Cheap Intelligence Raises the Price of What It Cannot Make

On 1 May 2023, Chegg's chief executive told investors that since March the company had seen "a significant spike in student interest in ChatGPT" and now believed "it's having an impact on our new customer growth rate". The next day its shares closed down 48.41 percent, CNBC reported. Chegg owned what most strategy decks would call a data moat: a homework-help service built on what its annual report called "millions of pieces of educational learning content". By the second quarter of 2025 its subscribers were down 40 percent on the year. In October 2025 it cut 388 roles, about 45 percent of its workforce, and said AI and reduced traffic from Google had led to "a significant decline in Chegg's traffic and revenue".

The library was large and proprietary. It failed as a moat anyway, because what it sold was an answer, and answers were the thing models had just made cheap. In Part 1 of this series I argued that Five Forces scores built for scarce thinking work overstate what expert headcount protects. Once intelligence is cheap for everyone, what is still worth building?

My answer is that value moves to whatever cheap models make more useful but cannot make themselves. A moat now is an asset that passes two tests: cheaper intelligence raises its value, and a rival cannot buy or copy it. Speed and owning your own model, the two answers I hear most often, fail the second test. Three assets can pass both: data no rival can collect because only your operations produce it, a seat inside the customer's daily work, and being the party that signs for the outcome.

I am writing for the chief strategy officer or business-unit head who has re-scored the forces and now has to decide what to fund. Part 1 ended with a longer list of what stays scarce, including trust, distribution and regulatory standing. Here distribution sits under position, and trust and regulatory standing under accountability.

Complements Gain, Substitutes Lose

The economics are old. In a 2016 Harvard Business Review article, Ajay Agrawal, Joshua Gans and Avi Goldfarb described machine intelligence as a fall in the cost of prediction and set out what follows when the cost of a basic input drops that far: "The value goes up for complements and down for substitutes." Their main example of a complement was human judgment. When the cost of prediction falls, they wrote, "demand for judgment rises."

Chegg sat on the wrong side of that line. Its answers did the job a model's answer does. Every fall in the price of model answers cut the value of the library. The same archive could have sat on the other side: used to ground a model, it becomes an input to cheap intelligence. So the first test applies to the use you plan to fund. That gives two questions to ask of anything a strategy team proposes to call a moat.

• Does cheaper intelligence make it worth more? A complement gains value as models get cheaper and more widely used. A substitute loses value.
• Can a rival buy it or copy it? If a competitor can buy the same thing from the same vendor, or rebuild it within a year, it may be useful, but it will not protect your returns.

An asset has to pass both. Plenty of valuable things pass the first and fail the second, and a strategy that funds only those buys capability with no protection attached.

Two Popular Answers That Fail

Speed

The answer I hear most is speed: decide faster and ship faster. It makes a poor moat, because every rival can buy the same models and copy the same practices. Porter made the point long before AI. In "What Is Strategy?" in 1996 he gave a section the heading "Operational Effectiveness Is Not Strategy" and warned that "competitors can quickly imitate management techniques, new technologies, input improvements, and superior ways of meeting customers' needs." Rivals who copy one another, he wrote, end up in "a series of races down identical paths that no one can win."

Speed earns its keep when it buys one of the three assets before a rival does: the first tool in a customer's workflow, or the first stream of operating data. Fund speed toward a named asset, and treat speed for its own sake as an operating cost.

Owning a model

The second answer is to own the model: build on open weights and fine-tune your own. That is sound purchasing. Part 1 showed that open models cap any one vendor's power, and that protects your margin from suppliers. It does not stop a rival, who can download the same open weights. A model of your own passes the first test only if it learns from something your rivals cannot get. Then the moat is that data.

Three Assets That Can Pass

Data only your operations produce

Data is the obvious candidate and the most overrated one. Andrei Hagiu and Julian Wright, writing in Harvard Business Review in 2020, wrote that "in most instances people grossly overestimate the advantage that data confers." Three of their seven questions do most of the work for a strategy team: whether the data is proprietary, meaning it "can't be purchased from other sources, easily copied, or reverse-engineered"; how fast its relevance depreciates; and whether data from one user improves the product for others.

Chegg's library failed the proprietary question from an unexpected direction: nobody copied it, but a general model could produce a close substitute on demand. Data that passes is usually a by-product of running the business, such as transactions, outcomes or what customers did after the advice, collected with the right to reuse it. A static archive, however large, is exposed. A stream that only your operations generate, and that improves what you sell, can hold.

A place in the customer's workflow

The second asset is position: being where the customer's work already happens, with the content and checks they rely on. In August 2023 Thomson Reuters completed its $650 million cash purchase of Casetext, whose products include CoCounsel, an AI legal assistant built on GPT-4. In June 2025 Harvey, an AI platform for lawyers, announced an alliance with LexisNexis to bring LexisNexis "primary law content, and Shepard's Citations" inside its platform. In both deals an AI newcomer ended up joined to an incumbent's case law and citation checks, one by sale and one by alliance. Cheap intelligence made those assets more valuable, because a model that drafts a brief still needs authoritative law to ground it and a citator to check it.

Position fails when the workflow moves. If customers start their work in a new tool, and that tool can license or rebuild the content, the incumbent's place goes with them.

Accountability for the result

The third asset is answering for the outcome. In February 2024 a British Columbia tribunal ruled against Air Canada after the chatbot on its website told a grieving passenger they could apply for a bereavement fare retroactively, which Air Canada did not permit. Air Canada had argued, in the tribunal's words, that "the chatbot is a separate legal entity that is responsible for its own actions." The tribunal called that "a remarkable submission" and held that "it should be obvious to Air Canada that it is responsible for all the information on its website" (Moffatt v. Air Canada). The award was C$812.02. It is one small-claims decision, but its reasoning is the one a deployer should plan for: the firm that deploys the model answers for what it says.

Customers pay for a party that checks the work and answers when it is wrong. That is a cost for a careless deployer and a source of margin for one that checks. A model can draft the audit opinion, the diagnosis or the engineering sign-off. It cannot hold the licence, carry the insurance or stand behind the result. This is the judgment Agrawal, Gans and Goldfarb expected to gain value. It protects incumbents against AI-native entrants more than against rivals that hold the same licence, and between those rivals the edge goes to whoever checks best. Accountability fails as a moat when a firm signs for work it has not checked.

The Strongest Objection

The best case against me is that I have only moved the problem. Martin Casado and Peter Lauten of Andreessen Horowitz argued in 2019 that "data is rarely a strong enough moat", because "the cost of adding unique data to your corpus may actually go up, while the value of incremental data goes down." Positions erode too: LexisNexis put its content inside the next platform. Casado and Lauten themselves put lasting defensibility elsewhere, in "verticalization, go-to-market dominance, post-sales account control, the winning brand", which is close to what I call position. The harder version: every position erodes, so keep moving.

I accept most of that. None of the three is permanent, and each section names how it fails. A short half-life argues for renewing a moat. It is a weak argument for relying on speed alone, since a firm that runs faster on rented tools ends up in Porter's race. One whose data, position or accountability gains value as intelligence gets cheaper has something to defend while it moves. Chegg also cuts against me in one respect: its content was proprietary and it still fell. That is why the first test comes before the second.

What to Fund Before the Next Review

Take your list of what stays scarce, from the Part 1 exercise, and put each item through the two tests. Fund what passes both, and write into the budget paper the condition that would make it fail. Treat what passes only the first test, such as faster cycles and better tools, as operating improvements, measured as costs and savings. Then look for any asset of yours that does the same job as a model's output, as Chegg's library did. That is the one to reprice first.

Sources and Changes

• Chegg, "Chegg Announces First Quarter 2023 Earnings", Form 8-K exhibit with prepared remarks, 1 May 2023; Form 10-K for 2022, 21 February 2023; second-quarter 2025 results, 5 August 2025; strategic review and restructuring release, 27 October 2025.
• Sarah Min, "Chegg shares drop more than 40% after company says ChatGPT is killing its business", CNBC, 2 May 2023.
• Ajay Agrawal, Joshua Gans and Avi Goldfarb, "The Simple Economics of Machine Intelligence", Harvard Business Review, 17 November 2016.
• Michael E. Porter, "What Is Strategy?", Harvard Business Review, November–December 1996.
• Andrei Hagiu and Julian Wright, "When Data Creates Competitive Advantage", Harvard Business Review, January–February 2020.
• Thomson Reuters, "Thomson Reuters Completes Acquisition of Casetext, Inc.", 17 August 2023.
• LexisNexis Legal & Professional, "LexisNexis and Harvey Announce Strategic Alliance", 18 June 2025.
• Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024.
• Martin Casado and Peter Lauten, "The Empty Promise of Data Moats", Andreessen Horowitz, 9 May 2019.

All sources checked on 6 October 2026.

Changes on 6 October 2026. New headline, summary and argument. The opening no longer continues a strategy-offsite story from Part 1: that story was a composed scene and is gone from both parts. Also removed: a board-meeting anecdote I cannot document; the Infinite Leverage Strategy Canvas and its quadrant placements, which no data supported; Value-on-Intelligence; the claim that an AI-native firm completes four decision cycles to a rival's one, and the speed comparisons behind it; the 20 percent budget and 30 percent revenue targets; a 70 percent exception-handling figure; and the 90-day plan built on them. The Hema example is gone. The earlier text said Alibaba had named it non-core, but Alibaba sold its other physical retail holdings and said it had no plan to sell Hema. The EU AI Act section and the IMF jobs estimate fell outside the new argument and were cut. On 4 October 2026, two charts built from undocumented advisory assessments, with their 91 and 9 percent figures, had already been removed.