01 Conversational AI at enterprise scale
Enterprise Insurance AI Chatbot
Designed, prototyped, and implemented an AI-powered chatbot and agent interface for a Fortune 500 commercial insurance company — trained on products, solutions, and agent directory data — integrating Vertex AI and LLM APIs to deliver intelligent, conversational user experiences at enterprise scale.
Fortune 500 Insurance
Simplifying commercial underwriting
One of the largest U.S. commercial property and casualty insurance companies. To write and manage complex business insurance policies, thousands of independent agents rely daily on massive catalogs of regional product rules, appetite guidelines, and directory databases.
With dozens of specialized industries to cover, finding specific policy guidelines quickly is a critical operational bottleneck.
The Problem
Lost time in the search for policy guidelines
Agents and underwriters spend significant portions of their days searching through dense PDF documents, legacy regional appetite matrices, and internal personnel directories to answer basic coverage questions.
This search friction creates quoting delays, increases tier-1 support volumes, and occasionally leads to agents abandoning quotes in favor of faster competitors.
Impact
Measurable gains in agent efficiency
Measured across 60 days post-launch with the target agent cohort. Search time and adoption from portal analytics; answer accuracy from in-session thumbs feedback collected across 2,400+ queries.
- Search Time
- -62% Avg. guideline lookup time — 12 min to ~4.5 min
- Tier-1 Deflection
- 34% Agent support queries resolved without human escalation
- Answer Accuracy
- 89% Of AI responses rated accurate by agents via in-session feedback
- Active Adoption
- 71% Of target agents actively using the tool within 60 days
Kick Off
Co-designing with domain experts
AI is only as good as the guardrails we define. To build a system capable of parsing complex policy rules without "hallucinating" terms, we established a daily feedback cadence with commercial insurance underwriters, prompt engineers, and independent brokers.
User Perspective
Understanding what agents face
We analyzed daily query workflows to map where manual search slowed down quoting speeds, categorized by task complexity.
“What is the maximum building limit for Retail Class 3 in Illinois?”
The Manual Friction
Finding this requires downloading a 120-page regional PDF appetite guide, scrolling to class guidelines, and cross-referencing the Illinois-specific state exception table. (Takes 5–10 mins)
The AI Chatbot Resolution
The chatbot retrieves the exact dollar limit and page reference instantly from the vector store search, resolving in under 2 seconds.
“Can we write General Liability for a local carpentry contractor that performs 15% roofing work in Indiana?”
The Manual Friction
Requires checking both the Carpentry class guidelines and the special Roofing exclusion rules in separate endorsement PDFs, then manually calculating if the 15% exposure exceeds regional limits. (Takes 15–20 mins)
The AI Chatbot Resolution
The chatbot cross-references both policy booklets, evaluates the exposure threshold, and displays the exact citation in the side drawer. (Resolves in 4 seconds)
“A client operates a multi-state food truck and catering business in Michigan and Ohio. They need General Liability with a Liquor Liability endorsement. Is this writeable, and who handles this region?”
The Manual Friction
Agents must verify multi-state licensing rules, catering classifications, and liquor endorsements across multiple product lines, often resulting in complex queries and long phone wait times. (Takes 45+ mins)
The AI Chatbot Resolution
Instantly checks policy exceptions for both states, outlines liquor rules, and displays a direct underwriter card with pre-filled case details for quick escalation. (Resolves in 8 seconds)
Strategic Focus
Focusing the model's capabilities
Because LLMs are broad, we had to narrow down the assistant's functional boundaries to focus on high-impact tasks while deferring more complex, transactional capabilities to future iterations.
NOW
Immediate release goals
- Policy RAG text lookup
- Direct policy page/section citations
- Regional underwriting contact routing
- Structured search auto-suggestions
LATER
Deferred backlog
- Interactive policy premium quoting
- Dynamic coverage endorsement additions
- Voice-activated dictation commands
Deliver commercial policy answers and regional underwriting directory details cleanly and reliably in under 10 seconds.
Prioritizing Effort
Optimizing the chat experience
Conversational interfaces can easily lead to open-ended confusion. We evaluated our prompt mechanics and UI helpers to ensure the system guided users toward successful queries.
Hypotheses developed
How we defined the MVP
Design Showcase
Conversational flows, grounded in reality
Conversational patterns require unique visual structures. We kept conversational threads contextualized with inline policy snapshots and contact tools so agents never hit a dead-end.
The Chat Interface
The primary interaction window. It features quick-select question bubbles, a real-time conversational field, and visual feedback selectors so users can rate answer quality on the fly.
Policy Assistant
Vertex AI Agent
Suggested Topics
- Suggested Topics — Chips that surface common policy updates based on recent regional guidelines, reducing typing effort.
- Direct Actions — Integrated positive/negative feedback nodes capture alignment data with zero workflow disruption.
Inline Source Citations
Clicking on any citation triggers a side-rail document drawer. This reveals a cropped visual preview of the corresponding policy rule page, guaranteeing accuracy without pulling agents out of the app.
- Interactive Document Inspector — Highlighted policy segments render dynamically alongside the chatbot canvas.
- Download Integration — Provides immediate direct link paths to official policy PDFs for broker reference documentation.
The Underwriter Finder
When a scenario requires direct underwriting consultation, the assistant converts the conversational request into a structured regional contact panel, displaying phone, email, and territorial guidelines.
Sarah Jenkins, CPCU
Lead Middle Market Underwriter, Great Lakes Region
Middle Market Property
Inland Marine Coverage
Ohio, Michigan, Indiana
- Contact Card Layout — Converts natural language searches for regional underwriters into clear, functional directory elements.
- Direct Routing — Integrated messaging actions minimize telephone handoff friction, allowing agents to instantly query local underwriting teams.
Design Operations
Maintaining design system consistency
We tracked model quality and design component metrics to ensure a smooth, predictable system rollout.
- Prompts Tested
- 125
- Model Iterations
- 18
- User Review Sessions
- 42
- UI Components
- 35+
Closing Reflection
Coexistence of technology and experience
Enterprise insurance workflows are deeply complex, and AI models cannot fully replace the years of experience commercial underwriters bring to risk assessment.
By treating AI not as an autonomous decision-maker, but as a contextual retrieval system, we built a tool that empowers brokers to quote policies quicker, safer, and with fewer errors. Supporting human expertise remains the most reliable vector for enterprise AI integration.