Huzzle Labs

Core insurance workflows

Insurance runs on a handful of document-heavy workflows: underwriting a risk, working a claim from first notice to settlement, and the actuarial work that prices and reserves for it all. Here's how that work gets done — step by step — and where AI fits. It's the same work InsureBench measures models on.

~70%
of underwriters' time is spent not underwriting — Accenture
>50%
of claims activities could be automated by 2030 — McKinsey
$80–160B
potential P&C fraud savings from AI by 2032 — Deloitte

Underwriting

The decision process by which an insurer evaluates a risk, decides whether to accept it, and at what price. The Insurance Information Institute defines it as "examining, accepting, or rejecting insurance risks, and classifying the ones accepted, in order to charge appropriate premiums."1

01

Submission & intake

Applicant or broker submits the risk, exposure, and loss history.

02

Risk assessment

Analyse the submission; pull data, scores, inspection/medical reports.

03

Apply guidelines

Check appetite, limits, and authority; sort into a risk class.

04

Pricing & rating

Apply rate factors to set a premium adequate for expected losses.

05

Accept / decline / refer

Decide, possibly with terms and exclusions, or escalate.

06

Quote & bind

Issue a quote; a binder gives temporary cover once accepted.

07

Policy issuance

Produce the formal contract, superseding the binder.

Personal vs. commercial: personal lines (auto, home) are high-volume and rules-based, so clean cases process straight through; commercial lines are document-heavy and judgement-based, with senior underwriters reserved for complex accounts.2

Underwriters spend ~40% of their time on administration and ~30% on sales support — leaving only about 30% for underwriting itself.

Accenture P&C Underwriting Survey3  ·  McKinsey projects >90% of simple-lines underwriting automated by 20304

See the underwriting AI benchmark for how InsureBench scores models on this work.

The claims lifecycle

A claim runs from the first report of a loss to final settlement and recovery. A claims adjuster "investigates, evaluates, and settles claims," determining whether the policy covers the loss and how much the insurer should pay.5

01

FNOL

First Notice of Loss is reported, triggering the process.

02

Registration & triage

Log the claim, categorise by severity, assign an adjuster.

03

Coverage verification

Confirm cover, limits, deductibles, and exclusions.

04

Investigation

Assess damage, liability, and possible fraud.

05

Reserving

Set a case reserve for the expected payout.

06

Settlement

Determine the payable value; approve, partially pay, or deny.

07

Subrogation

Recover from liable third parties and salvage.

08

Closure

Settle and close the file; audit later for leakage.

Where AI fits today. McKinsey estimates more than half of claims activities could be automated by 2030, with loss-adjustment expenses falling 25–30% through digitisation.6 On fraud, Deloitte projects AI could save P&C insurers $80–160 billion by 2032, against annual P&C fraud losses of roughly $122 billion.7 See the claims AI benchmark.

Actuarial work

Actuaries price risk, set reserves, and model capital. The work splits into three areas, each governed by standard, named methods.

Pricing & ratemaking

Setting the rate

Built from the pure premium method (loss per exposure unit, loaded for expenses and profit) or the loss ratio method, with frequency×severity analysis, trending, and credibility weighting. GLMs are the industry-standard multivariate technique.8

Reserving

Estimating ultimates

Losses are arranged in a development triangle and projected to ultimate via the chain-ladder method, the expected-loss-ratio method, or the Bornhuetter-Ferguson blend — producing the reserve for claims incurred but not reported (IBNR).9

Capital & exposure

Modelling solvency

Catastrophe models simulate rare, severe losses; frameworks like the NAIC's Risk-Based Capital and the EU's Solvency II (99.5% one-year VaR) set required capital.10 2024 insured nat-cat losses neared $137B.11

Where AI fits today. Machine-learning methods (gradient boosting, neural nets) increasingly complement GLMs by capturing non-linearities, while GLMs remain the explainability and regulatory baseline.12 See the actuarial AI benchmark.

Key terms

FNOL

First Notice of Loss — the initial report that triggers the claims lifecycle.

Combined ratio

Loss ratio + expense ratio. Below 100% is an underwriting profit; above 100% a loss.

IBNR

Incurred But Not Reported — reserves for losses that have happened but aren't yet reported or developed.

Development triangle

Cumulative losses by accident period against development age, used to project ultimates.

Chain-ladder

Reserving method projecting ultimates via age-to-age development factors.

Subrogation

The insurer's right, after paying a loss, to recover from a liable third party.

Binder

A legal agreement giving temporary evidence of insurance until the policy issues.

Endorsement

An amendment (rider) that adds, deletes, excludes, or changes coverage.

Sources

  1. Insurance Information Institute — Glossary: Underwriting. iii.org
  2. McKinsey — Redefining excellence in P&C underwriting (2021). mckinsey.com
  3. Accenture — Why underwriters don't underwrite much (Feb 2022). accenture.com
  4. McKinsey — Insurance 2030: the impact of AI (Apr 2018). mckinsey.com
  5. U.S. Bureau of Labor Statistics — Claims Adjusters (OOH). bls.gov
  6. McKinsey — Claims 2030: dream or reality? mckinsey.com
  7. Deloitte — Using AI to fight insurance fraud (Apr 2025). deloitte.com
  8. Casualty Actuarial Society — Basic Ratemaking (5th ed.). casact.org
  9. Bornhuetter & Ferguson — The BF reserving method. reference
  10. NAIC — Risk-Based Capital. naic.org
  11. Swiss Re Institute — sigma 1/2025, natural catastrophes. swissre.com
  12. Society of Actuaries — P&C pricing in the age of ML (Jan 2024). soa.org
Leaderboard opening 2026. Built by Huzzle Labs.
Get in touch about InsureBench