I recently built an OCR system specifically for mortgage underwriting, and the real-world accuracy is consistently around 96%. This wasn’t a lab benchmark. It’s running in production. For context, most underwriting workflows I saw were using a single generic OCR engine and were stuck around 70–72% accuracy. That low accuracy cascades into manual fixes, rechecks, delays, and large ops teams. By using a hybrid OCR architecture instead of a single OCR, designed around underwriting document types and validation, the firm was able to: • Reduce manual review dramatically • Cut processing time from days to minutes • Improve downstream risk analysis because the data was finally clean • Save ~$2M per year in operational costs The biggest takeaway for me: underwriting accuracy problems are usually not “AI problems”, they’re data extraction problems. Once the data is right, everything else becomes much easier. Happy to answer technical or non-technical questions if anyone’s working in lending or document automation.

My client uses an AI Agent (that was trained on Hatch) to talk to his customers. He wants information that the AI Agent gathers about the client such as name, service type etc. to land on his CRM (JobNimbus). How can this be done? Is it possible to gather data from phone calls to be transferred to the CRM?