RAG-Powered Agent Assist Tool for Insurance Contact Centers

allianz_direct · Insurance · 2024 · source Agentic: no · 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: Insurance contact center agent assist and policy-question answering

Allianz Direct needed contact center agents to answer policy terms-and-conditions questions faster and more accurately without replacing human agents or sending unreviewed GenAI responses directly to customers.

Constraints:

• The tool needed to assist human agents, not automate away customer conversations.
• Policy answers had to come from publicly available terms and conditions documents.
• Compliance confidence was higher because AI output went to an agent before the customer.
• The proof of concept had to improve accuracy over an existing GenAI tool and earn agent trust.

Architecture

Allianz Direct built a RAG-powered agent-assist tool on the Databricks Data Intelligence Platform using Mosaic AI. The system ingests product terms and conditions, retrieves relevant policy information for questions such as whether a family member can drive a car, and presents the answer to contact center agents in real time. Unity Catalog governs data and AI assets, AWS provides cloud infrastructure, Databricks notebooks support workflow changes, and human agents remain the final customer-facing decision point.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  terms-and-conditions-corpus["Terms and conditions corpus"] --> databricks-ingestion-workflow["Databricks ingestion workflow"]
  terms-and-conditions-corpus["Terms and conditions corpus"] --> rag-agent-assist["RAG agent assist"]
  databricks-ingestion-workflow["Databricks ingestion workflow"] --> unity-catalog["Unity Catalog"]
  databricks-ingestion-workflow["Databricks ingestion workflow"] --> rag-agent-assist["RAG agent assist"]
  unity-catalog["Unity Catalog"] --> aws-environment["AWS environment"]
  unity-catalog["Unity Catalog"] --> rag-agent-assist["RAG agent assist"]
  aws-environment["AWS environment"] --> mosaic-ai["Mosaic AI"]
  mosaic-ai["Mosaic AI"] --> rag-agent-assist["RAG agent assist"]
  rag-agent-assist["RAG agent assist"] --> contact-center-agent["Contact center agent"]
  contact-center-agent["Contact center agent"] --> customer-conversation["Customer conversation"]
  contact-center-agent["Contact center agent"] --> adoption-feedback-loop["Adoption feedback loop"]
  adoption-feedback-loop["Adoption feedback loop"] --> rag-agent-assist["RAG agent assist"]

Components

Data flow

Product terms and conditions are ingested into the Databricks environment and governed with Unity Catalog. When a customer asks a policy question, the contact center agent uses the GenAI agent-assist tool. The RAG application retrieves relevant terms and conditions, generates or surfaces the answer through Mosaic AI, and displays it to the human agent. The agent uses that information in the customer conversation, while proof-of-concept feedback and accuracy results guide further GenAI rollout.

Why this architecture

Allianz Direct chose an agent-assist RAG workflow because policy questions require accurate document grounding while the organization wanted humans to keep the customer relationship. Public terms and conditions reduced data-sharing risk, Unity Catalog provided governance, and keeping answers agent-facing gave compliance confidence while still improving speed and trust.

Models used: Databricks Mosaic AI · Databricks Agent Bricks Custom Agents · RAG agent assist

Scale

• The proof of concept improved answer accuracy by 10-15%.
  “demonstrated a 10-15% improvement in answer accuracy.”summary, stated
• The first production GenAI use case was built on the Databricks Data Intelligence Platform.
  “represents their first production GenAI use case built on the Databricks Data Intelligence Platform.”summary, stated
• The proof of concept is being rolled out across all Allianz Direct contact centers.
  “Field Corbett’s team rolls the proof of concept out across all Allianz Direct contact centers”article, stated
• The tool covers all Allianz Direct product terms and conditions documents.
  “The system ingests all of Allianz Direct's product terms and conditions documents”summary, stated

Alternatives considered (roads not taken)

Continue with the previous GenAI application

Rejected because: The Databricks proof of concept improved answer accuracy by 10-15% over the previous application.

Send AI responses directly to customers

Rejected because: Routing feedback to a human agent gave compliance and quality confidence.

Optimize primarily for call-time reduction

Rejected because: The stated goal was eliminating disliked back-office tasks so agents could spend more time with customers.

Build foundational AI infrastructure from scratch

Rejected because: Allianz already had Databricks as its data platform, making startup effort low.

Patterns

rag-retrieval (retrieval)

The agent-assist tool retrieves relevant insurance terms and conditions to answer policy questions.

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

AI output goes to a contact center agent rather than directly to the customer.

privacy-preserving-processing (governance)

The first use case uses publicly available terms and conditions, reducing data-sharing risk.

eval-driven-development (quality)

The proof of concept compared answer accuracy against the previous application and used agent trust/adoption as feedback.

tool-use-orchestration (architecture)

The contact center workflow orchestrates document retrieval, generated assistance, and human-agent response.

Tradeoffs

Assist agents rather than replace them

Chose Human-facing RAG assistance over Fully automated customer support.

Rationale: The goal was more meaningful customer conversations, not merely reducing call times.

Use public terms and conditions first

Chose Shareable policy documents over Sensitive or private customer data as the initial corpus.

Rationale: The public corpus reduced compliance risk and made the proof of concept easier to trust.

Build on existing Databricks foundation

Chose Databricks Data Intelligence Platform and Mosaic AI over A separate AI stack outside the existing data platform.

Rationale: Existing Databricks infrastructure reduced setup effort and kept data and AI workloads together.

Gotchas (what bit them)

Agent-assist success depends on human trust, not just accuracy metrics.

Root cause: Contact center agents will not use a tool that makes them second-guess answers or search multiple systems anyway.

Resolution: Improve accuracy and deliver information directly in the agent workflow so agents adopt it and suggest new use cases.

Customer-facing automation raises compliance confidence questions.

Root cause: Insurance policy answers affect customers and must be reliable.

Resolution: Keep the AI output agent-facing and use publicly available policy documents as source material.

Vendor/platform support can affect proof-of-concept speed.

Root cause: GenAI teams often need fast issue resolution and implementation guidance.

Resolution: Databricks engineers worked directly in Allianz Direct Slack channels during implementation.

Evaluation

Allianz Direct evaluated the proof of concept against the previous GenAI tool using answer accuracy, agent trust, adoption, and agent-suggested future use cases.

Metrics: answer accuracy uplift · agent trust · agent adoption · suggested additional use cases · implementation speed

Outcomes

• The Databricks/Mosaic AI proof of concept improved answer accuracy by 10-15% compared with the previous GenAI application.
• Contact center agents trusted the tool more and used it more often.
• The tool reduced the need to search multiple systems and second-guess policy answers.
• Allianz Direct planned to expand GenAI use cases across contact centers and company workflows.

Concrete anchors include a 10-15% accuracy uplift, all product terms and conditions documents, rollout across Allianz Direct contact centers, and the first production GenAI use case on the Databricks Data Intelligence Platform.