Bevaya AI Agents — AF Group

Today's Discussion

Agenda

AF Group & Bevaya — Claim Indexing
Presentation overview

01 Introductions

02 Bevaya Platform & Our AI Agents

03 Claims Workflows

04 Layers of Indexing

05 Demonstration

06 Customer Results

07 Build, Run, Govern with Bevaya

08 Our Engagement Model

09 Open Discussion

10 Next Steps

Our Solution

All-in-One Platform

AI Agents That Triage, Analyze, and Recommend

Claims Automation

Underwriting Automation

Policy Servicing Automation

Platform Capabilities to design, deploy, and govern AI

Build

Workflow Canvas Document Intelligence Pre-built AI Agents

Run

Workspaces Live Runs

Review

Human-in-the-Loop Grounded Explainability Analytics

Govern

Governed Automation Integrations Organization & Admin

Insurance-Native AI

InsurGPT™

A mosaic of specialized insurance AI models, trained on real insurance data.

Human-in-the-Loop

Your experts keep every decision. Bevaya surfaces its work in your team's workflow where a human call is needed.

Works Atop What You Have

Guidewire, OnBase, ImageRight, DuckCreek, and more. Integrates via API.

AI Assistant

An AI agent across the platform

Builds your workflows. Executes work alongside you. Explains every step.

Why isn't this straight-through?

Three rules failed.

Premium delta is above the $5,000 STP threshold, and two new drivers need underwriter review before…

See the AI Assistant +

Production Layers

From Prototype to Production.

Most AI projects fail after Layer 1. Bevaya ships — and operates — all the layers it takes to go live and stay live.

Build

Design any workflow. Deploy it fast.

Connect pre-built nodes — OCR, LLM extraction, HTTP, and human review — into production-grade automations on a visual canvas.

Node Types

Set Variables

HTTP

Code Sandbox

OCR

LLM Extraction

LLM Insights

Review

Trigger

Intake

›

OCR

Read

›

Extraction

Classify

›

HTTP

Policy lookup

›

Insights

Judgment

›

Review

Human

›

HTTP

Write back

Deterministic

AI Judgment

Human

30+ system connectors available at launch

300M+ real insurance documents used to train InsurGPT™

Run

Autonomy is a setting. Not a leap.

Set it per workflow, per step. Move it when your own numbers say you can.

Level 01

Assistive

Agent drafts. Person decides.

Every output goes to a person. Measure the agent against your team before you trust it.

Automation25%

Where every workflow starts.

Level 02

Supervised

Agent decides. Person approves.

The agent commits to an answer and shows its evidence. A reviewer confirms or corrects it.

Automation70%

Where most workflows live.

Level 03

Autonomous

Agent decides. Exceptions route out.

Clean work flows straight through. Anything below your confidence threshold routes to a named queue.

Automation95%

Where the volume pays off.

Live Runs Every step, every retry, every routing decision, streaming as it happens. Nothing to reconstruct later. CTP-9101

Govern

Every judgment, provable.

Adjuster_report_final.pdf Page 4 of 11

Evidence · Lines 12 to 18

Debris removal quoted separately at $3,200.
Not included in the structure estimate above.

Judgment Record CTP-9101 · Whitaker · tree loss

Agent Claim file summarization

Step Coverage gap check

Model InsurGPT™

Autonomy Supervised

Evidence p4, lines 12–18

Confidence

91%

Outcome Gap flagged — $3,200

Reviewed by S. Chen

Decided 14:12:07 UTC · r-88214

Grounded, not summarized.

Every answer points back to the page and the lines it came from. Click through and read the source.

Roles decide who can do what.

Build, run, approve, and change are separate permissions. Your admins set them, not us.

Logged, end to end.

Every run, every override, every version change. Exportable when audit asks.

Trusted By

Enterprise P&C Organizations

Deployed at carriers, TPAs, and brokers managing millions of documents a year across commercial lines, specialty, workers' comp, and personal lines.

120+ Active AI Agents

40+ Carriers, TPAs & Brokers

3 of Top 5 Carriers

3 of Top 10 TPAs

3 of Top 20 Brokers

Pre-Built AI Agent Library

Meet your AI workforce

Bevaya ships pre-trained agents for every major insurance workflow. Each is ready to deploy against your systems and document types — no lengthy model training required.

Underwriting

Submission Ingestion & Clearance In Prod

Ingests new submission documents — applications, schedules, and supporting files — extracting structured data, screening for completeness, and clearing submissions for underwriter review.
Learn more →

Loss Run Processing In Prod

Reads multi-year loss run reports, extracting incurred/paid totals, claim counts, and loss ratios by policy period for underwriting analysis.
Learn more →

Triage & Routing

Analyzes incoming submissions and correspondence, classifying by line of business, urgency, and complexity — then routes each item to the right underwriter, team, or downstream workflow automatically.
Learn more →

Rating & Quoting

Extracts structured submission data and applies rating logic to generate preliminary quotes — reducing manual re-keying, accelerating the quote-to-bind cycle, and surfacing risks that warrant underwriter review.
Learn more →

Policy Renewal Handling

Identifies renewal submissions, compares current to expiring policy terms, and generates renewal summaries with change flags for underwriters.
Learn more →

Policy to Policy Comparison

Compares two policy documents side-by-side — or one policy against many — highlighting coverage changes, endorsement additions, exclusion differences, and limit adjustments across the full set.
Learn more →

Claims

FNOL / FROI Setup Today's Demo

Processes First Notice of Loss and First Report of Injury documents, extracting key details and creating structured claim records directly in your system.
Learn more →

Claim Indexing Today's Demo

Reads incoming claim documents — medical records, bills, correspondence — and automatically indexes them into the correct claim file by document type and date.
Learn more →

Claim to Policy Comparison

Checks claim details against the in-force policy to verify coverage, flag potential exclusions, and surface mismatches before adjudication begins.
Learn more →

Legal Demands Identification & Extraction

Identifies legal demand letters, extracts demand amounts, deadlines, and liability details, and routes to the appropriate claims handler with urgency flags.
Learn more →

Medical Bill Identification & Extraction

Reads medical bills and EOBs, extracting procedure codes, billed amounts, and provider information for adjudication, audit, and cost review.
Learn more →

Invoice Payment Processing

Extracts vendor invoice data, matches against open claims or purchase orders, and routes for payment approval with discrepancy flags and audit trail.
Learn more →

Claim File Summarization

Generates structured summaries of entire claim files — chronology, reserves, coverage, key decisions — in seconds for handler review or litigation prep.
Learn more →

Policy Servicing

Endorsement Processing

Reads endorsement requests, extracts the requested policy changes, compares to current in-force terms, and generates processing summaries for policy service staff.
Learn more →

COI Creation

Reads certificate of insurance requests, validates coverage requirements against the in-force policy, and generates pre-populated COI drafts ready for review and issuance.
Learn more →

Premium Audit Processing

Extracts payroll, revenue, and classification data from audit worksheets, calculating exposure changes and flagging discrepancies for auditors to review.
Learn more →

Build Your Own AI Agent

Don’t see your workflow? We build agents for it.

Bevaya’s team works directly with your claims operations team to design, train, and deploy custom AI agents tailored to your specific workflows, document types, and system integrations — no generic solutions, just agents built for your exact process.
Learn more →

Claims Indexing · Future State Workflow

Classify documents, route claim files & set up FROIs

Physical mail, fax, and email arrive through AF Group’s intake channels and are classified by document type. Bevaya AI Agents route each document to the correct claim file, extract structured data, and set up First Reports of Injury — delivering decisioning output directly into Guidewire ClaimCenter.

Intake Channels

Output Destination

Bevaya AI Agents

InsurGPT™ powers every step

Email

Fax

Mail

Content Management

Extraction

details

Claim Matching

details

Analysis

details

Decisioning

details

Bevaya Platform

details

Your Team

Guidewire
ClaimCenter

Claim Indexing Intelligence

Solving for Indexing

Challenge

How We Solve It

1

Multiple document types in one PDF

A single package contains mixed document types that need to be separated and classified.

PSS — Page Stream Segmentation

Our proprietary model identifies and splits each document type within a mixed package — built and refined over years of claims data.

2

One document belongs to multiple claims

One document references multiple valid claim numbers and must reach each file.

Multi-Claim Routing Rules

Post-processing rules route the document to each claim file automatically, or flag it for user validation when confidence is mixed.

3

Missing claim numbers

Documents arrive with no claim number, blocking standard indexing.

Triangulated Identification

We triangulate policy number, claimant name, and date of loss to identify and index the correct file.

4

Incorrect claim numbers

The claim number on the document doesn't match the claimant or insured on file.

Name-to-Claim Validation

We validate the claimant name against the claim number at index time. Mismatches are routed to your team using the search logic adjusters use today.

5

Overlapping document types

Ambiguous or overlapping categories create classification conflicts.

Pre-Test Category Review

Our Customer Success team reviews your categories before testing and advises the optimal taxonomy based on industry knowledge.

6

Not all documents are equal

A single confidence threshold across all document types leads to errors.

Dynamic Thresholding

We apply distinct thresholds per document type — a medical bill and a demand letter are handled differently than returned mail.

7

Document routing

Some documents belong to a different unit or third party and need to be routed outside the standard claims flow.

Contextual Business Rules

We configure routing rules that mirror your team's process — special handling units, Underwriting, or external third parties.

Indexing Model · Document Coverage

Every Document Type, Already Mapped.

62 types across 7 categories — start with any subset and expand or reduce as your operation evolves.

Coverage at a Glance

62

document types

7

AI Agent Capabilities

Click any capability to learn more about how it works inside the Bevaya platform.

Multi-Modal Extraction

Reads text and visual layout together, so nothing gets lost in translation for complex document extractions. For many document types, accurate extraction requires understanding both what is written and how it is structured on the page.

Image Understanding

Understands and summarizes the content of standalone images — describing what's in a damage photo, interpreting diagrams, and surfacing visual evidence without any additional processing steps.

Live Web Access

Agents can pull in live, publicly available information from the web in real time, surfacing current context and data that goes beyond what the model was originally trained on.

Highlight Mode

Places visual markers directly on source documents to show exactly where each extracted value was found — works on unstructured and structured documents alike.

Page Stream Segmentation

Automatically splits a single document into multiple distinct documents based on their content — separating a combined submission into individual policies, endorsements, and attachments.

Request Missing Information

Automatically identifies required fields that are absent and sends a structured request back to the original sender to complete the submission — no manual follow-up needed.

Claim & UW System Integration

Native bi-directional integration with Guidewire, Duck Creek, Majesco, and other core systems — reading existing records, writing structured outputs, and triggering workflow actions via API.

3rd Party Application Integration

Validates extraction results against 3rd party applications and enriches them with external data — connecting to rating engines, data providers, and partner systems for enhanced accuracy.

Document Comparison

Compares multiple documents or source records against each other to surface anomalies, discrepancies, and inconsistencies — flagging mismatches that may indicate errors or fraud.

Calculations

Performs automated calculations on extracted values — totaling exposures, verifying premium computations, applying deductibles, and flagging mathematical inconsistencies in submitted data.

Custom Outputs

Generates structured output in any format your workflow requires — Excel, Word, PDF, JSON, XML, and more — populated with extracted and enriched data from the AI agent pipeline.

Team Chat Integration

Communicate with your AI agents through the channels you already use — Slack, Microsoft Teams, and email — receiving updates, asking questions, and triggering actions without leaving your inbox.

Real Time Assistance

Chat directly with your AI agents inside the Bevaya Platform to ask follow-up questions, request deeper analysis, and gain additional insights on any task or document — live, in context.

Task Creation

Automatically creates tasks in your claim or underwriting systems via API — assigning work, setting priorities, attaching extracted data, and triggering the right workflow the moment a document is processed.

Risk Analysis

Scores submissions and claims against configurable risk criteria — flagging high-frequency accounts, coverage gaps, missing documentation, and anomalies before they reach an underwriter or adjuster.

Claims AI · In Production

Claim Indexing: Proven at Scale

Four enterprise carriers · 45M+ annual pages. No-Touch Rate and Business Rule Triggers show the automation split — AI Agent Success confirms accuracy across all routed decisions.

Company

No-Touch Rate & Business Rule Triggers

Success Rate

No-Touch Rate

A

7.5M pages/yr

39 doc types

No-Touch

92%

Business Rules

8%

100%Success

92% No-Touch Rate

B

10M pages/yr

46 doc types

No-Touch

76%

Business Rules

22%

99%Success

76% No-Touch Rate

D

25M pages/yr

96 doc types

No-Touch

87%

Business Rules

13%

100%Success

87% No-Touch Rate

Customer results have a direct correlation to the customer's trust of the AI — reflected in their selected confidence threshold and the business rules they configure. Higher trust enables higher automation.

In Production

Comparative Customer in Production

Each field group graded independently across real production documents.

Extraction Field Group & Fields

Accuracy Today

Result

Routing Keys

Claim # · Policy #

0% Accuracy

Loss Detail

Date of loss · Loss location

0% Accuracy

Classification

Document type

0% Accuracy

4,997 annotated values graded across both joint review calls. 96.00% overall accuracy today across all field groups.

Project Overview

Configuration Through Testing in 4 Weeks

From design through UAT, AF Group goes live in 4 weeks. System connections and scoping are finalized before the clock starts — keeping configuration focused and on schedule.

Before the clock starts

System Connections Established

Scoping Finalized

Document Types Defined

Business Rules Documented

Phase

Week 1 Week 2 Week 3 Week 4 Week 5 Week 6 Week 7

Design & Workflow Review

2 wks

AI Agent Configuration

2 wks

UAT & Tuning

2 wks

Go-Live & Hypercare

Go-Live

Teams with document types and business rules documented before kick-off consistently hit the 4-week timeline.

AI Performance

InsurGPT™ — purpose-built for insurance accuracy.

An ensemble of specialized AI models trained on 300M+ real insurance documents. Benchmarked quarterly against frontier models — InsurGPT™ leads on every field category across underwriting, claims, and loss runs.

93%

Loss run accuracy

9+ points ahead of all frontier models

98–100%

Claims extraction

at 0.9 confidence threshold

300M+

Training documents

insurance-specific, expert-annotated

Loss Run Extraction — Claim-Level Fields Dec 2025

Field	GPT-4.1	GPT-5.0	Gemini 3.0	InsurGPT™
Indemnity Paid	66%	53%	81%	98%
Medical Reserves	86%	60%	91%	98%
Recoveries	97%	61%	91%	99%
Status	70%	46%	69%	96%
Claimant Name	88%	58%	84%	97%
Claim Number	83%	54%	76%	93%
Line of Business	91%	60%	86%	98%
Nature of Injury	53%	40%	72%	90%
Date of Incident	88%	58%	85%	97%
Total Paid	74%	52%	73%	88%

Underwriting Extraction — Overall

InsurGPT™

93%

Gemini 3.0 Pro

84%

GPT-4.1

81%

GPT-5.0

80%

Dec 2025 · 9 points above next-best model

Claims Extraction — Overall

InsurGPT™

95%

Gemini 3.0 Pro

81%

GPT-4.1

80%

GPT-5.0

54%

Dec 2025 · 14 points above next-best model

Full benchmark methodology and results at bevaya.ai/benchmarks

Capacity Gained

We handle the work. Your experts make the calls.

We remove manual, low-judgment work across insurance, freeing your team's capacity for the decisions that need their expertise.

Workflow

Share of effort Bevaya takes off your team

Capacity Freed

COI Generation

95%

Claims Doc Indexing

90%

Triage & Clearance

88%

FNOL / FROI

85%

Submission Intake

78%

Endorsement Processing

78%

Premium Audit Prep

73%

Legal Demand Intake

72%

Rating / Quoting

70%

Integrations

Works with the systems you already run.

Bevaya AI Agents connect directly to your claims systems, policy admin platforms, document management tools, and communication channels — orchestrating workflows across them, not just passing data between them.

Homegrown & Custom Applications

Built something in-house? Bevaya connects to custom applications via REST APIs, SFTP file exchange, and direct database connectors — no pre-built adapter required.

Custom REST API

OpenAPI / Swagger spec ingestion

RESTOAuth 2.0API Key

Import an OpenAPI 3.0 or Swagger spec — Bevaya auto-generates the connector
Support for OAuth 2.0, Bearer token, API key, and basic auth
Webhook receiver for event-triggered automation
Retry logic, rate-limit handling, and error routing built in

SFTP / File Exchange

Batch file ingestion and delivery

SFTPFile PollBatch

Poll SFTP directories on a schedule or event trigger
Handles PDF, TIFF, CSV, XML, EDI, and custom flat files
Archive, move, or delete source files post-processing
Delivery of structured outputs back to SFTP drop-zones

Trust & Governance

Security and compliance built in, not bolted on.

Designed for the most regulated industry in the world. Every deployment meets the data, privacy, and governance standards insurance carriers and TPAs require.

Data Ownership

Your data, always

Never shared with other customers or vendors

Bevaya never trains shared models on your data

Logical tenant isolation — enforced at data layer

Customer-managed keys available

SSO / SCIM provisioning · US data residency

Compliance

Regulated industry ready

SOC 2 Type II — annual third-party audit

HIPAA — covered entity ready

GDPR & CCPA / CPRA — state privacy controls

23 NYCRR 500 — NY DFS cybersecurity

AES-256 at rest · TLS 1.3 in transit

Deployment

Enterprise infrastructure

Microsoft Azure — primary, per-tenant env

Signed to live in days — Bevaya manages all infra

Minimal effort from IT for API access

Oversight

Audit-ready by design

Immutable audit logs on every AI action

Confidence scoring on every field extraction

Patented Grounded Explainability — source-cited

HITL review queues for low-confidence items

Workflow versioning & rollback

Certifications & Frameworks

SOC 2 Type II

HIPAA

GDPR

CCPA / CPRA

23 NYCRR 500

AES-256

SAML 2.0 / OIDC

US Data Residency