Every number on this site, with its source.
Where each figure came from, when it was published, and what it does not prove. Each article lists its own sources inside it.
6 figures, and the limit of each.
56%
of CEOs said AI has brought no significant financial benefit so far.
What it does not prove: 4,454 CEOs in 105 countries. What they report, not audited results.
40%
of desk workers received AI-made work in the past month that looked finished but was not. Each case took about two hours to sort out.
What it does not prove: 1,150 full-time US desk workers, self-reported. The hours are their own estimates, and BetterUp sells coaching.
2 hours
less on email each week, for workers who used an AI tool. Beyond that, the researchers found no change in how much work people did, or what kind.
What it does not prove: A six-month field experiment: 7,137 workers in 66 firms, one AI tool. The hours are email time only. A change not detected is not proof of none.
46%
of UK and Ireland business leaders saw little profit-and-loss impact from AI so far.
What it does not prove: 510 leaders, self-reported. “Little impact” is not “no value”.
50% vs 67%
Engineers who learned a new tool with AI help scored lower on a test of what they understood.
What it does not prove: 52 engineers, one task, tested straight after. A small, specific study.
80%+
of domain experts at one law firm dropped out of its AI innovation work.
What it does not prove: Two organisations over two years. A pattern, not a rate for all firms.
The research
The studies behind ABC and the book.
Research-informed, not proof: none of these studies tested the ABC Method directly.
Behind the three checks
- Kok, Mostert & de Lange, 2017Assumptions: The brain fills gaps before it looks.
- Reber & Schwarz, 1999Assumptions: Easy to read feels true.
- Abrahamyan et al., 2016Biases: We lean towards what we already chose.
- Sharma et al., 2023Biases: AI can learn to agree with you.
- Cowan, 2010Complexity: We can hold about four things at once.
- Kool et al., 2010Complexity: We prefer the easier task.
- Hollerman & Schultz, 1998Afterwards: Surprise is how we learn.
- Arnsten, 2009 · Edmondson, 1999Afterwards: Fear narrows thinking. Safety opens it.
Behind the story of the book
- Hiroshi IshiguroAndroid research on presence, behaviour and brain responses. The 2024 meeting is Amit’s own account; Professor Ishiguro’s work is the setting for the question, not an endorsement of the book.
- Masahiro Mori, “The Uncanny Valley” (1970)The original essay, in the authorised English translation. A proposed relation, drawn as a concept, not a measured curve.
- Anil Seth, “Conscious artificial intelligence and biological naturalism”, Behavioral and Brain Sciences (2025)Argues that intelligence is not the same as consciousness, and that real artificial consciousness is unlikely on current paths. The phrase “controlled hallucination” is from his book Being You (2021).
- Roshni Lulla, “From the Trolley Dilemma to Robot Sociopaths” (2026)Separates reading another’s feelings from sharing them, and asks what follows for AI.
- Christov-Moore et al., “Preventing antisocial robots: A pathway to artificial empathy”, Science Robotics (2023)A viewpoint arguing that empathy built only on its cognitive side can encourage harmful behaviour.
- Kosmyna et al., “Your Brain on ChatGPT”, MIT Media Lab (2025)A preprint with 54 participants, and 18 in the final session. Small, early and widely debated.
- Yiu, Kosoy & Gopnik, “Transmission versus truth, imitation versus innovation”, Perspectives on Psychological Science (2023)Large models as cultural technologies: strong at passing on what people know, weaker at discovering what is new. The coffee-sleeve test is from the preprint version (arXiv 2305.07666), where children aged three to seven chose the object that worked and the models mostly chose the one whose name went with the usual tool.
Widely quoted. Not printed here.
“95% of generative AI pilots fail”
It comes from one early report, The GenAI Divide by MIT Project NANDA (2025): 52 interviews, 153 survey responses and 300 public deployments, not peer-reviewed. It found no measurable profit-and-loss impact yet, which is not the same as failure.
“AI triples knowledge worker productivity across all roles”
The large speed-ups come from narrow tasks in test conditions, such as simple code or short copy. At work, review, compliance and fixing errors can cancel the gain.
Corrections
What we got wrong, and fixed.
Each entry keeps the earlier wording so readers can see what changed and why.
Klarna service-agent figure
Sourced to a "Klarna SEC Filing & Associated Press".Sourced to Klarna's own announcement of 27 February 2024, and restated as the company's claim about message volume rather than an audited outcome.
Why: The documents named did not carry the figure. It is a company statement, and the page now says so at the point of claim.
Microsoft Frontier description
Stated that "Microsoft is investing $2.5 billion".Restated as "Microsoft announced $2.5bn and more than 6,000 people".
Why: Microsoft would not say whether the sum is new or repurposed spending, or over what period. The original wording made a claim the source does not support.
CEO survey wording
Reported as "neither increased revenue nor lower costs".Corrected to PwC's published wording: "no significant financial benefit from AI to date".
Why: The original was our paraphrase presented as a quotation. Same finding, but the survey should be quoted as it was written.
Cohort size in this log
A previous entry stated the CEO cohort was corrected from N=1,200 to N=1,400.Entry withdrawn. The PwC cohort is 4,454, as the figures table states.
Why: The entry contradicted the audited figure on the same page. It described a correction that never took place, which is the exact failure this log exists to prevent.