Neural network underwriting

Category: AI in insurance · Reviewed by Matt Bartlett, Director · Founder · Last reviewed 2026-06-10

Neural network underwriting is the use of artificial neural networks — multi-layered models of weighted connections — to predict insurance outcomes, particularly from unstructured inputs such as text, images and time-series telematics. In United Kingdom insurance the technique has matured from a niche research tool into a production component for specific use cases where deep learning meaningfully outperforms generalised linear models and gradient-boosting machines.

Category: AI in insurance · Aliases: Deep learning underwriting, neural net pricing · Established: Classical neural networks since the 1980s; deep-learning era from 2012; UK insurance production use from c.2019 · Related: Machine learning underwriting, AI underwriting, Computer vision claims

Definition

A neural network consists of layers of “neurons” — weighted linear combinations of inputs passed through a non-linear activation function. Training proceeds by stochastic gradient descent on a loss function (Poisson deviance, cross-entropy, mean squared error). Deep neural networks (DNNs) have many layers; specialised architectures include:

• convolutional neural networks (CNNs) for images (e.g. motor damage, roof condition);
• recurrent neural networks (RNNs) and transformers for sequences and text (e.g. telematics traces, broker emails);
• tabular DNNs (e.g. TabNet, FT-Transformer) for structured insurance data.

In UK underwriting they are most often used where the input is unstructured.

Legal / Regulatory basis

• FCA Handbook SYSC 4 governance and SYSC 7 risk-control requirements apply to neural-network models with material customer impact.
• PRA SS1/23 Model risk management principles for banks (May 2023): although directly addressed to banks, UK insurance model-risk teams treat its tiering, validation and ongoing-monitoring expectations as directional best practice — particularly for opaque deep-learning models.
• FCA & PRA DP5/22 / FS2/23: explicitly call out the trade-off between performance and explainability, with neural networks at the more opaque end.
• Solvency II internal-model use, including documentation and use-test, would apply where neural networks feed technical provisions.
• UK GDPR Article 22 automated decision-making; Data Protection Act 2018.
• Equality Act 2010: proxy-discrimination risk demands bias testing of training data and outputs.

How it works in practice

1. Use-case selection — typically where structured-data baselines under-perform, e.g. damage assessment from photographs, behaviour signals from telematics, or text classification of broker submissions.
2. Data assembly — labelled training data; for images this may involve substantial labelling work.
3. Architecture choice — CNN for vision, transformer for text/sequence, tabular DNN for structured data; pre-trained backbones are common.
4. Training and validation — held-out and time-split validation; monitoring for over-fit; reproducibility through fixed random seeds and containerised training.
5. Explainability — Grad-CAM heatmaps for vision, attention visualisation for text, SHAP/Integrated Gradients for tabular DNNs.
6. Bias and robustness — adversarial testing and slicing across protected-characteristic proxies.
7. Governance — Tier 1 model-risk file under the firm’s framework, with senior-manager accountability under SMCR.

Common variations / Subsequent developments

• CNNs in motor claims powering platforms such as Tractable, Solera/Audatex and Mitchell Intelligent Open Shop (see Computer vision claims).
• CNNs in property underwriting using aerial / satellite imagery (CAPE Analytics, EagleView, Nearmap) for roof condition, swimming pools, debris and tree-canopy risk.
• Sequence models on telematics for driver-style scoring (see Telematics insurance in batch 12).
• Foundation-model fine-tuning — from 2023, UK firms have explored fine-tuning open-source large language and vision models on their own data, with retrieval-augmented generation overlays.
• Federated learning trials in the London market, allowing model training across syndicates without centralising data.

Example

A UK commercial property MGA uses a CNN trained on aerial imagery to flag survey priority on incoming submissions. The model returns a roof-condition score and visual evidence. Underwriters retain the final decision. The model is classified Tier 1 under the firm’s framework, with an annual independent validation, quarterly performance monitoring against ground-truth surveys, and a documented procedure for replacing the model if its calibration drifts.

See also

• Machine learning underwriting
• Computer vision claims
• Explainable AI insurance
• AI underwriting (batch 12 expansion)

References

• LeCun, Y., Bengio, Y. & Hinton, G., “Deep learning”, Nature, 521(7553), 2015.
• PRA, SS1/23 — Model risk management principles for banks, May 2023. https://www.bankofengland.co.uk
• FCA & PRA, DP5/22 — Artificial Intelligence and Machine Learning, October 2022.
• FCA & PRA, FS2/23 — Feedback Statement, October 2023.
• EIOPA, AI Governance Principles, June 2021.
• Bank of England & FCA, AIPPF Final Report, October 2022.
• UK GDPR; Equality Act 2010, https://www.legislation.gov.uk

This entry is part of the Apex Insurance Wiki. Last reviewed by Matt Bartlett on 2026-06-10. Next review: 2026-12-10.

Apex Insurance Brokers Limited. Authorised and regulated by the Financial Conduct Authority, FRN 724952. Registered in England and Wales, Companies House 07014570. This entry provides general information about UK insurance concepts and is not regulated advice. Consult your insurance broker on your specific position.