Why Carriers Need AI for Underwriting
Carriers face a compound underwriting problem that has worsened over the past several years. Submission volumes are growing faster than underwriter headcount. Loss ratios are drifting in targeted segments as external data signals -- climate exposure, litigation trends, supply chain costs -- change faster than traditional actuarial models can absorb. And regulators are beginning to scrutinize the algorithmic decisions that underwriters use to price and select risks.
AI underwriting tools do not replace underwriters. What they do is handle the parts of underwriting that are poorly suited to human judgment at scale: processing unstructured submission documents, cross-referencing external enrichment data, scoring risk against the carrier's historical loss experience, and routing submissions to the right underwriter based on complexity and appetite fit.
The AI underwriting market has matured considerably since the early experimental deployments of the mid-2010s. Today's platforms come with pre-built integrations, documented model explainability frameworks, and implementation track records at carriers that allow new adopters to evaluate real-world results rather than vendor promises. The question for most carriers is no longer whether AI underwriting tools deliver value -- it is which tools fit their lines of business, data environment, and regulatory posture.
For a broader view of how predictive underwriting works in practice, see the glossary entry on predictive underwriting and the companion piece on what is underwriting for foundational context. The risk scoring entry explains the specific techniques most platforms use to translate raw data into a usable decision signal.
Key Use Cases and Workflow
Submission scoring and triage. The most common entry point is automated scoring of inbound submissions. A submission arrives -- typically a mix of structured ACORD data, PDFs, spreadsheets, and email context -- and the platform extracts structured data, enriches it with third-party signals, and produces a score reflecting both loss likelihood and appetite fit. High-confidence, low-complexity submissions route to straight-through processing or a streamlined underwriter workflow. Complex or marginal risks route to senior underwriters with supporting data already assembled.
Appetite operationalization. Carriers define their appetite in underwriting guidelines, but those guidelines are rarely precise enough to drive consistent decisions at the submission level. Platforms like Federato translate appetite parameters into routing logic and scoring thresholds, so that appetite decisions made in the home office actually govern what individual underwriters quote. Without this layer, underwriting guidelines exist on paper but are applied inconsistently in practice.
Portfolio-level monitoring. Underwriting decisions aggregate into a portfolio. AI tools that monitor portfolio composition -- concentration by geography, industry class, limit profile -- allow chief underwriting officers to identify segments where pricing is drifting before the next actuarial review catches it. This is the use case most directly tied to loss ratio improvement. Segment-level deterioration that once took an annual review to surface can be detected in near-real time.
Pricing model refinement. Tools like Akur8 are specifically designed to help actuarial teams build and deploy ML-based pricing models faster than traditional GLM methods allow. The platform produces models that are interpretable using actuarial methods, which matters for state filing. Actuaries can validate outputs against their own judgment rather than treating the model as a black box.
Underwriter decision support. Generative AI tools like Sixfold work differently from scoring platforms. They read the full submission -- including unstructured narrative -- and produce a structured summary with a recommended risk narrative. The underwriter gets a synthesized view of the risk in minutes rather than spending time reading raw documents. This use case addresses underwriter time cost rather than model-based selection.
Data enrichment for submissions. Planck and similar tools enrich sparse submissions with external signals about the insured business before scoring begins. This is especially relevant for small commercial lines, where applicants often provide minimal information and the underwriter must evaluate a risk with limited data. External enrichment signals -- industry classification, business operational data, online presence -- fill gaps that the submission itself does not cover.
What to Look For
Data requirements. Most ML-based loss prediction tools require a minimum volume of historical claims data -- typically several years of policy and loss history at the line-of-business level. Carriers expanding into new lines or geographies may find that their training data is insufficient for those segments and need to evaluate how vendors handle thin-data scenarios. Ask vendors specifically what they do when carrier data is insufficient, and whether they blend industry data or benchmarks into the model.
Model explainability. Explainable AI is not just a regulatory requirement -- it is also an operational one. Underwriters who cannot understand why a model scored a risk a particular way will not trust the score, and an untrusted score gets overridden. Look for platforms that produce contributing-factor explanations at the individual submission level, not just aggregate model statistics. Regulators in Colorado, California, and Washington are currently the most active in scrutinizing AI underwriting decisions, and explainability requirements are likely to expand.
Policy admin integration. An AI underwriting tool that requires manual data re-entry into the policy administration system creates friction that limits adoption. Evaluate integration depth with your specific policy administration system, whether that is Guidewire PolicyCenter, Duck Creek, or another platform. API-based integrations that pass scored submissions directly to the policy admin workflow produce better adoption than standalone portals.
Implementation timeline. Realistic AI underwriting implementations run 6-18 months from contract to production, depending on data readiness, integration complexity, and organizational change management. Carriers that have underestimated this timeline have found themselves with delayed go-live dates and underwriter adoption challenges. Data preparation -- cleaning historical claims data, normalizing policy records, establishing data pipelines from core systems -- is typically the longest phase.
Risk scoring methodology. Different platforms take different approaches -- gradient boosting on historical loss data, network analysis of third-party enrichment signals, generative AI for document processing. The right methodology depends on the use case. Understand what each vendor is actually doing before evaluating score quality, and ask for validation methodology documentation rather than relying solely on vendor-provided accuracy statistics.
Gradient AI
Gradient AI uses ML-based loss prediction for commercial lines underwriting. The platform ingests historical claims data, enriches it with external signals, and produces loss probability scores with contributing factors that explain the score at the submission level. It is used by both carriers and MGAs, and is particularly strong in workers compensation, commercial auto, and general liability. Pricing is quote-based.
For a direct comparison of approaches, see Gradient AI vs. Planck. The two platforms address related problems -- Gradient AI focuses on loss prediction from historical data, Planck focuses on real-time external enrichment -- and are often evaluated together for small commercial use cases.
Federato
Federato is a portfolio-level underwriting management platform rather than a pure scoring tool. It handles submission intake, triage routing, and appetite operationalization -- the workflow layer that sits above individual risk scoring. Many carriers deploy Federato alongside a dedicated scoring tool like Gradient AI: Federato manages the workflow and routing logic, while the scoring tool provides the loss probability signal. They are more complementary than competitive, though Federato has some scoring capabilities and Gradient AI has some workflow features.
See the Federato vs. Gradient AI comparison for a breakdown of how the two platforms complement or compete depending on a carrier's specific needs.
Akur8
Akur8 is an actuarial pricing platform that uses ML to augment traditional ratemaking rather than replace it. The platform uses gradient boosting to identify rate inadequacy at the segment level and produces models that are interpretable through actuarial methods, which matters for state rate filings. It is strongest in personal lines pricing modernization -- carriers moving away from legacy GLM models to more granular, frequently updated rate plans. Actuarial teams use Akur8 alongside their existing ratemaking workflow rather than replacing it. Pricing is quote-based.
Cytora
Cytora focuses on risk digitization and submission routing. It converts unstructured submissions -- PDFs, emails, spreadsheets -- into structured data, then routes by carrier appetite. The approach differs from Federato's workflow platform and Gradient AI's scoring engine, though the three tools address related problems. Cytora is often evaluated by carriers whose primary pain point is the cost of converting unstructured submissions to structured data before any scoring can begin. See the Cytora vs. Federato comparison for specifics. Pricing is quote-based.
Sixfold
Sixfold uses generative AI to read commercial submissions and produce structured summaries and risk narratives. Rather than producing a numeric score, Sixfold synthesizes the submission -- including unstructured narrative documents, loss runs, and supplemental applications -- into a concise briefing for the underwriter. The use case is reducing the time underwriters spend reading raw documents before they can make a decision. For the comparison between Planck and Sixfold on small commercial, see Planck vs. Sixfold. Pricing is quote-based.
Planck
Planck specializes in small commercial business intelligence. The platform enriches submissions with real-time web signals and third-party data about the insured business -- industry classification, online presence, operational signals -- to supplement what the insured reports on the application. It is particularly useful for small commercial lines where underwriters have limited time per account and the submission data itself is often sparse. Pricing is quote-based.
Zelros
Zelros provides AI recommendations for insurance product distribution and underwriting. The platform is stronger in the European market where it has more carrier deployments, but it is growing its US presence. Use cases include underwriting recommendation engines and next-best-action tools for distribution. Pricing is quote-based.
Hyperexponential (hx Renew)
Hyperexponential is a pricing platform for complex and specialty commercial lines, including excess, casualty, and specialty. It supports underwriter-led pricing decisions with ML model guidance -- the underwriter retains judgment while the platform provides data-driven pricing signals. It is one of the stronger options for carriers in the London market and US specialty segments where underwriter judgment is central to pricing. Pricing is quote-based.
