The extraction accuracy on our highest volume document types was better than anything we tested during evaluation and it has only improved since we went live. Our operations team spent the first few weeks genuinely surprised by how little they needed to correct and that surprise fairly quickly turned into a new baseline expectation for how document processing should work. The confidence scoring on individual fields has also meant that human review time gets spent where it actually matters rather than uniformly across everything the platform processes. Review collected by and hosted on G2.com.
Getting the model trained on some of our older document formats that have inconsistent layouts took considerably more time and iteration cycles than we budgeted for upfront. Review collected by and hosted on G2.com.



