Short answer: underwriting automation is an allocation problem
Allocate judgement; do not replace it. Risk is concentrated, yet judgement is spent evenly across the queue. A triage engine reads, scores and routes each case to the smallest sufficient level of human judgement. In the modelled doctors' book, that decides 58 per cent of cases with no desk and cuts cost per decision by 40 per cent.
Last reviewed: 27 October 2026. Modelled analysis, not a forecast.
- Risk concentrates. About 1 per cent of US physicians account for 32 per cent of paid malpractice claims.
- Capacity is diluted. Underwriters still spend 35 to 42 per cent of their time on non-core work.
- Autonomy has a price. Automating a renewal pays only while slippage stays below 1.15 per cent.
Why is underwriting an allocation problem?
Because risk is concentrated while effort is not. In a US study of 66,426 paid claims, about 1 per cent of physicians accounted for 32 per cent of them, yet underwriters spend 35 to 42 per cent of their time on non-core work, and a first-come-first-served queue treats every case alike.[1][2]
| Finding | Figure |
|---|---|
| Share of paid claims from the most claim-prone 1% of physicians | 32% |
| Recurrence risk with two prior paid claims, against one | 1.97× |
| Recurrence risk with three prior paid claims, against one | 3.11× |
| Annual claim rate, psychiatry to neurosurgery | 2.6% to 19.1% |
Sources: Studdert et al., 2016 (66,426 paid claims, 54,099 US physicians); Jena et al., 2011 (40,916 physicians; 7.4% face a claim in any year).[2][3]
The desk is diluted too. Three surveys put non-core and administrative time at about 40 per cent in 2021, 41 to 43 per cent in 2024 and 35 per cent in 2025, and renewals dominate the work: Travelers retains 85 per cent of its business.[1][4]
How does a triage engine decide when not to decide?
It reads the submission, verifies what it can, scores the case and routes it. Routing is the decision that matters: each case goes to no desk, a junior's structured review or a senior underwriter. As Chow showed in 1970, the referral threshold is set by the cost of an error against the cost of a referral.[5]
- Verify at source. For UK doctors, the General Medical Council register of 313,829 licensed doctors is downloadable daily for a licence of £815 plus VAT.[6]
- Learn the router. A routing function trained for the job beats a bare confidence cut-off on the scorer.
- Calibrate on evidence. Modern models are overconfident, so set thresholds on measured performance on held-out cases, never on a model's stated confidence.[7]
The engine does skim the easy work, and that is the design: it frees senior underwriters for the cases that need 40 minutes.
What did life insurance learn from automating underwriting?
That insurers automate eligibility but route timidly, and that the miss rate can be measured. Gen Re's 2024 survey found 82 per cent of carriers running accelerated underwriting, yet only 17 per cent of eligible applications passed with no human touch. The Society of Actuaries put slippage at 10 to 15 per cent.[8][9]
The most useful export is the random holdout: a permanent random slice of automated approvals sent to a human for full re-underwriting, because the machine did not flag them. In the actuaries' study, 81 per cent of audited cases landed in the class the machine chose, and holdouts caught 1.5 times more tobacco non-disclosure than post-issue audits. A programme with a holdout knows its miss rate as a number. One without is guessing.[9]
What does a modelled book of 20,000 doctors show?
Triage decides 58 per cent of cases with no desk, concentrates senior attention on 12 per cent and cuts cost per decision by 40 per cent, from £25.75 to £15.34. The composite coverholder book has 20,000 policies and about £29.4m of premium, with 17,000 renewals and 3,000 new applications a year. It shows relative size, not a forecast.
Design A is the traditional desk, B adds extraction but humans still decide everything, and C is the triage engine with an 8 per cent escalation rate and a 5 per cent random holdout.
| At steady state | A: every case reviewed | B: extraction only | C: triage engine |
|---|---|---|---|
| Underwriting staff | 5.0 FTE | 3.0 FTE | 2.0 FTE |
| Decisions needing no desk | None | None | 11,658 (58%) |
| Total annual cost | £515,000 | £343,500 | £306,815 |
| Cost per decision | £25.75 | £17.18 | £15.34 |
Data Hatch model at an illustrative scale; platform, licences and inference are inside Design C's figure. Its other decisions go 5,372 to the junior desk, 2,300 to the senior desk and 670 to the random holdout.
The machine is not the expensive part. Reading 50,000 documents a year through a small-tier model costs about £150, while the two underwriters cost £220,000, which echoes our analysis that the model is never the expensive part. The two desks are loaded to about 70 per cent, so the book can grow roughly 43 per cent before anyone is hired.
Automation pays until slippage reaches 1.15 per cent. Moving a renewal to the automated band saves about £10.10 of review cost, against an expected annual loss of £882 (a £1,470 premium at a 60 per cent loss ratio). Senior-desk cases break even at 5.5 per cent, which is why complex cases stay human. The random holdout measures actual slippage continuously. Life insurance tolerates 10 to 15 per cent, but a doctors' book has no such subsidy.
What conditions must automated underwriting meet?
Three: the binder is encoded as rules, every decision is reconstructable, and the UK safeguards are met. Since February 2026 the Data (Use and Access) Act 2025 permits solely automated decisions with significant effects, provided the applicant is told, can make representations, can obtain human intervention and can contest the outcome.[10]
- Encode the binder first. A binding authority schedule already defines classes, limits, territories and referral triggers. The automated band sits strictly inside it.
- Keep a decision record. Store the inputs, model, threshold and appetite versions, the route taken and who reviewed it.
- Keep declines human. A wrong automated bind costs slippage on one policy. A wrong automated decline costs a customer and possibly a discrimination complaint.
- Mind health data. The permission narrows sharply for it, so route anything touching health information to a desk.
- Expect outcomes to be tested. The FCA has no AI rulebook because the Consumer Duty and senior manager accountability already apply.[11]
How do you start with a triage engine?
Start with renewals, run the holdout from day one and encode the binder before the models. Renewals are 85 per cent of the model book's decisions and arrive with a year of history, so they are the safest place to begin. Measure your own desk first.
- Renewals first. Define no material change conservatively. New business follows only after months of clean routing.
- Holdout from day one. Widen the band while measured slippage stays under 1.15 per cent, and narrow it the moment it does not.
- Instrument first. Record minutes per case, referral reasons and decision latency.
- Avoid the known failures. Thresholds set on model self-confidence, no holdout, automating the tail first, and thresholds without an owner.
The platform in our underwriting case study applies this approach for a Lloyd’s coverholder, and the same principle of a human on every decline runs through our claims product, Claim It Quick.
Frequently asked questions
What is a triage engine in insurance underwriting?
The pipeline that decides how much human judgement each application deserves. It reads, verifies, scores and routes each case. Clean renewals and in-appetite new business are decided inside the binder, marginal cases go to a junior desk, and complex risks go to a senior underwriter.
Can underwriting decisions be automated safely?
In the modelled doctors' book, 58 per cent of decisions complete with no desk, but every one stays inside the binder's encoded appetite. An 8 per cent escalation rate pulls flagged cases back to a desk, a 5 per cent random holdout measures what automation misses, and declines are never automated.
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 per cent reduction. The same book runs on two underwriters instead of five, loaded to about 70 per cent; the desks have headroom to grow.
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 it pays only while the excess error stays below 1.15 per cent. A permanent random holdout, with 5 per cent of automated cases re-underwritten by a human, measures that error.
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 if the applicant is told, can make representations, can obtain human intervention and can contest the outcome. Health data narrows the permission sharply, and declines stay with people.
References
Sources cited, in the order first relied upon. The US medical studies show structure, not UK rates.
- The Institutes and Accenture, P&C Underwriting Survey, 2021; Capgemini, World Property and Casualty Insurance Report 2024 (41–43%); Accenture and The Institutes, Underwriting Rewritten, 2025 (commercial P&C 35%).
- Studdert, D. M., et al., Prevalence and characteristics of physicians prone to malpractice claims, New England Journal of Medicine 374:354–362, 2016.
- Jena, A. B., et al., Malpractice risk according to physician specialty, New England Journal of Medicine 365:629–636, 2011.
- Travelers, Full Year 2025 results, January 2026: Business Insurance retention 85%.
- Chow, C. K., On optimum recognition error and reject tradeoff, IEEE Transactions on Information Theory 16(1):41–46, 1970.
- General Medical Council, workforce report 2024; register download service, licence price £815 plus VAT.
- Guo, C., et al., On calibration of modern neural networks, ICML 2017; OpenAI, GPT-4 technical report, 2023.
- Gen Re, US individual life accelerated underwriting survey, November 2024.
- Society of Actuaries, accelerated underwriting mortality slippage study, August 2024.
- 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.
- FCA, AI update, April 2024.
Last reviewed: 27 October 2026.
Modelled analysis. The figures describe a composite book, not a forecast for any organisation. Legal references reflect the position at the review date and are not legal advice.