Insurance risk pricing & underwriting

Pricing is the core actuarial task in insurance. Estimate the expected cost of a risk, add expenses and margin, and decide whether and on which terms to accept it. Generalised linear models (GLMs) have been the standard for decades. Today gradient boosting, explainable GAM-style models and tabular foundation models find non-linear effects and interactions faster. Pricing platforms then turn those findings back into transparent, fileable tariffs and push them to the rating engine within days. On the underwriting side, LLM assistants read submissions and medical evidence and prepare decisions. In the EU this happens under tight constraints: no gender-based premiums, fair-pricing expectations from EIOPA, and high-risk status under the AI Act for life and health pricing of natural persons.

Problem & value

• Today: pricing actuaries build GLMs per peril and line in tools such as SAS, Emblem or R. They review them in committee and hand them to IT to implement in the rating engine. A tariff cycle often takes months. Commercial and specialty underwriters price on spreadsheets and judgement. Life underwriters manually assess questionnaires and physician reports.
• AI value: better risk segmentation, which reduces anti-selection and improves the loss ratio. Tariff cycles get faster, models become consistent across lines and channels, more simple risks are accepted automatically, and underwriters spend their time on complex risks.
• Vendor-reported effects (not independently audited): Akur8 claims more than 350 insurers as customers and says time to market falls "from months to days" [2]. Earnix reports more than 100 insurers, rate creation in days instead of about six months, and quoting 12× faster than legacy engines [3]. hyperexponential says carriers writing more than $75bn of annual commercial P&C premium use its platform [4]. Test such claims on your own data before relying on them.

Variants

• Personal lines P&C (motor, home, liability): high-volume GLM/GBM pricing plus price optimisation within regulatory limits. Telematics-based motor pricing is a separate use case (PRED-030).
• Commercial & specialty / MGA: pricing models per account within an underwriting workbench (hyperexponential, Akur8, in-house Python). Submission intake upstream is DOC-024.
• Property with geospatial enrichment: hazard scores (flood, hail, wildfire, subsidence) and roof or building attributes from aerial imagery feed the rating factors. Long-horizon climate scenarios belong to PRED-036.
• Life & health (natural persons): automated underwriting rules engines from reinsurers, increasingly paired with ML mortality and morbidity models and LLM reading of medical evidence. This is the Annex III high-risk variant (see below).
• Reinsurance & portfolio pricing: experience and exposure rating, with ML used for loss-ratio prediction per cedent and segment.
• Brokers / comparison portals: predict acceptance and price ranges across carriers (advisory, not the carrier's tariff).
• Private use: not a private-user use case. Consumers only experience the result (quote price).

How it works

1. policy + claims + exposure data (data warehouse, IFRS 17 / Solvency II data lineage)
2. enrichment (geocoding & hazard zones, vehicle/building data, company registers)
3. feature engineering & data-quality checks
4. risk models: frequency (Poisson) × severity (Gamma) or Tweedie pure premium GLM/GAM baseline | GBM / EBM / tabular FM as challenger or to discover interactions
5. "glass-boxing": translate ML insights into tariff factors, smoothing, monotonic constraints
6. fairness & stability checks (no protected attributes or proxies, segment back-tests)
7. [optional] demand / retention models
8. constrained price optimisation (P&C, where allowed)
9. pricing committee + actuarial function sign-off ← human-in-the-loop
10. deployment to rating engine via API, A/B or shadow rating, monitoring & drift alerts underwriting: submission/application
11. rules + ML score
12. auto-accept | refer to underwriter (LLM assistant summarises evidence, drafts terms; underwriter decides)

• Model families: GLMs with Poisson/Gamma/Tweedie losses remain the tariff format of record. GBMs (LightGBM, XGBoost, CatBoost) and EBMs [8] serve as challengers and help find interactions. Combined actuarial neural networks (CANN) and the methods in Wüthrich & Merz's actuarial ML literature [10] bridge GLMs and deep learning.
• Human-in-the-loop: pricing committee approval of every tariff, actuarial function review, underwriter referral above authority limits, and manual review of every declined or loaded life/health case.

Data preparation

Which data: policy and exposure data, claims history with costs and causes, customer and risk attributes, third-party data (geo, vehicle, credit, telematics) and underwriting notes.

Where it lives: the policy administration system, claims system, broker portals, data warehouse, plus e-mails and PDF applications or surveys in a DMS or on file shares.

Typical gaps: exposures and claims not joined cleanly, changing product definitions over the years, free-text underwriting notes, scanned applications, inconsistent coding of causes, restricted attributes (health, gender) with unclear legal basis, and access rights spread across departments.

Steps: 1. Build a historical policy-claims table with consistent exposure periods (PREP-005, PREP-011). 2. Define target variables (frequency, severity, loss ratio) and harvest labels from claims records (PREP-020). 3. Extract structured fields from application PDFs and surveys with OCR/LLMs (DOC-014, DOC-018). 4. Agree a business glossary for risk factors (PREP-012). 5. Remove or justify protected attributes, then pseudonymise development extracts (PREP-022, PREP-025). 6. Reserve out-of-time test data (PREP-018).

Effort share: about 50-60 percent of the work. Buying a rating engine cuts this to onboarding and mapping your own data. Check readiness with GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (C): Pricing and underwriting are the insurer's core and a source of competitive advantage; sensitive policy and claims data and BaFin/EIOPA expectations favour in-house models on a controlled platform.
• SMB (A): Small insurers, MGAs and brokers should use established rating engines and risk-scoring services from specialist vendors rather than build models.
• Developer (C): Open-source actuarial and ML libraries (GLMs, gradient boosting, SHAP) run on HW1 and fit custom pricing work.

Path A — Buy (off-the-shelf)

• Pricing platforms (P&C): Akur8 (France, EU; transparent ML that outputs GLM-style tariffs, plus reserving and life modules) [2]; Earnix (pricing, rating engine, underwriting) [3]; Guidewire PricingCenter (cloud pricing and rating with GLM/GAM, GBM and API deployment; quantee.ai, formerly a Polish EU vendor, now redirects there) [5]; WTW Radar / Emblem (long-established actuarial suite).
• Commercial/specialty underwriting: hyperexponential (UK; pricing models and renewals in an underwriting workbench) [4]. Submission AI vendors are listed in DOC-024.
• Life/health automated underwriting: reinsurer rules engines such as Munich Re ALLFINANZ, Swiss Re Magnum and Hannover Re's hr | ReFlex. They are usually free or cheap when bundled with reinsurance treaties. All three vendors are EU or Swiss.
• Choose A when: you are a small or mid-sized insurer or MGA without a large data science team, you need fileable and explainable tariffs, and you want a rating-engine integration.

Path B — Build on cloud APIs

• ML platforms in EU regions: Azure Machine Learning, AWS SageMaker (Frankfurt), Google Vertex AI (EU), Databricks or Snowflake on EU regions; the STACKIT and IONOS clouds as sovereign alternatives.
• LLM for underwriting assistance: Mistral (EU), or Claude, GPT and Gemini via EU data residency. Typical tasks are summarising submissions and medical evidence, drafting referral notes and explaining rating factors. Keep the LLM out of the price calculation itself.
• Tabular FM API: Prior Labs offers a hosted TabPFN API (Germany) for quick challenger models.
• Choose B when: you have an in-house actuarial data science team and want flexibility on top of a managed ML platform.

Path C — Self-host (open source)

• Modelling: glum (QuantCo, BSD-3; fast regularised GLMs with Tweedie and monotonic constraints) [7], statsmodels, LightGBM, XGBoost, CatBoost, InterpretML EBM (MIT) [8], scikit-learn, R (glm, mgcv, gbm). Explainability via SHAP and partial-dependence plots.
• Tabular foundation models: TabPFN (Prior Labs GmbH). Version 3.5 accepts up to 1 million rows and 20,000 features. Weights from 2.5 onwards are released under non-commercial licences, and only TabPFN-2 is Apache-2.0-based [6], so commercial use needs a licence.
• MLOps: MLflow (registry), Feast (features), Evidently (drift), ONNX or a Python microservice for scoring behind the rating engine. An open-weight LLM (Mistral Small, Qwen3) on vLLM serves the underwriting assistant.
• Hardware: HW1. GLM/GBM training and scoring run on CPU servers. A GPU (HW2) helps only for TabPFN on large data or the self-hosted LLM assistant.

Path D — Custom training

This is effectively what pricing means: every insurer trains on its own claims experience, because market models do not transfer. The data needed is several years of policy-level exposure and claims, ideally with tens of thousands of claims per peril for stable GBMs. Smaller portfolios use credibility weighting, market or reinsurer base rates, or tabular foundation models. Fine-tuning an LLM is rarely worth it; retrieval over underwriting guidelines usually suffices.

State of the art (2026)

• Hybrid "transparent ML": the market converges on GBMs or automated GAM search for discovery, with output delivered as constrained, smoothed GLM/GAM tariffs that regulators and pricing committees can read. Akur8, PricingCenter and Earnix all emphasise this [2][3][5].
• Tabular foundation models (TabPFN-3.x) now handle datasets up to a million rows in context [6] and are strong quick challengers. Gradient boosting remains the production workhorse because of licensing, latency and auditability. Community benchmarks such as TabArena compare these model classes.
• Agentic actuarial and underwriting tools: vendors are adding agents that prepare model reviews, rate-change rationales and renewal repricing (Akur8 Agents, hyperexponential Renewals, Earnix AIOS) [2][3][4].
• Limits & failure modes: sparse large-loss data (severity tails), proxy discrimination through postcode or occupation, and concept drift from inflation and claims-cost shocks. Feedback loops also appear, because the portfolio you price is shaped by your previous prices. Price optimisation that charges loyal customers more is under supervisory scrutiny [9].

Shared resources & combinations

• Same tabular-ml stack and feature store: PRED-009 credit scoring, PRED-003 churn (renewal retention models feed the price optimisation), PRED-015 AVM (sums insured for property), PRED-037 claims severity and reserving (same claims data, other target), ANOM-003 claims fraud. One CPU-based MLOps platform (HW1) and one governed policy/claims data mart can serve all of them.
• Upstream data: DOC-024 submission intake (normalised SOVs and loss runs), PRED-030 telematics scores, PRED-036 physical climate risk and DATA-011 geospatial enrichment as rating factors.
• Downstream: OPT-004 dynamic pricing logic (demand × cost), DATA-027 experimentation for controlled rate tests, AGENT-014 claims handling, CONV-013 customer assistant explaining quotes.
• Governance: SEC-022 AI inventory and AI Act documentation. The same fairness, explainability and monitoring tooling also covers PRED-009 and HLTH-020.

Effort, cost & KPIs

• Path A: 3–9 months to put a platform into production, including data migration and rating-engine integration. Cost is a subscription per line or premium volume (vendor quote).
• Path B/C: 6–18 months for an in-house pricing stack with a team of 3–8 actuaries and data scientists. Compute cost is minor compared with people and data cost.
• Cost drivers: data quality and history, external enrichment data licences, rating-engine integration, model governance and documentation.
• KPIs: Gini/lift and deviance on holdout, actual-vs-expected by segment, loss ratio and combined ratio by cohort, conversion and retention, quote-to-bind time, tariff release cycle time, straight-through underwriting rate, referral rate, fairness metrics per protected-group proxy.

People & change

Involve early: management (strategy, risk appetite), the heads of underwriting and actuarial as process owners, IT and IT security, the DPO, legal/compliance (BaFin, Solvency II, discrimination law) and data owners for policy and claims systems. Include underwriters as end users from the first workshop; they know where models fail.

Works council: usually no, since the model rates applicants and risks, not staff. It becomes relevant when underwriter overrides or processing times are logged per person (§87(1) no. 6 BetrVG), and the council must be informed of AI-planned work processes (§90(1) no. 3). Seek an agreed purpose limitation.

Data protection: a DPIA (Art. 35 GDPR) is likely where health or other special categories are used or profiling is systematic; AVV with each data and model provider; involve the DPO. See GUIDE-005.

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy. Train underwriters and actuaries on model limits, drift and explanations (GUIDE-006).

Pilot and communication: run a new model in shadow mode on one product line and compare with current pricing (GUIDE-009).

Concerns: deskilling and automated rejections are addressed with referral rules, documented override rights and regular fairness reviews. Humans keep final authority on declines and exceptional risks (GUIDE-008).

Risks, regulation & compliance

• EU AI Act (classification: depends): Annex III point 5(c) makes AI systems for risk assessment and pricing of natural persons in life and health insurance high-risk [1]. That triggers Art. 9–15 obligations (risk management, data governance, documentation, human oversight, accuracy), registration and a fundamental-rights impact assessment by deployers (Art. 27). The Annex III obligations were originally due from 2 Aug 2026. Under the Digital Omnibus on AI (Reg. (EU) 2026/1744, in force since 27 Jul 2026), they apply from 2 December 2027 (as in PRED-009). P&C, commercial and reinsurance pricing are not listed in Annex III and are usually minimal risk. Art. 4 AI literacy still applies. Not prohibited under Art. 5, unless it becomes social scoring that uses unrelated behavioural data to the detriment of individuals.
• EIOPA: the Opinion on AI governance and risk management (6 Aug 2025) covers AI not classed as high-risk. It sets proportionate expectations on data governance, record-keeping, fairness, explainability, cyber security and human oversight under Solvency II and IDD [11]. The Supervisory Statement on differential pricing practices (16 Mar 2023) targets price-setting that treats customers unfairly, such as price walking [9].
• Anti-discrimination: CJEU Test-Achats (C-236/09) bars gender as a premium factor [12]. Avoid proxies for ethnicity, religion or disability (Directive 2004/113/EC, German AGG §§19–20).
• GDPR: pricing uses personal data. Health data (Art. 9) needs explicit consent or a legal basis. Fully automated declines or loadings fall under Art. 22. The CJEU SCHUFA judgment (C-634/21) held that a score which decisively shapes a decision can itself count as an automated decision [13]. Provide meaningful explanations, the right to human review, a DPIA and minimal external data.
• Sector law: Solvency II (actuarial function, model validation), IDD product oversight and demands-and-needs test, national supervision (BaFin/VAG in Germany), and DORA for ICT third-party risk with pricing SaaS.
• Mitigations: keep protected attributes out of model inputs but available for fairness testing, monotonic and plausibility constraints, challenger models, documented committee approval, drift monitoring, and humans deciding all adverse life/health outcomes.

Getting started

1. Inventory current tariffs, data sources and the rating engine. Classify each model under the AI Act (life/health natural persons = high-risk) and register it in the AI inventory (SEC-022).
2. Build a clean, versioned policy-and-claims modelling dataset with exposure, and reproduce the current GLM as a baseline.
3. Run a GBM/EBM challenger, or a trial of Akur8, PricingCenter or Earnix, on one line. Compare holdout Gini and actual-vs-expected, and check which interactions the baseline misses.
4. Define fairness, explainability and sign-off standards with the actuarial function and compliance before you deploy.
5. Deploy in shadow mode next to the existing tariff, monitor, then release via controlled rate tests.

Sources

1. EU AI Act, Annex III (high-risk use cases, point 5(c) life & health insurance) — https://artificialintelligenceact.eu/annex/3/
2. Akur8 — pricing, reserving, life and agent products; customer and time-to-market claims — https://www.akur8.com/
3. Earnix — pricing, rating, underwriting, AIOS; customer and speed claims — https://earnix.com/
4. hyperexponential — commercial P&C pricing and underwriting platform — https://www.hyperexponential.com/
5. Guidewire PricingCenter — https://www.guidewire.com/products/core-products/insurancesuite/pricingcenter-insurance-pricing-management-software
6. Prior Labs TabPFN repository (versions, limits, licences) — https://github.com/PriorLabs/TabPFN
7. QuantCo glum — high-performance GLMs in Python — https://github.com/Quantco/glum
8. InterpretML / Explainable Boosting Machines — https://github.com/interpretml/interpret
9. EIOPA Supervisory Statement on differential pricing practices in non-life insurance (2023) — https://www.eiopa.europa.eu/publications/supervisory-statement-differential-pricing-practices-non-life-insurance-lines-business_en
10. Wüthrich & Merz, Statistical Foundations of Actuarial Learning and its Applications (Springer, open access) — https://link.springer.com/book/10.1007/978-3-031-12409-9
11. EIOPA publishes Opinion on AI governance and risk management (6 Aug 2025) — https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en
12. CJEU C-236/09 Test-Achats — https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62009CJ0236
13. CJEU C-634/21 SCHUFA Holding (Scoring) — https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62021CJ0634
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj