Disruption theory is narrow on purpose, and most AI products fall outside it

In a 2007 BusinessWeek interview, quoted by Joshua Gans in 2012, Clayton Christensen said that the prediction of his theory was that Apple would not succeed with it. The phone, he reasoned, was a sustaining technology relative to Nokia, and the existing players were heavily motivated to beat it. The iPhone succeeded anyway. Eight years later, Christensen, Michael Raynor and Rory McDonald wrote in the Harvard Business Review that the 2007 iPhone had indeed been a sustaining innovation in smartphones, that its early success was likely explained by product superiority, and that its later growth was better explained as disruption of the laptop. Each statement can be defended on its own. Side by side, they show how much room the label leaves once the outcome is known.

My position is that disruption theory is narrow on purpose, and that the narrowness is where its use lies. It describes one mechanism by which incumbents lose, and it says little about every other way they can lose. Applied to AI, most of what gets called disruptive is not: a copilot inside an existing suite is the incumbent's own product getting better for the incumbent's own best customers, which is the definition of sustaining. The theory bites only where two conditions hold, a foothold among customers nobody was serving or among customers whose needs the incumbent overshoots, and an entrant that improves faster than those customers' needs grow. I also think its record as a forecaster is mixed enough that it should be used to classify a threat, and not to predict how the threat ends.

The theory makes one narrow claim about how incumbents lose

The 2015 Harvard Business Review article is the most careful statement of the theory by its authors, and it is stricter than the way managers use the word. Incumbents improve their products for their most demanding and usually most profitable customers, and in doing so they exceed the needs of some segments and ignore the needs of others. Entrants that turn out to be disruptive begin in those overlooked segments, often at a lower price. Chasing profit higher up, the incumbents tend not to respond vigorously. The entrants then move upmarket, and disruption has occurred once the incumbents' mainstream customers start buying from them in volume.

The article names two kinds of starting point. A low-end foothold is a set of customers who are overserved, for whom a good-enough product is enough. A new-market foothold is a set of nonconsumers, people priced out or otherwise unserved, as when personal copiers reached school librarians, bowling-league operators and other small customers who had made do with carbon paper or mimeograph machines while Xerox sold to large corporations. Mainstream customers, the authors add, do not switch just because something is cheaper. They wait until quality reaches their standard, and only then accept the lower price.

By these rules the authors declare that Uber is not disruptive to taxis, because it launched in a well-served market and its first customers already hired rides, and that Tesla is not disruptive either, because its foothold was a high-end segment the incumbents care about. They predict that an entrant attacking head-on with a better product will meet an incumbent that accelerates, and they report that in Christensen's disk-drive study only 6% of sustaining entrants succeeded. They also warn managers not to dismantle a profitable business out of fear, and they concede that the theory says very little about how to win once you are in the foothold market and that it will never explain everything about business success.

One more detail matters for AI. The authors describe the pace of a disruption as a function of how fast the enabling technology improves. In steel, Nucor needed more than 40 years to match the revenue of the largest integrated mills. Compaq reached parity with DEC in 12 years. A foothold can look harmless for a very long time, and a verdict reached in one year of data says almost nothing.

Its record as a forecaster is worse than the confidence it inspires

Four sources should lower anyone's confidence in using the theory to call outcomes. The authors of the 2015 article take the opposite view in their closing paragraphs, stating that empirical tests show the theory makes predictions about which fledgling businesses will succeed measurably and significantly more accurate. That passage does not say which tests they mean, so I cannot weigh it.

The first is the iPhone, above. A prediction from the theory in 2007 failed, and the entrant was reclassified later. Joshua Gans, an economist who has written about the theory repeatedly, made the point in 2012 that a product being sustaining or disruptive cannot by itself predict whether it succeeds.

The second is Jill Lepore's 2014 New Yorker essay. Dan Kennedy's account of it records her arguments that the disk-drive companies Christensen presented as casualties of disruption did quite well, and that his steel analysis ignored labor strife. Christensen's reply, as Bloomberg Businessweek reported it that June, was that she had read only the 1997 book and not the later refinements. On the iPhone he conceded that he had missed it. I have not been able to open the New Yorker text, so I rely on Kennedy's summary for her claims.

The third is the study by Andrew King and Baljir Baatartogtokh in MIT Sloan Management Review in 2015. They surveyed and interviewed 79 industry experts, by the count in the journal's published debate, about 77 cases that Christensen and Raynor had cited, and coded each case against four elements of the theory: a leading incumbent on a sustaining trajectory, overshoot of customers' needs, an incumbent able to respond that failed to exploit that capacity, and an incumbent that then floundered. Where experts disagreed, the authors gave the theory the benefit of the doubt. In Gans's summary of their results, 53 of the 77 cases had the first element, 17 the second, 47 the third and 48 the fourth, and only 7 had all four. I could not open the results table myself, so those counts are secondhand.

Gans objected to how the study was scored. The first element works as a selection filter and not as a test, he said, and overshoot is a sideshow because disruption does not need oversupply. Martin Bienenstock called the share of cases the theory explains meaningless, since it was never meant to explain every reason a leader declines. I think both points have force. They also concede the larger fact: seen from outside, a clean match between a real industry story and the full theory is the exception.

The fourth is Erwin Danneels, who in 2004 used Christensen's work to examine five issues, among them how disruptive technology is defined and what predictive use the theory has. I have read only his abstract, which says he challenges and integrates the existing theory and raises questions for research. I cite it because it shows the predictive question was already on the table two decades ago.

None of this shows the mechanism is false. The record supports a weaker use: the theory sorts situations better than it forecasts them, and sorting is how I apply it below.

Five questions sort a threat into sustaining, low-end or new-market

The following checklist is my own condensation of the 2015 article and of the elements King and Baatartogtokh coded. It is not a checklist the authors published, and it asks about the evidence available at the time of the decision.

1. Who buys the entrant's product first: the incumbent's best customers, overlooked low-end customers, or people who were buying nothing?
2. Is the product worse than the incumbent's by the mainstream's own yardstick on day one?
3. Are the incumbent's mainstream customers overserved, already holding more performance than they use?
4. Would a response hurt the incumbent's margins, channels or pricing, or can it be added to what it already sells?
5. How fast does the entrant improve, and what exactly does it have to reach to satisfy mainstream customers?

Read the answers together. If the first customers are the incumbent's best customers, the product is not worse, and the response can be bundled, the entrant is on a sustaining path, and the theory predicts a fight that the incumbent usually wins. If the first customers are nonconsumers or overserved, the product is worse by the mainstream's yardstick, and responding would cost the incumbent money, the conditions of the theory hold, and the fifth question tells you how long you have. Most real cases land between those poles, which is useful in itself, because the mixed answers tell you which assumption to test next.

Suite copilots improve the incumbent product within its existing business model

Microsoft announced on 21 September 2023 that Microsoft 365 Copilot would be generally available to enterprise customers on 1 November at $30 per user per month. On the company's April 2026 earnings call, it reported more than 20 million paid seats. Three months later, on 29 July 2026, it reported more than 30 million, with net seat additions more than doubling from the prior quarter. On that call Satya Nadella named NHS England, rolling Copilot out to 505,000 clinicians and staff, and KPMG, expanding across a workforce of more than 276,000 professionals.

Run the questions. The first buyers are the incumbent's existing enterprise customers, who already pay for the suite. The product is an add-on to the suite, so the question of whether it is worse than the incumbent's offer does not apply. Overshoot is not the issue either; the open matter is how much customers use it, which paid seat counts alone do not establish. The response question does not arise, because the incumbent is the entrant. And the finance chief, Amy Hood, said on that call that premium offerings, Copilot among them, drove growth in average revenue per user, which is what a sustaining innovation does: it lets the firm sell more to its most profitable customers at a higher price. The fifth question, the slope, is the only one left open, and it concerns how fast the product improves for existing users, not whether a newcomer is climbing. The same logic applies to other assistants that owners embed in their own products.

The theory classifies this as sustaining innovation; it does not establish that Microsoft is safe from competitors or that the product earns an adequate return. The live question here is adoption. Taking the January 2026 figure of more than 450 million paid commercial seats from Microsoft's fiscal Q2 earnings call, 30 million paid Copilot seats is roughly 6 to 7 percent of the base, my own arithmetic and not a company disclosure, nearly three years after general availability. Microsoft gave no newer seat total in July and said commercial seats had grown 6% over the year, so the denominator is an estimate and the share is approximate. Whether that is a slow start or a deliberate pace for a product with a premium price, I cannot tell from public data. The competitive questions that remain, such as what happens when the model supplier is also a rival, belong to a different framework.

Chegg fits half the pattern, and a search-engine change confounds the rest

Chegg is the case most often cited as an AI casualty, and it fits the theory partly. The Associated Press reported on 2 May 2023 that Chegg shares fell 48% to $9.08 after its chief executive said on a call that, since March, student interest in ChatGPT had spiked and was affecting new customer growth. Students who usually paid around midterms or finals were reluctant to subscribe when a free alternative existed. The stock had traded above $100 in early 2021.

Total net revenue was $766.9 million in 2022, $716.3 million in 2023 and $617.6 million in 2024, according to its 10-K filings, then $376.9 million in 2025, a fall of 39% in the company's February 2026 release. Its legacy Academic Services line brought in $154.0 million in the first quarter of 2024 and $54.9 million in the fourth quarter of 2025. Its restructuring plans of May and October 2025 cut about 640 employees, or 56% of the workforce then in place.

Apply the questions. The first buyers of the free chatbot were price-sensitive students, who are the low end of Chegg's market. On the second question, a general assistant was less reliable than answers checked by experts, and Chegg's own release says that its accuracy and quality keep retention high among those who remain. That pattern of the less demanding customers leaving first and the more demanding ones staying is what the theory describes. On the third question, whether a student who wants a quick draft is overserved by verified step-by-step solutions is my judgment and not a finding.

The mechanism breaks at the fourth question. The theory expects an incumbent that does not respond. Chegg announced CheggMate, built on GPT-4, in April 2023, so it responded within months, and it shrank anyway. The 10-K for 2025 adds a second driver: Google's expansion of AI Overviews from August 2024, which it says keeps users on Google results instead of sending them to Chegg. Its 2024 10-K reports that non-subscriber traffic fell 8% year over year in the second quarter of 2024, 19% in the third, 39% in the fourth and 49% in January 2025. So, by Chegg's own account, part of the traffic collapse traces to an incumbent search engine adding AI to its own product. That is a sustaining innovation by Google, and Chegg was a casualty of it. A free substitute and the loss of a distribution channel explain Chegg at least as well as the foothold-and-climb story, and I would not call it a textbook disruption.

Search is still an open case, and a year of revenue cannot close it

Michael Horn of the Christensen Institute argued on 3 June 2024 that the business model determines whether generative AI participates in a disruption. His article identifies Google as the large technology firm facing the clearest risk, on the reasoning that an AI answer product could start worse than search, would be hard for Google to monetize, and could improve quickly. Those are the conditions in the third, fourth and fifth questions.

The first question is partly answered by usage data. An NBER working paper from September 2025 by Aaron Chatterji and colleagues, six of whom list an OpenAI affiliation, analyzed a sample of consumer ChatGPT conversations. It found that adoption reached about 10% of the world's adults by July 2025, that non-work messages grew from 53% to more than 70% of use, and that practical guidance, information seeking and writing made up nearly 80% of conversations. Information seeking is Google's home ground, and that part is a head-on contest. Practical guidance and writing, for people who had no professional to ask, look more like a new-market foothold. The paper does not say which.

Now the response question. Alphabet's second-quarter 2026 release, filed on 22 July 2026, reports Google Search and other revenue of $63,271 million against $54,190 million a year earlier, up 17%, and says its AI features are driving query growth. Google defended the core head-on, which is what the 2015 article says incumbents do against head-on entrants, and revenue has so far not dropped. The condition under which the theory would bite is different: a foothold in nonconsumption that climbs into commercial queries, in a form the advertising model cannot copy. A year of revenue cannot show that, given Nucor's 40 years and Compaq's 12. I cannot settle whether practical guidance is a climbing foothold or an adjacent market that never reaches Google's mainstream, and the data in hand does not let anyone else settle it either.

The strongest objection says a theory this narrow explains almost nothing

Robin Chase raised it in an HBR response in January 2016, observing that the authors' definition counts Zipcar as disruptive and Uber as not, and arguing that the definition should be expanded. The worry is that a definition this strict leaves out the changes that hurt incumbents most. Ben Thompson made a related point in 2014, that the low-end half of the theory fits business purchasing and fails with consumers, who do not buy on measurable attributes alone. AI chatbots are consumer products, so that is the half most in doubt. And Horn's own conclusion, that AI is neither inherently disruptive nor inherently sustaining, makes the label depend on the business using it. The objection, then, is that a label which can excuse Uber, miss on the iPhone and flatter the incumbent in the Copilot case has no use.

My answer has three parts. First, the 2015 article itself treats Uber as an outlier and explains it by regulation, which left taxi companies little room to innovate. A sustaining label is therefore not a safety certificate. It says the incumbent has the usual counter-moves, and where it cannot execute them, as with Chegg's dependence on search traffic, it can lose anyway, for reasons that supplier power and the other forces in the earlier essay on five forces describe better.

Second, a narrow instrument is what stops two expensive errors. One is treating a competitor who woos your best customers as a disruptor and dismantling a profitable business in response, which the authors warn against. The other is ignoring a nonconsumption foothold because the entrant's product looks like a toy.

Third, the DeepSeek episode of January 2025 shows the instrument at work. DeepSeek's R1 paper, posted on 22 January, reported performance comparable to OpenAI's o1-1217 on reasoning tasks, with 79.8% on the AIME 2024 benchmark against 79.2%, and slightly below o1 on GPQA Diamond, and the weights were released openly. On 27 January Nvidia's stock fell 16.9%, nearly $600 billion of market value by TechCrunch's count. The reaction treated R1 as a disruptor. Applying the five questions gives a different reading. By my reading its customers are the same developers and companies that buy from the incumbents, and the paper claimed parity, not inferiority, so there was no overlooked low end for it to start in. That is price competition from an entrant attacking head-on, and the theory expects incumbents to defend by improving their own offerings. A one-day market move told us what investors feared, not what the process was.

This does not make the theory a forecaster. It is a way of asking what kind of fight you are in, and of writing the answers down so they can be wrong visibly. Before the next planning cycle, I would take each AI entrant on a leadership team's worry list, answer the five questions in a sentence each with the evidence, and revisit them every quarter against one number the theory cares about: whether the entrant's customers have started to look like the incumbent's mainstream.

Sources

1. Christensen, C. M., Raynor, M. E., McDonald, R. (2015). What Is Disruptive Innovation? Harvard Business Review 93(12), 44-53. Read through a PDF copy hosted on a university course site.
2. Christensen, C. M. (1997). The Innovator's Dilemma. Harvard Business School Press. Not opened; cited only through the 2015 article.
3. Christensen's 2007 BusinessWeek interview on the iPhone, as quoted in Gans, J. (2012). Clay Christensen on the iPhone: wrong about success but right about disruption. digitopoly.org, 2 July 2012.
4. Lepore, J. (2014). The Disruption Machine. The New Yorker, 23 June 2014 (not opened; site blocked), as summarized in Kennedy, D. (2020). When Clayton Christensen and Jill Lepore tangled over disruption theory. dankennedy.net, which also quotes Christensen's reply in Bloomberg Businessweek, 20 June 2014.
5. King, A. A., Baatartogtokh, B. (2015). How Useful Is the Theory of Disruptive Innovation? MIT Sloan Management Review 57(1), 77-90 (method read on the publisher's page; results counts from Gans, J. (2015). Disrupting the theory of disruption. digitopoly.org, 21 September 2015). Responses by Vazquez Sampere, Bienenstock, Zuckerman and Gans: Debating Disruptive Innovation, MIT Sloan Management Review (2015).
6. Danneels, E. (2004). Disruptive technology reconsidered: a critique and research agenda. Journal of Product Innovation Management 21(4), 246-258. Abstract only.
7. Thompson, B. (2014). Critiquing Disruption Theory. Stratechery. Chase, R. (2016). We Need to Expand the Definition of Disruptive Innovation. Harvard Business Review online, January 2016 (opening excerpt only).
8. Horn, M. B. (3 June 2024). What does Disruptive Innovation Theory have to say about AI? Christensen Institute. https://www.christenseninstitute.org/blog/what-does-disruptive-innovation-say-about-ai/.
9. Microsoft (2023). Announcing Microsoft 365 Copilot general availability, 21 September 2023. Microsoft fiscal Q4 2026 earnings call transcript, 29 July 2026. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4. Microsoft fiscal Q2 and Q3 2026 earnings calls: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q2 and https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3.
10. Alphabet Inc. (2026). Exhibit 99.1, second quarter 2026 results, 22 July 2026. https://s206.q4cdn.com/479360582/files/doc_financials/2026/q2/2026q2-alphabet-earnings-release.pdf.
11. Chegg, Inc. Form 10-K for fiscal 2024 and fiscal 2025; Chegg Reports 2025 Fourth Quarter and Full Year Financial Results, 9 February 2026. Associated Press (2 May 2023). Chegg's stock plunges on fears of competition from ChatGPT.
12. DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948 (v1, 22 January 2025). TechCrunch (27 January 2025). Nvidia drops 600bn off its market cap amid the rise of DeepSeek.
13. Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y., Wadman, K. (2025). How People Use ChatGPT. NBER Working Paper 34255. https://cdn.openai.com/pdf/a253471f-8260-40c6-a2cc-aa93fe9f142e/economic-research-chatgpt-usage-paper.pdf.
14. Illustrative arithmetic: 30 million Copilot seats divided by a base of about 450 million commercial seats is roughly 6.7%; The 39% fall in Chegg's total net revenue is $376.9 million against $617.6 million. Both are my calculations from the cited figures.