← Case studies

Case Study

Insurance-document automation for a specialty franchise lender

Proof-of-insurance verification cut from hours to under 30 minutes, with full audit lineage on every extraction.

The problem

A specialty lender serving the franchise finance market, a sub-100-person team that has originated well over $1B in loans across 200+ franchise brands, verified proof-of-insurance documents by hand. Every policy that arrived went through a loan officer reading a PDF and re-keying its fields into the origination platform.

Four things followed from that:

• Compliance exposure. Mis-keyed fields entered the system of record with no structured trace of who extracted what, or when.
• A bottleneck in underwriting. Verification took hours per loan, delaying approvals and borrower funding.
• Errors surfacing late. Mistakes surfaced at or after closing, when rework costs most.
• Senior time spent on data entry. Repetitive transcription consumed the hours loan officers should have spent on credit and underwriting judgment.

A loan officer could not tell a reliable read from an unreliable one. A clean scan and a bad fax carried the same implied confidence: none, stated.

What our team built

An event-driven extraction pipeline, triggered by the document upload itself, with no polling anywhere in the chain:

• Event trigger. A borrower uploads proof of insurance in the loan origination platform. A platform trigger publishes an event, which relays to the cloud event bus in real time.
• Serverless extraction. A function authenticates back to the origination platform over JWT bearer auth, streams the PDF via its REST API, and writes it to encrypted object storage with a structured metadata sidecar.
• Document AI parsing. Object-created events auto-ingest the file into the data platform in under 60 seconds; a scheduled task then runs document parsing plus an LLM extraction step that returns seven structured fields: policy number, carrier, effective date, expiration date, coverage amount, insured name, and coverage types.
• Confidence scoring. Each extraction carries a model confidence score, so a low-quality read routes to a human rather than entering the system at face value.
• Security posture. JWT bearer auth, server-side KMS encryption at rest, keyless IAM role trust with no static credentials, and least-privilege function permissions.
• Config-driven extensibility. New document types onboard through a configuration record, with no function or pipeline code changes.

The outcome

• Zero manual touchpoints. Loan officers no longer handle insurance documents after upload.
• Hours to under 30 minutes typical end-to-end latency, worst case under 90 minutes, inside the platform's one-hour SLA.
• Seven fields, structured and queryable, replacing free-form transcription.
• Complete audit lineage. Every extracted row traces to its source PDF by content-version ID and storage key, with no gaps.
• A reusable pattern. The same pipeline now extends to flood certificates, title insurance, and appraisals by configuration.

The boundary of this claim

We measured the latency and field-count figures from the delivered pipeline. Full lineage and confidence scoring replace an unaudited manual read. We make no claim about any specific audit outcome, and no client revenue figures.

Sizing basis: a specialty online lender in this segment typically runs 50–100 employees with $1B+ in cumulative originations and a network of 100–250+ partner brands.

Tools & technology

• Salesforce / nCino
• Apex + Platform Events
• Event Relay
• AWS EventBridge
• AWS Lambda (Python)
• Amazon S3 (SSE-KMS)
• Snowpipe
• Snowflake Document AI
• Snowflake Cortex
• JWT bearer auth

Not ready to book yet? Read another case study — legacy platform modernization or test-automation acceleration.