Maximizing the Performance of Investment Products with GenAI Knowledge Search

Venerable is a leader in the insurance industry, specializing in the acquisition and management of legacy variable annuity portfolios. The company focuses on maximizing the performance, value, and stability of these investment products through strategic asset management. With backing from experienced investors, Venerable leverages innovative technologies and financial expertise to ensure reliable and beneficial outcomes for all of their customers and stakeholders.

01 The Challenge

A fragmented internal knowledge base is a tax on every employee’s day

Before the engagement, roughly half of Venerable’s employees were satisfied with the search experience. Only 5% found what they needed on the first try. Time lost to hunting for policies, handbook answers, and engineering SOPs exploded across the organization — most painfully in customer-facing roles where the next right answer is a policyholder waiting.

The leadership wanted to use generative AI to remove the search tax. Not to replace the expertise. To make the expertise reachable.

02 The Approach

Narrow first: HR policies, employee handbook, engineering SOPs

Provectus opened with a discovery phase that mapped pain points, readiness, and technology landscape. The team and the client agreed to scope the first engagement tightly: search across HR policies, the employee handbook, and engineering Standard Operating Procedures. Narrow enough to ship. Real enough to measure.

An evaluation set was built from high-priority employee questions. Model options were compared on answer quality, cost, and fit. Each prototype iteration was reviewed by real employees, not a curated panel.

03 The Build

Claude 3 Haiku on Bedrock, LlamaIndex, a conversational UI

The assistant is built in Python with LlamaIndex, FastAPI, and Swagger. It handles PDF, DOCX, PPTX, and other formats. Core models are Amazon Titan Embeddings and Anthropic’s Claude 3 Haiku, both on Amazon Bedrock.

Model selection was a measured decision. Claude 3 Haiku cleared the quality bar on the question set at 95% accuracy, ran 12x cheaper than Claude 3 Sonnet, and ran 40x more cost-efficiently than GPT-4 Turbo. That cost curve is what makes continuous operation viable – the assistant is on, for every employee, every day.

The UI lets employees query in plain language and rate the answers they receive. Every rating feeds calibration.

04 The Results

90% reduction in time-to-answer, 95% accuracy on the production set

In the insurance industry, regulatory compliance, operational efficiency, and customer service quality determine a business’s success or failure. For companies like Venerable, efficient employee time and resource management is critical, and even a simple improvement in back-end and customer-centric processes can make a significant difference – for both the client’s business and its customers.

3 sec

Average response time

With source citation

The Intelligent Search Assistant answers 100% of the top ten employee questions correctly and 93% of the full sampled question set accurately. Average response time dropped 90% against the manual baseline.

Before the engagement, 5% of employees found what they needed on the first try. After, the large majority do – in seconds, with a link to the source document for verification.

05 What’s Next

The platform carries forward to the next use case

The LlamaIndex + Bedrock stack is documented for AWS deployment, Dockerized for local development, and deliberately shaped for reuse. Venerable continues the engagement on follow-on use cases that can sit on the same foundation.