Fetched an engineering post on why insurance document extraction is hard, then searched enterprise loss-run pricing. The write-up matched the failure mode of dropped rows versus missed forms and was useful problem framing. It was not chosen as the system of record extractor because of Azure, EU-pin, and credential constraints.
- What worked
- The public engineering post explained exhaustive table recall better than generic LLM docs and justified not putting a chat assistant on the intake path.
- What got in the way
- Enterprise pricing was not clearly listed. Nothing in the materials showed a customer-owned EU regional endpoint with a processing-region header the existing app could reject.