case study 01 · ai-driven data extraction
AI submissions extraction at scale
Commercial insurance runs on submissions: the broker packs of documents that describe a risk before it is quoted. At Concirrus these arrived as PDFs, Excel files, email bodies and attachments, sometimes handwritten, each with a different template and schema depending on the line of business: Aviation, Political Violence & Terrorism, Logistics, Construction. Extracting the underwriting data was manual, slow, and the bottleneck across all four lines.
The problem
This was not extract ten fields from a file. A single submission carries hundreds of data points across multiple file types with completely different schemas per line of business. A generic extractor does not survive contact with that variety, and the accuracy bar is contractual: underwriting decisions and pricing sit on top of this data.
What we built
An ingestion pipeline that receives submissions by email, normalises the documents, applies OCR where the source is image or handwriting, and runs extraction over the result. The design is agent-shaped: sense what the document is, decide which schema applies, extract with confidence, and flag anything below the confidence bar for a human.
The extraction runs on LLM APIs rather than custom models, so the system inherits model improvements instead of fighting them. Around 70% of submissions run end to end with no human touch. The rest flow through human-in-the-loop validation, which is a design decision, not a fallback: humans own the judgement where the cost of being wrong is highest.
Multi-tenant by design
Every client sees only their own data and their own line-of-business rules, on shared infrastructure. Extraction quality is measured per tenant and per line, because a model that nails Aviation documents can still stumble on Construction.
Cost as a first-class metric
LLM inference is not free, and at roughly a thousand documents a month per line it is not negligible either. I owned inference cost modelling for the pipeline and fed it into pricing and go/no-go decisions, so accuracy and unit economics were optimised together rather than traded.
The result
The pipeline moved submission processing from the critical path of the business to a managed, measurable capability: roughly a thousand submissions per month per line at 95% field-level accuracy, with humans reviewing only what the model flags.
Submission processing went from a manual bottleneck to a managed pipeline at 95% field-level accuracy, with humans reviewing only what the model flags.