Why leaders mistake a memorable story for market data — and a six-question test to catch it before it costs you.
Depending on which headline you read this year, AI is either transforming the global economy or failing 95% of the time. Both statements can point to real research. Neither one is lying. The problem is that they were never measuring the same thing.
One widely circulated figure traces back to MIT’s NANDA initiative, which found that roughly 95% of enterprise generative-AI pilots studied had not yet produced measurable P&L impact, as reported by Tom’s Hardware (2026). That is a statistic about corporate pilot programs. It says nothing about whether AI startups are surviving. Meanwhile, IdeaProof’s 2026 database catalogs 168 notable AI-related shutdowns — a real number, but one the firm explicitly frames as tracking newsworthy failures, not the mortality rate of the entire AI-startup population (IdeaProof, 2026). At the same time, Stanford’s 2026 AI Index reports that 1,953 new AI companies were funded in the United States in 2025 alone, backed by $285.9 billion in private investment (Stanford HAI, 2026).
Three real numbers. Three different denominators. And in most boardrooms, they get compressed into one sentence: “AI is failing.”
That compression has a name, and it costs companies money well beyond AI.
Why the Story Beats the Spreadsheet
In 1973, Amos Tversky and Daniel Kahneman described what they called the availability heuristic: people estimate how common or likely something is based on how easily examples come to mind, not on its actual frequency (Tversky & Kahneman, 1973). Vivid, recent, and emotionally charged events are easier to recall — so they feel more representative than they are.
A closely related pattern, base-rate neglect, compounds the problem. When people have a specific, compelling case in front of them, they tend to underweight the broader statistical frequency of the category it belongs to (Kahneman & Tversky, 1973). Put the two together and you get a predictable failure mode: a memorable anecdote enters the room, and the base rate never gets a vote.
Executives are not exempt. Reviews of professional judgment across management, finance, and medicine find that expertise reduces but does not eliminate heuristic reasoning (PubMed Central, n.d.).
When the Story Became the Strategy: WeWork
The clearest documented case of a growth narrative overrunning its own denominator is WeWork. In January 2019, SoftBank’s Vision Fund valued the company at $47 billion, one of the highest private valuations ever assigned to a U.S. venture-backed startup (Nguyen, 2023). The valuation wasn’t built on unit economics — WeWork’s 2018 filings showed $1.8 billion in revenue against a $1.9 billion loss (Nguyen, 2023). It was built on a story: that a company leasing office space and subleasing desks was a “tech-enabled, community-as-a-service platform.”
When WeWork filed to go public in August 2019, investors finally looked past the narrative to the denominator — the ratio of revenue to losses, the lease liabilities, the governance structure. The valuation collapsed to roughly $8 billion in six weeks, the IPO was withdrawn, and the CEO was gone within a month (Nguyen, 2023). Nothing about WeWork’s business changed in those six weeks. What changed was who was finally asking about the denominator.
The Boardroom Version of This Mistake
You don’t need an AI narrative or a $47 billion valuation to fall into this trap. It shows up in smaller, weekly decisions.
A CMO sees a competitor’s LinkedIn post hit 10,000 likes and asks why the team isn’t “doing more content like that” — without asking how many of the competitor’s other posts got 40 likes, or whether the viral one converted a single lead.
A CRO hears about two lost deals that mentioned price and concludes the company is “too expensive” — without checking what percentage of all losses were actually price-driven, and how that compares to last quarter.
A CEO reads that a well-known company killed its freemium tier and declares freemium “dead” — without asking how many comparable companies still run freemium successfully.
None of these observations are false. The salesperson really did lose the deal. The competitor really did go viral. That is exactly why the bias is dangerous: the anecdote is true, and it quietly becomes “usually true” without anyone deciding that it should.
A Six-Question Test Before an Anecdote Becomes Strategy
Before a memorable example turns into a resourcing decision, a hiring rule, or a pricing change, it should survive six questions:
- What exactly is the claim? “AI fails” is not testable. “Enterprise GenAI pilots frequently show no P&L impact within their first deployment cycle” is.
- What is the denominator? Name the population the statistic actually describes.
- Are the numerator and denominator measuring the same thing? This is where both the WeWork story and the AI-failure story break.
- How was the outcome defined? “Failure” can mean bankruptcy, a missed forecast, a canceled pilot, or a bad quarter — each producing a different rate.
- Why is this example memorable? Recency, drama, and personal experience inflate availability without changing the base rate.
- What would the complete dataset say? If you could see every outcome instead of only the ones that became a story, would the conclusion hold?
That last question is the discipline. It separates a leadership team that reacts to whichever example is loudest this week from one that builds its Revenue Architecture on evidence that actually predicts what happens next.
The Takeaway
The problem was never that stories are wrong. WeWork’s story was, for a while, true — revenue really was growing. The AI-failure headlines are true too, for the population they describe. The problem is letting a true, vivid story stand in for a denominator nobody checked.
Before your next strategy meeting turns “everyone says” into a budget line, ask the only question that matters: compared with what?
References
IdeaProof. (2026). 2026 startup failure report. IdeaProof.io.
Kahneman, D., & Tversky, A. (1973). On the psychology of prediction. Psychological Review, 80(4), 237–251.
Nguyen, B. (2023, November 7). WeWork’s rise to $47 billion — and fall to bankruptcy: A timeline. Forbes.
PubMed Central. (n.d.). Cognitive bias in professional judgment: A review.
Stanford HAI. (2026). The 2026 AI index report. Stanford Institute for Human-Centered Artificial Intelligence.
Tom’s Hardware. (2026). MIT NANDA study: 95% of enterprise generative-AI pilots show no measurable ROI.
Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5(2), 207–232.
