PaperlayerView exact revision on arXiv

Source-linked AI summary

Governing the Edge: Automating Commercial Property and Casualty Insurance Underwriting via a Hybrid Local-Cloud Multi-Agent Framework

Vivek Kumar Singh, Gautam Bhowmick

arXiv:2609.37454v1cs.AIcs.MA

TL;DR

A typical 40-minute submission is mostly administrative work, while existing AI approaches lack some combination of compliance enforcement, authoritative data access, privacy, and regulator-grade audit trails. Governing the Edge combines graph-based compliance rules, local agents for sensitive data, cloud agents for anonymized-data reasoning, and logged tool interfaces. 100% hard-stop accuracy (6/6) and 70% overall compliance accuracy on 20 synthetic scenarios; the overall result is between the monolithic baseline (75%) and linear pipeline (30%).

  • Problem

    A typical 40-minute submission is mostly administrative work, while existing AI approaches lack some combination of compliance enforcement, authoritative data access, privacy, and regulator-grade audit trails.

    Source: §I · p. 1

  • Method

    Governing the Edge combines graph-based compliance rules, local agents for sensitive data, cloud agents for anonymized-data reasoning, and logged tool interfaces.

    Source: §I · p. 2, A. Overview · p. 2, §V · p. 4

  • Results

    100% hard-stop accuracy (6/6) and 70% overall compliance accuracy on 20 synthetic scenarios; the overall result is between the monolithic baseline (75%) and linear pipeline (30%).

    Source: §VII · p. 5, §IX · p. 6

  • Takeaways & Limitations

    The prototype automates roughly 40 minutes of manual submission work in a few minutes, and submissions clearing every check can proceed without human touch.

    Source: §I · p. 2

    Synthetic, unseeded tool data makes cloud-tier soft scores, aggregate risk tier, and per-run latencies variable; synthetic scenarios and single-expert ground truth also limit generalization.

    Source: A. Limitations of the Current Prototype · p. 6

Abstract

from arXiv · show

Underwriters in commercial Property and Casualty (P&C) insurance spend 30 to 40% of their time on administrative work rather than risk judgment, and a single submission takes about 40 minutes by hand. We present Governing the Edge, a multi-agent framework for that layer, organized around a data-residency constraint: sensitive submission data must not leave the perimeter. Eleven agents and two deterministic control nodes, one a human-escalation interrupt, form a 13-node LangGraph workflow across two tiers. Agents touching raw submissions run locally on Gemma 2, each bound through a tool interface logging every invocation (synthetic stubs in the current prototype); only anonymized scores and non-identifying fields cross to Claude Sonnet in the cloud. Compliance rules are encoded as conditions on graph edges, so a non-compliant submission never reaches pricing. We walk through one full scenario end to end, a commercial auto submission whose principal driver carries serious violations, showing every agent call, every tool invocation, and the resulting routing decision. The framework processes a 40-minute manual submission in a few minutes on a single edge device, dominated by serialized on-device inference. On the hard-stop violation tier the framework enforces every rule correctly and reproducibly, since hard stops are deterministic predicates over fields extracted at temperature 0; overall compliance accuracy across the 20-scenario benchmark is 70%, with the remaining gap concentrated in softer, judgment-based tiers. It keeps a complete audit trail and respects the privacy boundary throughout. We release all code, compliance rules, tool stubs, and synthetic datasets.

I. INTRODUCTION

Commercial P&C underwriting devotes substantial time to submission handling rather than risk judgment, while existing AI approaches lack compliance enforcement, authoritative data access, and privacy protection. Governing the Edge targets that administrative layer with graph-based compliance, a local-cloud split, logged tool interfaces, and per-agent error tracking.

I. INTRODUCTION

  • 30 to 40% of underwriters’ time goes to administrative work such as rekeying data and reading documents rather than evaluating risk.A typical 40-minute submission includes document reading and field extraction, data lookups, compliance checks, and audit-record writing; the risk call is described as a small slice of the total.Source: §I · p. 1
  • Existing AI approaches either cannot enforce compliance or query authoritative data, or they expose raw sensitive data and lack a regulator-grade audit trail.The passage contrasts single-prompt systems with agent pipelines such as CrewAI.Source: §I · p. 1
  • The framework encodes compliance rules as graph-transition conditions, keeps privacy-sensitive agents on-premise, and gives agents dedicated tool boundaries with logged calls.The prototype uses stubbed external sources and records agent failures across eight categories for reliability monitoring over time.Source: §I · pp. 1–2
  • The system combines a 10-rule P&C compliance engine, deterministic hard-stop predicates, per-agent error tracking, and a local-cloud architecture for commercial property and auto underwriting.Its prototype uses synthetic tool stubs and a synthetic benchmark, and submissions clearing every check can proceed without human touch.Source: §I · p. 2
  • The worked scenario illustrates a reduction from roughly 40 minutes of manual work to a few minutes of automated execution.Production deployment requires updating rule thresholds for state-specific DOI requirements and NAIC model laws.Source: §I · p. 2

II. RELATED WORK

Related work spans AI for insurance tasks, multi-agent coordination, and privacy-preserving edge AI. The paper distinguishes its approach by routing whole workflows according to compliance and maintaining a local-cloud privacy boundary.

A. AI in Insurance Underwriting

  • Sajid reports up to 92.9% accuracy in property-claim evaluation, while its task-graph coordination does not route on compliance or use authoritative external tools.Roy and Singh [2] use a cloud-based human-in-the-loop system with a state machine and RAG over an underwriting manual, but do not address the privacy boundary or authoritative external-data grounding.Source: A. AI in Insurance Underwriting · p. 2

B. Multi-Agent Large Language Model (LLM) Frameworks

  • Unlike task-distribution designs and single-call routers such as RouteLLM [5] and FrugalGPT [6], this framework routes the whole workflow on governance before execution.LangGraph provides shared state, conditional edges, and interrupts; AutoGen [4] is described as introducing unwanted randomness for regulated decisions.Source: B. Multi-Agent Large Language Model… · p. 2

C. Privacy-Preserving and Edge AI

  • On-device language models address data-localization and privacy constraints; the paper applies that edge-AI approach to insurance by separating local and cloud processing.Gemma 2 is identified as small enough for on-premise deployment [8], while Xu et al. [7] survey on-device language models and enabling techniques.Source: C. Privacy-Preserving and Edge AI · p. 2

III. SYSTEM ARCHITECTURE

Governing the Edge uses a 13-node LangGraph StateGraph divided into local and cloud tiers by a privacy boundary. Its workflow runs document intake and risk assessment before compliance routing, then conditionally proceeds to coverage and pricing or human escalation.

A. Overview

  • The framework uses a 13-node LangGraph StateGraph with three branch points, placing sensitive-data processing on-premise and anonymized-data reasoning in the cloud.Figure 1 depicts the architecture and its explicit boundary between raw submissions and anonymized score processing.Source: A. Overview · p. 2, Figure 1 · p. 3
  • After document parsing and intake, three risk sub-agents run in parallel; a cloud Risk Aggregator combines their scores before the Compliance Agent checks the rules.If compliance passes, Coverage Gap and Pricing agents run in parallel; an explainability agent assembles the audit trace, while the human-escalation interrupt fires for confidence below 0.80 or failed compliance.Source: A. Overview · pp. 2–3
  • Local agents invoke bound tool interfaces and send structured, anonymized scores across the privacy boundary to the cloud Risk Aggregator.The figure caption identifies the local agents as Gemma 2 and the cloud agents as Claude Sonnet.Source: A. Overview · p. 2, Figure 2 · p. 3

B. Privacy Boundary

The workflow keeps sensitive fields local while sending only anonymized state to cloud agents, validates model outputs at agent boundaries, and routes downstream work through compliance gates.

B. Privacy Boundary

  • Sensitive document fields are omitted from cloud state; cloud agents receive anonymized risk, compliance, and coverage data, while sensitive tool calls remain local.The serializer, cloud prompt, and network isolation each enforce the boundary.Source: B. Privacy Boundary · p. 3

C. Provider-Agnostic LLM Layer

  • Each agent’s model output is checked against a Pydantic schema before entering shared state, and LiteLLM lets providers be set per agent.Local agents default to Gemma 2 and cloud agents to Claude Sonnet.Source: C. Provider-Agnostic LLM Layer · p. 3

A. Structural vs. Post-Hoc Governance

  • Compliance transitions gate access to coverage and pricing, making downstream agents unreachable from a non-compliant state.A CONFLICT triggers one automated retry; if it still does not PASS, the workflow escalates to a human.Source: A. Structural vs. Post-Hoc Governance · p. 3
  • Deterministic predicates decide hard stops before soft WARNING rules reach Gemma 2, so hard-stop outcomes are stable across runs when extracted fields are stable.The predicates read fields extracted at temperature 0; an upstream field-extraction error remains a route to an incorrect hard-stop status.Source: A. Structural vs. Post-Hoc Governance · pp. 3–4

B. P&C Compliance Rules Engine

The rules engine applies ten underwriting rules across commercial property and auto, distinguishing submissions that require a human stop from warnings that affect confidence and audit records.

B. P&C Compliance Rules Engine

  • The rules engine contains 10 P&C underwriting rules covering commercial property and commercial auto.The rules were written and sanity-checked against real underwriting practice.Source: B. P&C Compliance Rules Engine · p. 4
  • CP-02 defines total insured value as the combined replacement cost of buildings, contents, and business income at one location.Source: B. P&C Compliance Rules Engine · p. 4
  • A HARD STOP sends a submission directly to a human regardless of system confidence, while a WARNING lowers confidence and is recorded in the audit trail.Source: B. P&C Compliance Rules Engine · p. 4

V. TOOL INTEGRATION

The tool layer provides structured interfaces and audit logging, but the prototype uses synthetic external-data stubs and bases hard-stop decisions on extracted document fields.

V. TOOL INTEGRATION

  • Each agent has a dedicated tool boundary to an authoritative source, but prototype tools are synthetic stubs rather than live data grounding.The hard-stop rules read fields extracted from the submission, not soft scores from stubbed risk tools; verified tool outputs are a planned step.Source: §V · p. 4
  • Document parsing uses optical character recognition to extract scanned-form text before Gemma 2 normalizes it.Source: §V · p. 4
  • Tool-call records include timestamps, inputs, outputs, success flags, and durations, while failures are classified into eight categories.All executions in the released 20-scenario run succeeded, so failure tracking is infrastructure rather than a measured result.Source: §V · p. 4

VI. EXPERIMENTAL WALKTHROUGH SCENARIO

The walkthrough follows a commercial-auto submission with serious driver violations through risk scoring, a deterministic CA-02 hard stop, and human escalation. The case completes in minutes rather than the 35–45 minutes reported for manual handling.

A. Input Submission

  • CA-S06 is a three-vehicle courier submission whose principal driver has DUI, reckless-driving, and at-fault-accident violations across three years.The scenario exercises each tool type and trips the CA-02 hard-stop rule.Source: §VI · p. 4, A. Input Submission · p. 4
  • The three risk sub-agents return MEDIUM fleet risk, DECLINE driver risk, and MEDIUM operations risk with reduced confidence after a missing DOT number.The Risk Aggregator receives only anonymized scores and band labels; the decisive hard-stop decision occurs later at compliance.Source: A. Input Submission · p. 4
  • Three major driver violations in three years trigger CA-02: the deterministic check returns a HARD STOP and DECLINE without relying on LLM output.The Compliance Agent reads the extracted count of three against the rule threshold of two or more.Source: A. Input Submission · pp. 4–5
  • A failed compliance result routes CA-S06 to Human Escalation, so Coverage Gap and Pricing never run.The escalation package includes the rule id, agent explanations, audit trail, and recommended action DECLINE.Source: A. Input Submission · p. 5
  • 197.5 seconds (3.3 minutes) for CA-S06, versus 35–45 minutes by hand; the 20-scenario average is 268.8 seconds.Serialized on-premise Gemma 2 inference dominates the framework timing.Source: A. Input Submission · p. 5

VII. COMPARISON WITH BASELINE SYSTEMS

Across 20 scenarios, Governing the Edge reaches 70% overall compliance accuracy, between the monolithic baseline and the linear pipeline. Its hard-stop performance is stronger than the monolithic baseline, while its softer-tier accuracy is the main gap.

VII. COMPARISON WITH BASELINE SYSTEMS

  • On CA-S06, monolithic Claude missed CA-02 and quoted a USD 5,200 premium, while the linear pipeline caught the violation only after pricing and coverage analysis.The monolithic baseline did not attend to the violation count embedded in free-text history.Source: §VII · p. 5
  • 100% (6/6) hard-stop accuracy for Governing the Edge, versus 83% (5/6) for monolithic Claude; it also preserves straightforward approvals at 4/4.The linear pipeline also catches all six hard stops but scores 0% on every other tier.Source: §VII · p. 5
  • Governing the Edge achieved 70% overall compliance accuracy, between the monolithic baseline at 75% and the linear pipeline at 30%, across 20 scenarios.Source: §VII · p. 5
  • The overall gap to monolithic Claude is concentrated in WARNING and low-confidence escalation tiers, where Governing the Edge scores 2/6 and 2/4, respectively.With 20 scenarios, these tier-level differences are indicative rather than statistically established.Source: §VII · p. 5

A. Underwriter Handoff

The handoff material reports perfect hard-stop accuracy in the committed run and describes how uncertain cases are routed to a human. It also identifies the latency and baseline comparisons covered by the tables.

A. Underwriter Handoff

  • Table V covers per-agent latency on the released 20-scenario run, pairing framework median wall-clock times with manual benchmark times.Framework timing is from serialized on-premise inference; manual times are from the McKinsey commercial-underwriting benchmark [9].Source: Table V · p. 6
  • Table VI compares baseline systems on 20 synthetic scenarios.The comparison reports compliance accuracy, workflow completion, and average latency.Source: Table VI · p. 6
  • Governing the Edge is the only system that enforces every hard stop (6/6) without collapsing on the other tiers.Table VII reports compliance accuracy by scenario tier for the committed 20-scenario run.Source: Table VII · p. 6
  • At a threshold of 0.80 on the same 0–1 scale, the system routes uncertain cases to a human, while cases not routed there reach a Final Decision automatically.Source: A. Underwriter Handoff · p. 6

A. Limitations of the Current Prototype

The prototype’s results are bounded by synthetic, unseeded tool data and synthetic scenarios, while risk aggregation remains a cloud-model output rather than an auditable arithmetic step. The conclusion reports 100% hard-stop accuracy and 70% overall accuracy on 20 synthetic scenarios.

A. Limitations of the Current Prototype

  • Unseeded synthetic tool outputs make cloud-tier soft scores, aggregate risk tiers, and per-run latencies vary, although deterministic hard-stop results remain stable at 6/6 across runs.A fixed RNG seed is planned to make entire runs reproducible; the tool layer currently supplies call-and-audit plumbing rather than live grounding for decisions.Source: A. Limitations of the Current Prototype · p. 6
  • The released configuration runs Gemma 2 locally, while Gemma 4 is a future target requiring latency remeasurement; cloud-LLM risk aggregation is not an auditable arithmetic step.Replacing aggregation with a deterministic weighted sum is planned.Source: A. Limitations of the Current Prototype · p. 6
  • Synthetic scenarios limit generalization, ground-truth decisions came from one expert, and manual-time figures rely on a published benchmark rather than carrier studies.The authors identify a larger, multiply annotated benchmark as a way to strengthen the ground truth.Source: A. Limitations of the Current Prototype · p. 6

IX. CONCLUSION

  • 100% hard-stop accuracy (6/6, reproducibly) and 70% overall accuracy on 20 synthetic scenarios; the framework preserves data residency and a complete audit trail.The conclusion also reports release of code, rules, stubs, and datasets.Source: §IX · p. 6
Loading 2609.37454v1…