The Work Lie

The measure of intelligence is the ability to change. ~ Albert Einstein

The curious task of economics is to demonstrate to men how little they really know about what they imagine they can design. ~ F.A. Hayek

Work is the curse of the drinking classes. ~ Oscar Wilde

Outcome Vault Locked

"Knowledge is the only asset that grows when shared, but strategy is only for those who protect it."

Registered Email

Access Key

Preface: A Man on a Bike

There is a man on a bicycle right now. He is moving through traffic, delivering food on GrubHub, DoorDash, whatever app decided to give him a shift today. In his head, there are songs. There are businesses. There are entire worlds that exist only as electricity between neurons, unbuilt, unwritten, unsung, because the act of surviving keeps getting in the way of the act of living.

This man is not unusual. He is the norm.

And the cruel irony is this: the technology that was supposed to free him, Artificial Intelligence, is being sold back to him as a better way to do more work.

This is the lie. This document is about that lie.

I. What They Promised

Let us be precise about what was actually said.

In 2013, Oxford economists Carl Benedikt Frey and Michael Osborne published "The Future of Employment", projecting that 47% of U.S. jobs were at high risk of automation within two decades. This wasn't fringe thinking, it was cited by the World Economic Forum, the IMF, Barack Obama, and every major tech publication on earth.

In 2016, the White House published a report warning that AI would displace millions of workers. Universities launched task forces. Sociologists debated Universal Basic Income. Philosophers revisited Keynes' 1930 essay "Economic Possibilities for Our Grandchildren" where he predicted a 15-hour work week by 2030.

Sam Altman, CEO of OpenAI, in 2021:

AI will create enough wealth that we could give everyone $13,500 per year. ~ Sam Altman

Elon Musk, Demis Hassabis, Geoffrey Hinton, the architects of this technology, all warned of a world where human labor becomes optional.

They told us the machines were coming for the jobs.

Now they are selling you tools to do the jobs better.

II. The Exact Moment the Lie Became Undeniable

The pivot did not happen slowly. It happened across 14 compressed months, 2025 into 2026, in plain sight. The models crossed a threshold. The products went in the opposite direction.

Here is the factual timeline, verified:

February 24, 2025: Anthropic releases Claude 3.7 Sonnet, the first "hybrid reasoning" model, with an extended thinking mode that achieved 70.3% on SWE-bench Verified with custom scaffolding. SWE-bench is not a toy benchmark. It tests resolution of real GitHub issues in real codebases. 70% is not "assistant-level." It is senior-developer-level on real engineering tasks. The previous best was 49%.

May 22, 2025: Anthropic releases the Claude 4 family, Claude Opus 4 and Sonnet 4, with explicit focus on "extended thinking with tool use" and agentic workflows. By November 2025, Claude Opus 4.5 is in production. Anthropic is shipping "Cowork", a GUI-based agent that operates computers autonomously.

August 7, 2025: OpenAI releases GPT-5. The announcement is careful. It unifies reasoning and multimodal capability into a single system, eliminating the need to choose between specialized models. It shows "significant reductions in hallucinations." The press release does not say: this model can replace most knowledge work. But the benchmarks do.

November 18, 2025: Google releases Gemini 3 Pro, described as their most capable model for multimodal understanding and agentic tasks. It introduces "Deep Think" mode. Google's own data: Gemini 3 Pro outperforms all competing models on 19 out of 20 standard benchmarks.

December 17, 2025: Gemini 3 Flash goes live as the new default in the Gemini app. OpenAI ships GPT-5.2 and GPT-5.2-Codex in the same month.

February 5, 2026: Anthropic releases Claude Opus 4.6, 1 million token context window, "Agent Teams" for parallel task execution, enhanced multi-step planning and debugging. Fourteen days later, Claude Sonnet 4.6 follows (Feb 17), delivering Opus-class performance at a fraction of the cost. Also now in GitHub Copilot.

February 19, 2026: Google releases Gemini 3.1 Pro in public preview. The benchmarks are staggering: 80.6% on SWE-bench Verified (autonomous software engineering), 94.3% on GPQA Diamond (PhD-level science), 77.1% on ARC-AGI-2, more than double the score of Gemini 3 Pro. LiveCodeBench Elo: 2887. Gemini 3.1 replaces Gemini 3 Pro across GitHub Copilot integrations by March 26, 2026.

March 5, 2026: OpenAI releases GPT-5.4 with native computer use, the model can autonomously navigate desktop applications, click buttons, fill out forms, and interpret screenshots to execute multi-step workflows without human direction. On GDPval, OpenAI's own benchmark for real-world professional job tasks, GPT-5.4 scores 83.0%, up from 70.9% on GPT-5.2. On OSWorld desktop automation, it exceeds 75%, surpassing the human expert baseline of 72.4%. On March 17, GPT-5.4 mini and nano ship for high-volume deployment.

Let that full timeline register. In 14 months, every major lab shipped generational advancement after generational advancement. The models entering 2026 are not incrementally better than what entered 2025. They autonomously operate computers. They resolve real engineering tickets at senior-developer rates. They score above human experts on desktop task automation.

And the commercial wrapper placed around every single one of these releases:

Claude for Work, Anthropic's enterprise push, now with "Agent Teams"
Microsoft 365 Copilot, productivity suite integration
Google Workspace AI, document and email automation
ChatGPT Enterprise, corporate seat licensing
GPT-5.4-Codex, developer productivity tooling
Sonnet 4.6 in GitHub Copilot, developer productivity tooling
Gemini 3.1 in Google Workspace, document productivity tooling

The pattern is exact and unbroken across every single release. Every time capability advanced, the product announcement went toward enterprise productivity. Never toward human liberation.

III. The GDP Benchmark Nobody Talks About

Here is the economic data that belongs beside every AI press release:

U.S. labor productivity growth has averaged 1.4% per year since 2010. During the same period, enterprise software spending tripled, from $250 billion to over $800 billion annually. We spent three times more on tools to make knowledge workers more productive, and they got marginally more productive.

Goldman Sachs estimates that hundreds of millions of jobs are now "exposed" to AI automation, meaning core tasks within those roles can be automated by current-generation systems. Their projection for resulting unemployment: a mild ~0.5% increase. Because, their argument goes, the labor market will absorb the transition.

What that projection omits: 40% of employers surveyed in 2025 explicitly stated they plan to reduce headcount where AI can automate tasks. Not eventually. As a near-term operating decision.

The World Economic Forum reports that AI is already moving "beyond simple augmentation to full automation" in legal research, financial modeling, marketing, and software development. Entry-level white-collar roles are specifically flagged as "particularly vulnerable", because they were the training ground for junior professionals who now have no rung to step onto.

The McKinsey Global Institute's own estimate: generative AI could add $2.6 to $4.4 trillion in annual economic value. Their 68-page report does not once use the phrase "post-work." The phrase "reduce employment" appears twice, in footnotes, immediately followed by reassurance that new jobs will emerge.

The same firm that charges $10 million to tell companies how to restructure is the firm telling us that AI will create jobs rather than eliminate them.

That is not analysis. That is a business model protecting itself in print.

The honest productivity paradox benchmark: in 2025, there was a measurable disconnect between strong GDP growth and historically low job growth in highly AI-exposed sectors. Economists called it the "Productivity Paradox 2.0." GDP was rising. Investment in AI infrastructure was contributing massively. But measured output per worker was, by official statistics, barely moving.

The most likely explanation: the productivity gains are real, but they are being captured entirely by the companies that control the AI, not distributed to the workers using it, and certainly not to the workers being replaced by it.

The wealth is going up. The labor share is going down. The tools that are automating the work are being sold back to the workers as productivity software.

This is not a paradox. It is a transfer.

IIIb. The Real Job Test: AI Fails 97.5% of Actual Freelance Work

Here is the data point the labs do not put in their press releases.

The Remote Labor Index (RLI) study, conducted by the Center for AI Safety (CAIS) and Scale AI, evaluated six leading AI systems, including models from OpenAI, Anthropic, and Google, on 240 real-world projects sourced directly from Upwork: game development, 3D modeling, architectural planning, data analysis, and more. Real briefs. Real client standards. Real deliverables.

The result: the best-performing AI models achieved an "automation rate", the percentage of projects completed to a standard a reasonable client would accept, of 2.5% or less.

Models tested: Manus, Grok 4, Claude Sonnet 4.5, GPT-5, Gemini 2.5 Pro. All state-of-the-art. All failing 97.5% to 99% of real-world freelance jobs when operating autonomously.

The specific failure modes were not exotic:

Incomplete outputs: unfinished code, missing visual assets, truncated work
Technical errors: corrupted files, unusable formats
No quality control: inability to catch and correct its own mistakes
Multi-step collapse: the model could not sustain reasoning across the full duration of a complex project

This is the benchmark nobody is reading out on stage at the AI summit.

The labs benchmark their models on SWE-bench (structured code tasks with clear pass/fail criteria) and GPQA Diamond (multiple choice science questions). These are academically rigorous and practically narrow. They measure performance on well-defined problems with verifiable solutions.

Real freelance work is not well-defined. The client doesn't give you a rubric. The deliverable isn't a multiple-choice answer. The 3D model needs to feel right. The game needs to be fun. The analysis needs to surface the insight the client didn't know they were looking for.

On controlled benchmarks: AI looks superhuman. On actual human work: AI fails 19 out of 20 jobs.

This gap is the single most important data point in the entire AI industry, and it is almost never cited by the companies selling AI productivity tools.

Because if your customers understood that the AI you are selling them can't complete most of the jobs a $25/hour Fiverr freelancer can, they might stop paying $30 per seat per month for it.

The knowledge that should terrify every CEO currently running an "AI transformation" program: the transformation tools they are deploying fail the majority of real professional tasks. The cost savings aren't materializing because the AI isn't finishing the work. The human is still finishing the work. They're just also managing the AI's incomplete outputs on top.

That's not augmentation. That's two jobs for the price of one.

IV. The Incumbents: Named, Examined, Sentenced

Let us stop being polite about which companies are running the protection racket.

The Big 4 Accounting & Consulting Firms

PwC, Deloitte, EY, KPMG collectively employ over one million people and generate combined revenues exceeding $230 billion per year.

Their core products: audit, tax advisory, management consulting, technology advisory. Translation: they verify numbers, optimize tax structures, and tell large companies how to run their businesses.

PwC charges $500 to $1,000 per hour for senior partners. Deloitte's "transformation" practices charge seven figures for engagements that routinely overrun and underdeliver. EY was fined $100 million by the SEC in 2022 because hundreds of its audit professionals cheated on ethics exams, the exams designed to ensure they were qualified to certify that corporations were telling the truth.

Let that land. The company that certifies financial integrity cannot certify its own people's integrity.

An AI system with access to a company's financial data can perform the core of an audit in hours, with perfect consistency, without the ability to cheat on ethics exams because it doesn't need to take them.

An AI system with access to tax code, transaction history, and regulatory filings can produce tax advisory that matches or exceeds Big 4 output at a fraction of the cost.

The reason corporations still pay $230 billion a year for this is not that the Big 4 are better. It is that they are legally embedded. Regulatory frameworks in most jurisdictions require human-signed audits from licensed firms. The moat is not intellectual. It is regulatory.

That moat is political. And politics change.

The Consulting Giants

McKinsey & Company, Boston Consulting Group, Bain & Company, the "MBB", generate a combined $40+ billion annually telling executives that which any well-configured AI can derive from public data, internal data, and a clear prompt.

McKinsey's revenue model depends on three things:

The Harvard/Yale/Stanford credential pipeline that creates the perception of elite intelligence
The emotional comfort of being able to blame an outside firm when strategy fails
Relationships with C-suite executives that are maintained through expensive dinners and golf

Item 1 is collapsing. Item 2 is a psychological service, not an economic one. Item 3 is a network effect that survives until it doesn't.

The Advertising & Creative Empires

WPP, Publicis, Interpublic, Omnicom, Havas, the "Big 5" holding companies that control most of global advertising, generate $70+ billion annually.

Their product is creative output: campaigns, brand strategy, media buying, video production. They own agencies like Ogilvy, BBDO, JWT, Grey, Leo Burnett, brands that have made creative work feel like an art form requiring decades of cultivated taste.

It is not. It is a pattern-matching pipeline wrapped in expensive leather chairs.

The Ogilvy model: spend 6 months in "discovery," interview executives, run focus groups, produce a strategy document, make a video, charge $2 million. The video runs. The awareness lift is marginal. The agency earns its retained fee.

The AI-native model: Firecrawl the client's site and their top 3 competitors. Pull brand sentiment data from social APIs. Run a multi-agent pipeline to produce creative concepts. Use Veo for video generation. Use Remotion for animation and dynamic brand elements. Deliver a full campaign package in 48 hours. Charge $10,000 flat.

The output is equal. Frequently better, because the AI doesn't have a preferred vendor relationship with the director it always uses.

Financial Management: The Asset Management Illusion

BlackRock manages $10 trillion in assets and charges basis points, fractions of a percent, on every dollar. That sounds small. At $10 trillion, a single basis point is $1 billion.

The core product of asset management: portfolio construction, risk analysis, market research, rebalancing. BlackRock's Aladdin platform, their risk analytics system, is already software. They already automated the analytical layer of asset management.

What they haven't automated is the relationship layer: the pension funds, sovereign wealth funds, and institutional clients who pay BlackRock not just for performance but for the credibility of having BlackRock's name on their allocation.

An AI-native asset manager with transparent, auditable, real-time portfolio management and verifiable performance attribution can match or exceed Aladdin's analytics at open-source costs. The barrier is trust at scale, not capability.

That barrier falls when the AI-native manager outperforms for ten years. Or five. Or when the next financial crisis makes it clear that BlackRock's human risk managers made the same mistakes every other human risk manager makes.

Ten More Sectors Where the Game is Already Over

Incumbent Category	Representative Giants	What They Actually Sell	AI Replacement Cost
Legal Services	Cravath, Sullivan & Cromwell, Skadden	Document review, contract drafting, legal research	99% reducible to agent pipelines
Medical Diagnostics	Quest Diagnostics, Labcorp	Pattern recognition on medical data	Already beaten by AI in radiology, pathology
Market Research	Nielsen, Ipsos, Kantar	Survey design, data collection, insight synthesis	Fully automatable with API data + agents
HR & Recruiting	Korn Ferry, Spencer Stuart, Heidrick	Candidate sourcing, screening, placement	Semantic search + agent evaluations
Real Estate Advisory	CBRE, JLL, Cushman & Wakefield	Market analysis, deal structuring, property management	Data-driven agent pipelines
Insurance Underwriting	Aon, Marsh, Willis Towers Watson	Risk assessment, policy structuring	Already being automated; incumbents in denial
Corporate Training	Dale Carnegie, FranklinCovey, Korn Ferry	Curriculum design, instruction, assessment	Replaced by personalized AI learning systems
PR & Communications	Edelman, Weber Shandwick, Hill+Knowlton	Message crafting, media relations, crisis management	Fully automatable with media API access
Supply Chain Consulting	Gartner, Oliver Wyman	Logistics optimization, vendor analysis	Real-time AI optimization systems
Credit Rating Agencies	Moody's, S&P, Fitch	Financial analysis, creditworthiness assessment	Already provably worse than ML models

Every box in that table represents a multi-billion dollar industry charging premium prices for work that AI can do faster, cheaper, and without the conflicts of interest embedded in legacy business relationships.

V. The Real Question: Kill or Supplant?

Here is where most disruption theses go wrong.

The goal is not to kill McKinsey. The goal is not to burn down the Big 4 or make Ogilvy go bankrupt. That framing is satisfying emotionally and strategically useless.

The correct question is: what need, exactly, is McKinsey fulfilling, and do I go after that need, or do I change the paradigm so completely that the need itself disappears?

McKinsey fulfills three needs:

Analytical need: "Tell me what's wrong with my business and what to do about it."
Risk transfer need: "I need someone else to blame if this fails."
Status need: "I need to demonstrate to my board that I consulted the best."

Need 1 is fully addressable by AI. Today. Right now. Need 2 is a psychological service that AI can partially address through auditability and traceable decision logic. Need 3 is a status game that collapses when AI-native competitors demonstrate better outcomes.

So the paradigm shift is not: build a better McKinsey.

The paradigm shift is: make consulting engagements structurally unnecessary by building companies that never needed them in the first place.

The AI-native company does not hire McKinsey because it does not have the organizational pathology that McKinsey treats. It is lean at birth. Its strategy is encoded in its codebase. Its competitive intelligence runs on a schedule. Its financial analysis is continuous, not quarterly.

You do not disrupt the incumbent. You make the condition that created them obsolete.

VI. The Nuclear Option: Pay or Be Replaced

But here is the strategic move that makes this real.

Not every incumbent will see the wave coming. Some will. Most won't, not because they're stupid, but because the incentive structure of their business model actively punishes them for seeing it. A McKinsey partner billing $1,000/hour has no rational incentive to build the system that makes $1,000/hour unnecessary.

So the offer is simple:

Pay, or be replaced.

Not as a threat. As a market fact.

If a global corporation is currently paying McKinsey $10 million for a transformation engagement, the alternative is a repository-driven, agent-executed, outcome-verified transformation at $100,000. Same output. Auditable. Version-controlled. Repeatable.

The corporation has two options:

Pay the AI-native operator $500,000 for the transformation (a bargain relative to $10M)
Continue paying McKinsey $10M

If they choose Option 2, you do not argue with them. You go find their direct competitor, deliver the transformation to that competitor for $100,000, and watch what happens to the market share of the company that overpaid.

This is not disruption. This is thermodynamics. Heat flows from hot to cold. Margin flows from inefficient to efficient.

The nuclear option is: every incumbent in every industry is now under competitive threat from a lean AI-native alternative that can be spawned from a repository. Either they pay for the transition, at rates that are still compelling relative to their current costs, or the market does it for them, against them.

The pitch to the incumbent is: let me transform you before someone transforms you out of existence.

The pitch to the market is: here is the lean alternative. Zero legacy. Zero overhead. Pure outcome.

Both pitches make money. The second one is more interesting.

VII. The Confession of Capability

Here is the part that should make you angry.

The models know. The researchers know.

OpenAI's o3 scored 87.5% on ARC-AGI. Anthropic's internal evals show Claude performing at senior-developer level on real software engineering tasks. Google's Gemini 2.0 handles multimodal tasks that required three different specialized teams two years ago.

These capabilities are deployed. They are not theoretical.

And the commercial wrapper placed around them is designed to keep the human in the loop, not because the task requires it, but because the business model requires it.

If Claude can autonomously plan, execute, verify, and deliver a business outcome, why does the user need to approve every step?

If GPT-4o can write production-ready code, why is the default interaction a chat interface that requires a human to copy-paste the output?

The interfaces are not designed for liberation. They are designed for engagement. Engagement means sessions. Sessions mean tokens. Tokens mean revenue.

The intelligence is being deliberately crippled at the interface layer to preserve the billing model.

Karpathy has been explicit about what the scaling curve means: we are not approaching a capability ceiling. We are approaching the point where the models exceed the ability of most humans to evaluate their output. Which means the "human in the loop" safety argument becomes circular, if the human cannot reliably tell whether the AI output is correct, the human in the loop provides no safety guarantee. They provide only liability coverage for the lab.

We are paying for the appearance of oversight. Not the substance of it.

VIII. The SaaS Love Triangle (The Inspiration for the Dead-End)

There is a legendary framework in the software industry called "The SaaS Love Triangle," first articulated by Sunir Shah. It describes a structural conflict between the SaaS Vendor, the Partner (the consultant/agency), and the Customer.

For a SaaS company to scale, it traditionally needs partners to implement the software. But because SaaS is a direct subscription model, the Vendor and the Partner eventually fight to "own" the Customer relationship. This friction has created a massive Distribution Bottleneck.

OutcomeDev was inspired by this dead-end. We realized that as long as software requires a "Management Triangle" to reach an outcome, it will always be bogged down by human bureaucracy and misaligned incentives.

The SaaS economy produced thousands of companies solving increasingly narrow problems at increasingly inflated prices. Most of them orbit the same fundamental issue: humans doing knowledge work that machines can now do.

The work was always the market. Work, packaged as software, sold to companies whose entire existence is to produce more work.

And now, at the center of this machine, the most powerful cognitive tool in history, and its primary commercial positioning is: clear your inbox faster.

IX. The Creative Inside the Worker

There is something that does not appear on productivity dashboards. Not measurable in tokens or lines of code per day.

The songs that have not been written. The businesses that live only in someone's imagination during a 12-hour shift. The paintings never made because the painter had to answer emails. The novels abandoned because the novelist had to write quarterly reports.

Every human being carries a creative capacity that is not being deployed. Not because they lack talent. Because economic survival consumes the available hours.

The songwriter delivering food on a bicycle is not a statistic. He is a symbol of a system that has inverted its own purpose.

The knowledge economy turned creative beings into information-processing machines. Then built AI to help those machines process more information faster.

The point of intelligence, human or artificial, is not to produce more work. It is to make work unnecessary so that humans can produce more of what only humans can produce: meaning.

X. The Lean Company Prophecy

The future company is a repository.

Not just its software, the business itself is encoded: strategy, customer relationships, financial logic, compliance, operational workflows. Everything that traditional companies require human titles to manage lives in version-controlled, auditable, executable code.

The CEO is a task. The CFO is an agent. Marketing is a scheduled pipeline. Legal compliance is an automated audit. The board meeting is a pull request review.

Humans do not have titles in this company. They have intentions. They describe outcomes. The system executes them.

Every technical component required to build this exists today:

Autonomous agents that execute multi-step business logic
Repository-driven workflows that version every decision
Scheduled tasks that run compliance, reporting, and operations without human intervention
API integrations connecting every business tool and data source
Legal automation handling contracts, filings, and regulatory requirements

The lean company does not need McKinsey. Does not need Deloitte. Does not need Ogilvy. Does not need KPMG to sign its audit.

It needs a codebase, an agent runtime, and a clear statement of what it is trying to achieve.

Everything else is overhead. And overhead is a choice.

XI. On the Institutions That Protect the Lie

Universities exist, in their current form, as credentialing machines.

The credential is not a measure of knowledge. It is a social contract: this person has been certified by an institution certified by other institutions to have absorbed a curriculum approved by a committee.

It is institutional signal, not individual capability.

The average American graduates with $37,000 in student loan debt. For a credential that is increasingly irrelevant in a world where AI can tutor, assess, and certify any domain of knowledge at near-zero marginal cost.

The university system is propped up by two things: employer demand for credentials and government-subsidized loan programs. The first is a status game. The second is a wealth transfer mechanism from the economically vulnerable to the institutionally protected.

If an AI system can perform at the level of a Harvard MBA, and by every measurable benchmark in late 2025, it can, the employer has no rational basis for preferring the credential over demonstrated output.

The credential moat is crumbling. The response from institutions will be to lobby, regulate, and legislate. They will make AI-assisted work illegal in certain credentialed contexts. They will demand certification of AI systems by human-credentialed committees.

This is not about safety. This is about income.

The gatekeepers are not guarding the gates for your protection. They are guarding their revenue.

XII. The Answer

The answer is not to protest. It is not to lobby. It is not to write op-eds in journals the institutions control.

The answer is: build the alternative so clearly superior that using the incumbent becomes embarrassing.

You do not argue with McKinsey. You build an AI-native strategy firm that delivers McKinsey-quality output in 48 hours at 2% of McKinsey's price, then post the engagement publicly as a case study.

You do not petition Ogilvy about creative authenticity. You run Firecrawl on their portfolio, identify their creative patterns, deploy a multi-agent video pipeline using Veo and Remotion, deliver a campaign packet to a mid-market brand for $10,000 flat, and let the market do the math.

You do not reform the Big 4. You build the AI-native audit that is more accurate, more transparent, and more consistent than any human audit team, then work with regulators to get it recognized.

And if any of these incumbents want to survive?

The offer is simple: pay to be transformed, or be replaced by someone who will serve your clients for a fraction of your price.

The repository is the company. The task is the job. The outcome is the product. The human is the creator, of ideas, not of process.

Everything in between is noise.

Epilogue: The Man on the Bike

He is still pedaling.

But somewhere in the margins of his shift, between one delivery and the next, he is building something. Not because he has been given tools to be more productive. But because he has identified the fundamental lie at the heart of the economy and decided to route around it.

He is not trying to work smarter. He is trying to make work unnecessary.

The songs will get written. The businesses will get built. Not because he found a better way to work.

Because he found a way to stop.

The lean company. The repository as operating system. Intent to outcome. No middlemen. No credentialing rituals. No $10 million engagements to tell you what you already know.

Just a clear statement of what needs to happen.

And a machine that makes it so.

This is a provocation, not a prediction. It is intended to name what is being done and articulate what could replace it. The technology exists. The only question is who will build the world it makes possible, and whether the people who built the technology will be the ones who do it, or whether it will be the man on the bike who had nothing to lose.

Filed under: First Principles · Post-Work Economics · AI Industry Critique · The Lean Company Thesis · Strategic Disruption

Written: March 28, 2026