Underwriting

Personal lines insurance can achieve straight-through processing rates of 70 to 80 per cent. Specialty and commercial lines hover below 20 per cent. The gap between these numbers represents the opportunity and the challenge of applying AI to underwriting: the decisions that are easy to automate are already automated, and the ones that remain are genuinely hard.

What is underwriting?

Underwriting is the process of evaluating risk, deciding whether to accept it, and determining the price at which it is profitable to do so. In insurance, this means assessing an application, whether for motor, property, liability, or specialty cover, against the insurer's appetite, pricing the risk, and setting the terms of the policy. It is the core commercial function of an insurance company: the quality of underwriting determines profitability. The decisions made here flow directly into claims processing, where the terms and conditions set during underwriting shape how claims are assessed and settled.

The complexity varies enormously by line of business. A personal motor policy can be priced in milliseconds using a rating engine with hundreds of factors. A marine cargo policy for a shipment of lithium batteries through the Suez Canal requires a specialist underwriter who understands the cargo, the route, the vessel, and the geopolitical context. These are not the same problem, and they do not have the same AI solution.

The underwriting process for commercial and specialty lines involves ingesting a submission, typically a broker's presentation of the risk via documents, spreadsheets, and narrative, extracting the relevant data, assessing it against underwriting guidelines, pricing the risk, and communicating terms. Much of the underwriter's time is spent on data extraction and validation, not on the judgement that is their actual expertise.

The landscape

The demand for embedded underwriting APIs is growing rapidly. Insurtechs, e-commerce platforms, and banks want to offer insurance at the point of need, which requires underwriting decisions in real time. Traditional underwriting workflows, designed around email-based submissions and multi-day turnaround, cannot serve this demand. Insurers that can expose their underwriting capability as an API capture distribution that others cannot.

Climate risk is transitioning from a reporting requirement to an underwriting imperative. The PRA's expectations on climate risk management require insurers to integrate climate scenarios into their pricing and reserving. For property insurance, this means flood, windstorm, and wildfire risk models that reflect current and projected climate conditions, not just historical loss experience. For liability lines, emerging climate litigation risk is harder to quantify but increasingly material.

Actuarial modelling and underwriting are converging. Historically, actuaries set the technical price and underwriters applied commercial judgement. AI blurs this boundary: models can incorporate both actuarial risk factors and market intelligence, producing prices that reflect both the statistical risk and the competitive context. The organisational implications, who owns the pricing decision when the model spans both functions, are as significant as the technical ones.

How AI changes this

Submission ingestion is the highest-value, most immediately deployable application. AI systems extract structured data from broker submissions, PDFs, email attachments, spreadsheets, and loss runs, and populate the underwriting workbench automatically. This reduces the time an underwriter spends on data entry from hours to minutes per submission. The technology is a specific application of document intelligence trained on insurance-specific document types.

Automated triage routes submissions to the right underwriter based on the extracted data. A submission within appetite, within authority limits, and matching standard terms can be fast-tracked. One outside appetite can be declined immediately rather than sitting in a queue for days. The underwriter's time is redirected to the complex risks where human judgement adds value.

Pricing augmentation feeds additional data sources into the rating process. For property risks, satellite imagery, IoT sensor data, and real-time weather exposure. For liability risks, litigation trend analysis and regulatory change monitoring. For motor, telematics data that reflects actual driving behaviour rather than demographic proxies. These additional signals improve pricing accuracy, which means fewer underpriced risks and more competitive quotes on well-understood ones.

Portfolio-level analytics give underwriting leaders visibility into risk accumulation that was previously available only through periodic actuarial reviews. Real-time dashboards showing aggregate exposure by peril, geography, and line of business enable dynamic appetite management: tightening terms in areas of excess concentration and targeting growth where the portfolio is underweight relative to capacity.

What to know before you start

The data in broker submissions is the bottleneck, not the AI. If your submission intake process does not capture data in a structured, consistent format, your extraction accuracy will suffer. Work with your broker relationships to improve submission quality; the benefit is mutual.

Underwriting authority frameworks must be updated for AI. If your authority limits are defined in terms of human sign-off thresholds, you need to define equivalent thresholds for automated decisions. Who is accountable when the model accepts a risk that subsequently produces a large loss? The answer must be clear before deployment, and it must satisfy both the board and the regulator.

Risk assessment models used in underwriting are subject to model risk management expectations. The PRA's SS1/23 applies to insurance models as much as to banking models. Validation, monitoring, and governance are not optional. Build the model risk framework before you build the model.

Start with the line of business where the submission volume is highest and the decision complexity is lowest. Personal lines are already heavily automated, so the next tranche is typically SME commercial lines, where submission formats are relatively standardised and underwriting guidelines are well-defined. Specialty and treaty reinsurance underwriting, where judgement and relationship are paramount, is a longer-term target.