Multi-Agent AI Platform for Life Insurance Sales Acceleration

prudential · Insurance · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Life-insurance sales, underwriting, quoting, forms, and advisor operations

Prudential needed to reduce friction in a personalized, regulated life-insurance sales journey that can take 8 to 10 weeks, spans prospecting, quoting, forms, underwriting, issuance, and long-term policy management, and forces advisors through PDFs, spreadsheets, portals, calls, emails, and state/product-specific rules.

Constraints:

• Life insurance is highly personalized by age, lifestyle, health conditions, state, carrier, product, riders, and policy type.
• Incorrect forms or recommendations can delay or reject applications in a regulated financial-products environment.
• Medical underwriting, product recommendations, forms, quotes, and book-of-business management each have distinct data sources and logic.
• The system must preserve advisor-specific short-term and long-term context while authenticating users by OIDC ID.
• Complex medical underwriting cases still require human review rather than full automation.

Architecture

Prudential built Just Ask as an advisor intelligence fabric with a chat interface, non-deterministic intent-routing orchestrator, specialized agents for medical underwriting, quick quote, forms, product, book of business, and future illustrations, plus a shared platform layer. Raw forms, products, underwriting manuals, and other datasets are ingested through EventBridge into embeddings and a vector database. Each agent can access DynamoDB-backed chat history, short/long-term memory by advisor OIDC ID, the RAG pipeline, Redis local retrieval/cache, a GenAI gateway connected to OpenAI and Bedrock endpoints through an LLM model registry, Bedrock and custom guardrails, agent-core monitoring/observability, prompt management, and human/rule/LLM-as-judge evaluation.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  Apps["Advisor and wholesaler applications"] --> Orchestrator["Orchestration agent"]
  Orchestrator --> UW["Medical underwriting agent"]
  Orchestrator --> Quote["Quick quote agent"]
  Orchestrator --> Forms["Forms agent"]
  Orchestrator --> Product["Product agent"]
  Orchestrator --> Book["Book of business agent"]
  Orchestrator --> Illustration["Illustrations agent"]
  Orchestrator --> Memory["DynamoDB conversation memory"]
  FormsData["Forms and product data"] --> EventBridge["EventBridge ingestion"]
  UWData["Underwriting manuals and decisions"] --> EventBridge
  Policies["Policy history data"] --> Embeddings["Embedding and vector space"]
  EventBridge --> Embeddings
  Embeddings --> Vector["Vector database"]
  Vector --> Redis["Redis retrieval and cache"]
  Product --> Redis
  UW --> UWData
  Book --> Policies
  UW --> Gateway["GenAI gateway"]
  Quote --> Gateway
  Product --> Gateway
  Gateway --> Registry["LLM model registry"]
  Orchestrator --> Guardrails["Bedrock and custom guardrails"]
  UW --> Guardrails
  Product --> Guardrails
  Orchestrator --> Core["Agent core"]
  Core --> Eval["Evaluation framework"]
  Eval --> Human["Human review"]
  Core --> A2A["A2A layer"]

Components

Data flow

An advisor or wholesaler logs into an application with single sign-on and asks Just Ask a plain-language question. The orchestrator stores chat history and memory against the user's OIDC ID, interprets intent, and routes to the appropriate specialized agent. Agents retrieve embedded content from vector storage, use Redis for local retrieval/cache, call the GenAI gateway for generation/reasoning, apply Bedrock and custom guardrails, and return product, quote, form, underwriting, or book-of-business answers. Agent core collects monitoring, logging, alerting, prompt management, and evaluation signals, while future A2A work aims to let agents run independently in other Prudential systems.

Why this architecture

Prudential avoided a monolithic assistant because product selection, quoting, forms, underwriting, and book-of-business management each use different data sources, logic, complexity, and regulatory constraints. A single conversational entry point keeps advisor experience simple while the orchestrator handles routing to specialized agents. Shared memory, retrieval, guardrails, monitoring, prompt management, and evaluation layers let those agents reuse platform services without losing domain-specific behavior.

Models used: Small language models for underwriting reasoning · OpenAI endpoints · AWS Bedrock endpoints · AWS Bedrock guardrails · Cohere rerank · Semantic search algorithms · LLM-as-a-judge evaluation · Embedding models and vector database retrieval

Scale

• The US life-insurance coverage gap is described as close to $12T.
  “in US alone, this entire coverage gap in the market is close to $12 trillion,”transcript, stated
• A single life-insurance policy can take up to 8 to 10 weeks from first conversation to issuance.
  “A single policy can take up to 8 to 10 weeks, right from the first conversation all the way till issuing of the policy.”transcript, stated
• Advisors often run five to ten quote scenarios before presenting options to clients.
  “advisors usually run five to 10 scenarios before even presenting something to the customer to the client.”transcript, stated
• Manual underwriting inquiries can take 24 to 48 hours.
  “they would probably reach out to an underwriter manually. It's an email conversation. Usually, it takes like 24 to 48 hours today.”transcript, stated
• Forms or product/service-center inquiries can take 10 to 15 minutes.
  “If they want to get a product or like a form, again, like they have to call advisor support line or service center. It takes like 10 to 15 minutes.”transcript, stated
• Advisor portal credential workflows can take five to ten minutes.
  “they need to access advisor portal, which is like 5 to 10 minutes for credential alone.”transcript, stated
• The underwriting agent uses more than 700 medical underwriting documents or conditions and 20+ reasoning steps.
  “took all the 700 underwriting documents or more than 700 medical underwriting conditions that we have in place and trained them with past underwriting decisions”transcript, stated

  “came up with 20 plus reasoning steps and give back the decision back to the end customer.”transcript, stated
• The forms agent covers more than 600 forms.
  “Same thing with forms, more than 600 forms, they can actually get it in terms of like what state”transcript, stated
• The product agent covers more than 400 product content items.
  “same with product agent, more than 400 plus product content is available to them.”transcript, stated
• The first 12 weeks processed 1,800 messages, 1,200 queries, 900+ financial planners, and 550+ organizations.
  “in just over 12 weeks, the platform ended up processing 1,800 messages. It handled 1,200 distinct queries.”transcript, stated

  “we reached 900 plus unique financial planners out there across 550 plus external organizations.”transcript, stated

Alternatives considered (roads not taken)

Build one monolithic AI assistant for the entire life-insurance sales journey

Rejected because: Prudential found multiple problem domains with different data sources, logic, and complexity, so it built an agentic platform with specialized agents.

Make advisors pick the correct tool or agent manually

Rejected because: Just Ask uses a plain-language chat interface and an orchestrator that routes queries by intent to the right agent.

Fully automate every medical underwriting case

Rejected because: The medical underwriting agent handles many situations, but complex conditions still use human-in-the-loop oversight.

Use generic platform guardrails only

Rejected because: Prudential combines Bedrock guardrails with custom guardrails at orchestrator and agent levels to satisfy regulated insurance-function requirements.

Keep agents only inside a central chat interface

Rejected because: The roadmap is an A2A layer where agents can work on their own and be embedded inside other systems.

Patterns

multi-agent-orchestration (architecture)

Just Ask uses an orchestrator to route advisor questions to specialized medical underwriting, quick quote, forms, product, book-of-business, and future illustration agents.

rag-retrieval (knowledge)

The product and agent workflows ingest forms, products, and underwriting manuals into embeddings and a vector database, then retrieve relevant information with semantic search and reranking.

agent-memory (state)

The platform stores chat history and short/long-term conversation context in DynamoDB against each advisor's OIDC ID.

safety-policy-enforcement (governance)

Prudential applies Bedrock guardrails and custom per-agent/orchestrator guardrails to control information in a heavily regulated insurance environment.

observability-tracing (operations)

Agent core centralizes monitoring, observability, logging, alerting, and prompt management for the production agent ecosystem.

eval-driven-development (quality)

The platform evaluates agents with human-in-the-loop, rule-based, and LLM-as-judge methods and reports 90%+ accuracy for most agents.

human-in-the-loop-gate (governance)

Complex medical underwriting remains human-reviewed even as common underwriting questions are automated.

context-engineering (context)

The system explicitly manages short-term and long-term advisor memory, conversation retention windows, and context grounding for long policy relationships.

Tradeoffs

How to expose multiple specialized capabilities to advisors.

Chose A single conversational interface with intent-routing orchestration. over Separate advisor tools for product, quote, forms, underwriting, and book-of-business tasks..

Rationale: Advisors can ask in plain language while the system routes to the right specialized agent behind the scenes.

How to split the platform into agents.

Chose Purpose-built agents per problem domain. over One monolithic AI solution..

Rationale: Different insurance tasks have different data sources, logic, complexity, and compliance needs.

How much medical underwriting to automate.

Chose Automate common underwriting reasoning but retain human-in-the-loop for complex conditions. over Fully autonomous underwriting for every case..

Rationale: The system reaches high reported accuracy, but complex medical conditions still need human oversight in a high-stakes domain.

How to evaluate production agents.

Chose A multi-method evaluation framework: human-in-the-loop, rule-based, and LLM-as-judge. over A single evaluation method for all agent types..

Rationale: Different agent behaviors require different checks, and the source frames evaluation as necessary for scaling and measuring accuracy.

How to evolve beyond chat-only access.

Chose Build toward an A2A layer and reusable agents embedded in other Prudential systems. over Keeping all agents accessible only through the Just Ask chat layer..

Rationale: The roadmap treats agents as reusable services that can work on their own and be embedded where workflows need them.

Gotchas (what bit them)

A monolithic AI assistant would blur distinct insurance-sales problem domains.

Root cause: Product, quoting, forms, and book-of-business tasks have different data sources, logic, and complexity.

Resolution: Prudential built an agentic platform with specialized agents behind a single chat experience.

Medical underwriting has too many condition interactions for a simple lookup.

Root cause: Pricing depends on medication type, diabetes control, other medical conditions, and hundreds of underwriting conditions.

Resolution: The underwriting agent uses underwriting documents, past decisions, models, and 20+ reasoning steps, with humans kept for complex conditions.

Wrong forms can restart the long application process.

Root cause: Forms vary by product, policy type, and state-specific rules.

Resolution: The forms agent retrieves the right form by state and product through natural language.

Insurance agents need continuity across very long customer relationships.

Root cause: Life-insurance policies can span decades, and advisors may revisit clients years after prior interactions.

Resolution: The platform preserves advisor-specific memory/context and grounds book-of-business recommendations in historical policy data.

Regulated insurance advice cannot rely only on model behavior.

Root cause: The information that can be provided is constrained by compliance and legal requirements.

Resolution: Prudential uses responsible AI, Bedrock guardrails, and custom guardrails at orchestrator and agent levels.

Evaluation

Prudential evaluates Just Ask through agent-core monitoring/observability, logging and alerting, prompt management, human-in-the-loop evaluation, rule-based evaluation, LLM-as-judge evaluation, reported 90%+ accuracy for most agents, financial-professional verification against human underwriters, and production adoption/usage metrics over the first 12 weeks.

Metrics: 90%+ accuracy for most agents · 1,800 messages in just over 12 weeks · 1,200 distinct queries · 900+ unique financial planners · 550+ external organizations · 100+ successful quotes · 4,500 human hours saved · 250+ forms interactions · 2,500+ minutes saved · 175% forms-agent adoption growth · 4x retention improvement · 50%+ organizational reach expansion

Outcomes

• Just Ask reached real adoption across Prudential's distribution network within 12 weeks.
• The platform processed 1,800 messages and 1,200 queries from 900+ financial planners across 550+ organizations.
• The quoting agent produced 100+ successful quotes and saved about 4,500 human hours.
• The forms agent resolved 250+ interactions and saved 2,500+ minutes.
• Forms-agent adoption grew 175%, retention improved 4x, and organizational reach expanded over 50%.
• Advisors reported they did not want to return to the old way of working.

The source describes a $12T US coverage gap, 8-10 week policy cycles, 20-30 year policy lifetimes, 150,000 target financial professionals, and first-12-week production usage across 900+ planners and 550+ organizations.