Summary of findings
One doctor in a hundred accounts for roughly a third of all paid malpractice claims. Yet in most underwriting operations the neurosurgeon with three prior claims and the dermatologist with a spotless record join the same queue, wait for the same desk, and take the same slice of the same underwriter's day. That is the real inefficiency in underwriting: not that humans decide, but that scarce judgement is spent evenly across a book whose risk is anything but even. This article treats underwriting automation as what it actually is, an allocation problem, and works through the machinery that solves it: a triage engine that reads every application and renewal, scores it, and decides how much human judgement it deserves.
Judgement is the scarce input, and it is spent flat.
Surveys have asked underwriters how they spend their time since 2008, and the answer barely moves: about 40% went on non-core and administrative work in 2021, and still 35% in 2025. Meanwhile the deals get smaller and more numerous, with US surplus lines transactions growing at nearly twice the rate of premium in 2025. More applications per desk, the same hours in the day, and no mechanism that matches attention to risk.
Risk concentrates; a queue treats it as uniform.
In a study spanning a decade of national payment data, 1% of US physicians accounted for 32% of paid malpractice claims, and a doctor with three prior claims carried a 24% probability of another within two years. Annual claim rates span roughly sevenfold between psychiatry and neurosurgery. A first-come-first-served queue prices none of this into its allocation of attention.
The theory of when not to decide is 56 years old.
The reject option dates to 1970: a classifier should decline to decide when its confidence falls below a threshold set by the cost of an error against the cost of a referral. The modern literature extends it to deferring to the right expert among several, which is precisely what routing an application to the right underwriter is.
Life insurance already ran the experiment, and it exposes the gap.
82% of surveyed US life carriers operate accelerated underwriting, yet only 17% of eligible applications pass through with no human touch. The industry automates eligibility, then routes timidly. Where the miss rate of automation is actually measured, through random holdouts, it comes out at 10–15% and is managed rather than feared.
Modelled on a doctors' book, triage frees the desk and the arithmetic holds.
On a composite book of 20,000 medical indemnity policies, a triage design decides 58% of cases with no desk involvement, concentrates senior attention on the 12% that carry the judgement, and cuts cost per decision by 40%, from £25.75 to £15.34. The model, and every assumption behind it, is in the appendices.
Autonomy is conditional, and the conditions are knowable.
An automated bind is lawful, auditable, and safe under specific conditions: a delegated appetite encoded as rules, a reconstructable record for every decision, a permanent random holdout that measures what automation misses, and the safeguards UK law now attaches to solely automated decisions. Declines stay human, by design as much as by law.
Where underwriting capacity actually goes
Start with the desk itself. The Institutes and Accenture have surveyed underwriters since 2008, and in 2021 the survey found underwriters spending roughly 40% of their time on non-core and administrative tasks: rekeying data, chasing missing information, moving submissions between systems. Capgemini's separate 2024 survey put administrative work at 41–43% of underwriter time. By the 2025 edition of the Accenture survey, commercial lines stood at 35%, down from 38% three years earlier. Seventeen years of digitisation programmes, and a third of the profession's working life still goes on work that requires no judgement at all.
McKinsey reached the same range from its own client observations: 30 to 40% of underwriting time spent on administrative tasks. Whichever survey you prefer, the shape is identical. Underwriting capacity is expensive, finite, and heavily diluted before a single risk decision is made.
Sources: The Institutes/Accenture P&C Underwriting Survey 2021 (n=434, ~40% on average); Capgemini World Property and Casualty Insurance Report 2024 (n=201 underwriters, 41–43%, midpoint shown); Accenture/The Institutes Underwriting Rewritten 2025 (n=430, commercial P&C 35%). Full citations at Appendix D.
The workload, meanwhile, is growing in exactly the wrong shape. US surplus lines premium grew 7.8% in 2025 while transaction counts grew 14.1%: more risks, smaller tickets, each one still expecting a decision. Delegated books, where a coverholder underwrites high volumes of small risks under an agreed appetite, feel this hardest, and delegated arrangements now carry roughly 40–45% of premium at Lloyd's, more than $26bn a year; premium written through US managing general agents grew at about twice the pace of the wider market in 2025 (reference 9).
The people side is tightening too, and here it is worth clearing away a statistic that refuses to die. The claim that half the insurance workforce will retire within a few years, leaving 400,000 unfilled jobs, circulates with a different deadline in every retelling and traces back to no primary source at all. The verifiable facts are less theatrical and more useful: roughly a quarter of the US insurance workforce is aged 55 or over, and the US Bureau of Labor Statistics projects underwriter employment to decline 3% over 2024–34, a fall it attributes directly to automated underwriting software. Underwriting is one of very few professional occupations where the official projection has pointed down, for that reason, for over a decade of editions.
Two more facts complete the picture of the desk. First, the money: Lloyd's, the world's specialty marketplace, ran a 35.6% expense ratio in 2025, 26.1% acquisition and 9.5% administration, up on the prior year even in a near-record-profit market. More than a third of every premium pound goes on distribution and running the machine. Second, the mix: a mature book is mostly renewals. Travelers, one of the largest US commercial lines writers, retains 85% of its business and writes new premium equal to only about 12% of the book. The bulk of underwriting work, in other words, is re-deciding something the organisation decided last year, about a customer it already knows.
A desk diluted by administration, workload growing faster than premium, headcount flat to falling, expenses over a third of premium, and a book dominated by repeat decisions. That is the capacity side. The next question is what the demand side, the risk itself, actually looks like.
Risk concentrates; effort does not
If risk were spread evenly across applicants, a flat queue would be a defensible design. It is not spread evenly, and in medical professional indemnity, the line this article uses as its worked example, the concentration has been measured with unusual precision.
Studdert and colleagues analysed 66,426 paid claims against 54,099 US physicians in the National Practitioner Data Bank, the most complete record of malpractice payments linked to individual practitioners, and published the results in the New England Journal of Medicine in 2016. Approximately 1% of physicians accounted for 32% of all paid claims. Among doctors with any paid claim, 84% had only one; the 16% with two or more carried a third of the claims between them. And the past predicted the future with uncomfortable strength: against a doctor with one prior paid claim, a doctor with two had roughly double the recurrence risk, a doctor with three had triple, and that third claim carried a 24% probability of another within two years.
| Prior paid claims | Recurrence risk, against one prior claim | What it means at the desk |
|---|---|---|
| 1 | Baseline | 84% of claimant doctors stop here |
| 2 | 1.97× | Roughly double the risk of another paid claim |
| 3 | 3.11× | A 24% probability of a further paid claim within two years |
Source: Studdert et al., New England Journal of Medicine, 2016: 66,426 paid claims against 54,099 physicians, US National Practitioner Data Bank, 2005–14. Recurrence figures are hazard ratios against physicians with a single prior paid claim. Reference 11, Appendix D.
Specialty stratifies risk just as sharply. Jena and colleagues, in an earlier New England Journal study of 40,916 physicians, found 7.4% of all doctors faced a malpractice claim in any given year, but the rate ranged from 2.6% in psychiatry to 19.1% in neurosurgery, a spread of roughly sevenfold. By age 65, three quarters of doctors in low-risk specialties had faced at least one claim; in high-risk specialties, 99% had. And most claims go nowhere: 78% closed with no payment, which makes the line as much about defence cost and process as about indemnity itself.
Source: Jena et al., New England Journal of Medicine, 2011: 40,916 physicians, 233,738 physician-years, one large US liability insurer. 78% of all claims closed with no payment. Reference 12, Appendix D.
Both studies are American; no UK equivalent of the National Practitioner Data Bank exists, and NHS Resolution publishes claims data only in aggregate, which is itself a fact worth knowing about this market. But the structure travels. UK medical negligence operates at serious scale, with £3.1bn paid out in 2024/25 against a £60.3bn provision for future claims, and the same drivers, specialty and history, dominate any rating conversation a UK underwriter would recognise.
Put the two sections together and the mismatch is stark. Capacity is flat, diluted, and spent roughly evenly per case. Risk is concentrated, skewed, and largely predictable from a handful of fields: specialty, claims history, scope of practice. An allocation this poor would not survive an afternoon in any other operations function. What fixes it is a machine whose defining skill is knowing which cases it must not decide.
The machine that knows when not to decide
A triage engine is an agentic pipeline that sits between the inbox and the desk. It does four things to every application and renewal, and then makes the one decision that defines it.
First it reads: a language model extracts the proposal form, the disclosures, and the attachments into a structured record, whatever the format the broker or applicant sent. Second it verifies and enriches: registry lookups confirm what can be confirmed at source rather than trusted from the form. For UK doctors this is unusually rich ground. The General Medical Council's register of 313,829 licensed doctors (the 2023 count), including registration status, specialty, and public fitness-to-practise history, is downloadable daily; the licence costs £815 a year. Underwriting-grade data, machine-readable, sitting in the open. Third it scores: risk models turn the verified record into an assessment against the book's appetite. Fourth it routes, and this is the decision that matters: does this case need no desk at all, a junior's structured review, or a senior underwriter's full attention?
Data Hatch. The routing decision, not the extraction, is the load-bearing step: it allocates the book's scarcest resource. The random holdout and the human decline path are governance features, not afterthoughts; both return in Sections 6 and 7.
The temptation is to treat that routing step as a product feature. It is better understood as the application of a 56-year-old body of theory, because the discipline of deciding when not to decide has a name, and proofs.
In 1970, C. K. Chow showed that the optimal recognition system rejects a case, handing it to a human, whenever its confidence falls below a threshold, and that the threshold is not a matter of taste: it is set by the ratio between the cost of a wrong decision and the cost of a referral. Modern selective prediction turns the same idea into an operating dial. You do not ask how accurate the model is; you choose the error rate you are willing to tolerate, and the system earns as much coverage, as many automated decisions, as it can while guaranteeing that rate. Later work sharpened two points that matter at the desk. A learned routing function beats a naive confidence cut-off, so the router deserves its own model rather than a threshold on the scorer's output. And the newest results, on deferring to multiple experts, formalise exactly our problem. The system learns two things at once: when to hand a case to a human, and which human, deciding by whether the junior desk or the senior desk has the better expected outcome on this specific case, net of cost.
Data Hatch illustration of the selective-prediction trade-off; the curve's shape is a mechanism, not a measurement of any specific model. The dashed line is the break-even error rate derived in Section 6 and Appendix C: automate up to the point where the marginal error cost meets the referral cost saved, and no further.
One warning from the same literature deserves a card on the wall of any underwriting platform team: modern models are overconfident. Deep networks systematically report more certainty than they possess, and the post-training that makes large language models pleasant to talk to makes their calibration worse, not better. A model's stated confidence is an opinion about itself. So the routing threshold is never set on self-reported confidence; it is set on measured, task-level performance, calibrated on held-out cases and monitored continuously. The dial must be connected to reality, not to the model's self-esteem.
It is fair to ask, at this point, whether the machine is simply skimming the easy work and leaving the hard part exactly where it was. It is, and that is the design. The easy work is most of the volume, and in Section 2 it was consuming the majority of an expensive desk. Skimming it is what funds the judgement: the senior underwriter stops being a queue-clearing mechanism and starts spending 40 minutes on the cases that need 40 minutes.
Two reference points calibrate how good the routing actually has to be. The first is humbling for the status quo. A study of 5.3 million US emergency department visits found the standard human process mistriaged 32.2% of patients, mostly by over-caution: 28.9% overtriage against 3.3% undertriage. Triage quality is measurable, the human baseline is not zero-error, and well-designed systems deliberately buy safety with over-referral. The second is cautionary in the other direction: in anti-money-laundering screening, industry figures put false positives at 85–95% of alerts. A router with uncalibrated thresholds does not allocate human judgement; it buries it. Both failure modes are governed by the same object, the threshold, and Section 6 prices it.
The experiment life insurance already ran
None of this needs to be taken on faith, because an adjacent line of insurance has been running the experiment at scale for a decade. US life insurers built accelerated underwriting: algorithmic pathways that let clean applications skip medical evidence and, in principle, the underwriter. The results, published by the reinsurers and the actuarial profession, are the best public evidence on what happens when an industry automates decisions on individual human risk.
Adoption is nearly universal; automation of the actual decision is not. Gen Re's 2024 survey of 38 carriers found 82% operating an accelerated workflow, yet only 17% of eligible applications passed through with no human touch at all; 83% still saw an underwriter somewhere. Munich Re's earlier market survey put the industry's average straight-through rate at 21% of applications, while established programmes, per Swiss Re, run around 75% through their automated engines, and up to 90% in some markets. The gap between the average and the leaders is not model quality. It is routing confidence: the willingness, or not, to let the system decide the cases it was built to decide.
| Finding | Figure | Source and year |
|---|---|---|
| Carriers with an accelerated workflow | 82% | Gen Re survey of 38 US life carriers, 2024 |
| Eligible applications fully untouched | 17% | Gen Re, 2024: the routing gap |
| Industry-average straight-through rate | 21% | Munich Re market survey, 30+ carriers, 2022 |
| Established programmes' share through automated engines | 75–90% | Swiss Re, provider-published figures |
| Mortality slippage on the automated path | 10–15% | Society of Actuaries, 2024 |
| Automated risk class confirmed on audit | 81% | Society of Actuaries, 2024, random holdouts |
| Tobacco use undisclosed on automated path | 40% | Society of Actuaries, 2024 |
| Tobacco non-disclosure caught by holdouts against post-issue audits | 1.5× | Society of Actuaries, 2024 |
Sources at Appendix D, references 28–31. Personal lines P&C goes further still: McKinsey reports that in its experience up to 95% of policies may pass straight through at leading carriers. No equivalent public benchmark exists for small commercial lines, and Datos Insights' 2023 study found many insurers do not systematically track their own straight-through rates at all.
The most valuable export from life insurance is not a rate but an instrument. The Society of Actuaries' 2024 slippage study quantified exactly what the automated path misses: risk classes come out 10–15% lighter than full underwriting would have produced, 81% of audited cases land in the same class the machine chose, and 40% of tobacco users go undisclosed when nobody checks. The instrument that produced those numbers is the random holdout: a permanent, random slice of automated approvals routed to a human for full re-underwriting, not because the machine flagged them but precisely because it did not. Holdouts caught 1.5 times more tobacco non-disclosure than after-the-fact audits. An automation programme that runs one knows its own miss rate as a number; a programme without one is guessing.
Closer to home, the market our worked example lives in has quietly crossed its own threshold. Lloyd's has licensed algorithmic underwriting since 2020, when Ki, the first algorithmically driven syndicate, began quoting follow lines in minutes. The Lloyd's Market Association's 2024 study of what it calls enhanced underwriting, built on interviews covering 77% of the market's premium, found firms expecting augmented underwriting to grow around 60% a year, and not one respondent expecting the trend to decline. Meanwhile Lloyd's itself sunset Blueprint Two, its central digitisation programme, in early 2026 and refocused on standards, which settles where modernisation will actually happen: inside individual carriers, coverholders, and managing general agents, one book at a time. Surveys this year suggest only about a fifth of commercial insurers yet use AI triage on submissions. The tooling exists, the market structure now points the right way, and the discipline, measured routing under an explicit appetite, is the part still missing. So we model it.
A modelled book of 20,000 doctors
Consider a composite coverholder underwriting medical professional indemnity for individual practitioners in the UK: 5,000 consultants and surgeons in private practice, 8,000 GPs insuring their private and non-NHS work, and 7,000 aesthetics and allied practitioners. That is 20,000 policies and roughly £29.4m of premium, a blended £1,470 per policy. It is a recognisable book. Since the state indemnity schemes absorbed NHS GP negligence in 2019, the private market for doctors has been exactly this: private practice, aesthetics, and the medico-legal residue, sold as regulated insurance in a market the government's own unfinished reform of discretionary indemnity keeps nudging towards insured products.
The book is renewal-heavy, as Section 2 predicts: 17,000 renewals and 3,000 new applications a year, 20,000 decisions in all. Three operating designs compete to make them. Design A is the traditional desk: every case reviewed, 45 minutes for new business, 20 for a renewal, which comes to 7,917 hours a year, five underwriters. Design B adds extraction: the machine reads and pre-fills, humans still decide everything. Design C is the triage engine of Section 4, with four bands: renewals with no material change auto-renew; standard new business inside the binder's appetite auto-binds; light-change renewals and mid-band new business go to a junior desk with a structured checklist; material changes and the genuinely complex go to the senior desk with the evidence already assembled. An 8% escalation rate moves flagged cases up a band, and a 5% random holdout sends untouched cases to a desk anyway, to measure what the machine misses.
Data Hatch model. The four routes sum to exactly 20,000: 70% of renewals qualify as no material change and 50% of new business as in-appetite, an 8% escalation rate moves flagged cases up a band, and a 5% holdout samples the automated stream. Composition and sensitivities at Appendix C; the untouched share ranges from 44% to 66% as the no-material-change rate varies from 50% to 80%.
The economics fall out cleanly, and the full derivation of every line is in Appendix C. Design A needs five underwriters and costs £515,000 a year to run, £25.75 per decision. Design B needs three and costs £17.18 per decision: worthwhile, though what it proves is that extraction alone leaves every decision on a desk. Design C needs two underwriters, one senior and one mid-level, and total running costs of £306,815 a year: £15.34 per decision, 40% below the traditional desk, with the platform, evaluation harness, registry licences, and inference all inside that figure.
Inference deserves a word of its own. Reading 50,000 documents a year through a small-tier model costs about £150. The two underwriters cost £220,000. The machine reading is a rounding error; the human judgement is the entire budget, which is precisely why pointing it at the right 12% of cases is worth more than any saving on the reading. Readers of our July analysis of AI cost structures will recognise the shape: the model is never the expensive part.
| At steady state | Design A | Design B | Design C |
|---|---|---|---|
| Underwriting staff | 5.0 FTE | 3.0 FTE | 2.0 FTE |
| Payroll, fully loaded | £505,000 | £302,500 | £220,000 |
| Platform, evaluation, and licences | £10,000 | £40,000 | £85,815 |
| Model inference | None | £1,000 | £1,000 |
| Total annual cost | £515,000 | £343,500 | £306,815 |
| Cost per decision | £25.75 | £17.18 | £15.34 |
| Median time to decision | Days | Days, faster reading | 58% instant; the rest routed same day |
Data Hatch model at an illustrative scale; derivation of every line, including the phasing arithmetic and staffing steps, at Appendix C. All three designs start identically; renewal triage goes live around month 4 and new-business triage around month 10. New-business automation barely moves the cost line, because two people is the floor for judgement in any design; what it buys is speed and headroom, not payroll.
Notice what the model says about growth. Design C's two desks are loaded to about 70% with today's volumes; the book can grow roughly 43% before anyone is hired. Under Design A the same growth means recruiting two more underwriters into a market that is not producing them. Speed compounds the advantage: 58% of decisions return instantly rather than joining a multi-day queue, and for the doctor renewing cover, or the aesthetics practitioner who cannot work without it, an instant certificate is the product improvement, not a by-product.
The objection that matters: is this how books rot?
Every underwriter reading this has the same objection ready: auto-binding is adverse selection with better plumbing. The clean-looking application that a human would have paused on gets waved through, and the book decays quietly until the loss ratio announces it. The objection deserves arithmetic, not reassurance, and the arithmetic is the threshold's.
Moving a renewal from the junior desk to the automated band saves about £10 of review cost. The expected loss on that policy is £882 a year, a £1,470 premium at a 60% loss ratio. So the automation pays if and only if the excess error introduced by not reviewing, the slippage, stays below £10.10 divided by £882: 1.15%. That single number is the whole governance question. Below it, review was costing more than it caught; above it, automation is eating the book. The senior desk's complex cases price out differently, a 5.5% break-even, which is exactly why they stay human. And the 5% random holdout is what measures the actual slippage continuously, the same instrument that let life insurers state their miss rate as a number instead of an anxiety. The life industry tolerates 10–15% slippage because skipping medical evidence transforms its acquisition economics. A doctors' indemnity book enjoys no such subsidy, and its threshold must be run far tighter, which the arithmetic above is for.
Data Hatch model: net saving equals the £10.10 review cost avoided minus slippage multiplied by the £882 expected annual loss per policy. New business reviewed by the junior desk breaks even at 1.72%, and senior-desk complex cases at 5.5%, which is the arithmetic case for keeping them human. Derivation and sensitivity to loss ratio and premium mix at Appendix C.
What do the two remaining underwriters do all day? Everything the queue used to crowd out. They own the thresholds and review the holdout results. They handle the 2,300 cases a year that genuinely need judgement, with the file assembled before it arrives. They talk to brokers about the referrals worth having, and they steer the portfolio: appetite, pricing adequacy, and the emerging-risk questions, a two-desk version of what McKinsey calls the underwriter as portfolio manager. The consultancy's own warning from 2019 bears repeating here: firms that forced model output to override underwriting judgement watched performance deteriorate and skills atrophy. The triage design is the opposite arrangement. The machine does the reading and the sorting; the judgement stays where the binder always said it must.
The conditions of autonomy
A coverholder cannot simply decide to let software bind risks. Autonomy has conditions: contractual, evidential, and legal. The striking thing about 2026 is how concrete they have become.
The binder is already a machine-readable appetite
Delegated authority runs on binding authority agreements, and the market-standard wordings define the delegation with the precision of a schema: classes of business, maximum limits, territorial scope, premium income limits, rating guides, excluded risks, and the triggers that require referral back to the managing agent. That schedule is a risk appetite expressed as rules, which is to say it is halfway to being code. Encoding it is the first condition of autonomy: the automated band exists strictly inside the binder's lines, and anything that approaches an edge routes to a person. The Lloyd's Market Association's own analysis of enhanced underwriting warns of the accountability vacuum when firms cannot say with certainty that a specified appetite will be met in all cases, and warns that every tier of delegation strains audit trails further. The answer to both warnings is the same artefact: a decision record.
| Recorded per decision | What it contains | Who relies on it |
|---|---|---|
| Inputs | The extracted application, every registry response, and the verification results, as at decision time | Auditors; the applicant, on request |
| Versions | Model, prompt, threshold, and appetite-rule versions that produced the decision | The platform team; the managing agent's audit |
| Appetite line | The specific binder provisions the case was decided inside | The managing agent; Lloyd's coverholder oversight |
| Route and rationale | The band, the score, the reasons, and any escalation on the way | Underwriters reviewing referrals; conduct reviews |
| Human involvement | Reviewer identity and what they changed, where a desk touched the case | The accountability chain under SM&CR |
| Applicant rights | The information given, and the route to representations, human intervention, and contest | The applicant; the ICO, under Articles 22A–22D |
Data Hatch. The record answers, for any decision and at any later date: what was decided, on what data, under which authority, by which version of which model, and with what recourse. Monthly bordereaux, the market's minimum reporting standard, contain a fraction of this; the engine generates it as a by-product of deciding.
The law now permits automated decisions, and prescribes their safeguards
UK law on automated decision-making was rewritten this year, and most commentary has not caught up. The Data (Use and Access) Act 2025 replaced the old Article 22 of UK GDPR with new Articles 22A to 22D, in force since February 2026, and flipped the default: solely automated decisions with significant effects, and refusing or pricing someone's insurance is such a decision, are now permitted, provided the controller supplies specific safeguards. The applicant must be told about the decision, be able to make representations, be able to obtain human intervention, and be able to contest the outcome. The regulator's draft guidance adds teeth to the human-involvement test: spot-checking is not sufficient, and the reviewing human must have the competence and the authority to change the decision. A rubber stamp does not launder an automated decision into a human one.
One provision bites specifically for a medical book. Where a solely automated decision rests on special category data, and health data is special category, the permissive regime narrows sharply: broadly, explicit consent or legal authorisation is required. A doctors' indemnity engine is designed around that constraint from day one, by building the automated band on registration, specialty, claims history, and practice profile rather than health data, and by routing anything that engages health information to a desk. The EU AI Act, for completeness, does not reach this book at all: its high-risk insurance category covers life and health insurance pricing, not professional liability, and a UK coverholder deploying in the UK sits outside its territorial scope besides. Citing it as a blocker is a category error, though its neighbouring guidance, like EIOPA's 2025 opinion on AI governance in insurance, is a sensible benchmark for supervisory expectations: proportionate governance, documented data lifecycles, explainability calibrated to the audience, and accountability that stays with the insurer whoever built the system.
The two design consequences are worth stating plainly, because they are where law and prudence agree. Declines and non-renewals stay human. The asymmetry is economic as well as legal: a wrong automated bind costs slippage on one policy, while a wrong automated decline costs a customer, their lifetime value, and potentially a discrimination complaint, and the safeguards regime makes contested declines expensive. Let the machine say yes inside the appetite; let only people say no. And the reviewing human must be able to overrule the machine, which is an organisational design fact, a trained desk with authority, not a checkbox.
The regulator will look at outcomes, and has form
The FCA has declined to write an AI rulebook, deliberately: existing frameworks, the Consumer Duty and senior manager accountability chief among them, already apply to an automated book. Anyone who doubts the regulator's willingness to intervene in algorithmic outcomes should recall that it re-engineered the entire home and motor pricing market in 2022 after finding systematic price-walking, and paused roughly 80% of the GAP insurance market in 2024 over fair-value data. Portfolio-level intervention is not hypothetical. An automated book should expect to evidence, continuously, that its outcomes are fair as well as fast, which is one more use for the holdout and the decision record.
Drift is a when, not an if
A final condition comes from the machine-learning literature rather than the law. An underwriting book drifts by construction: the cycle turns, the mix shifts, brokers learn. Worse, automation creates its own feedback loop, because the book the engine binds today is the data the next model trains on, and applicants adapt to what the machine accepts; the 40% tobacco non-disclosure figure from the life studies is adversarial drift in the wild. The controls are the unglamorous ones the standards prescribe: monitored calibration, drift detection on inputs and outcomes, defined triggers for retraining or narrowing the automated band, and the holdout as the permanent measurement of ground truth. Thresholds are not settings; they are living commitments with an owner.
Putting this into practice
Everything above is a model, and a real book will differ in the places models cannot see. The route from here to a running triage engine is short, but it has an order, and the order is most of the risk management.
Start with renewals, because that is where the volume and the safety are.
Renewals are 85% of the model book's decisions, arrive with a year of history attached, and admit a conservative definition of no material change: same specialty, same scope of practice, clean claims year, clear register check, turnover inside the band. Detecting that reliably is a far easier problem than underwriting a stranger, and it converts the majority of the queue on its own. New business follows only once renewal routing has run clean for months.
Run the holdout from the first day, and size the automated band by its results.
The 5% random holdout is not an audit formality; it is the instrument that turns slippage from an anxiety into a number, and the number is what earns each widening of the band. The break-even from Section 6, around 1.15% on renewals, is the ceiling the measured slippage must stay under. Widen while it holds, narrow the moment it does not, and let the thresholds move only at scheduled reviews, with an owner, never silently.
Encode the binder before the models, and keep declines human.
The appetite rules from the binding authority schedule are the walls of the automated band; the scoring model only chooses positions inside them. Building the walls first keeps every automated decision inside delegated authority by construction, gives the managing agent an audit trail it can test, and leaves the decisions the law and the economics both want humans making, declines, non-renewals, and anything touching health data, on a desk with authority to differ from the machine.
Instrument the desk you have before switching on the engine.
The model in Section 6 runs on assumptions a fortnight of measurement replaces: minutes per case by type, the true no-material-change rate, referral reasons, decision latency. Measured first, they become the baseline every later claim of improvement is judged against, and the routing distribution stops being a guess. The escalation rate and the override rate then become the two numbers management actually watches, just as the escalation rate governed routing in our July analysis.
The failure modes are known in advance
- Thresholds set on model self-confidence. Language models are poorly calibrated about their own certainty, and post-training worsens it. Thresholds belong on measured task performance, or the dial is connected to nothing.
- No holdout. A programme without a random holdout does not know its miss rate; it has an opinion about it. This is the single cheapest control to run and the first one dropped.
- A router with no learned rejector. A bare confidence cut-off on the scoring model is provably weaker than a routing function trained for the job, and it is the difference between triage and the anti-money-laundering pattern of drowning humans in false referrals.
- Automating the tail first. The complex cases are where models are weakest and errors dearest; their break-even slippage is 5.5% against the renewals' 1.15% precisely because senior minutes are expensive. Automate the middle of the distribution, never its edges.
- Thresholds without an owner. A threshold that nobody owns gets moved in a sprint, for throughput reasons, and discovered in a loss review. Governed objects have names attached.
- Treating the binder as documentation. If the appetite is not encoded, the engine's honesty about operating inside it is unverifiable, and the accountability vacuum the market already warns about opens underneath the programme.
Triage exists because the 1% exists. A book where one doctor in a hundred carries a third of the claims does not need more underwriters; it needs its underwriters pointed at that hundredth. The machine's job is the pointing. The judgement, as the binder always insisted, stays with the people licensed to hold it.
That is the routing problem, and it is solvable with instruments that already exist: a register that costs £815 a year, theory that has been in the literature since 1970, an audit instrument borrowed from life insurance, and a binder that was always, quietly, a specification. What it needs from the organisation is the discipline to treat autonomy as something the system earns in measured increments, case by case, threshold by threshold. The desks stay. The queue goes.
Frequently asked questions
What is a triage engine in insurance underwriting?
The pipeline that decides how much human judgement each application deserves. It reads the submission into a structured record, verifies what can be verified at source (for UK doctors, the GMC register), scores the result against the book's appetite, and routes the case. Clean renewals and in-appetite new business are decided automatically inside the binder; marginal cases go to a junior desk with a structured checklist; material changes and genuinely complex risks go to a senior underwriter with the evidence already assembled.
Can underwriting decisions be automated safely?
Safety is a property of the design, not the model. In the modelled doctors' book, 58% of decisions complete with no desk involvement. But every automated decision stays inside the binder's encoded appetite, an 8% escalation rate pulls flagged cases back to a desk, a 5% random holdout measures what automation misses, and declines are never automated. The machine says yes inside the rules; only people say no.
How much does automating underwriting save?
In the model, cost per decision falls from £25.75 on a traditional five-person desk to £15.34 with triage, a 40% reduction; the platform, evaluation harness, registry licences, and inference are all inside that figure. The larger effect is capacity: the same book runs on two underwriters instead of five, the desks are loaded to about 70%, and volumes can grow roughly 43% before anyone is hired.
How do you stop automation from quietly degrading the book?
With a threshold and an instrument. Automating a renewal saves about £10 of review cost against an £882 expected annual loss, so automation pays only while the excess error it introduces stays below 1.15%. A permanent random holdout, 5% of automated cases re-underwritten by a human, measures that error continuously. Widen the automated band while measured slippage holds under the ceiling; narrow it the moment it does not.
Is fully automated underwriting legal in the UK?
Yes, since February 2026, under conditions. The Data (Use and Access) Act 2025 replaced Article 22 of UK GDPR with Articles 22A to 22D. Solely automated decisions with significant effects are permitted, provided the applicant is told about the decision, can make representations, can obtain human intervention, and can contest the outcome. Health data narrows the permission sharply, which is why the modelled engine routes anything touching it to a desk, and declines stay with people throughout.
Appendices
AThe formal spine of triageOpen +Close −
The routing decision in Section 4 rests on four results from the machine-learning literature. This appendix states them plainly; the papers are in Appendix D, references 19 to 27.
The reject option (Chow, 1970). For a classifier with a cost ce per error and cr per rejection (referral to a human), the optimal rule rejects whenever the maximum posterior probability falls below a threshold determined by the ratio cr/ce. The error-reject trade-off is monotone with diminishing returns: each additional referral buys a smaller reduction in error. In Section 6's terms, cr is £10.10 of junior review and ce is the expected cost of a slipped decision, which is how the 1.15% break-even is a Chow threshold in pounds.
Selective classification (Geifman and El-Yaniv, 2017). Fix a target risk, the error rate you will tolerate on the cases the system decides, and maximise coverage, the share it decides, subject to guaranteeing that risk with high probability. This inverts the usual question: accuracy is chosen, coverage is earned. Exhibit 5 is this trade-off drawn as a curve.
Learned rejection and deferral (Cortes et al., 2016; Mozannar and Sontag, 2020; Verma et al., 2023; Mao et al., 2023). A confidence cut-off on the scoring model is suboptimal unless that model is already perfect; the rejector deserves to be learned jointly or as a second stage on top of a fixed scorer. Learning to defer generalises this by modelling the downstream human: defer when the expert's expected performance on this case, net of cost, beats the model's. The multiple-expert extensions assign each deferred case to the best-placed expert, which is precisely routing between a junior and a senior desk. The two-stage results matter practically: an existing scoring model can keep its job while the routing layer is trained around it.
Calibration (Guo et al., 2017; OpenAI, 2023). Modern neural networks are systematically overconfident, and reinforcement learning from human feedback degrades calibration further, a result reported in the GPT-4 technical report itself. Post-hoc calibration, temperature scaling at minimum, plus continuous monitoring of expected calibration error on live decisions, is a precondition for any threshold meaning what it says.
Limitations. These results are stated for classification with known costs; underwriting adds adversarial behaviour, delayed loss signals, and portfolio effects that the drift controls in Section 7 exist to manage. The theory sets the shape of the machinery, not a licence to skip the measurement.
BBook composition and cost assumptionsOpen +Close −
The book is a composite, sized to be recognisable rather than to describe any client. Substituting measured volumes reproduces every figure in Section 6.
| Segment | Policies | Average premium | Premium |
|---|---|---|---|
| Consultants and surgeons, private practice | 5,000 | £3,500 | £17.5m |
| GPs, private and non-NHS work | 8,000 | £700 | £5.6m |
| Aesthetics and allied practitioners | 7,000 | £900 | £6.3m |
| Total | 20,000 | £1,470 blended | £29.4m |
Premium levels are stated model assumptions. No reliable public rate card exists for UK medical indemnity, an opacity that is itself a feature of this market; the figures are set to be conservative against available market commentary (Appendix D, references 16 and 18), and the arithmetic is linear in them. Decisions per year: 85% of the book renews (17,000) and 15% is written new (3,000), consistent with large-carrier renewal shares (Appendix D, reference 10). Mid-term adjustments are excluded for simplicity; including them raises every design's volume and improves the triage case, since adjustments skew heavily routine.
Staff and systems
| Input | Value | Basis |
|---|---|---|
| Senior underwriter, fully loaded | £120,000 | London market professional indemnity mandates advertised at £80,000–£110,000 base plus bonus in 2025; loaded at roughly 1.3× base for employer costs and overheads (Appendix D, reference 43) |
| Mid-level underwriter, fully loaded | £100,000 | £60,000–£80,000 base range, same basis |
| Assistant underwriter, fully loaded | £65,000 | Junior band, same basis; the weakest public data of the three, so the model leans on it least |
| Productive hours per year | 1,650 | Standard planning figure net of leave, training, and administration |
| Review times: new business, renewal | 45, 20 minutes | Design A assumption; falls to 30 and 10 with extraction (Design B). Junior structured review 10–15 minutes; senior review 25 minutes for a material-change renewal and 40 for complex new business; holdout re-underwrite 20. All are assumptions a fortnight of desk measurement replaces (Section 8). |
| Triage platform share | £60,000 | Annualised share of a shared data-and-AI foundation: hosting, pipelines, and a part-share of an engineer. Consistent with the cost structure in our July analysis. |
| Evaluation and observability | £25,000 | Evaluation harness, monitoring, calibration and drift dashboards |
| Registry licences | £815 | GMC register download service, annual licence, published price |
| Model inference | ≈£150, held at £1,000 | 50,000 documents at 1,500 input and 500 output tokens on a small-tier model, $0.004 per document, converted at $1.33 to £1; held at £1,000 in the tables as deliberate headroom |
Known limitations. Staffing is modelled in whole and half heads with a two-desk floor in Design C, since judgement, holiday cover, and referral capacity do not fractionalise. Underwriting assistants' broader workload, credit control, bordereaux, and broker administration, is excluded from all three designs equally. The loss ratio used in the threshold arithmetic (60%) is an assumption; Appendix C varies it. Design A's five desks assume no overtime and no backlog, which flatters it.
CDerivation of the modelled resultsOpen +Close −
Routing volumes (Exhibit 7)
Renewals: 17,000, of which 70% qualify as no material change (11,900 automated), 20% show light change (3,400 to the junior desk), and 10% material change (1,700 to the senior desk). New business: 3,000, of which 50% is in-appetite (1,500 automated), 30% mid-band (900 junior), 20% complex (600 senior). Escalations: 8% of each automated band re-routes on validation flags, 952 renewals and 120 new, all to the junior desk. Holdout: 5% of automated cases (670) are re-underwritten at random. Untouched cases: 11,900 + 1,500 − 952 − 120 − 670 = 11,658, which is 58.3% of 20,000. Junior desk: 3,400 + 900 + 1,072 = 5,372 (26.9%). Senior desk: 2,300 (11.5%). The four routes sum to exactly 20,000.
Hours, staffing, and cost per decision (Exhibit 8)
Design A: 3,000 × 45 + 17,000 × 20 = 475,000 minutes = 7,917 hours = 4.8 FTE raw, staffed at five (two senior, two mid, one assistant): payroll £505,000, systems £10,000, total £515,000, £25.75 per decision. Design B: 3,000 × 30 + 17,000 × 10 = 260,000 minutes = 2.63 FTE raw, staffed at three (one senior, one and a half mid-level, half an assistant): £302,500 payroll, £41,000 systems, £343,500, £17.18 per decision. Design C: junior minutes (3,400 × 10 + 900 × 15 + 952 × 10 + 120 × 15) plus senior minutes (1,700 × 25 + 600 × 40) plus holdout (670 × 20) = 138,720 minutes = 2,312 hours = 1.40 FTE raw, staffed at two (one senior, one mid): payroll £220,000, systems £86,815 including £1,000 inference headroom, total £306,815, £15.34 per decision. Saving against Design A: 40.4%. Loading: 2,312 hours against 3,300 available = 70%, so the book absorbs roughly 43% growth before hiring.
The 24-month path assumes extraction live from month 4, renewal triage ramping over months 4–9, and new-business triage over months 10–14; transition-period staffing is approximated in half-head steps, so the plotted line smooths lumpy real-world hiring. All three designs start at Design A's £25.75. Design C passes £17.40 at month 6 and reaches £15.34 at month 9; new-business automation after month 10 leaves cost flat, because two desks are the judgement floor, and shows up instead as speed and headroom.
Threshold economics (Exhibit 9)
Expected annual loss per policy: £1,470 blended premium × 60% loss ratio = £882. Review cost avoided by automating a renewal: 10 junior-desk minutes at £100,000 over 1,650 hours = £10.10. Net saving per automated renewal at slippage s: £10.10 − s × £882, which crosses zero at s = 1.15%. New business at the junior desk: 15 minutes, £15.15, break-even just over half a point higher at 1.72%. Senior-desk complex cases: 40 minutes at the senior rate, £48.48, break-even 5.5%. Sensitivity: at a 50% loss ratio the renewal break-even is 1.37%; at 70% it is 0.98%. On the consultant segment alone (£3,500 premium, expected loss £2,100) the renewal break-even tightens to 0.48%, and on the GP segment (£700, £420) it relaxes to 2.40%: the threshold should differ by segment, which is one more argument for a learned router over a single cut-off.
Untouched-share sensitivity
Holding all else constant and varying the no-material-change rate from 50% to 80% moves the automated renewal band from 8,500 to 13,600 and the untouched share of all decisions from 44% to 66%. Varying in-appetite new business from 30% to 70% moves the untouched share by roughly ±3 points; renewals dominate, which is why Section 8 starts there.
DReferencesOpen +Close −
Listed roughly in the order the article draws on them, with background sources and late additions at the end. Market and pricing references are point-in-time, established August 2026.
- The Institutes and Accenture, P&C Underwriting Survey, October 2021: n = 434 US underwriters; roughly 40% of time on non-core and administrative tasks for the average underwriter, 45% in personal lines.
- Accenture and The Institutes, Underwriting Rewritten, 2025: n = 430 senior underwriting executives across 11 countries; commercial P&C non-core time 35%, down from 38% three years prior; AI-based capability usage under 15%, expected near 70% within three years.
- Capgemini, World Property and Casualty Insurance Report 2024: n = 201 underwriters; 41–43% of time on administrative activities.
- McKinsey & Company, From art to science: the future of underwriting in commercial P&C insurance, February 2019: 30 to 40% of underwriting time on administrative tasks; loss ratio varies up to 28 percentage points between top and bottom quintile insurers while expense ratios vary 2 to 4 points; the black-box caution cited in Section 6.
- WSIA stamping office data via The Insurer, January 2026: 2025 surplus lines premium $90.3bn, up 7.8%; transaction counts up 14.1%.
- Lloyd's of London, delegated authority market resources, accessed August 2026: delegated arrangements approximately 40–45% of premium income, over $26bn; the range reflects differing vintages in Lloyd's own materials.
- US Bureau of Labor Statistics, Occupational Outlook Handbook, 2024–34 projections: insurance underwriter employment projected to decline 3%, attributed to automated underwriting software; roughly 8,200 openings a year, all replacement. The widely circulated retirement statistic discussed in Section 2 has no traceable primary source; the aged-55-plus share (~25%) derives from industry analyses of BLS workforce data.
- Lloyd's of London, Full Year Results 2025, March 2026: expense ratio 35.6% (acquisition 26.1%, administration 9.5%); combined ratio 87.6%.
- Conning, US MGA study, July 2026: 2025 MGA direct premium $102.6bn, up 12%, roughly twice the wider market's growth.
- Travelers, Full Year 2025 results, January 2026: Business Insurance retention 85%; new business approximately 12% of segment net written premium (derived from reported figures).
- Studdert, D. M., Bismark, M. M., Mello, M. M., Singh, H., and Spittal, M. J., Prevalence and characteristics of physicians prone to malpractice claims, New England Journal of Medicine 374:354–362, 2016.
- Jena, A. B., Seabury, S., Lakdawalla, D., and Chandra, A., Malpractice risk according to physician specialty, New England Journal of Medicine 365:629–636, 2011.
- NHS Resolution, Annual Report and Accounts 2024/25, July 2025: £3.1bn paid across clinical schemes; £60.3bn provision; 14,428 new clinical claims and reported incidents.
- NHS Resolution, Clinical Negligence Scheme for General Practice, from 1 April 2019; NHS Wales General Medical Practice Indemnity, same date.
- Department of Health and Social Care, Appropriate clinical negligence cover consultation, 2018–19; summary of responses December 2022; no implementing legislation as of August 2026.
- Hansard, GP indemnity costs, HC Deb 15 March 2017: average in-hours GP indemnity £5,200 (2010) to £7,900 (2016).
- General Medical Council: The state of medical education and practice in the UK, workforce report 2024 (313,829 licensed doctors in 2023); register download service, published licence price £815 plus VAT, updated daily.
- American Medical Association, Policy Research Perspectives, February 2025, analysing Medical Liability Monitor rate survey data: 2024 manual premiums for the same $1m/$3m cover range roughly thirtyfold by specialty and geography.
- Chow, C. K., On optimum recognition error and reject tradeoff, IEEE Transactions on Information Theory 16(1):41–46, 1970.
- Cortes, C., DeSalvo, G., and Mohri, M., Learning with rejection, Algorithmic Learning Theory, 2016.
- Geifman, Y., and El-Yaniv, R., Selective classification for deep neural networks, NeurIPS 2017.
- Madras, D., Pitassi, T., and Zemel, R., Predict responsibly: improving fairness and accuracy by learning to defer, NeurIPS 2018.
- Mozannar, H., and Sontag, D., Consistent estimators for learning to defer to an expert, ICML 2020.
- Verma, R., Barrejón, D., and Nalisnick, E., Learning to defer to multiple experts, AISTATS 2023.
- Mao, A., Mohri, C., Mohri, M., and Zhong, Y., Two-stage learning to defer with multiple experts, NeurIPS 2023.
- Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q., On calibration of modern neural networks, ICML 2017.
- OpenAI, GPT-4 technical report, 2023, figure 8: post-training degrades calibration relative to the base model.
- Gen Re, US individual life accelerated underwriting survey, November 2024: 38 carriers; 82% with accelerated workflows; 17% of eligible applications processed without human touch.
- Munich Re Life US, accelerated underwriting market survey, 2022: industry-average straight-through processing 21% of applications.
- Society of Actuaries, accelerated underwriting mortality slippage study and monitoring practices, August 2024: 10–15% slippage; 81% class agreement on random holdout; 40% of tobacco users undisclosed; random holdouts catch 1.5 times more tobacco non-disclosure than post-issue audits.
- Swiss Re, underwriting automation programme materials: around 75% of applications through automated engines on average for established programmes, up to 90% in some markets. Provider-published figures.
- McKinsey & Company, How data and analytics are redefining excellence in P&C underwriting, September 2021: in the authors' experience up to 95% of personal lines policies may undergo straight-through processing; digitised underwriting worth 3 to 5 points of loss ratio at best-in-class carriers; executives cite human capital as their scarcest resource.
- Datos Insights (Aite-Novarica), Straight-through processing in underwriting and claims: 2023 update: many insurers do not systematically track straight-through rates; commercial and specialty automation rising through narrowly defined product parameters.
- Sax, D. R., et al., Evaluation of version 4 of the Emergency Severity Index in US emergency departments for the rate of mistriage, JAMA Network Open, 2023: 5,315,176 emergency department encounters; 32.2% mistriaged (28.9% overtriage, 3.3% undertriage).
- Anti-money-laundering screening false-positive rates of 85–95% are widely cited industry figures rather than a single study, and are used here only to indicate the failure mode's direction and scale.
- Lloyd's Market Association, The growth of enhanced underwriting in the Lloyd's market, November 2024: interviews and surveys covering 77% of Lloyd's 2023 premium; augmented underwriting growth expectations around 60% a year; no respondents expecting decline. Ki Syndicate 1618 received Lloyd's permission to underwrite algorithmically in October 2020. The accountability and audit-trail warnings in Section 7 are from the LMA's chapter in ICLG Insurance and Reinsurance 2026, Navigating the risks of enhanced underwriting.
- Lloyd's of London, 2026–30 strategy, March 2026: transition away from Blueprint Two towards market standards and incremental modernisation.
- Data (Use and Access) Act 2025, section 80, inserting Articles 22A–22D into UK GDPR, in force 5 February 2026; ICO draft guidance on automated decision-making and profiling, consultation closed May 2026, final guidance pending at the time of writing.
- Regulation (EU) 2024/1689 (AI Act), Annex III point 5(c), as amended by Regulation (EU) 2026/1744: high-risk classification covers risk assessment and pricing of natural persons in life and health insurance; Annex III obligations deferred to 2 December 2027.
- EIOPA, Opinion on artificial intelligence governance and risk management, August 2025.
- FCA: AI update, April 2024; general insurance pricing practices remedies (PS21/5), in force January 2022; GAP insurance sales pause, February 2024. Bank of England and FCA, Artificial intelligence in UK financial services, November 2024: 75% of surveyed firms using AI; 34% reporting complete understanding of the systems they use.
- Gama, J., et al., A survey on concept drift adaptation, ACM Computing Surveys 46(4), 2014; Sculley, D., et al., Hidden technical debt in machine learning systems, NeurIPS 2015; NIST, AI Risk Management Framework 1.0, January 2023.
- London market underwriter compensation: specialist recruiter mandates advertised 2025 (professional indemnity underwriters £60,000–£100,000 base by seniority, with senior mandates at £80,000–£110,000); fully loaded multiplier 1.3–1.5× base per standard UK employer-cost conventions. Indicative, not a salary survey.
- Sollers Consulting survey, reported June 2026: around 40% of insurers use AI in underwriting, and roughly 20% of commercial insurers use AI to triage incoming submissions. Industry survey, reported at second hand.
References 31, 35, and 44 are provider-published or industry-circulated figures and are labelled as such where used. The two US New England Journal studies are bridged to the UK explicitly in Section 3; no UK equivalent dataset exists at physician level, which constrains any UK rating model and is acknowledged rather than hidden.
ETerms usedOpen +Close −
| Term | Meaning as used in this article |
|---|---|
| Coverholder | A firm authorised to enter into contracts of insurance on behalf of a Lloyd's syndicate under a binding authority agreement. |
| Binding authority, binder | The agreement delegating underwriting authority, whose schedule defines classes, limits, territories, rates, exclusions, and referral triggers: the appetite, expressed as rules. |
| Triage engine | The pipeline that reads, verifies, scores, and routes each application or renewal to the smallest sufficient level of human judgement. |
| Straight-through processing | A decision completed with no human involvement at any step. |
| No material change | A renewal whose risk-relevant facts are unchanged within defined tolerances: same specialty and scope, clean claims year, clear register check, turnover in band. |
| Escalation rate | The share of automated-band cases re-routed to a desk on validation flags. The first number a routing programme should watch. |
| Override rate | The share of desk-reviewed cases where the human changes the machine's recommendation. The second number. |
| Random holdout | A permanent random sample of automated decisions re-underwritten by a human to measure slippage. Distinct from an audit, which investigates after the fact. |
| Slippage | The excess error on the automated band relative to full human underwriting: the quantity the holdout measures and the threshold prices. |
| Reject option, deferral | The formal machinery by which a classifier declines to decide and hands a case to a human, optimally at a threshold set by the cost ratio of error to referral. |
| Coverage | In selective prediction, the share of cases the system decides itself at a chosen error tolerance. |
| Calibration | Agreement between a model's stated confidence and its actual accuracy. Modern models are overconfident by default, so thresholds are set on measured performance. |
| Bordereau | The periodic report of risks, premiums, and claims a coverholder submits to the market: monthly, batch, and far thinner than the decision record in Exhibit 10. |