Product
How InsurGPT Reads Messy Insurance Documents
Handwritten notes, faxed scans of scans, thousand-page demand packages. Two minutes inside the reading engine that turns them into answers you can defend, and the benchmark numbers behind it.
Video not playing? Switch video player.
About this video
Every AI agent stands on whether it can read real insurance correspondence: handwritten notes, faxed scans of scans, nested tables, and thousand-page demand packages with records out of order. This two-minute video shows how InsurGPT™, Bevaya's insurance-specific model, reads them.
InsurGPT is a mosaic, not a single model: eight specialized models across four layers, each doing one narrow job. Page stream segmentation finds where each document in a 1,014-page bodily injury demand package starts and stops and routes every piece to the model built for it. Where most pipelines flatten a page into text and let columns collapse, InsurGPT preserves the 2D structure, so every value keeps its cell and every cell keeps its pixels. A verifier model then argues with the extraction, rejecting answers that make no insurance sense before anyone sees them.
In Bevaya's July 2026 benchmark on 346 underwriting samples, that architecture delivered 93% macro accuracy on underwriting field extraction, nearly eight points above the best frontier model, and returned 53% of documents with every field right versus 27% for the best frontier model.
Key takeaways
Not a model, a mosaic
InsurGPT is eight specialized models across four layers: perception, routing, reasoning, and adjudication, each doing one narrow job.
Page stream segmentation splits the pile
A 1,014-page bodily injury demand package is split into its documents and each piece is routed to the model built for it.
Every value keeps its cell
The 2D structure of forms and tables is preserved, and a verifier model rejects answers that make no insurance sense, like a closed claim carrying an open reserve.
Measured against the frontier
93% macro accuracy on underwriting extraction and 53% of documents returned with every field right, versus 27% for the best frontier model (Bevaya benchmark, July 2026).
Full transcript
Every agent stands on one thing. Can it actually read the correspondence that drives your processes? Not a clean PDF. The real ones. Handwritten. A faxed scan of a scan. A table nested inside a table. And InsurGPT™ isn't a model, it's a mosaic. Eight specialized models across four layers. Each doing one narrow job, arranged so that the answer beats anything one model produces alone.
A bodily injury demand lands as one PDF. A thousand pages, notebook marks, records out of order. Page stream segmentation finds where each document starts and stops and sends every piece to the model built for it. Frontier models truncate at their context limit and never tell you. Or they just flat out give up.
Most pipelines flatten a page into text, columns collapse, and a closed claim inherits an open reserve. Ours keeps the two dimensions. Every value keeps its cell, and every cell keeps its pixels.
Then a second model argues with the first. It knows a closed claim carries no reserve, and that paid dollars can't exceed total incurred. It rejects the answer and sends it back before you ever see it.
93% macro accuracy on underwriting extraction, nearly eight points above the best frontier model. But the number that clears a queue is this one. Documents that come back with every field right. For us, 53%. For them, 27%.
Frequently asked questions
How accurate is InsurGPT at insurance document extraction?
In Bevaya's July 2026 benchmark on 346 underwriting samples, InsurGPT reached 93% macro accuracy on underwriting field extraction, nearly eight points above the best frontier model, and returned 53% of documents with every field right versus 27% for the best frontier model. Full tables are published at bevaya.ai/benchmarks.
What kinds of insurance documents can it read?
The documents that drive real insurance processes: handwritten notes, faxed scans of scans, nested tables, ACORD forms, loss runs, medical records and bills, and thousand-page bodily injury demand packages with no bookmarks and records out of order.
How is this different from sending documents to a single large model?
Frontier models truncate at their context limit without saying so, and most pipelines flatten pages into text, collapsing the column structure that gives values their meaning. InsurGPT uses eight specialized models across four layers: page stream segmentation splits a package into its documents, extraction preserves the 2D structure of every form, and a verifier model rejects answers that violate insurance logic before a person sees them.
What is page stream segmentation?
It is the layer that finds where each document inside a combined PDF starts and stops. A demand package arrives as one file containing a demand letter, police report, medical records, imaging reports, and itemized bills; segmentation separates them and routes each to the model built for that document type.
Related resources
Get new sessions first
New webinars, customer stories, and product videos from Bevaya.
- New webinars and interviews as soon as they go live
- Customer stories from insurers putting AI Agents into production
- Product walkthroughs of new Bevaya capabilities
