 10.9 → 12.1 (2024 → 2025) |
Under the one label "flattening", the actual paths split four ways: structural-reorganization (Google, Meta, OpenAI), layoff-driven (Amazon, Microsoft, X, Block, Coinbase), AI-substitution (Salesforce), and always-flat (Nvidia, Anthropic). The last column is not decoration: every entry opens its source directly, and it flags which conclusions are cross-verified and which still rest on a single report.
| Company | Core action | Quantified target / result | Signature line | Method & cost | Evidence strength |
|---|---|---|---|---|---|
| AI teams merged (Brain+DeepMind, Gemini App → GDM) + management-role cuts + the 2026.08 leadership shake-up | manager/director/VP down 10% (two years); the precise measure: managers with fewer than 3 reports down 35% | "Fewer layers, team reorganization" | Mostly reorgs, with limited layoffs; risk: frequent reorgs breed morale swings | Official / firsthandGoogle Blog · 2026-08 management roles and AI-org changes; Single report / datasetCNBC (2025-08-27) managers with fewer than 3 reports down 35% | |
| Amazon/AWS | IC:manager ratio +15%, bureaucracy mailbox → then two more layoff rounds 2025.10–2026.01 | 1,500 complaints → 455 process changes; still followed by ~30,000 corporate layoffs | "Fewer managers = fewer layers"; later attributed to "culture", not AI | First flattened by widening spans, then laid off at scale anyway; cost: visible tension between narrative and action | Official / firsthandAmazon · Andy Jassy memo manager ratio and the anti-bureaucracy program; Independently corroboratedCNBC (2026-01-28) · The Seattle Times layoff scale |
| X / xAI | Physically deleted the organization: ~80% of employees gone, engineers reporting straight to Musk | 7,500 → about 1,500 people | "extremely hardcore" | Extreme compression; costs: ad-revenue decline, brand and compliance risk | Independently corroboratedThe Verge · CNBC (2025-03-28) people and org changes; no official org-structure account |
| Nvidia | Always flat: ~60 direct reports, almost no 1:1s, no long-term plans | ~60 direct reports to the CEO (his own figure; media accounts 36–60) | "Move at the speed of light" / "extreme co-design" | Culture-borne flatness; cost: high intensity, attrition of those who cannot adapt | Official / firsthandLex Fridman Podcast #494 · HBR IdeaCast · Jensen Huang Huang's own interviews; direct-report counts vary by source (36–60) |
| Meta | MSL merged then split, three Reality Labs rounds, the AI-native pod pilot | ~21,000 laid off in 2023; ~10,000 more across three rounds in 2026 | "Flatter is faster" / "one very talented person is enough" | Mass layoffs + repeated reorgs; structural stability in doubt | Official / firsthandMeta · Year of Efficiency · Meta IR · Q4 2025 transcript layoffs and org changes; Single report / datasetEngadget (Reuters) · 2026-08-26 Project OT (secondary accounts of a single Reuters report) |
| Microsoft | 15,000+ laid off across the year in the name of "span of control" | 2025 layoffs 15,000+ / 228,000 employees | "Fewer layers, fewer managers" | Real layoffs reaching IC engineers; accused of cost-cutting under the flattening banner | Independently corroboratedCNBC (2025-05-13) · CNBC (2025-07-02) layoff scale; Single report / datasetBusiness Insider (2025-05) the internal "span of control" target |
| Salesforce | Froze engineering hiring; agents replacing support roles | AI does 30–50% of internal work; customer support down from 9,000 to about 5,000 | "I need less heads" | Headcount tracks agent capacity directly; sales roles actually growing | Official / firsthandFortune · Marc Benioff interview the CEO's own interviews; the work-share figure is his own account |
| Block | One big layoff in 2026.02 → the theory essay in 2026.03 | ~40% laid off (~4,000 people); stock closed up ~18% the next day | "From Hierarchy to Intelligence" | Layoffs first, theory after; the DRI "responsibility without authority" worry has no named source yet | Official / firsthandBlock · From Hierarchy to Intelligence the company essay; Independently corroboratedForbes (2026-04-01) · Fortune (2026-04-02) the layoff and reorg backdrop |
| Coinbase | Abolished the "pure manager" role in favor of player-coach | ~14% laid off; hierarchy compressed to ≤5 layers below the CEO | "player-coach" + AI-native pods | The most aggressive layer-slimming in the sample; cost: dense compression of frontline staff | Official / firsthandBrian Armstrong · X the CEO's own statements; Single report / datasetFortune (2026-05-05) reorg details |
| OpenAI | Created a CEO of Applications; dual heads on the research side | COO/CFO/CPO moved to report to the new CEO | "Increase focus on research, compute and safety" | The structure lasted only ~14 months before changing; no single successor on the applications side | Official / firsthandOpenAI · Fidji Simo announcement · Sam Altman · X the appointment and reporting lines; Independently corroboratedCNBC (2026-07-09) · TechCrunch (2026-07-09) the later adjustments |
| Anthropic | Research-engineering integration + RSP-institutionalized safety process | headcount range 2,300–5,600 (no consistent measure); the CEO has exactly 1 direct report | — | Deliberately small teams; cost: doubts about whether scale can keep growing | Official / firsthandAnthropic · RSP v3.0 · Lenny's Podcast · interview transcript safety process and team practices; Independently corroboratedRevelio Labs · SaaStr (2026-04) headcount estimates; Single report / datasetFortune (2026-06-18) the CEO reporting structure |
The Real Practice: How AI-Era Companies Actually Run
Removing layers does not automatically make an organization faster. How decisions travel, who gets to make the first draft, where the rules live, and who owns verification and release all have to be redone together. "Shipping" here does not mean bypassing professional review: it means the person who understands the problem best uses agents to produce a working version first, with product, design and engineering then reviewing, verifying and releasing.
- Decision chains get shorter by moving communication into collective settings (Nvidia, Meta's core team) — not by communicating less.
- The right to make the first draft is shifting to whoever understands the problem best. After interviewing a dozen-plus fast-growing startups, Anthropic distilled the practice as Everyone ships; professional review and controlled release still follow.
- Rules move out of managers' heads into shared environments (convention files, Skills, design systems); with the "pure manager" abolished, those who remain must build.
Put communication in collective settings instead of filtering it layer by layer
Nvidia: Huang describes ~60 direct reports and no routine one-on-ones — "no conversation has just one person in it; we pose a problem and everyone solves it together — that is extreme co-design". Official / firsthandLex Fridman Podcast #494
Meta: Zuckerberg directly runs a core team of about 25–30, with "no fixed-cadence one-on-ones; I am talking with these people all the time". Official / firsthandStripe Sessions 2025
Google: from 2026.08, Gemini models, frontier research, the Gemini App and developer teams sit under one SVP reporting directly to Pichai. Official / firsthandGoogle Blog · 2026-08
Coinbase: at most 5 layers below CEO/COO, with leaders allowed 15+ direct reports (permitted, not required). Single report / datasetFortune (2026-05-05)
Block: Dorsey calls the ideal "no layers at all — all 6,000 people reporting to me", planning to compress from 5 layers to 2–3 within a year. Official / firsthandSequoia Podcast · transcript · 2026-04
Anthropic (the counterexample): CEO Dario Amodei has exactly 1 direct report (a Chief of Staff), with all other executives reporting to President Daniela Amodei. Single report / datasetFortune (2026-06-18)The common thread: what gets shortened is not headcount but the number of hand-offs between the front line and the decision-maker; "flat" by itself is not the only answer.
Give 0 → 1 prototyping to whoever understands the problem best
AI-native startups: an official guide on Claude's site, written by Michael Segner from interviews with a dozen-plus fast-growing startups, distills "the person closest to the problem ships the first version" as working principle number one: Everyone ships. Crosby's lawyers edit the product directly, Parahelp's non-technical staff submit UI changes, and Heidi's CEO frames it as solving the "broken telephone" problem. Official / firsthandClaude by Anthropic · 2026-08-20
Anthropic's Claude Code team: lead Boris Cherny's words — "everyone on the team writes code: product managers, engineering managers, designers, finance, data scientists". Official / firsthandLenny's Podcast · interview transcript
Coinbase: some teams pilot AI-native pods — a few people plus multiple agents. Single report / datasetFortune (2026-05-05)
Meta Reality Labs: the ~1,000-person developer-tools team is also reported to be piloting AI-native pods. Single report / datasetBusiness Insider (2026-03)
Meta Project OT: per Reuters, the AI-native reorg had assessed cutting some teams by up to 60% and its second wave was canceled when output fell short; internal data showed code changes +220% year over year, but shipped features only +36%, and major incidents +40%. Single report / datasetEngadget (Reuters) · 2026-08-26Writing more code is not delivering more value — which is exactly why "trust, but verify" exists.
Move the norms out of managers' heads into a shared environment
AI-native startups: invariant architecture rules and safety boundaries go into CLAUDE.md (a project convention file the agent reads at every start); team practices are shared as Skills (reusable process instructions); MCP (an open protocol connecting tools and data) or CLIs plug agents into real systems. Official / firsthandClaude by Anthropic · 2026-08-20
Amazon: the 16 Leadership Principles are tied directly to performance reviews; the "bureaucracy mailbox" logged ~1,500 complaints and drove 455 process changes. Independently corroboratedFortune (2025-09-17) · CNBC (2025-09-16)
Anthropic: Responsible Scaling Policy (RSP) v3.0 writes the safety process into institution — RSO and CEO approval, board and Long-Term Benefit Trust oversight. Official / firsthandAnthropic · RSP v3.0
Coinbase: CEO Brian Armstrong said the company gave engineers one week to adopt AI coding tools. Official / firsthandBrian Armstrong · X
The "pure manager" is abolished; those who stay must build
Google: managers with fewer than 3 direct reports fell 35% in a year, many moving to non-managing specialist roles (Individual Contributor, IC). Single report / datasetCNBC (2025-08-27)
Block: deleted the word "manager" from the hierarchy in favor of three roles — IC, Directly Responsible Individual (DRI), and the ship-while-leading player-coach; a DRI takes named ownership and can pull in resources ad hoc, with no standing management layer. Official / firsthandBlock · From Hierarchy to IntelligenceSingle report / datasetBlock Q2 call / Yahoo Finance · 2026-08-05
Coinbase: abolished the "pure manager" role; managers must also carry real delivery work. Single report / datasetFortune (2026-05-05)
Meta Reality Labs: the ~1,000-person developer-tools team is reported to be piloting AI Builder / AI Pod Lead / AI Org Lead. Single report / datasetBusiness Insider (2026-03)
Across companies: Chief of Staff postings more than doubled in a year, with salaries reaching up to $400K — though Ladders puts the more common high end at $250–300K, with only occasional cases above. Single report / datasetBloomberg (2026-07-16)
The traditional delivery chain transcribes the problem from role to role; the AI-native path lets the person who understands the problem best produce a working version first, which professional teams then review, complete and release. What shrinks is the distance from idea to verifiable value — not the professional division of labor.
How to read this figure: left to right there are still four gates — the change is only at the start: the first draft goes from "transcribed requirements" to "a working version, directly". Professional review and controlled release remain.
Everyone ships: opening up 0 → 1, not abolishing specialization
After interviewing a dozen-plus fast-growing startups, Anthropic distilled the practice as Everyone ships: whoever understands the problem best can use agents to produce a working version first; product, design and engineering review still follows.
Hand the mechanical 80% to agents
"Agents take over the mechanical 80% of the lifecycle, and engineers spend their time on the cases that genuinely need judgment." First-pass code review, bug triage and routine migrations belong in this bucket.
Trust, but verify
"You cannot automate a process unless you have reliable means to monitor and verify the outcome." Tests, code review and permission gates are not abolished — they become the precondition for decentralized delivery.
Build for rebuilding
"Model capability keeps shifting under these teams' feet, so almost nothing is treated as permanent." Processes and tools are designed to be redone at any time, not to be done right once.
Prototype → dogfood → productize
"Building with AI is how these startups make disruptive AI products — the flywheel at the core of their process." Get the idea running, use it yourselves first, then decide whether to productize.
Where Did Middle Managers Go: Five Paths
Once the layers are gone, where did the people go? Combining official accounts from Google / Meta / Amazon / Coinbase with reporting from BI / CNBC / Guardian / Fortune, we group the publicly traceable destinations of former middle managers into five paths. One distinction first: whether a capability still has value, and whether a role or a person survives a reorganization, are not the same question.
- What gets compressed first is the middle layer's "information routing": pure relaying is the easiest to replace, while frontline managers who understand the business and its customers, make judgment calls and execute personally remain scarce.
- The five paths are not mutually exclusive exits but a sorting funnel: the same managers may go IC first, then double as player-coach, with a few entering the Chief of Staff track.
- But who stays is not a pure test of ability: cost targets, business priorities, reporting relationships and internal politics all shape who survives a reorganization.
The five destinations are not mutually exclusive exits but a sorting funnel; for the individual, the test reduces to one question — is your core work "relaying" or "judging"?
None of the five paths has published share data; band width reflects our judgment of relative prominence, not statistics
Table view
| Destination | Evidence |
|---|---|
| Down to IC | Google: many managers stayed as ICs |
| Retitled: AI Builder / AI Pod Lead / AI Org Lead | Meta RL (single source), Block (IC/DRI/player-coach) |
| Transformed: orchestrators directing AI | HBR: verifying AI output, coaching teams |
| Up: Chief of Staff | Postings 2×+ in a year (Revelio) |
| Out | Amazon / Meta / X / Block layoffs |
35% and ≤5 layers describe the force with which layers were removed; $400K and 2× suggest the need for coordination did not disappear and may be concentrating into a few highly paid roles — but with no base or methodology disclosed, they stand only as a signal of rising demand. What was removed is the layer — the need to coordinate remains.
Cutting layers is meant to save management cost — but a company that removes its middle layer must fill the coordination gap with tools and named-ownership mechanisms, or it is simply shifting the cost onto whoever remains.
Relaying up and down, summarizing and filtering, frontline coaching, layered gatekeeping — this work does not vanish with the role
Cross-team coordination may concentrate in a few hands: Chief of Staff postings doubled in a year (2×+), salaries up to $400K (common high end $250–300K); managers run more people, frontline coaching shrinks
The player-coach builds while leading — criticized by name in a Fortune column as "misunderstanding the value of good managers"; the Block DRI "responsibility without authority" worry has no named source yet, recorded as observation
In Meta's experiment code changes +220%, user-visible features only +36%, major incidents +40% — surging output bought no shipped value
"Who pays" is our synthesis, not official company language; the numbers here are shorthand for relative change, not same-basis comparison.
Demoted sideways into an IC
Google: when 10% of management roles were cut in 2024.12, a spokesperson was explicit that some "transitioned to individual contributor roles"; by 2025.8 the precise figure was out — managers with fewer than 3 reports down 35%, many staying on as ICs. Meta: Zuckerberg's Year of Efficiency language was literally to "ask many managers to become individual contributors". Amazon: internal guidance shows the transition often comes with more direct reports, fewer senior openings, even pay cuts — going IC is not a lateral move; it carries a real price.
New title, redefined duties
2025–2026 brought outright rewrites of the manager role: parts of Meta Reality Labs retitled to AI Builder / AI Pod Lead / AI Org Lead, organized into small pods by outcome rather than function; Block (Dorsey) deleted the word "manager" in favor of IC / DRI / player-coach (still rolling out as of 2026.08); Coinbase (Armstrong) ran the fullest version in May 2026 — about 14% laid off, the pure-manager role abolished, everyone moved to "player-coach", the hierarchy compressed to at most 5 layers below the CEO, and some teams piloting AI-native pods. The common thread: organizations no longer keep management roles that only coordinate and never deliver; managers must also build.
Becoming a Chief of Staff
Part of the middle layer did not disappear — it was moved next to the decision-makers after the slimming. Bloomberg (2026.07), citing Revelio Labs, reports Chief of Staff postings more than doubled over the prior year; salaries reach up to $400K, though Ladders puts the more common high end at $250–300K, with only occasional cases above. Organization-design commentators call the boom "the bill for the Great Flattening" — but the public reporting discloses no base, geography or precise growth rate: it shows demand for the role is rising, not that the rise was caused by AI or flattening, nor that these roles have effectively absorbed the coordination work that was removed.
From "coordinator" to "orchestrator"
For those who stay in management, the job itself changed: less relaying, summarizing and progress-chasing (now AI's work), more coaching, strategy translation and building the conditions for human-AI collaboration — an upgrade from coordinator to orchestrator. The ironic reality: the managers who were just cut are now the ones tasked with driving AI adoption — they are precisely the execution handle AI rollout depends on.
The role disappears; they leave
Google's "role eliminations", Meta's Year of Efficiency and its three 2026 rounds, X's ~80% reduction, Amazon's ~30,000 layoffs in 2025.10–2026.01 — all show that some roles had no transition exit and simply left the organization.
What's eliminated is the function, not the person
The five paths show that what gets compressed first is the middle layer's "information routing": relaying up and down, summarizing and filtering, chasing process.
Capabilities still scarce: understanding the business and its customers, cross-domain judgment, leading a team while executing personally.
Two things to keep separate: a capability having value does not make a role safe. Reorganizations are not pure tests of ability — cost targets, business priorities, shifts in power and reporting lines, and internal politics all shape who stays.
Being laid off cannot be read backwards as proof the capability had no value — nor the reverse.
How the Industry Reads This Shift
Beyond what companies themselves are doing: how analysts and academia read this shift.
- Gartner forecasts that through 2026, 20% of organizations will use AI to eliminate more than half of their current middle-management positions; Morgan Stanley treats organizational flatness as a modelable financial variable.
- The shared judgment of management scholarship (HBR, Gallup) is that the middle role is being reshaped, not eliminated — and is now carrying the extra burden of verifying AI output and coaching teams.
- The criticism is explicit: flattening taken too far stunts talent growth and weakens leadership development; the player-coach model has drawn named criticism.
The four lenses answer different questions and do not form one causal chain: forecasts ask whether structure will change, capital markets price what the change is worth, management research tracks who catches the work, and the critics test whether flattening has gone too far.
One shift, four scales of observation
Do not blend forecasts, financial models, management research and criticism into one conclusion; first see what question each is answering.
Will org hierarchies change?
Gartner's forecast: through 2026, 20% of organizations will use AI to eliminate more than half of current middle-management positions.
What is the change worth?
Morgan Stanley, using Amazon as the sample, modeled management compression as $2.1–3.6 billion of annual cost headroom.
Who catches the work that was removed?
HBR, Gallup and MIT TR Insights track manager load, coaching quality, and the new processes human-AI collaboration requires.
Will flattening go too far?
BI, Fast Company, Fortune and the Guardian point to talent growth, leadership development and the limits of player-coach.
The structural forecast
"Through 2026, 20% of organizations will use AI to flatten their organizational structure, eliminating more than half of current middle management positions." (annual prediction, IT Symposium/Xpo, 2024.10) Widely cited by the Guardian, Forbes, Fast Company and others, it has become the industry's benchmark call.
The financial lens
Taking Jassy's "IC ratio +15%" as the assumption, it modeled roughly 13,800 trimmable Amazon manager positions, freeing $2.1–3.6 billion a year (2024.10 research note, via CNBC) — the first time Wall Street treated "organizational flatness" as a modelable financial variable.
Management scholarship and industry research
HBR's "How AI Is Redefining Managerial Roles" (2025.07) and "AI Adoption Is Overloading Your Middle Managers" (2026.06) share one judgment: the middle role is being reshaped, not eliminated, and now carries the extra burden of verifying AI output and coaching teams. Gallup's 2026.01 data shows the average US manager's direct reports rose from 10.9 to 12.1, while manager quality explains about 70% of the variance in employee engagement. MIT Technology Review Insights (commissioned research with Ema) argues that when the "actor" is an AI agent operating across systems at machine speed, both the organization and the technology stack designed for human workflows need rebuilding.
"The Great Flattening may have gone too far"
Business Insider's May 2025 overview "Big Tech is crushing middle managers. Some fear the great flattening has gone too far" warns that badly managed flattening stunts talent growth and weakens leadership development; Fast Company (2026.06) cites a Korn Ferry survey in which 41% of employees say their company has already cut layers, and marshals Gallup data to argue "the data says don't"; a Fortune column (2026.05) criticizes the player-coach model by name as "misunderstanding the value of good managers"; the Guardian (2026.05) records the view from managers who were cut. The rapid growth in Chief of Staff postings is what the org-design blog Business Model Analyst calls "the bill for the Great Flattening".
Conclusions & Outlook
- Routing-type management roles are being replaced by models and agents — a structural, not cyclical, change; but the new structures themselves keep iterating, with no stable end state.
- The new architecture = fewer layers × wider spans × human-AI mixed teams; the costs are coordination vacuums and "responsibility without authority", and rising Chief of Staff demand is a signal that coordination cost may be relocating rather than disappearing.
- For Chinese companies: build the information infrastructure before cutting layers; transform management roles rather than simply cutting them; use verifiable quantitative metrics to make flattening measurable and auditable.
Middle management is AI's first structural casualty
Google cutting 10% of management roles over two years and 35% of sub-3-report managers in one; Amazon's IC ratio +15%; Microsoft's 15,000+ layoffs; Gartner's "20% of organizations will cut over half their middle layer" — the evidence converges on one conclusion: routing-type management roles are being replaced directly by models and agents. This is not cyclical layoffs; it is a paradigm switch in organizational technology.
New architecture = fewer layers × wider spans × human-AI blend
The emerging standard shape: 3–5 layers; dozens of direct reports to the CEO (the Nvidia / Meta core-team pattern); managers carrying 12–15+ reports (Gallup mean 12.1, Coinbase allowing 15+); AI agents embedded in teams (Salesforce's 30–50%). The minimal unit of "organization" shifts from "person/manager" to "person + agent cluster".
Risk: a coordination vacuum and "responsible but powerless"
Flattening is not free: wider spans mean less frontline coaching (BI's "gone too far" warning); rising Chief of Staff demand signals coordination cost relocating rather than disappearing (base and methodology undisclosed, so not proof of causation on its own); player-coach / DRI models have drawn named criticism (the Fortune column: "misunderstands the value of good managers"), and Meta's Project OT shows surging code output does not necessarily convert into shipped value. A company that cuts its middle layer must fill the coordination gap with AI tools plus named-ownership mechanisms — or it is just shifting the cost onto individuals.
The new structures are themselves iterating fast
No "new organizational form" is a stable end state. OpenAI's research/applications dual-CEO structure lasted about 14 months; Meta's MSL was split 7 weeks after formation, followed by three rounds of contraction in H1 2026; Amazon slid from "flattening without layoffs" to "two rounds, 30,000 people". Flattening is a continuous organizational experiment, not a destination reached once.
Implications for Chinese enterprises
① Build the "information infrastructure" before cutting layers — Amazon backstopped with the bureaucracy mailbox plus process governance; ② transform management roles rather than simply cutting them: managers either move up into strategy/resource orchestration or down into coaching/cross-team coordination; ③ establish verifiable, auditable metrics of the management-role-ratio kind (see Google's precise "fewer than 3 direct reports" definition) so flattening is measurable and acceptance-testable rather than sloganized layoffs; ④ keep "motive" and "action" consistent in external communication — Amazon's official denial that its layoffs relate to AI or cost is a cautionary example: muddled attribution erodes public trust.
From "cutting layers" to "changing how it runs"
① In your organization, how many roles have "relaying" rather than "judging" as their core work? ② Does the person who best understands the customer's problem have the authority and the tools, today, to produce the first working version? ③ Where are your rules written — in managers' heads, or in a shared environment everyone can invoke (convention files, the design system, the toolchain)? ④ Each time you remove a layer, do you install verifiable gates (tests, review, permissions) to replace the checking that layer used to do?
Video sources
20 publicly viewable videos/interviews, first-hand or close secondary material on each company's org changes — 12 carried over from the first edition (title/channel/date re-verified), 8 new in this revision.
Video sources
(A) Firsthand executive interviews and official video · 11 items
Leading the teams behind the world's fastest AI — Jensen Huang
Nvidia's management philosophy
Open on YouTube
Amazon CEO Andy Jassy on Agility, AI Strategy, and the Changing Role of Managers
Amazon's IC ratio and anti-bureaucracy reform
Open on YouTube
How Sundar Pichai is rethinking Google for the AI era
The rebuild of decision rights before Google's 2026.08 reorg
Open on YouTube
The history and future of AI at Google, with Sundar Pichai
Background on the evolution of Google's AI organization
Open on YouTube
Google CEO Sundar Pichai on Gemini, Self-improving AI, and World Models
The technical-organization logic after the DeepMind merger
Open on YouTube
AI will change the workforce, says Amazon CEO Andy Jassy
Amazon's position on AI and workforce structure
Open on YouTube
The Leadership Principles Explained by Andy Jassy (full film)
Background on the 16 Leadership Principles entering performance reviews
Open on YouTube
Scott Answers: Is AI Replacing Middle Management? | Office Hours
The industry-critic perspective
Open on YouTube
(B) Supplementary interviews and reporting · 8 items
Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution (#494)
Huang details the "extreme co-design" philosophy: no 1:1s, ~60 direct reports
Open on YouTube
Head of Claude Code: What happens after coding is solved | Boris Cherny
How Anthropic's Claude Code team runs research-engineering as one
Open on YouTube
A conversation with Mark Zuckerberg
Zuckerberg's 25–30 person core-team philosophy, compared with the Huang model
Open on YouTube
Dario Amodei — "We are near the end of the exponential"
Anthropic's scaling logic: compute investment vs team size
Open on YouTube
Amazon targets middle managers in mass layoffs, memo suggests more cuts coming as AI thins Big Tech
How Amazon's 2025.10 layoffs aimed squarely at middle management
Open on YouTube
Meta AI Chief Wang on Winning the Race in AI
The post-split state of MSL — firsthand material from Alexandr Wang
Open on YouTube
The Man Behind Google's AI Machine
The baseline before Hassabis stepped down as GDM CEO
Open on YouTube
Marc Benioff Predicts Half of Conversations Will be With AI Agents Next Year
The primary source for the "I need less heads" quote
Open on YouTubeReferences
All are publicly accessible news reports, official company statements or first-hand interviews (YouTube sources are listed separately in Chapter 10 and not repeated here); items marked "primary" are official company/individual primary sources. Grouped by company/topic for cross-checking against Chapter 4.