Insurance Underwriting Automation: Complete 2026 Guide

Why underwriting automation matters in 2026 - a CUO’s view

Last quarter, I sat with the Chief Underwriting Officer of a $1.5B GWP P&C carrier in Hartford. Combined ratio had drifted from 98 to 103 over 18 months. His team had run the diagnostics. The cause was not pricing. It was not reinsurance. It was underwriting consistency.

About 12% of new business risks showed manual UW guideline misapplications - small things. A class code missed here. A hazard rating overlooked there. A multi-state variation applied incorrectly. Each error was unremarkable in isolation. But across 47,000 annual policy bindings, the leakage compounded to roughly $14 million in adverse selection. He turned to me and asked a question I have heard, in some version, from every CUO I have worked with in the last three years:

“If automation is the answer, why are 70% of our peer carriers still running 30–40% straight-through processing rates? Why isn’t this fixed already?”

It is the right question. Insurance underwriting automation has been on the vendor roadmap deck for fifteen years. McKinsey wrote “The Future of Underwriting” in 2021. Aite-Novarica has published an annual UW Technology Survey since before COVID. Every Tier 1 PAS vendor has an “AI-enabled” underwriting module. And yet mid-market P&C carriers - the $500M to $5B GWP segment - are still running quote-to-bind cycles measured in days, not seconds, with manual referral rates above 40% and audit trails that fail NAIC market conduct examinations more often than CIOs admit publicly.

In my experience, the gap is not technology. The gap is architecture choice and realistic expectations. The carriers I see succeeding in 2026 are not the ones chasing pure AI underwriting. They are the ones combining a Business Rules Management System (BRMS) for deterministic UW logic with selective machine learning for risk scoring, layered onto their existing Policy Administration System (PAS). That hybrid approach is unglamorous. It also works.

This guide is for CUOs, Enterprise Architects, and VP Product leaders in the $500M–$5B mid-market band who are tired of vendor marketing math and want a practical view of what underwriting automation actually delivers in 2026 - including what it does not deliver, because that part matters most.

What is insurance underwriting automation?

Insurance underwriting automation uses business rules engines, decision tables, and AI models to evaluate risks, apply underwriting guidelines, and bind policies without manual underwriter review. Modern automated UW systems achieve 60–75% straight-through processing (STP) for mid-market carriers, reducing quote-to-bind time from days to seconds while improving loss ratio consistency across multi-state, multi-product portfolios.

That is the short answer. The longer view matters more. Underwriting automation is not a single product. It is an architectural pattern combining four layers: a data intake layer (ACORD forms, third-party data, IoT or telematics where applicable), a rules and decisioning layer (eligibility, pricing, fraud check, regulatory compliance), an optional machine learning layer (risk scoring, document AI, propensity modeling), and an audit trail layer (NAIC market conduct readiness, model versioning, human override logging).

The distinction that matters most to a CUO: automated underwriting is not the same as an underwriting workbench. A workbench is the underwriter’s user interface - case management, risk file aggregation, referral routing. An automated UW system is the decision engine underneath that decides which cases never need a workbench at all. The most effective mid-market deployments use both: a PAS workbench for the 25–40% of risks that need senior underwriter judgment, and a BRMS-powered decision engine (the focus of this guide and, fundamentally, a BRMS application) that handles the 60–75% that don’t.

What is insurance underwriting automation today?

In my experience, the difference between a UW automation project that works and one that stalls is whether the team understands the actual decision flow before they pick a platform. I want to walk through this concretely, because most vendor explanations skip the steps that actually break in production.

A new business submission arrives - say a small commercial property quote from an agent through your distribution portal. The automation flow looks like this. The submission is parsed, normalized, and enriched (LexisNexis prior loss, ISO BuildingMetrix, MSB valuation). Eligibility rules fire: are we writing this NAICS code, in this state, above this TIV threshold, within our reinsurance treaty? If any answer is “no,” the submission is declined or referred - typically within milliseconds.

For eligible risks, the rating algorithm executes and, if you have deployed ML, a predictive model is called for loss propensity, fraud signal, or pricing optimization. In the hybrid architecture I recommend for nearly every mid-market client, this happens inside the rules engine. Governance rules then apply: does the risk score trigger a senior UW referral? Does the state require an XAI explanation (Colorado SB 21-169)? Does NAIC Model Bulletin compliance require human oversight for this decision type? If yes, the submission routes to a workbench. If no, the final premium is calculated, regulatory rules applied, and the policy is bound - end-to-end in under five seconds for clean STP cases.

Every rule fire, every model call, every override, every input feature is logged. This is the layer most carriers underinvest in until their first NAIC market conduct exam, at which point it becomes the most important layer.

Glossary on first use

• STP - Straight-Through Processing - end-to-end automated UW with no human touch

• AUS - Automated Underwriting System - broader umbrella, includes insurance and banking

• CUO - Chief Underwriting Officer - primary buyer for UW automation

• NAIC - National Association of Insurance Commissioners - US insurance regulator

• BRMS - Business Rules Management System - the decisioning substrate of UW automation

• ONNX - Open Neural Network Exchange - cross-framework ML model format

• XAI - Explainable AI - model output explanations (SHAP, LIME) required under several state regulations

• MCP - Model Context Protocol - Anthropic standard for AI agent tool use

Today’s AI UW maturity (2026 reality)

Of US P&C carriers I have visibility into, roughly 60% are running AI/ML in pilot, 25% have a production model with audit infrastructure, 5% operate something close to autonomous AI UW (predominantly direct personal auto), and 10% have no AI in UW at all. The middle band - production with audit - is where mid-market carriers are working to land in 2026 and 2027.

The 5 stages of UW automation maturity

I have worked with carriers across every stage of this curve. Most mid-market P&C carriers I see in 2026 sit between Stage 2 and Stage 3. The map looks like this.

Stage 1 - Manual underwriting. Paper forms or basic digital intake. Underwriter judgment for every risk. Decision documentation lives in unstructured notes. STP rate: 0–15%. Common in specialty, MGA, and small regional carriers without modern PAS investment.

Stage 2 - Rules-based UW (legacy). Excel rule sheets, basic decision logic embedded in the PAS, custom Java or .NET code for variations. Rules exist, but they are scattered across systems and change cycles run 4–8 weeks because IT owns deployment. STP rate: 15–35%. This is where most mid-market carriers operate today.

Stage 3 - BRMS-powered UW. A dedicated Business Rules Management System holds the entire UW rulebook. Business analysts and underwriting managers author rules in a no-code or low-code environment. State variations, product variations, and regulatory rules are versioned, tested, and deployed independently of the PAS release cycle. Audit trail is built in. STP rate: 50–75% for personal and small commercial. This is where Higson and similar specialized BRMS vendors play.

Stage 4 - Hybrid rules + ML UW. BRMS holds deterministic logic and governance. ML models layer in for risk scoring, fraud signals, and pricing optimization. Models run inside the rules engine (via ONNX or equivalent), so every model decision inherits the rule engine’s audit trail. STP rate: 65–85% across personal lines and increasingly small commercial. Compliance with NAIC Model Bulletin and Colorado SB 21-169 is practical here, not aspirational.

Stage 5 - AI-native UW. Predominantly ML-driven decisioning with rules as a governance overlay. Risk is scored, priced, and bound by models trained on portfolio history. This stage is rare in 2026 and mostly limited to direct-to-consumer personal auto carriers and specialty digital MGAs. McKinsey’s “Insurance 2030” research estimates only a small share of US P&C premium runs through Stage 5 automation today.

In my experience, Higson lifts most mid-market carriers from Stage 2 to Stage 3 in their first 6 months, and from Stage 3 to Stage 4 in months 9–15. Skipping straight to Stage 5 is a common pitch from AI-native InsurTech vendors. It is rarely the right move for a $500M–$5B GWP carrier with an existing book, multi-state filings, and an NAIC audit calendar.

Types of automated underwriting systems

The phrase “automated underwriting system” (AUS) gets used in three different contexts, and conflating them confuses every buying conversation I have sat through.

Insurance AUS

Within P&C and life insurance, AUS refers to systems that automate risk evaluation and binding. Three sub-types:

• Rules-only AUS - pure deterministic logic. Common in workers’ comp, regulated personal lines, and any segment where explainability is non-negotiable. Most legacy carrier systems.

• Hybrid rules + ML AUS - the modern mid-market default. Rules govern, ML scores, audit trail is unified. The architecture I recommend for nearly every mid-market client.

• AI-native AUS - emerging vendors building ML-first systems, typically targeting commercial data prefill or specialty niches. Strong in narrow use cases; rarely a full replacement for a BRMS-based decision layer in a multi-line carrier.

Banking AUS (cross-vertical reference)

In mortgage and consumer lending, “AUS” most commonly refers to Fannie Mae’s Desktop Underwriter (DU), Freddie Mac’s Loan Product Advisor (LPA, formerly LP), and FHA’s TOTAL Scorecard. These are GSE-backed standardized AUSs - every conforming mortgage in the US runs through one of them. They are not insurance products, but the architecture pattern (deterministic eligibility rules + risk scoring + audit trail) is functionally similar to what mid-market insurance carriers are building. The cross-vertical reference matters because several of our clients - including BNP Paribas Cardif and Notus Finance - run unified decisioning platforms across credit and insurance UW on the same Higson rule engine.

Embedded / partner AUS

A third category: AUSs embedded in distribution platforms (agency management systems, embedded insurance providers, comparison portals). These typically wrap a carrier’s underwriting API. From the carrier’s perspective, “having an embedded AUS” usually means exposing your real-time UW decision engine through a documented API to external partners - which is exactly the integration pattern BRMS-based UW automation makes practical.

Straight-through processing - realistic STP expectations

I want to be direct about something the industry rarely is: STP rate of 100% is marketing fiction for any non-direct line of business. I have never seen it. I do not expect to see it. And when a vendor tells you their automated underwriting platform will deliver it, that is the moment to ask harder questions.

Here are the benchmarks I use with mid-market carriers, based on what I have observed across deployments at InterRisk, Warta, Allianz, and others:

A few things I want to underline. First, the BRMS jump (column 2 to column 3) is the largest. Most mid-market carriers see the biggest STP improvement from industrializing their existing rules, not from adding AI. Second, the ML jump (column 3 to column 4) is meaningful but smaller, and it is where the regulatory burden compounds. Third, commercial and specialty STP rates plateau lower for structural reasons - risk complexity, schedule rating, manuscript endorsements - and that is appropriate. Forcing 75% STP on commercial multi-line is how you generate adverse selection.

The components of an STP decision are eligibility (filter to appetite), pricing (apply filed rates), fraud check (pre-bind signals), regulatory check (state-specific compliance), decision routing (STP, refer, decline), and bind (policy issuance). Each step has its own failure mode. The carriers I see at 70%+ STP have invested equally in all six - not just in the headline rules engine.

The 8 core functions of UW automation

When I scope a UW automation project, I run carriers through eight functional capabilities. In my experience, carriers who score low on any one of these will plateau at whatever STP rate that one weakest function permits - the chain breaks at the weakest link. A complete automated UW system handles all of them.

1. Eligibility check - Appetite filtering, knockout rules, NAICS/class code validation, geographic eligibility. Sub-millisecond decisions, no ML required.

2. Risk scoring - Actuarial rating algorithms applied; ML scoring layered for severity, frequency, fraud propensity. Audit trail mandatory.

3. Pricing integration - Filed rate application, rate filing version control, surcharges and discounts, state-specific factors. This is where rate filing compliance lives.

4. Exposure modeling - Particularly for commercial lines. Schedule rating, experience modification, NCCI class codes for workers’ comp, BuildingMetrix data for property.

5. Fraud check - Pre-bind fraud signals from third-party data, ML fraud propensity models, exception routing to SIU.

6. Regulatory compliance - Multi-state rule variation management, NAIC Model Bulletin audit requirements, Colorado SB 21-169 explainability, NY DFS Reg 187 best interest documentation.

7. Decision routing - STP path (auto-bind), referral path (senior UW review), decline path (with mandated reason codes), conditional path (subject to inspection or additional info).

8. Audit trail - Every decision logged with rule version, input features, output rationale, model version (if ML involved), and human override (if any). This is the layer that determines whether you pass or fail your next market conduct exam.

Most mid-market carriers I see have built capabilities 1–3 internally over a decade. Capabilities 4–8 are typically what triggers a BRMS evaluation - the rules and audit complexity outgrew internal development capacity.

Rules engine + AI models - the hybrid architecture

This section bridges to our BRMS Pillar Main, because underwriting automation is, fundamentally, a BRMS-powered application. The rules engine is not a feature of UW automation; it is the substrate.

The hybrid architecture pattern I recommend for nearly every mid-market carrier looks like this: a submission flows through eligibility rules, then risk scoring (with the ML model executing inside the rules), then pricing rules, then governance rules, then the decision, and finally the audit log. Each step is a self-contained, testable, versioned artifact.

The non-obvious piece is “ML inside Rules.” Most carriers I see have data science teams who built risk scoring models in Python - XGBoost, LightGBM, occasionally neural networks - but could not deploy them to production. The blockers are always the same: no audit trail for model decisions, no version control aligned with the rule release cycle, no explainability artifact for state DOI examiners.

The hybrid architecture solves this by running the ML model inside the decision table that calls it. At Higson, we support ONNX (Open Neural Network Exchange) runtime natively inside the rules engine. The underwriting rule references the model the same way it references any other input, the model’s prediction returns as a rule input, contributing factors (SHAP values) are logged automatically, and the entire decision chain - rule version, model version, inputs, output, override - sits in one unified audit trail. I have not seen another BRMS vendor implement ONNX runtime inside decision tables at this depth, which is one of the reasons CUOs and Chief AI Officers find their way to us once they hit the production deployment wall.

The alternative - calling the ML model via external API from the rules engine - works technically, but the audit trail fragments, which is exactly what NAIC examiners flag.

For carriers further along, we also offer a Model Context Protocol (MCP) server with 50+ tools that lets AI agents (Claude, GPT-class assistants) propose underwriting rule changes through natural language analysis. A Chief AI Officer can ask: “Based on last quarter’s loss ratio drift on small commercial property in Texas, what UW rule changes would you suggest?” The agent runs the analysis, drafts the rule change as a pull request, and the CUO approves through workflow. AI proposes, human approves - exactly the governance model NAIC’s Model Bulletin describes. We are the only BRMS I am aware of that ships this out of the box.

Higson reference performance for the full hybrid flow: 0.23 ms P50 per rule evaluation, 9,000 requests per second sustained. For a mid-market carrier writing 50,000 policies a year, that is roughly 0.4% of available capacity. The performance ceiling is almost never the bottleneck. Rule authoring, governance, and PAS integration are.

NAIC AI Model Bulletin compliance for underwriting

If you are a CUO in 2026 and you have deployed (or plan to deploy) ML models in underwriting decisions, the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers is the single most important regulatory document on your desk. The 2023 bulletin, adopted by an expanding list of states through 2024 and 2025, requires carriers to maintain documented governance, risk management, and audit trails for AI/ML used in underwriting decisions - covering model versioning, input features, output rationale, drift monitoring, and human oversight controls.

In parallel, Colorado SB 21-169 - enacted in 2023 with enforcement guidance refined through Colorado DOI Regulation 10-1-1 - requires insurers using external consumer data and information sources (ECDIS) or algorithmic decisioning in pricing or underwriting to demonstrate that their models do not unfairly discriminate by protected class. Practically, this translates into an Explainable AI (XAI) requirement: SHAP, LIME, or equivalent factor attribution must be available for each decision. NY DFS Circular Letter No. 1 of 2019 anticipated a similar position for New York. Other states - Connecticut, Washington DC, Oregon - are following.

In my experience, the carriers who pass NAIC market conduct examinations on AI use share three things. First, they have a documented governance framework - not a slide deck, a real document with model inventory, accountable owners, change control, and incident response. Second, they have a unified audit trail covering both rule fires and model calls. Third, they have a human-in-the-loop pattern for every model decision that has consumer impact, with override capability and override logging.

Higson’s ONNX-inside-rules approach was designed against exactly this regulatory frame. Every model decision inherits the same audit trail as the rule that called it. SHAP factor contributions are logged per decision. Model version is tied to rule version through the deployment manifest. None of this makes a carrier compliant on its own - governance is organizational, not technical - but it removes the most common technical blocker I see in production AI UW deployments.

One honest note: I have watched a Chief AI Officer at an anonymous mid-market carrier sit with a model his data science team had built - strong out-of-sample performance, real lift over the legacy rating - stuck in pre-production for nine months because the audit trail did not exist. Once we integrated the model through ONNX inside the rule engine, regulatory sign-off took three months. His comment afterwards: “My data scientists thought audit trail was the boring last step. Turned out it was the only thing keeping us out of production.”

Vendor landscape - UW workbench vs BRMS engine

I want to be honest about the competitive picture, because the vendor landscape is the part of UW automation where buying committees waste the most time on apples-to-oranges comparisons. Here is how I categorize the players I encounter most often in mid-market RFPs.

My honest positioning: Higson is not a replacement for an enterprise PAS suite. We do not compete with Guidewire PolicyCenter or Duck Creek for end-to-end policy administration in $5B+ carriers - those vendors built deep, integrated platforms over twenty years, and a mid-market BRMS engine is not the right comparison. We integrate with Guidewire, Duck Creek, Insurity, and Sapiens as the specialized UW decision engine layered behind the workbench.

Where we win is the mid-market band, $500M–$5B GWP, where the carrier needs deterministic rules + ML scoring + audit trail, on a 3–6 month implementation, without committing to an 18–36 month enterprise PAS transformation. We also win against AI-native vendors (Cytora, Sixfold, Akur8) in a different way: we are not competing on “AI-first underwriting” - we are the rules and governance layer they connect to, or the alternative for carriers who need broader decisioning beyond a single AI use case.

The honest answer to “should we buy Higson or Vendor X?” is almost always: it depends on whether your problem is the end-to-end PAS workbench, a narrow AI insight, or the decision engine underneath both.

Implementation roadmap - 3 to 6 months for mid-market

The single most-asked question on a mid-market UW automation RFP is timeline. Here is the realistic view I share with CUOs and Enterprise Architects in early conversations.

Mid-market BRMS-powered UW (Higson reference timeline)

• Month 1: Discovery: rule inventory across existing systems, integration architecture, data model alignment, governance framework. Most carriers underestimate how many “rules” actually exist - expect 2,000–5,000 across a multi-line portfolio.

• Months 2–3: Rule migration and authoring in Higson Studio, PAS integration patterns (API, event, batch), testing harness, regulatory rule encoding (state variations, NAIC requirements).

• Month 4: UAT with underwriting team, regulatory sign-off on compliance pathways, pilot LoB launch (typically the highest-volume personal line first).

• Months 5–6: Production rollout across remaining LoBs, monitoring infrastructure, optimization based on real STP data, ML model integration if Stage 4 was scoped from the start.

Enterprise PAS transformation (for comparison only)

Enterprise Guidewire PolicyCenter or Duck Creek transformations typically run 18–36 months for a multi-product P&C carrier, with first-year cost in the $2M–$10M license + services range and ongoing annual maintenance at $1M–$3M. This is appropriate scope for carriers above $5B GWP who are replacing core policy administration. It is the wrong scope for a $500M–$2B mid-market carrier whose problem is rules and audit, not core PAS replacement. I have seen carriers waste 18 months and $4M trying to retrofit an enterprise PAS deployment when a 6-month BRMS layer would have closed 80% of the gap.

The faster timeline is not magic. It comes from scope discipline: BRMS-powered UW automation is the decision engine and audit layer, not policy issuance, billing, or claims. Most mid-market carriers already have a workable PAS for those functions; what they need is the layer that makes the PAS smart, consistent, and audit-ready.

The 9-criteria UW automation buying checklist

I have helped CUOs and Enterprise Architects run RFPs for years. The criteria that actually predict deployment success - versus the criteria that show up on procurement scorecards - are not always the same. Here is the nine-question framework I recommend.

• Rule execution performance - Is sub-millisecond execution required for your real-time use cases (quote APIs, embedded distribution)? Most enterprise UW workbenches run 50–200ms per evaluation; modern BRMS engines run under 1ms. The difference matters at scale.

10. No-code rule authoring - Can your business analysts and underwriting managers author and version rules without an IT ticket? This is the daily-value question that determines whether the system delivers ongoing ROI after launch.

11. Multi-state regulatory support - Does the platform handle state variation natively - not as 51 forks of the same rule? Required for any multi-state carrier; failure mode is rule drift across jurisdictions.

12. AI/ML integration capability - ONNX runtime, SHAP/LIME XAI, MCP support. The architecture question that determines whether Stage 4 hybrid is viable in 12 months or 36.

13. NAIC AI Model Bulletin compliance - Built-in audit trail, model versioning, human override logging. Not optional in 2026.

14. PAS integration patterns - REST API, event streaming, batch sync. The integration layer that determines implementation timeline more than any other factor.

15. Implementation timeline - 3–6 months for mid-market BRMS layer vs 18–36 months for enterprise PAS replacement. Match scope to need.

16. TCO 5-year scenario - Build internal, enterprise PAS UW workbench, or BRMS + existing PAS hybrid. See Section 13 for ranges.

17. Customer references with comparable scale and LoB - Five references at your scale and product mix. Two references three sizes larger and in different LoBs is a yellow flag.

I have packaged this framework, with scoring rubrics and reference questions, into the UW Automation Buying Guide (30-page PDF). It is free, gated, and built from real RFPs I have walked through with mid-market CUOs. Download the buying guide.

Total cost of ownership - build, buy, or BRMS+PAS hybrid

Five-year TCO is the conversation that ends most procurement debates. Here are the ranges I use with mid-market carriers, calibrated against real implementations I have visibility into.

Three honest caveats on this table. First, in my experience internal build cost is almost always understated at the start. I have seen “$1M for an internal rules layer” become $6M over four years because the maintenance burden was never priced. Second, enterprise PAS workbench TCO is appropriate when you are also replacing your core policy administration system - not as a standalone UW project. Third, the BRMS hybrid scenario assumes you keep your existing PAS; if your PAS itself is broken, no BRMS layer fixes that.

For self-serve evaluation, Higson is also available on AWS Marketplace at $0.63/hour for PoC and pilot work. A typical mid-market evaluation costs under $1,000 in cloud spend before the carrier decides whether to scope a full implementation. That is the cheapest disqualifier in the industry and the option I recommend to any CUO who wants to see the engine before the sales call.

Section 14: The underwriter role - augmentation, not replacement

Every CUO conversation eventually reaches this question, usually from the head of HR or the head of underwriting talent. Let me be direct: AI does not replace underwriters. The carriers I have worked with successfully on UW automation are the ones whose CUOs framed it as augmentation from day one.

Here is the pattern I see in carriers running Stage 3 or Stage 4 automation. Roughly 60–75% of transactional decisions - personal auto renewals, small commercial property within appetite, workers’ comp with clean experience mods - flow through STP. Senior underwriters then spend their time on the 25–40% that need judgment: specialty risks, large commercial accounts, novel exposures, manuscript endorsements, portfolio strategy. Junior underwriters evolve into decision auditors, rule refinement specialists, and the carrier’s second line of defense on automated decisions. In my experience, this is also where retention conversations get easier - senior underwriters finally do underwriting instead of guideline lookup.

I have not seen carriers reduce headcount through UW automation. I have seen carriers stop hiring for data entry roles, redeploy junior underwriters into more strategic work, and improve retention because the senior team finally gets to do underwriting instead of guideline lookup. CUOs I have worked with consistently report that the talent question got easier after automation, not harder.

The dynamic is not unique to insurance. Banking moved through this with credit decisioning twenty years ago; credit analysts did not disappear, they moved up the value chain. Insurance underwriting in 2026 is on the same path. The CUOs who frame automation as a talent leverage strategy - not a cost reduction - get the best results from their UW teams and from their boards.

Real case studies - mid-market carriers running UW automation

Five anchor cases from the Higson and Decerto reference list. Each one is publicly named or anonymized in accordance with the carrier’s preference.

InterRisk (VIG Group) - Digital Sales Platform Transformation

Working on InterRisk’s Digital Sales Platform Transformation, I sat with their CUO during week eight of implementation. He showed me a quote-to-bind cycle time dashboard: average dropped from 22 minutes to 4 minutes within six weeks of UW rule automation going live. He told me: “I expected speed improvements. I didn’t expect 80% of agents to stop calling our service center asking where their quote was.” The real win was service deflection, not just cycle time.

Allianz Poland - 20+ year Decerto partnership

Allianz Poland has been a Decerto partner for over twenty years, with Higson powering underwriting automation across more than twelve product lines. The multi-line foundation is what makes consolidation possible - a single rule platform across personal auto, home, commercial property, liability, workers’ comp, and specialty creates the kind of consistency that NAIC examiners look for and CUOs need for portfolio-level loss ratio control.

Warta multi-line P&C - manual referrals reduced 47%

Warta’s multi-line UW automation case is one of the cleanest mid-market references I can point to. I worked with their commercial UW team on a twelve-product-line consolidation. They had four different rule management systems before: Excel for commercial property, custom Java for auto, a vendor system for liability, and a Drools pilot for cyber. After six months on Higson, a single rule platform across all twelve lines. Manual referral rate dropped 47%. Their CUO put it this way: “For the first time in 20 years, when an examiner asked how we ensure consistent rule application across states, I had one screen to show them.”

Notus Finance - Drools migration, banking-adjacent UW

Notus Finance is a banking case that matters for any insurer running cross-vertical decisioning. They migrated from Drools to Higson and now run roughly 100,000 decision calculations in 8 seconds on their lending decision engine. The pattern matters: the same engine that handles loan eligibility can handle insurance UW eligibility - different rules, same execution semantics. Several of our mid-market insurance clients model their architecture on banking precedents like this.

BNP Paribas Cardif - unified credit scoring and insurance UW

BNP Paribas Cardif runs a unified rule platform spanning credit scoring and insurance underwriting on the same Higson engine. For bancassurance carriers and any insurer with embedded finance partnerships, this pattern is increasingly common: one decisioning substrate, multiple regulated decision types, shared audit trail and governance. The cross-vertical reference is a useful one for any CUO evaluating long-horizon platform consolidation.

‍

FAQ

Q1. What is insurance underwriting automation and how does it work in 2026?

Insurance underwriting automation uses business rules engines, decision tables, and (often) machine learning models to evaluate risks, apply underwriting guidelines, and bind policies without manual review. In 2026, the dominant mid-market architecture is hybrid: a BRMS holds the deterministic rule logic (eligibility, pricing, regulatory compliance) while ML models layer in for risk scoring, with everything sharing a unified audit trail.

Q2. How does automated underwriting work step by step?

A submission is parsed and enriched with third-party data; eligibility rules filter against appetite; risk scoring is applied (rating algorithm plus optional ML model); governance rules check whether human review is required; pricing rules apply filed rates; the decision is routed (auto-bind, refer, decline, or conditional); and every step is logged for audit. End-to-end, this runs in under five seconds for clean STP cases.

Q3. What are the different types of automated underwriting systems available?

In insurance: rules-only AUS (deterministic logic), hybrid rules + ML AUS (the modern mid-market default), and AI-native AUS (ML-first, typically narrow use cases). In banking/mortgage: GSE-standardized AUSs (Fannie Mae DU, Freddie Mac LPA, FHA TOTAL). Embedded AUSs (distribution-platform-integrated) are a growing third category for carriers exposing UW APIs to partners.

Q4. What is the difference between manual and automated underwriting in practice?

Manual underwriting routes every submission to an underwriter for case review, with decisions documented in unstructured notes. Automated underwriting routes only the cases that require human judgment (typically 25–40% of mid-market volume) to underwriters; the rest are decided by rules and (optionally) ML models, with structured audit trails for every decision.

Q5. What is a realistic STP rate for a mid-market P&C carrier in 2026?

60–75% with a BRMS-powered UW layer, 65–80% with hybrid rules + ML. Mid-market commercial multi-line plateaus lower (50–70%) for structural reasons. Anything above 85% mid-market should be treated with skepticism; STP rate of 100% is marketing fiction for any non-direct line.

Q6. How long does it take to implement underwriting automation for a mid-market carrier?

A BRMS-powered UW layer typically deploys in 3–6 months for a mid-market carrier with an existing PAS. An enterprise PAS UW workbench transformation (Guidewire PolicyCenter, Duck Creek Underwriting) runs 18–36 months and is appropriate when the carrier is also replacing core policy administration.

Q7. How much does insurance underwriting automation cost over five years?

Build-internal: $4–$12M five-year TCO. Enterprise PAS UW workbench: $8–$25M. Mid-market BRMS + existing PAS hybrid: $0.5M–$2M. For self-serve evaluation, Higson is available on AWS Marketplace at $0.63/hour for PoC work.

Q8. How does a rules engine support insurance underwriting automation?

A rules engine (BRMS) is the substrate of UW automation. It executes deterministic logic (eligibility, pricing, regulatory checks), governs ML model decisions through wrapping rules, and produces the unified audit trail that NAIC market conduct examinations require. UW automation without a BRMS is custom code with hidden maintenance debt; with a BRMS, rules are versioned, testable, and authored by business users.

Q9. Is AI underwriting NAIC-compliant in 2026 and how do you maintain audit trails?

AI in underwriting is permissible under the NAIC Model Bulletin (2023, updated 2024–2025) when carriers maintain documented governance, model versioning, input feature logging, output rationale (XAI), and human oversight controls. Compliance is achievable but governance is organizational, not purely technical. Architectures that run ML models inside the rules engine (ONNX runtime) make the audit-trail requirement practical rather than aspirational.

Q10. Can you automate commercial multi-line underwriting for mid-market carriers?

Yes, with realistic scope. Commercial multi-line UW automation typically achieves 50–65% STP at the BRMS layer and 55–70% with ML scoring - lower than personal lines because of schedule rating, exposure modeling, and manuscript endorsements. The pattern that works: automate the 14–20 standard risk profiles that cover 90%+ of submissions, and route the genuinely unique risks to senior underwriters.

‍

Talk to Higson

If you are a CUO, VP Underwriting, or Enterprise Architect at a $500M–$5B GWP P&C carrier evaluating underwriting automation in 2026, here are the three ways most useful conversations start with us.

Schedule a 30-minute underwriting automation demo. See Higson Studio (no-code rule authoring), the ONNX runtime inside rules, the MCP server, and the decision audit trail. Live, with your questions, with no slides. Book a demo.

Try Higson on AWS Marketplace ($0.63/hour PoC). Self-serve evaluation. Spin up the engine, run real UW rules against test data, decide whether the architecture fits before talking to anyone. Start on AWS Marketplace.

One last thing. The mid-market UW automation conversation in 2026 is not about replacing underwriters and it is not about chasing AI for AI’s sake. It is about industrializing the rules and audit layer you already need, exposing your decision engine through documented APIs, and making your senior underwriters the strategic asset they are supposed to be. The carriers I have worked with who treat it that way - InterRisk, Allianz, Warta, Notus, BNP Paribas Cardif - are the ones whose CUOs sleep well during NAIC examination season.

Related reading

• What is a Rules Engine and Why It Matters in 2026

• ‍Automated Underwriting Systems for Lenders

• Straight-Through Processing in Insurance

• Business Rules Management System

‍

Sources cited and recommended reading

• NAIC - Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (2023; state adoption 2024–2025).

• NAIC - Market Conduct Annual Statement guidance, underwriting examination procedures.

• Colorado Department of Insurance - SB 21-169 (“Restrict Insurers’ Use of External Consumer Data”) and Regulation 10-1-1.

• NY DFS - Circular Letter No. 1 (2019) on the use of external consumer data and information sources in underwriting for life insurance.

• NIST - AI Risk Management Framework (AI RMF 1.0).

• McKinsey - “Insurance 2030: The Impact of AI on the Future of Insurance.”

• McKinsey - “The Future of Underwriting” (annual research series).

• Aite-Novarica Group - P&C Underwriting Technology research (annual).

• Celent - “AI in P&C Underwriting” and Underwriting Workbench vendor research.

• Forrester - The Forrester Wave: Digital Insurance Platforms; The Forrester Wave: AI Decisioning Platforms.

• Gartner - Hype Cycle for Insurance; research on Decision Intelligence Platforms.

• AM Best - Methodology Papers on underwriting criteria in financial strength evaluation.

• Bain & Company - Insurance Customer Loyalty research (cycle-time and NPS impact).

• Deloitte - “Cognitive Underwriting” and annual Insurance Industry Outlook.

• Fannie Mae - Selling Guide (Desktop Underwriter documentation).

• Freddie Mac - Single-Family Seller/Servicer Guide (Loan Product Advisor documentation).