AI Disruption vs AI Hype: How to Tell the Difference

Many claims that AI will "disrupt" an industry use the word as a synonym for "change a lot." That makes them impossible to test. The original theory of disruptive innovation is narrower: it describes a specific path a competitor takes, and it gives you questions you can check against evidence. Here is that definition, the signals that separate real adoption from demos, and a checklist you can run on any claim.

What disruption means in Christensen's sense

In the December 2015 Harvard Business Review article "What Is Disruptive Innovation?", Clayton Christensen, Michael Raynor and Rory McDonald restated the theory in one sentence: disruption "describes a process whereby a smaller company with fewer resources is able to successfully challenge established incumbent businesses."

The mechanism runs in a particular order:

1. Incumbents keep improving their products for their most demanding, most profitable customers. In doing so they overshoot what some customers need and ignore others entirely.
2. An entrant gets a foothold in one of two places. A low-end foothold serves overserved customers with a "good enough" product, often at a lower price. A new-market foothold turns nonconsumers into customers. The authors' example is personal copiers in the late 1970s, which served small organizations that had been priced out of Xerox's machines.
3. The incumbent, chasing higher margins, tends not to respond vigorously.
4. The entrant improves and moves upmarket. Disruption has happened when mainstream customers adopt the entrant's offering in volume.

Three details from the article do most of the work when you apply it to AI:

• Disruption is a process, not a product. When Netflix launched its mail-order DVD service in 1997, it did not appeal to most of Blockbuster's customers. It became a threat because of the path it followed, eventually reaching Blockbuster's core customers through streaming.
• Success is not part of the definition. In the authors' words, "Not every disruptive path leads to a triumph, and not every triumphant newcomer follows a disruptive path." They argue that Uber, for all its growth, was not disrupting the taxi business, because it built a position in the mainstream market first instead of starting from a foothold.
• Technologies are rarely disruptive by nature. The article says it is rare that a technology or product is inherently sustaining or disruptive. What matters is how a business uses it.

Why "AI is disruptive" is usually the wrong sentence

That last point makes "AI is disruptive" a category error. A model is not disruptive. A business path can be. The testable version of the claim names who is using AI, to serve which customers, with what business model, against which incumbent.

Run that test and many AI launches turn out to be sustaining. When an established vendor adds a model to a product its customers already pay for, those customers get a better version of what they had. That can be valuable, but it is not disruption. And when an entrant tries to win those same customers head-on with a better AI product, the theory predicts the incumbents will speed up their own innovation and either beat the entrant back or acquire it. The HBR piece cites Christensen's disk drive research, where only 6% of sustaining entrants succeeded.

A disruptive AI path looks different. Picture an entrant selling a stripped-down, AI-run version of a professional service to small customers who never bought the full-service version, at a price the incumbent would not bother to match. The incumbent's best clients would call it inferior. That is the pattern to watch: not whether it is impressive today, but whether it improves fast enough to satisfy the mainstream while keeping its cost advantage.

What hype looks like

Hype has a recognizable shape: it talks about capability and skips the customer.

• A demo stands in for a market. A demo shows what a system did once, under chosen conditions. It says nothing about error rates, cost per task, or who pays.
• Growth is offered as proof of disruption. That is the exact error the HBR authors warn against.
• Urgency without a mechanism. "Disrupt or be disrupted" skips the question of which customers move and why. The authors say incumbents should respond to disruption that is actually occurring, but should not overreact by dismantling a still-profitable business.
• Counts without a denominator. "Used by thousands of teams" does not tell you what share of a market uses it, how often, or for what.

Adoption numbers depend on who was asked

Adoption data is the natural check on hype, but read the fine print. Stanford's 2026 AI Index reports that organizational AI adoption rose in 2025 "up to 88% of surveyed organizations." The US Census Bureau's Business Trends and Outlook Survey found that overall AI use hovered between 17% and 20% of businesses from December 2025 to May 2026, rising to 37% among firms with 250 or more employees. The Census Bureau also broadened its question in November 2025, from use "in producing goods or services" to use "in any business function."

Both figures can be accurate. The AI Index figure counts the organizations in the survey it draws on. The Census figure is a nationally representative view of US businesses, including firms with four or fewer employees. And each defines "use" in its own way. Before you lean on an adoption number, find out who was asked and what counted.

Stronger signals than "uses AI":

• Repeat use without a mandate. People keep using the tool after the pilot ends and nobody is tracking them.
• Budget moves. Spending shifts from an experiment line to an operating budget with a named owner.
• Workflow redesign. Handoffs, staffing plans or job descriptions change around the tool.
• Price response. Incumbents cut prices, unbundle, or launch a cheaper tier aimed at the entrant's segment.
• Procurement shifts. Mainstream buyers, not just early adopters, put the entrant on their shortlist. Mainstream customers adopting in volume is the point at which, by the HBR definition, disruption has occurred.

Read the cost curve and the quality curve separately

Disruption depends on two curves: the cost of delivering a given level of performance, and the level of performance mainstream customers need.

On cost, the 2025 AI Index reports that the inference cost for a system performing at the level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024. Falling cost at a fixed quality level is the kind of change that opens low-end footholds: tasks that were not worth automating at the old price become worth it at the new one.

On quality, the same report shows scores rising by 18.8, 48.9 and 67.3 percentage points on the MMMU, GPQA and SWE-bench benchmarks in a single year, while models still struggled with complex reasoning benchmarks such as PlanBench. Benchmark gains are real, but a benchmark is not your customer's requirement.

So plot the claim, even roughly:

• Define "good enough" in the mainstream customer's terms: error rate, turnaround time, liability, and how the output gets checked.
• Place the AI offering against that line today, along with its price.
• Check again each quarter.

If the offering is already better and costs more, it is sustaining. If it is cheaper but not improving on the dimension customers care about, it may stay a niche. If it is cheaper, below the line, and closing the gap, you are looking at a possible disruptive path.

A checklist for any AI disruption claim

Run these eight questions before you accept or dismiss a claim. They apply the ideas above to AI and overlap with the six-question test the Christensen Institute publishes:

1. Who is the incumbent, and who are its most profitable customers?
2. Is the entrant starting from a low-end foothold, a new-market foothold, or neither? Neither usually means sustaining.
3. Would the incumbent's best customers call the new offering inferior today?
4. Is the business model different, or only the technology? The HBR authors note that disrupters often build business models very different from those of incumbents.
5. Does the incumbent have a reason not to respond, such as margins, channel conflict or cannibalization?
6. Is there adoption evidence beyond pilots: repeat use, budget, workflow change, procurement?
7. Does the cost advantage hold at the quality level customers need, measured on a real task?
8. What result in the next 12 months would prove the claim wrong?

The last question is the fastest filter. A claim that cannot name the evidence that would disprove it is a slogan, and you can set it aside until someone can.