Industries that learn — and the one that won't

A working hypothesis: industries do not improve because the people in them get smarter. They improve because they are *structured to learn*. The most visible improvement curve of the last forty years belongs to software, and software's secret is not the cognitive horsepower of its workforce. The secret is that software has, by accident of its substrate, every property a learning system needs — and an open-source culture that ratchets those properties further with every passing year. If you want to know why adult safeguarding cannot self-improve, the right move is not to study safeguarding more carefully. It is to study the industries that do learn, identify what they share, and notice that safeguarding has none of it.

This essay does that, and then asks the question the comparison forces: why does adult social care not have an Air Accidents Investigation Branch?

A note before starting. Institutional learning structure is the property this essay argues is the single most under-appreciated factor in why some industries improve and others don't. It is not the only factor. Technological tractability, measurability of outcomes, capital concentration, and strong professional norms all contribute too. The institutional argument should not be read as denying any of those — only as identifying the structural feature that is most clearly absent from English adult safeguarding and most clearly correctable by reform. Necessary, in other words, but not on its own sufficient.

What software actually does

The achievements of the last four decades of software are easy to underestimate because they happened in plain sight. A web browser in 2026 renders typography, runs animations, handles encrypted payments, streams 4K video, and translates the page into any of ninety languages — and the entire thing was given away free. The Linux kernel runs roughly two-thirds of the world's servers, every Android phone, most embedded devices, and the Mars rovers; it was written by volunteers and is also free. Git, the version-control system that organises this whole edifice, was written by Linus Torvalds in ten days. Python — the language much of modern science now runs on — is maintained by a foundation that costs less to operate than a single mid-sized council's communications department.

None of this is the consequence of unusual intelligence. The people writing this software are intelligent, but no more so than the consultants and policy advisers staffing failed government programmes at ten times the cost. The difference is structural. Software has, all at once:

GitHub did not invent these properties; the open-source culture had them informally since the 1980s. What GitHub did was give them a substrate — a single global place where the entire improvement engine could run continuously. The result is one of the most powerful distributed learning systems humans have built outside science itself. Software also accumulates technical debt, regresses in reliability across many domains, and depends substantially on capital subsidy from corporate sponsors; the institutional learning argument captures one important property of why software improves, but it is not the only property at work.

The interesting question is whether any of this is transferable. Some of it clearly is.

The other learning industries

The pattern of self-improvement, once you know what to look for, recurs in a small number of places. None matches software fully, but each captures enough of the six properties to produce a visible improvement curve over decades.

Other domains that capture some of the pattern are worth a brief mention: Formula 1 and elite motorsport have fast feedback and cheap-enough iteration in a heavily measured environment, and the FIA has developed a substantial safety regime, but the analytical comparison is weaker than aviation or anaesthesia because the FIA both regulates and investigates, and the cross-era performance comparisons depend on changing regulations, track layouts, fuel rules, and tyre eras in ways that make headline lap-time claims unreliable. The improvement is real; the institutional fidelity to the AAIB model is partial.

Each of these maps onto the six properties to a different degree. Aviation has slower feedback than software (an accident is rare, by design) but the formal investigation system compensates by making the rare failures count enormously. Anaesthesia and surgery have fast feedback but expensive iteration — you cannot run an A/B test on anaesthesia mortality — so they invested in the apparatus of formal failure investigation and the parallel technological standardisation. Toyota's innovation was retrofitting the worker-can-stop-the-line mechanism onto a mass-production system that had previously suppressed it.

The unifying observation: industries learn when their failures are investigated by a body operationally independent of the people responsible for them, and the findings are made public and propagated across the whole sector. Every one of the cases above has some version of this. Software's bug tracker is the lightest-touch instance; aviation's AAIB is the heaviest. Without the independent investigation property, the improvement curve flattens.

The HSSIB precedent

Before turning to safeguarding, one further institutional fact matters. The AAIB model has, in the last decade, been deliberately transplanted into UK healthcare. The Healthcare Safety Investigation Branch was established in 2017 as an arm's-length body and became the statutory Health Services Safety Investigations Body (HSSIB) on 1 October 2023, with the same operational independence, statutory powers of compulsion, and safe-space testimony rules that the AAIB has had since 1915.^[9] HSSIB is, in legal and operational terms, an exact transplant of the AAIB model into healthcare. The transplant has happened. It is operational. It investigates patient-safety incidents that involve exactly the kind of ambiguous human-system complexity that critics of the AAIB-for-safeguarding proposal sometimes invoke as a reason the model cannot transfer.

The relevance is direct. If the AAIB structural features can produce learning in commercial aviation (where the events are technical-causal), in anaesthesia (where the events are clinical), and now in NHS healthcare (where the events involve contested values, incomplete information, and ambiguous human relationships) — then the case that the model cannot work in adult safeguarding has to identify what specifically about safeguarding makes it different from healthcare. That case has not, to date, been made.

The industries that do not learn

The negative cases are at least as illuminating. They tend to have one or two of the six properties but never all six, and the missing properties cluster around independent failure attribution.

The cost of the alternative architecture

The Grenfell architecture of accountability is what happens when a sector lacks the apparatus to learn from its near-misses and earlier failures. The Grenfell Tower Inquiry cost over £200 million across its seven years.^[11] The Metropolitan Police's parallel criminal investigation, announced in May 2026 as being ready to send files to the Crown Prosecution Service in respect of fifty-seven individuals and twenty companies, has cost a further £150 million to date.^[16] No criminal trial is expected before 2027 or 2028, eleven years after the fire. The accumulated cost — the public-inquiry bill, the police investigation bill, the future court costs, the £2 million replica of the tower being built so a jury can understand the building, the civil claims, the cost of replacing relevant cladding nationally, and the personal cost to seventy-two families and the survivors — is the cost of an industry that did not have the apparatus to learn from Lakanal in 2009, or from the earlier tower-block fires, or from the manufacturers' product-certification anomalies that an independent ongoing investigation would have surfaced.

This is what the architecture of accountability looks like in scale terms when relied on as the sector's learning mechanism. It is enormously more expensive than the AAIB-equivalent alternative, enormously slower, focused on retrospective blame rather than continuous learning, and produces its conclusions for one disaster at a time. The Building Safety Regulator created under the Building Safety Act 2022 is structured as a regulator rather than as an investigator — the regulator-versus-investigator split that aviation has between the CAA and the AAIB, and that healthcare has between the CQC and HSSIB, the built environment still does not have. The lesson generalises. For adult safeguarding, the equivalent accumulating cost is harder to see because it does not happen in a single newsworthy moment — but the structural mechanism is the same, and the Casey Commission's decision about whether the new National Safeguarding Board has AAIB-fidelity structural features will determine whether adult social care continues to accumulate that distributed cost at the current rate for the next decade.

One useful counterexample: the National Patient Safety Agency

It is worth dwelling briefly on a case where the form of an AAIB-style learning system was attempted in UK healthcare and did not produce the expected effect. The National Patient Safety Agency was established in 2001 as a special health authority of the NHS, charged with collecting incident reports through the National Reporting and Learning System (NRLS), promoting a no-blame reporting culture, and issuing safety guidance.^[13] By the time of its abolition in 2012 under the Health and Social Care Act, the NRLS had received roughly 10 million reports. Yet the NPSA never produced the kind of measurable improvement curve that anaesthesia or aviation did.

The reason is instructive. The NPSA had the form of an AAIB equivalent — incident reporting, blame-free framing, central analysis — but it lacked the AAIB's key structural features: operational independence (it sat inside the NHS, not outside it), statutory powers of compulsion, and the legal protection of testimony for safety-purpose use only. NPSA reports were not protected from use in litigation. Staff submitting incident reports could not be assured of confidentiality. The body's recommendations had no statutory propagation requirement. The NPSA had the apparatus of learning without the structural conditions that make learning happen.

The lesson for any new investigation body, including any adult safeguarding equivalent, is precise: form alone is not enough. The six structural features — operational independence, statutory powers, safe-space testimony, family-initiated investigation rights, sector-wide propagation, system-not-individual focus — are not optional. The NPSA's failure is the empirical demonstration of what happens when an institution adopts the appearance of the model without its load-bearing structural commitments. HSSIB, by contrast, was designed with those commitments in legislation and is producing the kind of cross-sector learning the NPSA could not.

Score the learning industries against the six properties and the pattern is stark.

● present ◐ partial ○ absent

Adult safeguarding scores zero of six. Feedback is slow: a wrongly-investigated family's harm takes years to surface, often only via the Local Government Ombudsman after the council's two-stage internal complaints process has been exhausted. Iteration is rare: statutory regimes change once a decade, and even then on a glacial parliamentary cycle. Reversibility is absent: once a safeguarding referral is made, the family it concerns cannot undo it, and the record persists in council systems for years. Failure attribution is blocked by individual confidentiality and institutional self-protection. Learning is not shared: every council reinvents its own Multi-Agency Safeguarding Hub, its own thresholds, its own training. And the culture fuses the work and the worker through the defensive practice incentive structure — to criticise the procedure can read as criticising the people executing it, which makes the criticism feel personal and is therefore deflected.

The AAIB question

Adult safeguarding does, formally, have an investigation mechanism. Under section 44 of the Care Act 2014, every Safeguarding Adults Board is required to commission a Safeguarding Adults Review when a vulnerable adult dies or experiences serious harm and partner agencies could have worked more effectively to protect them.^[14] The mechanism was the architects' good-faith attempt to import the AAIB principle. It does not work, and the reason it does not work is structural.

The fundamental problem is that the body conducting the investigation is the body being investigated. Each Safeguarding Adults Board is constituted of the local authority, the police, the NHS, and the partner agencies whose conduct the SAR is examining. The independent author appointed to write the report is selected by the SAB, paid by the SAB, briefed by the SAB, and dependent on the SAB for future commissions. The published report is signed off by the SAB. The recommendations are addressed back to the same SAB, who decides how to implement them and reports on their own progress.

Compare this with how the AAIB investigates an aviation accident. The Air Accidents Investigation Branch is operationally independent of every UK airline, every airport, the Civil Aviation Authority, and the Department for Transport (to which it nominally reports). Its inspectors have statutory power to enter premises, take possession of evidence, and compel testimony. Its reports are published in full, and the system is deliberately structured to prioritise safety learning over liability attribution, with statutory protections around witness statements and safety-purpose evidence designed to allow pilots and engineers to give honest evidence without fear of disciplinary or legal consequence.^[2] Recommendations are issued to manufacturers, airlines and regulators across the entire industry. The AAIB has no enforcement power, but its findings are taken seriously because the alternative is to be the airline that ignored a published safety recommendation and then suffered the next accident.

Every one of those structural features is absent from the Safeguarding Adults Review system. No operational independence: SARs are commissioned by the very bodies being reviewed. No statutory powers of compulsion: SAR authors must rely on the partner agencies' willingness to disclose. No blame-free testimony regime: frontline workers giving evidence to an SAR are doing so within the same disciplinary apparatus they will face afterwards. No cross-sector propagation: SAR findings stay locally, occasionally summarised in the LGA's national thematic analyses,^[15] but with no statutory duty on any other SAB to read or act on them. No structural protection against the report being used in liability proceedings: SARs are routinely cited in inquests and civil claims, which structurally incentivises every party to disclose as little as possible.

The result is a regime that produces the form of an aviation-style learning system while producing little of its substantive output. SARs are conducted; the national library is maintained; thematic analyses are published. Practice does not visibly improve.

What an AAIB for adult safeguarding would look like

An ASCIB constituted along these lines would not be cheap. The AAIB operates on a budget of roughly £12 million a year with around 80 staff to oversee an industry that carries 270 million UK passengers annually. HSSIB, the closest functional comparator, operates at roughly £15 million annually. An ASCIB would likely be of similar or somewhat larger scale, because adult safeguarding's case volume — 640,000 concerns and 185,000 section 42 enquiries a year in England alone — is vastly higher than either aviation accidents or NHS serious-incident equivalents. Unlike the AAIB, which can investigate substantively every serious aviation accident, an ASCIB would have to triage, select exemplary cases, and work thematically — investigating illustrative cases and patterns rather than every individual failure. The case-selection mechanism is part of the design; HSSIB has already worked through analogous questions for healthcare. The cost is non-trivial but small relative to the £25 billion England spends annually on adult social care, and small relative to the social cost of the failures the body is designed to identify.

The obvious objections, briefly

"This is what the CQC is for."

The Care Quality Commission inspects providers and (since the 2022 Health and Care Act) assesses local authorities. It does neither in the AAIB mode. CQC inspections are scheduled, broad-spectrum, and structured around regulatory compliance categories rather than focused around specific failures. Its reports rate authorities on a four-point scale and rarely identify named causal mechanisms. It has no statutory blame-free reporting regime, no power of family-initiated investigation, and its recommendations are framed at the level of the inspected body rather than the sector. It is a regulator, not an investigator. Aviation has both — the CAA regulates, the AAIB investigates — and healthcare now has the same split, with the regulator (CQC) and the investigator (HSSIB) operating as separate bodies. The distinction is not accidental.

"This is what the Ombudsman does."

The Local Government and Social Care Ombudsman adjudicates complaints from individual residents. It does not investigate systemic patterns, has no statutory powers of compulsion, can recommend remedies only at the individual level, and cannot find against named workers. Its remit is restorative justice for the complainant, not safety improvement for the sector. It is closer in design to a small-claims tribunal than to the AAIB.

"Safeguarding is too varied to learn from in this way."

Aviation accidents are also enormously varied — engine failures, weather events, pilot error, structural fatigue, ATC failures, terrorism, runway incursions. Healthcare incidents, which HSSIB now investigates, are if anything more varied still. The AAIB's methodology has produced a coherent improvement curve over six decades, and HSSIB's early operation suggests it is doing the same in healthcare, because variation in cases is not, in fact, an obstacle to learning. Pattern-recognition across heterogeneous cases is what investigators are trained to do. Adult safeguarding is not categorically harder than healthcare.

"The political will doesn't exist."

This is largely true, and not really an objection — it is a separate fact about the world. The argument here is not that ASCIB is politically achievable in the current parliament; it is that the absence of an ASCIB-like body is a sufficient structural explanation for the regime's failure to self-improve, and that without one no amount of training or guidance or restructuring will change the trajectory. The political will to fund another Munro-style review will probably appear after the next high-profile failure. The question is whether the next review proposes the right thing.

The most striking feature of the comparison is how undramatic the proposal is. Nothing in the ASCIB design above is novel, untested or experimental. Every component has been operating in aviation for decades, has been transplanted into anaesthesia, into rail safety (the RAIB), into marine accidents (the MAIB), and most recently into healthcare via HSSIB. The Department for Transport has, in fact, an entire ecosystem of independent investigation branches across its sectors, and the Department of Health and Social Care has now joined them. Adult social care — a sector whose annual safeguarding failures, when fully counted, exceed by orders of magnitude the total fatalities in the entire history of UK commercial aviation — remains without an equivalent body.

The corporate-stupidity diagnosis from the earlier essay in this series identified the institutional mechanism by which the safeguarding regime tends to produce poor outcomes from cognitively capable people. The comparative data brief established that the outcomes are real, measurable, and concentrated in identifiable councils. This piece names what is missing: not better training, not new guidance, not another Care Act, but the simple structural apparatus every learning industry has been built around, and which adult social care alone has been allowed to do without.

Software and aviation are not smarter. They are structured to learn. Safeguarding is structured to defend. The same humans, given the same problems, produce wildly different outcomes depending on which structure they are working inside. Until the structure changes, the outcomes will not.

Sources

1. [S] Boeing Commercial Airplanes (2025). Statistical Summary of Commercial Jet Airplane Accidents, Worldwide Operations 1959–2024, 54th edition, April 2025. The Boeing Statistical Summary is the standard industry reference for commercial jet accident rates and has been published annually since 1969. Headline figure: accident rates declining from about 50 per million departures in 1959 to low-single-digit rates by the 2000s. Boeing's own copy: https://www.boeing.com/content/dam/boeing/boeingdotcom/company/about_bca/pdf/statsum.pdf . Skybrary archive: https://skybrary.aero/articles/boeing-annual-summary-commercial-jet-airplane-accidents .

2. [S] UK Air Accidents Investigation Branch (AAIB) — official gov.uk page: https://www.gov.uk/government/organisations/air-accidents-investigation-branch . The AAIB operates under the Civil Aviation (Investigation of Air Accidents and Incidents) Regulations 2018, which implement Regulation (EU) No 996/2010 as retained UK law, including the safe-use protections for safety-purpose evidence.

3. [S] Aviation Safety Reporting System (ASRS), operated by NASA for the FAA: https://asrs.arc.nasa.gov/ . UK CHIRP (Confidential Reporting Programme for Aviation and Maritime): https://chirp.co.uk/ . Both schemes operate on the AAIB principle that confidential blame-free reporting produces safety information that would not otherwise surface.

4. [R] The historical literature on anaesthesia mortality is summarised in Bainbridge, D., Martin, J., Arango, M., Cheng, D., and the Evidence-based Peri-operative Clinical Outcomes Research Group (2012), 'Perioperative and anaesthetic-related mortality in developed and developing countries: a systematic review and meta-analysis,' The Lancet 380(9847): 1075–1081. The reduction from roughly 1 in 10,000 to lower rates over the 1980s-2000s period is well-documented, with the precise figures sensitive to definitional choices about anaesthesia-attributable versus perioperative mortality.

5. [S] Anesthesia Patient Safety Foundation (APSF) — official site: https://www.apsf.org/ . APSF was founded by Ellison "Jeep" Pierce in 1985 with the explicit mission of "no patient shall be harmed by anesthesia." Its methodology was modelled on aviation's safety-investigation approach.

6. [R] World Health Organization (2009), WHO Surgical Safety Checklist: https://www.who.int/teams/integrated-health-services/patient-safety/research/safe-surgery . Developed under Atul Gawande's leadership. See also Gawande, A. (2009), The Checklist Manifesto: How to Get Things Right, Metropolitan Books.

7. [R] Ohno, T. (1988), Toyota Production System: Beyond Large-Scale Production, Productivity Press; Deming, W. E. (1986), Out of the Crisis, MIT Press; Womack, J. P., Jones, D. T., and Roos, D. (1990), The Machine That Changed the World, Rawson Associates. The canonical academic-and-industry foundations of the Toyota Production System and lean manufacturing.

8. [R] Silver, D. et al. (2017), 'Mastering the game of Go without human knowledge,' Nature 550: 354–359. The AlphaGo Zero paper documenting the rapid advance over the prior human Go theory tradition.

9. [S] Health Services Safety Investigations Body (HSSIB) — official site: https://www.hssib.org.uk/ . HSSIB became a statutory body on 1 October 2023 under Part 4 of the Health and Care Act 2022. About page including the safe-space protections: https://www.hssib.org.uk/about-us/ . The statutory provisions for protected material, with criminal offences for unauthorised disclosure, are at sections 117–119 of the Health and Care Act 2022.

10. [R] Education Endowment Foundation (EEF), Teaching and Learning Toolkit: https://educationendowmentfoundation.org.uk/education-evidence/teaching-learning-toolkit . The EEF maintains the most rigorous current synthesis of education-intervention evidence for the English context. Structured literacy and explicit phonics teaching are among the interventions with the strongest evidence base.

11. [S] Grenfell Tower Inquiry — official site: https://www.grenfelltowerinquiry.org.uk/ . Phase 1 Report (October 2019) and Phase 2 Report (September 2024). The Phase 2 Report (chair Sir Martin Moore-Bick, with panel members Thouria Istephan and Ali Akbor) is a 1,700-page document identifying multiple distinct causal factors: the "systematic dishonesty" of Arconic, Kingspan and Celotex; Studio E's "cavalier attitude to fire safety regulations"; the Royal Borough of Kensington and Chelsea's building control failures; the Kensington and Chelsea Tenant Management Organisation's "pattern of concealment" and the recharacterisation of fire-safety-concerned residents as "militant troublemakers"; the "bonfire of red tape" political climate from 2010 onward; and the financial-considerations decision by RBKC's director of housing to slow installation of self-closing fire doors despite London Fire Brigade warnings. Contemporaneous summary in The Guardian, 4 September 2024: https://www.theguardian.com/uk-news/article/2024/sep/04/grenfell-report-blames-decades-of-government-failure-and-companies-systematic-dishonesty . The total inquiry cost was reported at over £200 million.

12. [S] The Lakanal House Fire (2009) — Coroner's recommendations following the inquest, summarised in the Independent Review of Building Regulations and Fire Safety (Dame Judith Hackitt, 2018): https://www.gov.uk/government/publications/independent-review-of-building-regulations-and-fire-safety-final-report .

13. [S/R] National Patient Safety Agency (NPSA) — Wikipedia overview with primary sources: https://en.wikipedia.org/wiki/National_Patient_Safety_Agency . Establishment 2001; abolished 1 October 2012 under section 281 of the Health and Social Care Act 2012: https://legislation.gov.uk/ukpga/2012/7/notes/division/5/10/4 . Research evaluation: Donaldson, L. J., Panesar, S. S., and Darzi, A. (2014), 'Patient-Safety-Related Hospital Deaths in England: Thematic Analysis of Incidents Reported to a National Database, 2010–2012,' PLoS Medicine 11(6): e1001667. https://pmc.ncbi.nlm.nih.gov/articles/PMC4068985/ .

14. [S] Care Act 2014, section 44 (Safeguarding Adults Reviews): https://www.legislation.gov.uk/ukpga/2014/23/section/44 . Statutory guidance: Department of Health and Social Care, Care and Support Statutory Guidance, regularly updated.

15. [S] Local Government Association, national SAR thematic analyses, available via the LGA safeguarding adults page: https://www.local.gov.uk/our-support/our-improvement-offer/care-and-health-improvement/safeguarding-adults . The Social Care Institute for Excellence (SCIE) also maintains an analysis of national SAR findings: https://www.scie.org.uk/safeguarding/adults/reviews/.

16. [J] Dodd, V. and Murray, J. (19 May 2026). 'Police to seek criminal charges against 77 companies and people over Grenfell fire.' The Guardian. https://www.theguardian.com/uk-news/2026/may/19/grenfell-fire-police-criminal-charges-companies-individuals . Scotland Yard announced that files would be sent to the Crown Prosecution Service in respect of fifty-seven individuals and twenty companies, with charging decisions expected by June 2027 and no trials likely before 2028. The Metropolitan Police investigation has cost £150 million to date. The lead investigator was Garry Moncrieff. Offences under consideration include corporate manslaughter, gross negligence manslaughter, fraud, health and safety offences, and misconduct in public office.

Compiled · May 2026 · For research and journalism use