Johnson Lambert Reinvents the Auditing Process in Insurance with Generative AI

Johnson Lambert is a CPA and consulting firm that has spent 35 years building its reputation on audit quality for insurance entities, nonprofits, and ERISA-qualified benefit plans. Partners from the firm first ran a GenAI workshop with Provectus in 2021.

01 The Challenge

Sixty to eighty hours per audit – and most of those hours were manual document processing

A single financial risk audit at Johnson Lambert could run from 60 to 80 hours. The significant part of that time was not spent on judgment. It was auditors converting PDF reports into CSV, extracting tables, and tracing values across the report corpus to check the numbers.

“It took auditors between 60 to 80 hours to complete a single financial risk audit, which diverted time and resources away from client-facing work.” · David Fuge · CIO at Johnson Lambert

The bottleneck was the pipe, not the people or their judgement. The firm needed a way to remove manual work from the audit workflow without weakening the record of the audit.

02 The Approach

Start narrow. Ship fast. Measure against the manual baseline.

Provectus started with discovery sessions specific to the Johnson Lambert audit methodology – not a generic workflow study. The team reviewed an audit firsthand: which tables matter, where tie-outs fail, what the reviewer needs to see before signing.

From that: a narrow first-pass target – table extraction and validation from unstructured PDF reports. Broad enough to move the hour count. Narrow enough to ship a prototype in two months. The agreement was measured: the prototype had to clear an efficiency gate against the manual baseline before production.

03 The Build

Amazon Bedrock plus tuned LLMs plus a reviewer UI for the auditors

The pipeline runs OCR on Amazon Textract (or, PyPDF/PyPDFium), then selects a proper model, hosted on Amazon Bedrock, to process table content for normalization and reference resolution. Table values, names, and cross-references land in a vector store that supports semantic search across reports.

Rigorous model evaluation led to a multi-model approach, with models from Amazon, Anthropic, and Google used for different tasks across the workflow. These included Amazon Nova Pro and Nova Lite, Anthropic Claude 3.5 Haiku and Claude 3.7 Sonnet, and Google Gemini 2.0 Flash and Gemini 2.5 Pro.

The solution’s auditor-facing layer is a thin UI:

• Upload a PDF
• See the extracted tables alongside the original
• Validate or correct values
• Export CSV for the downstream audit tool

Every correction made fine-tunes the model.

Amazon Bedrock was chosen to host the models for substitutability. Today the workload runs on Claude 3.7 Sonnet. Tomorrow it can move to another foundation model without a pipeline rearchitecture. The backend is an AWS Step Functions orchestrator that the Johnson Lambert engineers own.

04 The Results

From eighty hours to twelve. Twenty percent efficiency gain across the book.

80 h → 12 h

Per financial-risk audit

Measured against the pre-engagement baseline

The firm reported a 20% audit-efficiency improvement across the book and a 50% reduction in document-processing time. Auditors moved from doing tie-outs to reviewing them – the work they were hired to do.

The two-month prototype was the first gate. Production came next. The capacity that opened on the auditor side is now booked to new engagements; the firm is taking on more work without adding auditors.