Rick Becker - WebApp Portfolio

Executive Summary

Ace is a fully transactional AI agent for Acme Insurance’s E-Bike Insurance platform. Customers can get a quote, buy a policy, look up their account, make changes, and process payments entirely through natural conversation — no forms, no dropdowns, no “step 3 of 5.”

Transactional AI Agent

Not a FAQ chatbot. Handles the full insurance lifecycle through conversation.

Tiered Authority Model

Gives the agent autonomy on safe actions and routes premium-affecting changes to human underwriters.

Provider-Agnostic LLM

Designed for future flexibility. Swap Claude for OpenAI or Gemini without rewriting.

7-Phase Playbook

Applicable to any insurance product, carrier, or distribution model.

Production Considerations

Addresses regulatory compliance, E&O exposure, recordkeeping, and security.

The Problem

Insurance has the highest online quote abandonment rate of any industry. 84% of people who start an insurance quote never finish it (ProPair.ai / Drips.com, 2022-2025). Primary causes: lengthy forms, confusing questions, poor mobile experiences.

The average auto insurance quote requires 20-40 form fields. Even simple specialty products require 10-15 fields across multiple steps. Every field is a chance for the customer to leave.

Quote-to-policy conversion rate: 10-20%, some as low as 5% (ProPair.ai, 2025). 78% of customers choose the company that responds first (LeanData via Drips.com).

84%

Quote Abandonment

Industry-wide

20-40

Form Fields

Average auto quote

10-20%

Conversion Rate

Quote to policy

78%

Choose First Responder

Speed wins

Thesis

A customer who will never complete a 15-field form will happily answer five questions in a conversation. Conversational interfaces eliminate form fatigue — the single biggest source of quote abandonment.

What We Built

Ace is embedded as both a floating chat widget (all pages) and a dedicated full-page experience at /chat. It connects to the same rating engine, policy database, underwriting rules, and payment processor as the web app — a single source of truth.

Capabilities

Capability	How It Works
Quoting	Collects info conversationally, calls rating engine, presents premium breakdown — no form.
Purchasing	Walks through payment (CC or ACH), processes transaction, issues policy — same conversation.
Policy Lookup	After identity verification (email + last 4 of phone), retrieves full policy details.
Policy Changes	Non-premium changes confirmed and executed instantly. Premium-affecting changes explained with estimated impact and routed to underwriter.
Knowledge Base	Coverage explanations, eligibility rules, FAQ answers, billing questions in natural language.
Escalation	Creates support ticket with full conversation attached so customer never repeats themselves.

Key Decisions

Identity & Personality

Decision	Choice	Rationale
Name	Ace	Short, memorable, ties to Acme brand.
Personality	Friendly and warm	Reduces intimidation. Insurance is confusing.
Tone	Conversational, clear, no jargon	Target user doesn't speak insurance.
Language	English only	POC scope.

LLM Strategy

Decision	Choice	Rationale
Primary Provider	Claude (Anthropic)	Consistent with how platform was built. Strong tool-use.
Architecture	Provider-agnostic abstraction	Interface wrapping LLM provider. Can swap to OpenAI, Gemini without rewriting.
Fallback	Graceful degradation	If API down: shows fallback message with form/phone option. Never blank screen.
Vendor Lock-in	None by design	Enables A/B testing between providers or switching.

UI & Placement

Decision	Choice	Rationale
Chat Widget	Floating on all consumer pages, collapsed by default	Always available but non-intrusive.
Dedicated Page	Full-screen at /chat	For complex interactions — full quoting, policy changes.
Persistence	Survives page navigation and return visits	Stored in Firestore. Needed for escalation handoff.
Feedback	Thumbs up/down every conversation	Minimum viable quality signal.

The Authority Matrix

This is the most important design decision — only an insurance SME can make it. It determines what the agent does autonomously, what requires confirmation, and what requires a human.

Four levels of agent authority

Full Autonomy

Zero risk, no financial/coverage impact

Generate a quote
Answer product questions and FAQs
Explain coverage types and rating methodology
Provide general product information

Confirm + Execute

Low risk, reversible, no premium impact

Update phone number or email address
Change mailing address within same county
Process a scheduled payment

Route to Human

Premium impact, coverage change, or regulatory sensitivity

Change coverage type
Swap the insured e-bike
Change rider info that affects rating (age, county)
Cancel a policy
Reinstate a lapsed policy

Redirect Entirely

Out of scope, different department

File a claim → claims team
Legal questions → licensed professional
Complaints → customer service management
Fraud reporting → SIU

Golden Rule

If a change affects the premium, a human reviews it before it takes effect.

How Premium-Affecting Changes Work

Ace identifies the change and explains what it means.

Ace runs the rating engine with the proposed change.

“This change would adjust your annual premium from $350.75 to $488.00 — an increase of $137.25 pro-rated for the remaining term.”

Ace submits the change request for underwriter review.

“I’ve submitted this for review. You’ll hear back within 24 hours.”

Guardrails

Hard Rules

Rule	Rationale
No guaranteed pricing	Quotes always "valid for 30 days, subject to underwriting approval." Protects against rate disputes.
No claims handling	Claims involve coverage determination, reserve setting, legal exposure. Always redirects.
No legal advice	General coverage info only. Recommends licensed agent for specific situations.
No competitor discussion	Acme products only. Never compares, references, or disparages competitors.
No hallucination	Only references Acme products, current rating tables, actual policy data. If unsure: connects to human.
Conversation limit	Maximum 50 messages. Extended conversations redirected to prevent circular loops.

Escalation Path

Triggers when the customer asks for a human, Ace fails after 3 attempts, or the request is out of scope.

Creates a support ticket with the full conversation transcript — the human sees everything. No “can you repeat your issue?”

Privacy & Security

Measure	Implementation
Identity Verification	Email + last 4 of phone before accessing policy data. Quoting needs no auth.
PII Minimization	Policy reference IDs sent to LLM, not full records. Minimizes PII in API calls.
No PII Display	Never shows full payment info, SSN, or sensitive data in chat.
Conversation Storage	All conversations in Firestore. Customer informed at start.
Transparency	Identifies itself as AI at beginning of every conversation.

Measuring Success

Key Performance Metrics

Metric	What It Answers	Target
Quote-to-bind conversion (Ace vs. forms)	Does conversational quoting convert better? The money question.	Ace should beat forms by 20%+
Resolution rate	% completed without escalation	>80%
Cost per conversation	API tokens vs $8-12 call center	<$0.50
Customer satisfaction	Thumbs up/down ratio	>85% positive
Escalation rate & reasons	Where does Ace fall short?	<15%
Average conversation length	How many messages to complete task?	5-8 messages for quote

Architecture

Five-layer architecture from user interface to data persistence.

User

Customer / Agent

Chat UI

Widget + Dedicated Page

Conversation Service

History, Sessions, Auth, Escalation

LLM Abstraction

Provider-Agnostic Interface

Tool Registry

Quote, Lookup, Change, Pay, Info, Escalate

Existing Services

Rating Engine, Policy DB, Payments, UW Rules

Firestore

Policies, Quotes, Users, Conversations

Layer Descriptions

Layer	Responsibility
Chat UI	Widget + dedicated page. Manages display, input, session state.
Conversation Service	Message history, session persistence, identity verification, escalation.
LLM Abstraction	Provider-agnostic interface. Currently wraps Claude. Adding providers requires one interface implementation.
Tool Registry	Functions: generateQuote, lookupPolicy, requestChange, processPayment, getProductInfo, escalateToHuman. Each calls existing services.
Firestore	Policies, quotes, users, conversations, transactions. Single source shared with web app.

Implementation Playbook

Seven phases from definition to iteration. SME defines phases 1-2. AI builds phases 3-5. You measure and iterate 6-7.

Define the Agent

SME

Identity, personality, tone, scope, authority, guardrails, escalation, privacy. SME work.

Design the Authority Matrix

SME

Map every action to authority level. Only the insurance operations person can draw these lines. Single most important decision.

Build the Tool Layer

AI Build

Define functions agent calls. Each maps to existing service. Single source of truth.

Write the System Prompt

AI Build

Agent's brain. Structured: identity, personality, knowledge, tools, authority, guardrails, escalation.

Build the UI

AI Build

Chat widget + dedicated page. Conversation persistence. Feedback.

Measure Everything

Iterate

Six metrics minimum. Headline: does conversational quoting convert better than forms?

Iterate

Review escalations weekly. Review low-satisfaction. Expand tools. A/B test prompts.

Production Considerations

Estimated effort: 2-4 weeks compliance/legal, 1-2 weeks engineering. Parallel with agent build. No architectural changes needed.

Regulatory & Disclosure

State regulators are actively developing AI rules. Colorado SB 21-169 and similar legislation set the precedent.

Must identify as AI upfront. Multi-state compliance matrix needed.

Testing for unfair discrimination and disparate impact is required.

E&O Exposure

Conversational interfaces introduce ambiguity that forms avoid. Coverage representations carry risk.

Must recognize loss history disclosures.

Must stay on the information side of the advice vs. information line — only licensed agents can recommend specific coverage in many states.

Recordkeeping & Retention

Regulators require 3-7 year retention of customer communications. AI conversations ARE customer communications.

Must be searchable by customer, policy, and date.

Complete audit trails of every tool call with timestamps, inputs, and outputs.

Security & Adversarial Input

Prompt injection defense.

Social engineering to bypass identity verification.

Data exfiltration of rating tables and underwriting rules.

PII handling in LLM API calls.

What Comes Next

Roadmap

Horizon	Product	Description
Now	Consumer Ace	Quoting, purchasing, policy service. Prove the model.
Next	Agent Ace	Same capabilities for retail agents. Different authority level.
Then	Admin Ace	Underwriters/admins via natural language. "Show me all referred quotes."
Future	Voice + Proactive	Phone (Twilio). Renewal reminders. Coverage recommendations. Portfolio insights.

Progression

AI built the software → AI operates within the software → AI runs the business. Each step is achievable with current technology. The question is organizational willingness, not technical capability.

Sources

1.ProPair.ai (2025) — Insurance abandonment rate of 84%
2.RiskAdvisor / Drips.com (2022) — Abandonment rates and conversion benchmarks
3.Fenris Digital (2024) — 400% quote completion improvement through data prefill
4.FlowForma (2025) — Manual processes cost up to 40% more in operational expenses
5.LeanData via Drips.com — 78% choose company that responds first
6.Voxco via Drips.com — 88% demand more personalization
7.Colorado SB 21-169 — AI governance and disclosure in insurance