Chapter 1Execution Approaches Zero

1. Your Most Expensive Purchase Is on Sale and Clearance #

Open your company's income statement. Over the past two decades, no matter the industry, the biggest expense was almost always the same: labor. Cut one level further and the bulk of that cost is often execution, not judgment. Those hours go to the same work: turning a decision into documents, code, reports, contracts, design drafts, customer-service conversations, and marketing copy. Nobody thought this was a problem, because the entire business world rested on the same assumption: execution is expensive and must be sourced, organized, and supervised with care.

That assumption is failing, and it is failing at a measurable rate.

Stanford University's Human-Centered Artificial Intelligence Institute (Stanford HAI) tracked a curve in the AI Index annual report. The inference cost to reach GPT-3.5 capability (MMLU benchmark 64.8) fell from $20 per million tokens in November 2022 to $0.07 in October 2024—about one two-hundred-eightieth of the original (Stanford HAI, 2025). The shape is worth more attention than the endpoint: $1.80 in August 2023, $0.18 in June 2024, then further decline at similar multiples, year after year. This is not a one-time sale. It is a ramp that keeps sloping down. You did not miss the bottom. You are standing on a conveyor belt that is still moving down. a16z calls this LLMflation: the price of a fixed capability level falls roughly tenfold each year. Epoch AI's task-level estimates are more aggressive, with a median decline of fiftyfold per year; looking only at data after January 2024, the median rises to two hundredfold per year. The report uses medians rather than averages because different tasks range from ninefold to nine hundredfold per year, and averages get hijacked by extremes.

What collapsed is the price per unit of capability. The latest flagships are still not cheap; list prices fall only about five- to tenfold a year, far slower than the "equal capability" line. What matters to an organization is never the flagship price. It is the price of the ability to finish the job. An analysis that took a three-person team a week two years ago costs close to zero to invoke at that level today. Everyone can buy a model good enough for the work. The gap is who sees it first: what took a team yesterday is a single call today.

The curve also covers only the digital world. The physical and regulatory boundary comes later.

For a principal leader the meaning is plain: your most expensive purchase used to be human time, and now it is on sale and clearance. When a core factor of production falls by an order of magnitude every year, every institution designed around the old price—however elegant—has to be recalculated. The steam engine redrew the factory; the shipping container redrew trade. This time it is execution itself.

Push the ledger further and you hit something deeper than "cost reduction." Three-year contracts, five-year budgets, and headcount plans struck at today's cost all rest on a price that shrinks by an order of magnitude every year. Long-term pricing of execution will be systematically high. Planning should take the slope of the ramp, not today's price. "Wait until the technology matures" has nowhere to wait: the ramp is continuous, there is no station called "mature," and every year you wait, competitors compound another year of advantage. Money freed by falling execution, if it only flows back to the income statement, sells off the window for organizational evolution. It should not be booked as profit. It is migration budget, and it should go to the two things still rising in price: standards and context.

2. The Other Base: Knowing Where to Look #

If execution alone got cheaper, the shock would be smaller. What makes this a geological shift is the second base being pulled away at the same time: the cost of knowledge acquisition.

How much have organizations paid to "know"? Senior people cost more than newcomers, and a large part of the premium buys not judgment but knowing where to look, whom to ask, and what the industry norm is. Databases, industry reports, consultants, training courses all buy the same thing: shortening the distance from not knowing to knowing. That distance once supported several trillion-dollar industries. Publishing, consulting, and vocational education were, at bottom, arbitrage on knowledge-acquisition cost.

The search-engine era collapsed that distance once: from "know that person" to "find that document." The AI Q&A era collapses it another order of magnitude: from "find that document" to "ask that question." Retrieval, screening, reading, summarizing, and cross-checking are becoming internal steps in a single call, no longer billed to you separately.

The two bases hold up the old organizational structure. Execution was expensive, so work was finely divided, scheduled, and supervised. Information was scarce, so it moved through hierarchies, was priced by seniority, and synced in meetings. Execution approaching zero looks like an efficiency headline. It is organizational geology. Remove one base and the structure can lean for a while. Remove both at once and every beam above has to be recalculated.

One company ran that calculation for everyone, using its own life.

3. The First Specimen: A Company That Sold Both Bases #

Chegg is an American online education company that went public in 2013. The business, in one sentence: a paid answer library plus on-demand experts. Students pay up to $19.95 a month for pre-written textbook answers and a network of experts available on demand.

What it sold was exactly the two bases being pulled away. The answer library sold the product of execution; the expert network sold knowing where to look and whom to ask. Chegg did not lack AI, and the organization was not a mess. What stood directly under the zeroing curve was the business model itself.

Remote learning during the pandemic pushed it to a peak: in February 2021 the stock was $113.51, market cap about $14.7 billion. Twenty-one months later, ChatGPT launched.

On the May 2023 earnings call, CEO Dan Rosenzweig acknowledged that ChatGPT was eroding subscriber growth. The company withdrew full-year guidance; the stock fell 48% that day. There has been no real rebound since. Cumulative subscriber losses exceed 500,000. Q4 2024 subscribers were 3.6 million, down 21% year over year, full-year revenue down 24%. There were 441 layoffs in 2024, about a quarter of staff; another 22% cut in May 2025; another 45% in October, with announcements citing "the new realities of AI." Market cap bottomed around $156 million—a 99% drop from the peak (Chegg filings; WSJ; CNBC).

The sharpest detail is not the fall. It is a comparison. Chegg itself uses AI. Filings show that with AI, capital spending on content production fell 56% year over year while user questions rose 2% in the same period. They were not bad at using the technology. Victims and beneficiaries use the same tools; the difference is which side of the zeroing curve the business model stands on. For Chegg the question was never whether the tools worked. It was whether what it sold was still scarce. If your business model is to sell execution, AI is not your tool. It is your substitute.

This is not an "AI destroys everything" horror story. Chegg is a clean specimen: a business with almost no physical layer, no regulatory buffer, built purely on execution plus knowing—what happens when both bases are pulled away. Your company is probably not Chegg. Several departments almost certainly run daily work that looks like its business model.

Look inward and the mini-Cheggs usually take three forms.

Report factory: roles that regularly produce weekly reports, monthly reports, and analysis briefs. If the value proposition is "turn data into readable prose," that is an internal answer library.

Human search desk: legal precedent lookup, investment-research comparables, procurement vendor comparison. If the moat is "knowing where to look," it sells Chegg's second service.

Template output line: contract first drafts, bid frameworks, asset revisions. Clear input, clear output, execution by template—the form the zeroing curve absorbs fastest.

These three forms should not lead straight to layoffs. That reads the ledger too shallowly. They should lead to repricing. What becomes valuable in these roles is shifting from output to release: judging whether work passes, and by what standard to let it go.

4. Boundaries of the Claim: Where Zeroing Holds #

"Zeroing" is one of the most misread words in this book. Without clear conditions, the claim slides into slogan.

Execution zeroing applies first and mainly to digital knowledge work: writing, code, analysis, design, customer service, legal documents, reports, translation. What they share is digital inputs and outputs, with quality checkable in the digital world. Chegg's answer library, Duolingo's course translation, marketing's first-draft copy all sit in this set.

Physical execution has not zeroed. A wall on a construction site, a cut on an operating table, last-mile delivery—AI has not made them much cheaper. Regulatory processes have not either: audit, compliance, licensing, cost structures set by institutions, not by technology. Construction sites and operating rooms do not refute a claim about knowledge work. The collapse speed of knowledge work does not scare every industry equally.

Inside and outside the boundary are not insulated. Every physical execution is wrapped in a digital shell: scheduling, documentation, decisions, quality checks, settlement. That shell is knowledge work, and zeroing is pushing inward along that interface. A construction company's bricklaying cost has not fallen, but drawing development, quantity takeoffs, progress reporting, and contract review are approaching zero. A hospital's surgery has not zeroed, but records, imaging pre-screening, and follow-up notes are. Speed differs; direction is the same.

This book's claims default to holding within this boundary. If your organization sits mainly outside it, the timetable is looser, but the direction is unchanged. Competitors will evolve first on the digital shell, then meet you in the physical world with saved cost and shorter cycles.

5. Organizational Archaeology: Why Every Line Exists #

Put the org chart on the table and ask a question most principal leaders rarely ask seriously: what problem was each line, box, and process invented to solve?

The answer is uncomfortable. Almost everything was designed for "people do things slowly and know less."

A person can effectively manage only so many direct reports—textbooks say 7±2, in practice rarely more than 10—and can hold even less information in real time. Organizations therefore grow as trees: information rolls up from leaves to root, compressed and retold at each layer; instructions roll down from root to leaves, translated and amplified at each layer. Hierarchy looks like power hunger. It is an engineering compromise for "people know less." When there is no better way to sync, a tree is the best way to connect ten thousand people. The cost is in every textbook: loss at every relay, delay at every rollup. Organizations paid that coordination tax for a century because there was no alternative.

When judgment is scarce, approval is a rationing window. Matters line up to use a few vetted people. Every node is running the same calculation: the cost of trial and error exceeds the cost of waiting. When execution was expensive, that almost always held. A bad plan that burned three person-weeks was worth two days in the director's inbox.

Weekly reports, standing meetings, performance reviews, alignment sessions do another job: they batch-sync context scattered across minds, using human time—the most expensive medium in the organization. Ten people in a one-hour meeting costs ten person-hours, and what gets synced still decays with attention. When context lived only in heads, the meeting room was the only highway.

Managers cannot see everyone's work, and they cannot measure "value" directly, so KPIs become stand-ins: pieces, hours, lines, tickets, response time. You measure what you can measure, then hope it correlates with what you care about. That is not foolishness. It is what poor observation forces.

In their time, these four were optimal solutions. Mockery is easy and cheap. The people who designed them made the right engineering call under their constraints. A theme this book will return to: standards have boundaries; when the environment shifts, yesterday's optimum becomes today's liability. The hardest organizational act is admitting that.

With both bases pulled away, the old ledger has to be recalculated. Frontline context can appear losslessly and instantly to anyone, yet the tree still pays coordination tax on information that is no longer scarce, and the tax rate has not changed. Redoing a plan went from three person-days to three seconds; trial and error is no longer costly, but the gate still charges the old price—two days in queue now costs orders of magnitude more than the risk it intercepts. What meetings sync could live in shared human–AI context, queryable anytime and never decaying, yet the organization still spends everyone's same hour on sync a machine can do in real time. When AI can generate unlimited "output," pieces, lines, and hours can all be maxed out, and the assumed correlation between output numbers and real value breaks.

Every line on the org chart was drawn to manage expensive human execution. Once execution is free, those lines become decoration one by one.

Picture a pyramid. Execution cost and knowledge-acquisition cost are both pulled away at once, and the superstructure hangs in mid-air. Not everything suspended is scrap. Three things hang there with no taker—exactly what AI cannot take: choosing the goal, adopting the standard, and unifying context.

Figure: org-chart lines were drawn for expensive human execution; what hangs are the three things AI cannot take.

6. Industry Microscope: Translation #

Macro curves numb, and Chegg can be dismissed as bad luck. Turn the microscope to a full industry: translation, where execution zeroes hardest.

It is almost a pure sample of digital knowledge work: text in, text out, quality checkable digitally, no physical buffer, no license moat. If zeroing has a ground zero, it is here.

Ground zero already has first-hand job records. In late 2023, Duolingo cut about 10% of external contractors in two waves (August and December), mainly course translation and writing roles; the company confirmed to Bloomberg that GPT-4 could already generate translation and course content. Those who remained changed shape: one or two per team, titles shifted to "content curators"—review AI output, then release.

That is the claim in micro-slice: executors become reviewers. Machines take output; what stays on humans is judgment—whether it passes and by what standard to release. A four-person translation team becomes two curators. Execution hours vanish; judgment hours remain. That shape shift will replay in every category of digital knowledge work.

Two company types stand on the same collapse band. One is like Chegg: a business model built on selling execution and selling "knowing," bases pulled, business zeroed. Another made the opposite bet. Transn, a long-established Chinese language-services company, sits at the center of the collapse band. It did not train translators to work faster. That is high jump on a sinking floor. It rewrote the company into another form: a Chief AI Officer (CAIO), an AI Native decision committee, a rule that no meeting proceeds without a runnable demo, and "energy gold" so internal AI apps grow by market rules. Founder He Enpei's line works as a footnote: "Rather than wait for employees to become AI experts, let the organization grow AI capability."

That sentence does not bet on people learning tools. It bets on organizational form changing. The popular narrative says AI is here, run training fast. This chapter says the problem was never employee tool fluency. It was every line of the org still drawn at the old price of expensive execution. In industries where the execution layer is absorbed first, the survivors who moved earliest were not those who executed faster. They were those who admitted earliest that execution is no longer valuable, and rewrote themselves as standards plus context.

Credit where due. Transn's mechanisms are cross-reported in multiple outlets and hold up; operating results—whether revenue and profit improved because of them—have no third-party data yet, so this chapter cites mechanisms only, not outcomes. Chegg's side is filing-grade and complete. The two are different industries; what compares is mechanism, not performance: both in digital knowledge work, one sells "execution + knowing," one rewrites it as a cost item. A same-caliber loser inside translation—a translation company that held to execution-mode and declined—has not been found yet; until then, "self-transformers live better" stays a hypothesis to test. This chapter's claim does not rest on that comparison. Cost curves and Chegg are independent evidence chains.

7. The Opposite of "Thinking Is Free": Why Judgment Gets More Expensive #

In popular talk, some summarize this wave as "AI makes thinking free." This book says the opposite, and that opposite is the starting point for everything that follows.

What actually got cheaper? Two things: how to do it (execution) and where to look (knowledge acquisition). Writing code, building reports, checking cases, finding precedents—all are approaching zero. Together they are still not thinking. Thinking has a core left: what is worth doing—choosing one option among infinite possibilities, excluding all others for it, and bearing the cost of that exclusion. AI has not discounted that core by a cent.

If anything, it got more expensive. Scarcity is relative. When execution took three months, mediocre judgment could hide behind the long stretch of work; when results arrived, nobody remembered who decided. When execution takes three hours, judgment stands naked in the result: decide in the morning, see outcomes by afternoon; within a week the whole company can see whose judgment was worth something and whose was only rank echo. The cheaper execution gets, the larger judgment's share of total cost—and the higher the relative price of getting it wrong. When "how to do it" and "where to look" both get cheap, the only thing left expensive is "what is worth doing."

Chegg fits the same structure: not short on execution (twenty years of an answer library), not short on knowledge access (an expert network)—short on judgment about what was worth doing around November 2022. When what you sell starts going free, the answer is not "sell harder." Judgment errors always cost; execution zeroing just compressed the accounting period from years to quarters.

If AI someday fully decides what is worth doing for you, the company is effectively theirs. That is a definition, not a technical limit. As long as the company is yours, "what is worth doing" stays on your ledger—and now it is the only line still rising in price.

So the sharper question surfaces: execution is cheap, tools are affordable to all—why have companies that bought every AI tool not seen order-of-magnitude efficiency gains? Money spent, tools deployed—where did the dividend promised in the curve go?

The next chapter answers.

What to Do Monday Morning (principal-leader view) #

Draw your company's seismic zone map in four steps:

1. Mirror first: Start with the Chegg question—how much revenue sells "products of execution" or "knowing where to look"? That slice is the red zone of the business model. Most companies are not Chegg, but almost every one has a product line or department that is a mini-Chegg.
2. List execution work: Have each department list every "pure execution" role and step—clear input, clear output, digitally verifiable quality. For each, mark how much today's AI can do; mark red if it can do 80% or more.
3. Audit org-chart lines: For every line on the chart, ask: does this manage execution or judgment? Lines that manage execution lose their reason as the red zone grows; lines that manage judgment are the skeleton of the future org.
4. Count approval gates: Pick the longest approval flow and ask at each node: is the risk this gate intercepts still more expensive than the gate itself today? The nodes where the answer is no are idle approval gates.

This map does not require cutting anything today—it is the base map for what follows. Individuals and teams can ask too: how much of my work sits in the red zone? Does my value hang on execution or judgment? If the team went from four people to two "curators," would I be one who stays?

Chapter Acceptance Self-Check (against chapter acceptance standards) #

1. Claim restatable in one sentence ✓, and a corollary of the core claim (cost zeroing → structural failure).
2. Whiteboard framework ✓ (pyramid with both bases pulled, three suspended items above).
3. External comparison and data ✓: winner side Transn (mechanisms verified) + Duolingo job evidence; loser side Chegg (filing-grade verified); caliber differences stated explicitly; statistics report medians (Epoch AI). Same-caliber loser inside translation still to be added, marked as hypothesis to verify.
4. Nine quotable lines ✓.
5. "What to Do Monday Morning" four principal-leader steps + personal note ✓.
6. Fluency ✓: revised primarily through whole-sentence rewriting and English breath under current prose-standard; no Five Prohibitions or aside violations; sources in in-text author–year or end-of-sentence parenthetical cites.