P&C insurers are investing in AI at an accelerating rate. BCG’s AI Radar shows that AI spending as a share of revenue will triple in 2026 across the property and casualty insurance industry. (Source: Boston Consulting Group, “BCG AI Radar: Insurance,” bcg.com, 2025)
Yet only 38 percent of P&C insurers are generating value at scale from AI in core workflows. The gap between investment and return is not a technology problem. BCG’s analysis is explicit: the issue is not technical. It is strategic and organizational. AI dropped into legacy operating models designed for human-led execution produces incremental improvements at best. The insurers capturing material returns are redesigning core processes around AI capabilities, not adding AI to processes built for humans.
Underwriting is where this redesign matters most. It is the core revenue-generating function of a P&C insurer. The process determines which risks to accept, at what price, and on what terms. Getting it faster, more accurate, and more consistent has a direct impact on combined ratio, premium growth, and competitive positioning.
This guide explains what AI is actually changing in P&C underwriting, the specific capabilities being deployed, the data infrastructure that enables them, and the realistic picture of where the industry stands today.
The Traditional P&C Underwriting Process and Its Constraints
Traditional P&C underwriting is a document-heavy, labor-intensive process. It has several structural constraints that limit speed, consistency, and scalability. A submission arrives, typically an ACORD form, an application document, or a broker email with attachments.
An underwriter manually reviews the submission, extracts relevant risk information, searches multiple internal and external data sources to supplement the application data, applies judgement to assess the risk, determines pricing, and issues a quote or declination.
This process has several well-documented problems. Submissions are processed sequentially by individual underwriters, creating bottlenecks. Manual data extraction from unstructured documents introduces errors.
Underwriting decisions vary across individual underwriters for similar risks. That inconsistency creates pricing leakage in both directions. Time-to-quote extends to days or weeks for complex risks, damaging competitive positioning in broker markets where speed of response correlates with win rates. The data problem is central.
Underwriting in P&C insurance is fundamentally driven by unstructured documents: property descriptions, business narratives, prior loss histories, inspection reports. These resist systematic processing. The diversity of document formats across brokers and submission channels compounds the challenge.
Where AI Is Delivering in P&C Underwriting
Submission Intake and Triage
The first AI capability to reach production scale in P&C underwriting was automated submission triage. Machine learning classifies incoming submissions, extracts key risk attributes, and routes submissions to the appropriate underwriting queue without manual intervention.
Natural language processing models extract structured data from unstructured submission documents. That includes property class codes, coverage limits, loss history summaries, and business descriptions. The same information that an underwriter would spend 20 to 30 minutes manually extracting can be extracted and structured in seconds.
AI-assisted intake and pre-fill is now cutting time-to-quote by 30 to 40 percent for standard risks, according to BCG. (Source: Boston Consulting Group, “AI in Insurance Underwriting,” bcg.com, 2025). For broker-distributed business where time-to-quote is a significant competitive variable, this translates directly into higher bind rates on target risks.
Risk Scoring and Predictive Pricing
Predictive analytics models are now effectively universal in P&C pricing.
According to WTW’s 2026 Advanced Analytics and AI Survey, close to 80 percent of P&C insurers rely on advanced rating and pricing models. An additional 11 percent plan to implement them, making predictive rating models standard from 2026. (Source: WTW, “2026 P&C Insurance Advanced Analytics and AI Survey,” wtwco.com)
These models use a combination of traditional rating variables and alternative data sources. Traditional variables include property characteristics, prior loss history, and geographic factors. Alternative data sources include satellite imagery, building inspection data, weather history, and IoT sensor data.
The practical result is more accurate pricing at the individual risk level. This reduces adverse selection, where the insurer consistently prices risks incorrectly and attracts the worst risks while losing the best ones. It improves portfolio loss ratios over time.
Straight-Through Processing for Standard Risks
Straight-through processing (STP) is the automated endpoint for low-complexity, standard risks.
A submission is ingested, evaluated, priced, and a policy issued without human underwriter involvement. AI models identify submissions that fall within clearly defined underwriting appetite, exhibit no flags that require human review, and can be priced accurately by the automated model. For these risks, the system generates a quote and can bind automatically.
STP rates vary significantly by line of business and risk complexity.
For personal lines and standard commercial lines (small business BOP, commercial auto for standard operators), STP rates of 60 to 80 percent are achievable. For complex commercial risks, STP rates are lower. But even reducing the human underwriting burden on 40 percent of volume frees significant capacity for risks that genuinely require underwriting judgement.
Fraud Detection in Submissions
Submission fraud is a material source of adverse loss experience for P&C insurers. It includes misrepresentation of risk characteristics, prior loss history, property occupancy, and business operations.
AI models trained on historical loss data and known fraud patterns can flag suspicious submissions. Flags include anomalies, unusual combinations of risk attributes, inconsistencies between declared information and external data sources, and prior loss history that does not match submitted information.
These flags route submissions for enhanced review rather than processing them through standard channels. According to WTW’s survey, 33 percent of carriers currently use advanced analytics for fraud detection. This figure is expected to reach 65 to 70 percent within two years. (Source: WTW, “2026 P&C Insurance Advanced Analytics and AI Survey,” wtwco.com)
Portfolio Monitoring and Underwriting Appetite Management
AI analytical tools give underwriting leadership real-time visibility into portfolio composition, concentration risk, and loss experience by segment. This replaces the quarterly or annual review cycle with continuous monitoring.
Predictive models identify emerging loss trends before they show up in incurred loss data. That enables proactive underwriting appetite adjustments. Teams can tighten coverage terms, adjust pricing, or exit segments earlier in the development of a deteriorating portfolio trend.
BCG’s analysis describes this as portfolio intelligence. AI gives underwriting leaders the visibility to make proactive portfolio management decisions rather than reactive responses to already-developed loss trends.
The Role of LLMs in Underwriting
Large language models are a more recent addition to the P&C underwriting toolkit. Over half of survey respondents in WTW’s 2026 survey already use LLMs, with another 29 percent planning to adopt them within the next year. (Source: WTW, “2026 P&C Insurance Advanced Analytics and AI Survey,” wtwco.com)
The primary application is document understanding. LLMs are significantly better than earlier NLP approaches at interpreting complex, context-dependent language in underwriting documents. Consider a commercial property application that describes a building as “a two-storey joisted masonry structure with ground floor retail and residential above.” It requires contextual understanding to map to the correct property class code.
Earlier rule-based NLP approaches struggled with the variability of how risk characteristics are described in natural language. LLMs handle this variability far more robustly. LLMs are also being used for policy language review. They flag contract terms that deviate from standard policy language, identify ambiguous coverage provisions, and generate coverage summaries for brokers and policyholders in plain language rather than policy contract language.
Underwriting assistants powered by LLMs allow underwriters to query risk information, summarise submission documents, and access underwriting guidelines conversationally rather than navigating multiple disparate systems. The time savings from reducing system navigation and document search free underwriters to spend more time on analytical work that requires professional judgement.
AI Across the P&C Underwriting Workflow
| Workflow Stage | Traditional Approach | AI-Enabled Approach | Measurable Impact |
| Submission intake | Manual review of email and forms; manual data entry into systems | NLP and LLM extraction of risk attributes from any document format | 30 to 40 percent reduction in time-to-quote; lower data entry errors |
| Submission triage | Underwriter reads and routes each submission manually | ML classification routes to correct queue; flags complex risks | Underwriter capacity freed from routine triage; faster response times |
| Risk assessment | Underwriter judgement supplemented by manual data lookup | Predictive models score risk; external data auto-enriches submissions | More consistent pricing; improved loss ratio on standard risks |
| Quote generation | Manual calculation; manual document generation | STP for standard risks; AI quotes with human approval for complex | 60 to 80 percent STP on eligible volume; significant capacity release |
| Portfolio monitoring | Quarterly portfolio reviews; lagging loss metrics | Real-time portfolio dashboards; predictive loss trend detection | Earlier detection of deteriorating segments; proactive appetite management |
| Fraud screening | Manual anomaly review; claims-side detection only | ML flags anomalies at submission; cross-reference with external data | Pre-bind fraud reduction; lower claims cost on fraudulent policies |
| Renewal underwriting | Annual renewal review cycle; manual re-pricing | Automated renewal scoring; flagging of changed risk conditions | Reduced renewal processing cost; better detection of risk deterioration |
The Data Infrastructure P&C AI Requires
The 62 percent of P&C insurers that are not generating value at scale from AI are predominantly constrained by data infrastructure gaps, not AI capability gaps.
Unified Data Architecture
P&C underwriting AI requires access to data that currently exists in multiple disconnected systems.
These include policy administration systems, claims management systems, rating engines, third-party data providers, broker portals, and internal underwriting guidelines.
AI models that score submission risk require several inputs at once:
- Policy data from the PAS system.
- Claims history from the claims system.
- Property data from external providers.
- Underwriting guidelines from internal knowledge management.
Integrating these sources into a coherent data layer accessible to AI models is the foundational data engineering challenge.
Unstructured Document Processing
P&C underwriting is fundamentally driven by unstructured documents. ACORD forms, broker submissions, inspection reports, prior loss statements, and business descriptions are the primary inputs to the underwriting process.
Extracting structured data from these documents requires NLP and LLM models trained specifically on insurance document types. Generic document extraction tools under perform on insurance-specific documents. The quality of extraction directly affects the quality of downstream risk assessment.
External Data Integration
Modern P&C underwriting risk models use a wide range of external data to supplement application information.
Sources include:
- Property characteristics from building data providers.
- Weather and catastrophe exposure from climate risk analytics.
- Geospatial data from satellite imagery.
- Business credit scores and financial stability indicators.
- IoT sensor data for commercial telematics.
Integrating these external sources into a unified data platform is a significant data engineering undertaking.
Each source has different APIs, different data models, different refresh frequencies, and different coverage across geographies and risk classes. Without proper integration, AI risk models are constrained to applicant-provided information, which is systematically less reliable than verified third-party data.
Model Governance and Explainability
Regulatory requirements in insurance create specific constraints on AI model use in underwriting.
Many jurisdictions require that underwriting decisions be explainable to the applicant. A declination or a pricing differential must be traceable to identifiable risk factors, not attributed to an opaque model output.
This creates two requirements:
- Model governance: AI underwriting models must be audited, validated, and monitored for discriminatory patterns.
- Explainability: Models must produce outputs that can be translated into underwriter-reviewable risk factor summaries, not black-box scores.
The data governance infrastructure that supports AI model governance is required infrastructure for compliant AI underwriting, not optional enhancement.
That includes training data lineage, model version control, monitoring for model drift, and output logging for audit.
The Organizational Challenge
BCG’s 10-20-70 framework for AI transformation is directly relevant here. (Source: Boston Consulting Group, “What It Takes to Achieve Full AI Potential in Insurance,” bcg.com, 2024)
Algorithms account for just 10 percent of the required effort. Technology and data account for 20 percent. The remaining 70 percent is people and processes. That 70 percent covers upskilling underwriters, redesigning workflows, managing change, and building the governance structures that allow AI to operate responsibly in a regulated context.
Experienced underwriters are legitimately concerned that AI removes the analytical work that makes the role intellectually engaging. BCG’s recommendation is to position AI as a tool that removes the most tedious aspects of the job. That includes manual document processing, data lookup, and routine submission handling.
Underwriters can then focus on complex risk assessment, broker relationship management, and portfolio strategy. Insurers that frame AI as a productivity amplifier for underwriters, rather than a replacement for underwriting expertise, tend to achieve higher adoption and better outcomes.
The best-performing implementations involve underwriters in model development, use AI outputs as inputs to underwriter decisions rather than replacing those decisions for complex risks, and provide underwriters with tools that make their expertise more valuable.
Final Thoughts
AI is transforming P&C underwriting. Predictive rating models are already universal. Submission automation and LLM-powered document processing are reaching production scale. STP is becoming standard for standard risks. The insurers generating value at scale are those that have redesigned their underwriting processes around AI capabilities.
They have also managed the organizational change, treating underwriters as partners in the transformation rather than subjects of it. For the 62 percent investing without yet generating returns at scale, the bottleneck is almost always data infrastructure and process redesign, not AI capability. The models are available. The data platforms and the process architecture are what make them work.
For insurance data teams building or modernising the data platforms, data pipelines, and model governance infrastructure that P&C AI underwriting requires, Data Pilot’s data engineering and strategy consulting helps insurers build the data foundations that make AI underwriting operationally reliable and regulatory-compliant.