The Handover Document Is the Worst Way to Leave Thirty Years Behind

Two things are walking out of the building with you, and only one of them writes down.

The mechanics, in order:

• Documented knowledge survives writing. Procedures, contacts, thresholds, the quirks of a system nobody has rewritten since 2011. A handover document carries those adequately.

• Judgement does not survive writing. Why a rule was overruled, which supplier was never trusted and on what evidence, which quiet signal means a job is about to go wrong.

• Judgement comes out incident by incident, under questioning from another person. One real job you finished recently, walked through in order, probed at every point where you chose.

• The highest-yield question is what would have had to be different for you to choose the other way. Then ask what a competent beginner would have done instead.

• Structured prompts beat free recall by half again. In a 2014 study of surgeons, free recall produced 44% of a 46-step gold standard and structured prompts produced 66%.

• Two experts beat one by a wide margin. No single programmer in the Chao and Salvendy study could report more than 41% of their own diagnostic actions. Six of them, aggregated, reached 87%.

• Record audio rather than notes, and start while you are still doing the job. Speech leaves less room to tidy, and tidying is how the decisions get lost.

Michael Polanyi wrote the reason in one line, about driving: "We can know more than we can tell... The skill of a driver cannot be replaced by a thorough schooling in the theory of the motorcar" (quoted in Autor, 2014). Skill becomes automatic in order to become fast, and automatic processes stop reporting back to the part of you that makes sentences. Ikujiro Nonaka built the standard organisational model on the same split: what a firm knows is created "through a continuous dialogue between tacit and explicit knowledge" running through four patterns of interaction, rather than through one act of writing things down (Nonaka, 1994).

Take one page of your handover document: escalate to the regional manager if the variance exceeds eight percent. The page does not say where eight came from, whether the conditions that set it still hold, or what it cost the one time you went on the number alone. The conclusion survived. What produced the conclusion did not.

Trying harder does not fix this. A different kind of question does, asked by somebody else about one job you actually did.

Maura Sullivan and colleagues at the University of Southern California videotaped three expert surgeons teaching a cricothyrotomy in 2010, then sat the same three down for cognitive task analysis interviews. Three further surgeons produced a gold-standard list of 46 steps, and both sets of transcripts were scored against that list. Academic Medicine, May 2014:

• Teaching the procedure, experts omitted an average of 71% of clinical knowledge steps (10 of 14).

• They omitted 51% of action steps (14 of 27) and 73% of decision steps (3.6 of 5).

• For action steps, they described how to do the thing only 13% of the time.

• Inside the interviews, free recall produced 44% of the steps on the list (20 of 46). With structured elicitation prompts, that rose to 66% (31 of 46).

The same expert. The same interview. The same procedure. Half again as many steps, once somebody started asking properly (Sullivan et al., 2014, PMID 24667500). Prompting recovered 22 of the 56 points free recall left on the table, which is not everything.

Those were surgeons, and surgeons teach for a living. Your handover document will be written by somebody who was never trained to teach at all.

The recovered material also trains people better. A meta-analysis in BJS Open in November 2021 pooled twelve studies and found trainees taught from cognitive task analysis had better procedural knowledge (SMD 1.36, 95% CI 0.67 to 2.05) and better technical performance (SMD 2.06, 95% CI 1.17 to 2.96) than those taught conventionally (Edwards et al., 2021, PMID 34904648). Tofel-Grehl and Feldon found a large overall effect, Hedges's g = 0.871, with effect sizes varying substantially by method and training context (Tofel-Grehl & Feldon, 2013). A 2022 systematic review in Pilot and Feasibility Studies found 81 articles from 13 countries published between 1993 and 2019, most often in surgical and critical care settings, the critical decision method and cognitive task analysis interviews the most used forms (Swaby et al., 2022, PMID 35260195).

Expert knowledge is not lost when the last practitioner dies. It is lost decades earlier, when nobody asks them a specific question about a specific piece of work.

The W76 life extension programme began in 2000, and the United States National Nuclear Security Administration needed a material called Fogbank. The GAO report, GAO-09-385, published 2 March 2009, put the failure plainly. NNSA "had lost knowledge of how to manufacture the material because it had kept few records of the process when the material was made in the 1980s and almost all staff with expertise on production had retired or left the agency". NNSA spent 69 million dollars addressing Fogbank production problems. Plans to produce the first refurbished W76 weapon slipped from September 2007 to September 2008.

They had records, a budget no workshop will ever see, and no competitor. They still lost the process, because what they kept were records of the product rather than the reasoning of the people who made the material. Donald MacKenzie and Graham Spinardi had set out the general risk a decade earlier: nuclear weapons design rests on tacit knowledge that can be lost along with the people who hold it (MacKenzie and Spinardi, 1995).

If that apparatus could lose a manufacturing process because the people who held it retired, your employer's shared drive is not going to hold your thirty years.

Start building your legacy Free build, 50 memories, no card required.

Nobody invented this to sell software. Researchers worked it out on people who make hard calls fast.

John Flanagan published the critical incident technique in Psychological Bulletin in 1954 (Flanagan, 1954). Gary Klein, Roberta Calderwood and Donald MacGregor extended it into the critical decision method in IEEE Transactions on Systems, Man, and Cybernetics in 1989, and their probes name what your handover document lacks: they "elicit aspects of expertise such as the basis for making perceptual discriminations, conceptual discriminations, typicality judgments, and critical cues" (Klein, Calderwood & MacGregor, 1989). Their subjects are why the method travels: urban and wildland fireground commanders, tank platoon leaders, structural engineers, design engineers, paramedics, computer programmers.

Laura Militello and Robert Hutton published Applied Cognitive Task Analysis in Ergonomics in 1998, calling the fuller method "resource intensive" and "of limited use to design practitioners" and offering three streamlined interviews instead. The handbook is Crandall, Klein and Hoffman's Working Minds, MIT Press, 2006. None of it is proprietary, and you can run one on yourself this afternoon.

Forty minutes, on your own, recording audio rather than notes.

1. Pick one incident, not a topic. A day when you made a call a competent person three years into the job would have got wrong. Say the date out loud.

2. Talk it through in order, start to finish, before you analyse anything. Do not tidy the story.

3. Go back to the beginning. At each decision point: what did I see that nobody else in the room saw, and what would have had to be different for me to choose the other way. The second question is the highest yield in the method.

4. Then: what would my successor have done here, and what would that have cost.

5. Stop. Do the next incident another day.

Topics are where expertise goes to die. At the level of a topic your knowledge and a textbook's look identical, and the differences appear in one job on one day. Any version of how do you know when it is ready opens the most doors, and the first answer is almost always that you just know, so ask instead what a beginner calls finished that is not finished. A fault you can name is a fault you can teach.

The wrong-call question is the one people resist, and one non-routine incident is the unit the critical decision method inherited from Flanagan. Most published clinical work went the other way: of the 80 studies in the 2022 review, 64 examined routine or typical events and only 16 challenging or rare ones (Swaby et al., 2022).

One incident at real depth runs thirty to ninety minutes, and thirty to fifty incidents covers most of what a successor hits in their first two years. Twenty-five to forty hours, then, four months at two hours a week. That arithmetic is planning experience, not a research finding. The longer question set is at how to interview yourself.

Everything above assumes a named successor. Some readers have none, and the skill is in their hands.

A craft is classed by Heritage Crafts as critically endangered partly where there is "no mechanism to pass on the skills and knowledge". The test is the mechanism, not the surviving objects. The 2025 edition, published 12 May 2025 with 20 new additions, assessed 285 crafts and featured 165 on the Red List: 72 critically endangered, 93 endangered. Five are recorded extinct in the UK within the past generation. Hand-stitched cricket ball making. Gold beating. Lacrosse stick making. Mould and deckle making. Mouth-blown sheet glass making.

The list is not all obituaries. No craft has become extinct in the UK since the 2023 edition, Heritage Crafts introduced a resurgent category in 2025, and a glassblower funded by the charity is reviving mouth-blown flat glass, a craft thought lost the year before. A record of the how is part of coming back, though a record is not an economics fix. The 2025 Red List named rising operational costs, a lack of structured training and market pressure as the forces putting UK crafts at risk, and a skill dies because nobody can make a living at it more often than because nobody wrote it down.

Words carry decisions and nothing else. They do not carry grip, sequence, angle or pace, so film the hands and mumble your way through. They do not carry the visual standard, so photograph the rejects, because a standard is a boundary and a boundary needs both sides.

An apprentice wants one answer at eleven at night with a piece on the bench. Heritage bodies want the craft recorded as intangible cultural heritage, and UNESCO's Convention for the Safeguarding of the Intangible Cultural Heritage, adopted 17 October 2003 and in force since 20 April 2006, treats traditional craftsmanship as a domain and defines safeguarding to cover documentation and transmission. UNESCO's Living Human Treasures programme, running from 1993 until the Convention took effect, recognised individuals for their accomplishments and their willingness to pass their knowledge on. Heritage Crafts reported in May 2025 that the UK government had ratified the 2003 Convention, one of the last countries in the world to do so.

Whoever pays for the capture usually owns the output, so settle this first.

A document you wrote at work belongs to your employer, and that is fine: a document does not sound like you and cannot answer an unanticipated question. A cloned voice keeps working after you have gone, can be used for sentences you never said, and the person it sounds like is you (how to protect your voice from AI cloning has the failure modes).

Four things in writing before anybody pays for anything. Who holds the source material, the audio and the transcripts, because every model built later is downstream of that. Who is permitted to run a voice built from it, and answer to whom. Whether the voice may be used beyond answering questions from named colleagues. And what happens on the day you leave, including whether you can revoke.

The clean structure is that you own the recordings and your employer holds a licence for a defined purpose and term. The messy structure is that they commission the sessions as a work product, and in five years a version of your voice is onboarding graduates at a company you left (who owns your voice clone has the detail). If you developed the method itself as an employee, under a research grant or a commercial agreement, rights in the method may not be yours either, which is a contract question for a lawyer and nothing here is legal advice.

One constraint if you or your employer operate in the EU. From 2 August 2026, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) requires that AI systems intended to interact directly with people be designed so those people are informed they are dealing with an AI system, unless that is obvious, and requires deployers of deepfake audio, image or video to disclose that the content is artificially generated or manipulated, clearly and distinguishably, at the latest at the time of first interaction. Your firm cannot quietly put a version of your voice in front of staff. What Article 50 requires covers scope and penalties.

Nobody reads a handover. A queryable record does not ask them to read one.

The realistic use is narrow and under pressure: a job is two weeks late, the client has gone quiet, and there is a decision to make by four o'clock. Your successor asks one question in the words already in their head and gets back the time this happened to you, what you saw, what you did, and what the wrong version cost. No apprentice searches two hours of video for the ten seconds that hold the answer.

Then there is the stupid question, asked six times, at no social cost, because nobody in their first year of a senior role wants to ask a peer why the eight percent rule exists. Disagreement becomes legible too: when your successor decides against your recorded reasoning, they know what they are overruling, and change your rules on purpose rather than by drift.

Now the limits. One person is not a domain. Chao and Salvendy asked six expert programmers to work through hard troubleshooting tasks and explain what they had done, and no single expert could report more than 41% of their own diagnostic actions, 53% of their debugging actions or 29% of their interpretations. Aggregated, all six reached 87%, 88% and 62% (Chao & Salvendy, 1994, as reported in Clark et al., 2008). Use two experts where you can.

Not all of your intuition is skill. Kahneman and Klein agreed that trusting expert intuition depends on how predictable the environment is and whether the person had a real chance to learn its regularities, and that how confident an intuition feels is not a reliable indicator of accuracy (Kahneman & Klein, 2009). Your recorded reasoning is weakest against new regulation, a new competitor, or a market that stops behaving as it did for three decades, and nothing here replaces an apprenticeship.

This is what we build at Afterlife.ai™, and the honest version of the pitch is narrow. A guide called Idy asks about one specific incident, listens, and follows up on the parts you skipped, choosing the next question from what you just said. Each session produces memories, discrete and editable. Your voice is cloned from your own recordings, so an apprentice hears the explanation in the voice that goes with the hands. Idy will not push back the way a colleague who knows your trade does, hearing the sentence you glossed over and refusing to move on, so the strongest sessions are the ones where somebody sits beside you and Idy holds the structure. The video of your hands stays outside the product, because the pressure through a handle does not come back through a voice model.

What the sessions build is a "Persona": a version of you assembled from what you actually said, answering from your own memories, never inventing what you never told them. Personas talk the way you talk rather than read out a coverage report, so a thin area sounds like a short answer. You nominate who gets access, one person at a time, granted and revoked by you while you are alive, and Executor Lock covers release if you can no longer make the call: seven days, evidence verified, stoppable throughout, and nothing releases to anyone you did not name.

A last thing. You cannot explain why you overruled the model on a particular job without saying who was in the room and what you were like at the time. Judgement is stored inside episodes, and the episodes have you in them. Nobody in your family will ever ask for the handover document. Some of them will want the hours of you explaining why you did things the way you did.

You are doing the work anyway. Own the recordings and you get both.

Start building your legacy Free build, 50 memories, no card required.

How long does capturing my expertise actually take?

Twenty-five to forty hours of talking for a thorough capture, spread over three to four months, with one incident running thirty to ninety minutes. That figure is a planning estimate from practice rather than a research finding. Expect the first session to feel slow, because the format stops feeling like an exam once you realise you are telling work stories with follow-up questions attached.

Who owns the result if my employer pays for it?

Whoever the contract says, which is why you settle the question before the first recording. The clean structure is that you own the audio and the transcripts and your employer holds a licence for a defined purpose and term. Voice recordings are not like documents, because a cloned voice keeps working after you leave and can be used for things you never agreed to. Get the revocation right in writing.

Can my successor tell when there is nothing recorded on a topic?

Not reliably, so plan around that. Concrete facts have to come from your recorded memories, and Personas will not manufacture a specific answer nobody gave them. They are also built to talk like you rather than to narrate their own coverage, so a thin area reads as a short answer. Hand your successor a list of what you covered and treat everything else as unrecorded until you record it properly.

What does it cost to start capturing my expertise?

Nothing. The free build gives you 50 memories, requires no card, and does not expire, which is enough to run several real sessions and find out whether being questioned pulls anything out of you that a document did not. The interview method on this page costs nothing at all if you run it with a friend and a phone, and you should do that first. See the plans for what the paid tiers add.

What happens if the company running this shuts down?

The durable asset is the recordings and the transcripts, not the model, because a model can be rebuilt from good source material. Keep your own copies of the raw audio, the video of the hands and the stills. You can export your recorded memories as a data file, which carries the text rather than the media, so the media copies have to be yours. That is the right question to put to every vendor in this category, including us.

Can I delete all of this later?

Deleting your account is yours to trigger without a support conversation and without giving a reason, and it removes your memories, your recordings and your stored audio. The voice model held at the synthesis provider is a separate matter that the delete button does not reach today, so ask us where that stands before you record if that matters to you. If an employer holds a licence to a copy, that copy is governed by the licence you signed, which is one more reason to sort ownership first.

• How to interview yourself: the full free question set.

• The things only you know how to do: when you doubt your working knowledge counts.

• Who owns your voice clone: the ownership detail.

• What an AI trained only on me actually is: a general model against one answering from your recordings.

• Who controls your identity after you die: consent, revocation, control of a likeness.

• What makes a voice clone safe or unsafe: the external risks, plainly.

Every number here comes from a primary source read directly, and all of them are listed below with a DOI, a PMID or a URL.

What is not settled: the effect sizes vary a lot by method and setting, so treat them as a strong signal and not a constant. The Chao and Salvendy percentages are quoted as reported in Clark and colleagues' 2008 review chapter, because the publisher does not release the 1994 abstract. And the training studies measured humans taught from elicited material, not apprentices questioning a model built from that material, which is newer and unmeasured.

Found an error? Report the mistake to us and we will check the source and correct the page.

1. Maura E. Sullivan, Kenneth A. Yates, Kenji Inaba, Lydia Lam and Richard E. Clark, "The use of cognitive task analysis to reveal the instructional limitations of experts in the teaching of procedural skills", Academic Medicine, vol. 89, no. 5 (May 2014), pp. 811-816. DOI 10.1097/ACM.0000000000000224. PMID 24667500. https://pubmed.ncbi.nlm.nih.gov/24667500/

2. Thomas C. Edwards, Alexander W. Coombs, Bartosz Szyszka, Kartik Logishetty and Justin P. Cobb, "Cognitive task analysis-based training in surgery: a meta-analysis", BJS Open, vol. 5, no. 6 (9 November 2021), zrab122. DOI 10.1093/bjsopen/zrab122. PMID 34904648. https://pubmed.ncbi.nlm.nih.gov/34904648/

3. Lizzie Swaby, Peiyao Shu, Daniel Hind and Katie Sutherland, "The use of cognitive task analysis in clinical and health services research: a systematic review", Pilot and Feasibility Studies, vol. 8, article 57 (8 March 2022). DOI 10.1186/s40814-022-01002-6. PMID 35260195. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8903544/

4. Autor, D. H. (2014). Polanyi's Paradox and the Shape of Employment Growth. NBER Working Paper No. 20485. https://doi.org/10.3386/w20485 (source of the verbatim Polanyi 1966 quotation, "We can know more than we can tell")

5. Michael Polanyi, The Tacit Dimension, University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/T/bo6035368.html

6. Nonaka, I. (1994). A Dynamic Theory of Organizational Knowledge Creation. Organization Science, 5(1), 14-37. https://doi.org/10.1287/orsc.5.1.14

7. Klein, G. A., Calderwood, R., & MacGregor, D. (1989). Critical decision method for eliciting knowledge. IEEE Transactions on Systems, Man, and Cybernetics, 19(3), 462-472. https://doi.org/10.1109/21.31053

8. Laura G. Militello and Robert J. B. Hutton, "Applied cognitive task analysis (ACTA): a practitioner's toolkit for understanding cognitive task demands", Ergonomics, vol. 41, no. 11 (November 1998), pp. 1618-1641. DOI 10.1080/001401398186108. PMID 9819578

9. Beth Crandall, Gary A. Klein and Robert R. Hoffman, Working Minds: A Practitioner's Guide to Cognitive Task Analysis, MIT Press, 2006. DOI 10.7551/mitpress/7304.001.0001

10. Flanagan, J. C. (1954). The critical incident technique. Psychological Bulletin, 51(4), 327-358. https://doi.org/10.1037/h0061470

11. Tofel-Grehl, C., & Feldon, D. F. (2013). Cognitive Task Analysis-Based Training: A Meta-Analysis of Studies. Journal of Cognitive Engineering and Decision Making, 7(3), 293-304. https://doi.org/10.1177/1555343412474821

12. Chao, C.-J., & Salvendy, G. (1994). Percentage of procedural knowledge acquired as a function of the number of experts from whom knowledge is acquired for diagnosis, debugging, and interpretation tasks. International Journal of Human-Computer Interaction, 6(3), 221-233. https://doi.org/10.1080/10447319409526093 (the 41/53/29 and 87/88/62 percentages quoted on this page are as reported in source 21, the publisher does not release the abstract)

13. Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515-526. https://doi.org/10.1037/a0016755

14. United States Government Accountability Office, Nuclear Weapons: NNSA and DOD Need to More Effectively Manage the Stockpile Life Extension Program, GAO-09-385, 2 March 2009. https://www.gao.gov/products/gao-09-385

15. Donald MacKenzie and Graham Spinardi, "Tacit Knowledge, Weapons Design, and the Uninvention of Nuclear Weapons", American Journal of Sociology, vol. 101, no. 1 (July 1995), pp. 44-99. DOI 10.1086/230699

16. Heritage Crafts, "Craft skills remain under threat with 20 new additions to the Red List of Endangered Crafts", 12 May 2025. https://heritagecrafts.org.uk/our-stories/red-list-2025-launched/

17. Heritage Crafts, "Categories of risk", Red List of Endangered Crafts, checked 30 July 2026. https://heritagecrafts.org.uk/redlist/categories-of-risk/

18. UNESCO, Convention for the Safeguarding of the Intangible Cultural Heritage, adopted 17 October 2003, in force 20 April 2006. https://ich.unesco.org/en/convention

19. UNESCO, "Living Human Treasures: a former UNESCO programme", established 1993. https://ich.unesco.org/en/living-human-treasures

20. EU AI Act, Article 50, Regulation (EU) 2024/1689. Applies from 2 August 2026. https://artificialintelligenceact.eu/article/50/

21. Richard E. Clark, David F. Feldon, Jeroen J. G. van Merrienboer, Kenneth A. Yates and Sean Early, "Cognitive Task Analysis", chapter 43 in Handbook of Research on Educational Communications and Technology, 3rd edn (2008). Source for the reported Chao and Salvendy (1994) figures. https://hpttreasures.wordpress.com/wp-content/uploads/2018/11/cta_chapter_2008.pdf

22. Heritage Crafts, news release, "Craft skills remain under threat with 20 new additions to the UK's Red List of Endangered Crafts", May 2025. Source for the 285 assessed, 165 listed, 72 critically endangered and 93 endangered figures, for the resurgent category, and for the UK ratification of the 2003 UNESCO Convention. https://heritagecrafts.org.uk/wp-content/uploads/2025/05/NEWS-RELEASE-Red-List-of-Endangered-Crafts-2025-v2.pdf

Last reviewed: 30 July 2026. On a 90-day recheck cycle: the Article 50 application date, the Heritage Crafts Red List (revised every two years) and the UNESCO ratification position.