1. Why folklore costs money
A made-up number in a slide deck isn't harmless. It sets budgets ("plan for a 30% drop"), schedules ("hypercare lasts 30 days"), design rules ("learning styles") and expectations ("AI makes developers 55% faster"). When reality doesn't match, the project looks like it failed when it was the number that was wrong. CM-EVD-01
The rule the Database follows is simple: don't repeat a statistic you can't source, and when you can source one, keep it attached to its population, method and year. Each claim below appears in the Database's retired-claims lists, with the entry that replaces it. CM-EVD-01 · FAM-EVD-01 · PRM-RTS-02
2. Learning and training
| Stop quoting | What it actually rests on |
|---|---|
| 1. "Only 10% of training transfers to the job." | A 1982 rhetorical opener, re-cited as data. |
| 2. Learning styles (visual, auditory, kinesthetic) as a design rule. | Matching teaching to a preferred style doesn't improve learning. |
| 3. "Attention spans are 8 seconds." | No meta-analytic support, and no basis for fixed "brain-based" module lengths. |
| 4. "People remember 20% of what they read and 80% of what they see." | No research basis for these figures. |
| 5. "70-20-10" as a budget rule. | Executives recalling what developed them. A reminder, not a measurement. |
| 6. "It takes 21 days to form a habit." | No supporting evidence was found. |
| 7. "The forgetting curve means 90% is lost in a week." | Ebbinghaus studied one person learning nonsense syllables. |
| 8. "VR training is 4x faster" or "75% retention". | Uncontrolled vendor claims; overall, VR matches conventional training. |
| 9. "Microlearning improves retention by X%." | No direct comparisons support it. |
| 10. "Gamification increases retention by up to 40%." | A vendor blog with no study cited. |
| 11. "The 10,000-hour rule." | An average from one music study, turned into a threshold. |
| 12. "Our training returned 543%." | Under the ROI method's own conservative rules, the same example returns about 71% in the first year. |
Say instead: "Practice from memory, spaced over time, with feedback, has strong evidence. We'll check transfer on the job with supervisors and records, not a satisfaction survey." LSC-MYTH-01 · LSC-MYTH-03 · LSC-EVD-02 · GOL-TRF-03
3. Change, go-live and decisions
| Stop quoting | What it actually rests on |
|---|---|
| 13. "70% of change efforts fail." | Traced to five sources in 2011; none held valid, reliable evidence. Success rates depend on definitions. |
| 14. "Hypercare lasts 30 days." | No source supports a fixed length. |
| 15. "One champion per 20 users." | No evidence supports any champion ratio. |
| 16. "Plan for a 30% productivity drop at go-live." | Trade-magazine opinion; studies range from no long-term loss to a loss that never recovered. |
| 17. "Overlearning prevents skill decay." | Contested. Schedule refreshers anyway. |
| 18. "94% of problems are the system, 6% the people." | Deming's personal estimate, not a measurement. Keep the principle, drop the number. |
| 19. "Decide when you have 70% of the information." | No controlled evidence, and only ever meant for reversible choices. |
| 20. "Pre-mortems make predictions 30% more accurate." | The study found about 30% more reasons for failure; accuracy wasn't measured. |
| 21. "Half of all decisions fail." | The source could not be found. |
| 22. "Unsafe acts cause 90% of injuries." | A framing Texas City used before 2005; investigators found systemic failures. |
Say instead: "We'll end hypercare when tickets are back to forecast and there are no open severity-1 defects. We'll measure our own productivity dip by role, against a baseline." GOL-HYP-01 · GOL-HYP-03 · QMS-LSS-04 · DGN-FRM-03
4. Projects and quality
| Stop quoting | What it actually rests on |
|---|---|
| 23. "98% of megaprojects overrun by 80% and 20 months." | No traceable primary source. |
| 24. "9 in 10 projects overrun." | That's megaprojects, on a stated baseline; about half of large projects come in on budget. |
| 25. "Design-build delivers 102% faster." | A square-feet-per-month speed ratio for buildings, not projects finishing in half the time. |
| 26. "$177 billion lost to non-optimal time." | A 2018 vendor survey extrapolation: directional only. |
| 27. "41% of the construction workforce retires by 2031." | Found only in secondary sources; check the primary labour tables first. |
| 28. "The 1-10-100 rule." | Traces to a 1992 book with no data; versions differ. |
| 29. "Rework is 30% of construction." | A report figure; commentators settle near 5% direct. |
| 30. "60–70% of Lean Six Sigma programmes fail." | Old, with varying definitions of failure. |
| 31. "Operations are 80% of lifecycle cost" or "1:5:200". | Couldn't be substantiated when examined; real ratios are lower and vary widely. |
Say instead: "Here's how comparable projects actually went, from named sources, and here's what our own nonconformances cost last year." CM-RSK-02 · QMS-COQ-01 · QMS-COQ-03 · QMS-RWK-03
5. Software, AI and knowledge
| Stop quoting | What it actually rests on |
|---|---|
| 32. "AI makes developers 55% faster." | One greenfield lab task. Workplace trials show more modest gains, and one showed experts slowing down. |
| 33. "The 10x programmer." | Mixed conditions and languages; at most about 5 to 1 is down to individuals. |
| 34. "Elite performers deploy 973 times more often." | Year-specific cluster ratios from 2021; the tiers have since been retired. |
| 35. "Only 16% of software projects succeed." | A proprietary method with criticized definitions of success. |
| 36. "Expert personas and 'think step by step' make AI more accurate." | Controlled 2025 tests found little or no gain on current reasoning models. |
| 37. "It passed the engineering exam, so it can do engineering." | An aggregate score hid zero correct answers in structural design. |
| 38. "Grounded AI doesn't hallucinate." | Grounded commercial legal tools were wrong 17 to 33% of the time. |
| 39. "People spend 1.8 hours a day searching for information." | A 2012 extrapolation, repeated mostly by vendors. |
| 40. "Knowledge-Centered Service cuts costs by X%" or "AI deflects X% of tickets". | Member- or vendor-reported; no independent controlled study. |
Say instead: "We'll measure the speed-up with telemetry on our own work, and test the assistant on our own questions before we trust it." SWD-AI-01 · SWD-AI-03 · SWD-MET-02 · PRM-EVD-02 · PRM-EVD-06 · KIN-TWN-04
6. Check a number yourself
You don't need a research team to catch most folklore. Five questions do most of the work:
- Who measured it? A peer-reviewed study, an auditor, a trade body, a vendor or nobody? Vendor and member-reported figures are claims until someone independent checks them. FAM-EVD-01
- What exactly was measured? "30% more reasons for failure" isn't "30% more accurate"; "attendance" isn't "space utilization". PRM-RTS-02
- In whom? Students, physicians, one music academy, megaprojects? Effects found in classrooms are usually smaller in complex field work. LSC-EVD-02
- Can you reach the primary source? If every citation points to another citation, stop. CM-EVD-01
- Compared with what? A gain with no comparison group, or a baseline taken in the worst month, tells you little. DGN-BSL-01
When a number fails these questions, you rarely lose the argument by dropping it. "We don't know the industry rate, so here's ours" is more persuasive to a sceptical executive than a famous statistic they've already heard debunked.
The habits. Don't repeat a statistic you can't source. Keep every number attached to its population, method and year. Prefer your own baseline to an industry average. Treat vendor figures as claims. Say "associated with", not "proven", when that's all the evidence supports.