A Pension Venture Studio Unlocked 80% Doc AI in 6 Weeks

The venture studio of a major North American pension plan (hundreds of billions in assets under management) was evaluating an insurance-focused venture whose viability hinged on one technical question: can AI reliably extract structured data from the industry's messiest documents? Statement-of-value files, sprawling Excel sheets, and OCR-heavy PDFs had to be parsed and understood before the studio could justify committing capital to a full build. The studio rarely outsources development, making this a targeted feasibility bet with a hard decision point at the end.

Lazer delivered a six-week pilot structured around honest feasibility testing. The first three weeks were heavy AI R&D: testing models, processes, and libraries against real customer data rather than curated samples. Table-transformer-style approaches failed outright on the OCR-heavy inputs. Anthropic's models significantly outperformed the alternatives on accuracy and were selected as the pipeline's parsing engine, with Weights & Biases running structured evaluations throughout.

Critically, Lazer designed the framework so the underlying AI model can be swapped as the studio learns from customer testing, an architecture decision that protects the venture against model-market shifts. The pipeline runs on Python with Temporal handling workflow orchestration on AWS, provisioned with Terraform and packaged with Docker, with endpoints exposed to support frontend work. The pilot closed with 80% model accuracy on insurance OCR data and a prototype the studio could take directly into market conversations.

Lazer scoped the engagement as a six-week feasibility pilot with a hard decision point at the end. The first three weeks were dedicated to AI research and development: the team tested competing models, processing approaches, and libraries directly against the studio's real customer data rather than curated samples. Table-transformer-style methods were tried first and failed outright on the OCR-heavy inputs, a result the team reported honestly rather than working around. Anthropic's models were evaluated alongside alternatives using a Weights & Biases harness that scored candidates on accuracy, and Claude was chosen as the parsing engine on the evidence. Anticipating that the model market would keep shifting, Lazer deliberately architected the framework so the underlying model could be swapped as the studio learned from customer testing. The remaining weeks went to engineering: a Python pipeline with Temporal handling workflow orchestration, provisioned on AWS with Terraform, packaged with Docker, and shipped through GitHub Actions, with API endpoints exposed to support frontend work and a prototype the studio could take into market conversations.

Venture studios, insurers, and financial institutions that need a rigorous, time-boxed feasibility answer on document AI before committing to a full product build.