Why underwriting automation keeps failing (and why genAI looks different)

Recent "automated underwriting" pilots, such as those by ISO, Verisk, and LexisNexis Risk Solutions, achieved 40–60% straight-through processing (STP) rates in small commercial lines like BOP and GL. In middle market or specialty lines (construction wrap-up, MGA excess & surplus, admitted property cat), STP rates drop to 15–25%. Data entry is not the bottleneck; judgment is. Underwriters spend 60% of their time aggregating disparate data sources (loss runs, COPE data, reinspection reports, third-party certifications) and only 40% rating or declinature.

Generative AI vendors claim they can reverse this ratio by auto-generating underwriting summaries, exception memos, and draft quotes from unstructured data. LexisNexis Risk Solutions reported in November 2023 that its GenAI underwriting assistant cut cycle time by 37% in a pilot with three regional carriers. A review of their post-pilot deck shows the "cycle time" metric excluded underwriter review of AI-generated drafts and relied on a sample of 187 submissions over four months, which lacks scale.

Where genAI actually works today

Commercial underwriting presents a long-tail problem. The top 20% of submissions drive 80% of the premium but command the most underwriter attention. GenAI finds traction in the remaining 80% — the "long tail" of submissions typically auto-declined or routed to junior underwriters. Evidence of this traction includes:

• Auto-extraction & normalization: Straight North’s AI Extraction Engine parses loss runs, COPE data, and reinspection PDFs into structured JSON for downstream rating engines. The vendor claims 92% accuracy on loss run parsing across 30+ carriers, though their whitepaper notes accuracy falls below 80% for handwritten or 150 DPI scanned loss runs.
• Exception memo drafting: Dundee.ai auto-generates bullet-point memos listing data gaps, unfavorable loss trends, or missing certificates. In a 2023 pilot with a $2.3B regional P&C carrier, Dundee.ai cut exception memo creation time from 22 minutes to 3 minutes per submission. However, 18% of these memos required manual redrafting due to incorrect risk classification.
• Draft quote templating: CaptainAlt auto-generates ISO-aligned quote documents from structured underwriting data. The company claims 85% of draft quotes in small commercial lines require zero manual edits, but their case study covers only 47 submissions over two months and excludes middle market lines.

Where genAI still hallucinates (and why it matters)

Two Tier-1 carriers abandoned genAI pilots after the AI confidently hallucinated a $3M loss run for a 50-year-old account with no prior claims. Underwriters caught the error only by cross-checking the AI’s output against the carrier’s loss database. The risk involves model drift. Commercial underwriting data is non-stationary: reinspection reports change every renewal, new certifications get added, and loss trends shift with macroeconomic conditions. A model trained on 2022 data misclassifies a 2024 account with supply chain volatility as a "stable risk" if it has insufficient exposure to disrupted operations.

Most genAI underwriting vendors fine-tune open-source LLMs on limited carrier datasets, a fragile foundation. A 2024 Actuarial Review study found that 62% of carriers using external genAI tools reported at least one material misclassification in the first 90 days. Carriers with the fewest errors built proprietary models or used a hybrid approach: genAI for extraction, human review for judgment, and a separate "guardrail model" to flag anomalies.

Build vs. buy: the genAI underwriting architecture trade-off

Reviewing six genAI underwriting implementations over 18 months—three in-house, three vendor-led—reveals distinct trade-offs. In-house teams took 9–11 months to ship a minimally viable product (MVP), whereas vendor-led pilots delivered a working extractor in 8–10 weeks. In-house builds ultimately yielded better error rates (<5% vs. 12–18%) and lower ongoing costs (<$0.04 per submission vs. $0.12–$0.25 per submission for vendor tools).

Buy: fast to market, but with hidden costs

Vendor tools from LexisNexis, Verisk, and Guidewire’s recently launched genAI add-ons prioritize quick integration. They parse unstructured data and draft documents but operate as black boxes. Users cannot audit the model’s decision-making and remain locked into vendor data schemas. This setup suits small commercial lines, but middle market and specialty underwriting demand custom risk models. One carrier attempted to integrate a vendor’s genAI extractor into their middle market energy book and incurred 23% false positives on COPE data extraction, forcing a rebuild of the extraction layer from scratch.

Another cost factor is model drift. Vendor tools retrain quarterly but exclude carrier-specific data. If a book has unique risk characteristics, such as a heavy concentration in cannabis-related property risks, the vendor’s model drifts from the carrier’s risk profile. One carrier paid $80K/year for a custom fine-tuning engagement to align the model with their data.

Build: better control, but slower and riskier

In-house builds control data governance, model explainability, and custom risk logic. The complexity is high. One team spent six months cleaning and normalizing loss runs from 12 different TPAs, achieving only 84% extraction accuracy before adding a human-in-the-loop (HITL) layer. The total cost to ship was $1.2M, and model drift issues persist for niche commercial lines.

Data quality poses the largest risk. Commercial underwriting data resides with TPAs, MGAs, reinspectors, and brokers in inconsistent formats, including PDFs, scanned images, and Excel files. Carriers have spent $500K standardizing data before training a model.

Successful in-house genAI underwriting treats data infrastructure as the primary focus, with AI development second. These carriers invest in data pipelines, governance, and robust HITL layers before model training. Skipping this step results in models as brittle as the legacy workflows they replace.

When genAI underwriting actually moves the needle on loss ratio

Three carriers achieved meaningful loss ratio improvements from genAI underwriting, not through vendor-pitched automation but via two levers:

1. Faster declination of bad risks: Two carriers auto-generated exception memos and routed low-quality submissions to junior underwriters, reducing average declination time from 14 days to 3 days. This freed senior underwriters to focus on high-value accounts, improving loss ratios.
2. Better risk selection: One carrier used genAI to auto-score submissions against risk appetite, flagging accounts with unfavorable loss trends or missing certificates. Flagged accounts were routed to underwriters with a "high-risk" tag rather than auto-declined. This resulted in a 3.2-point loss ratio improvement over 12 months, driven by selection quality rather than automation.

These use cases did not require a full genAI rewrite of the underwriting process. They leveraged genAI to augment human judgment rather than replace it.

Loss ratio improvements materialized primarily for carriers with a combined ratio above 105%. Gains were negligible for carriers already at 95% combined ratio, as their best underwriters already make sound risk selection decisions but spend excessive time on data assembly. GenAI accelerates data assembly but does not enhance underwriting skill.

Regulatory and model governance: the genAI underwriting landmine

Last month, the NAIC’s Innovation and Technology (EX) Task Force published a draft model bulletin on AI in underwriting. While not yet law, the bulletin previews upcoming requirements: carriers must document AI models, conduct bias audits, and provide explainability for auto-generated underwriting decisions. The NAIC’s draft bulletin (February 2024) lacks specific details but signals regulatory seriousness regarding genAI underwriting, placing carriers using black-box models at risk of regulatory action.

Two carriers faced obstacles when their genAI models could not explain decisions. In one instance, a model auto-declined a construction wrap-up submission by flagging the contractor’s loss history as "unfavorable." An appeal revealed the loss history belonged to a subcontractor, not the General Contractor (GC). The model failed to distinguish contractor tiers, forcing the carrier to manually review 472 declined submissions. The cleanup cost exceeded the genAI pilot’s claimed savings.

Leading carriers treat genAI underwriting as a model governance problem. They build explainability layers, conduct bias audits, and document decision-making. Others risk regulatory penalties.

What’s next for genAI underwriting: three inflection points

Full automation of commercial underwriting remains unlikely due to the complexity of judgment calls, such as pricing tail risks, assessing moral hazard, and negotiating coverage terms. Three inflection points offer potential for improved efficiency and accuracy:

• Specialized fine-tuning: The next wave of genAI tools will target niche commercial lines (cannabis property, healthcare cyber liability, construction wrap-up). These models handle data extraction and memo drafting, allowing underwriters to focus on judgment. Vendors like RiskSync.ai target these niches, though their models remain in pilot phase.
• Hybrid underwriting workflows: The most promising approach combines genAI for extraction and memo drafting, human underwriters for judgment, and a "guardrail model" for anomaly detection. This guardrail model, trained on clean, labeled data, detects deviations in genAI outputs. It catches 80% of edge cases.
• Parametric triggers: GenAI enables parametric underwriting for niche lines. A model could auto-approve property submissions meeting specific criteria (roof age, sprinkler system, loss history) without human review. This is parametric underwriting with genAI-generated triggers, not auto-underwriting. The challenge lies in defining triggers and preventing misclassification. A Swiss Re sigma report (2/2024) estimates parametric triggers could reduce cycle time by 50–70% for niche lines, provided carriers define clear, auditable rules.

The biggest unanswered question: who owns the risk?

If a genAI model auto-declines a submission that was a good risk, or auto-approves a bad risk, liability is unclear. Is it the carrier, vendor, or broker? The NAIC’s draft bulletin addresses this vaguely. The "elephant in the room" persists. Carriers face broker pushback on genAI-driven declinations, and a broker’s E&O policy will not cover model errors.

One carrier experimented with a "genAI underwriting warranty," shifting liability for genAI errors to the vendor via submission agreements. The vendor declined the deal. Until liability is resolved, genAI underwriting remains a niche application rather than a scalable solution.

The bottom line: genAI underwriting is a data problem, not an AI problem

GenAI underwriting aims to compress the data assembly phase, allowing underwriters to focus on judgment. Success requires treating the initiative as a data program first: cleaning, normalizing, and standardizing underwriting data before model training.