Almost none of the elapsed time in an origination is work. It is waiting, for a document to arrive, for a check to be performed, for an exception to be routed to somebody with authority, for a second person to review what the first person did. The actual processing time is a fraction of the cycle time, which is why adding staff produces less improvement than anyone expects.
Document handling is where the waiting concentrates. A file arrives incomplete, which is discovered by a person reading it, who then requests what is missing, which arrives days later and is also incomplete. Each of those cycles is a customer experience as well as a cost, and in a competitive origination market the cycle time is frequently the reason a deal goes elsewhere.
The control environment is not the problem, though it is usually blamed. Dual review, exception routing and audit trails exist because an examiner will ask, and removing them to go faster is a trade nobody sane makes. The waste is not in the controls themselves; it is in the manual assembly of the evidence that the controls were followed, and in the human effort spent re-reading documents to check things a machine can check reliably.
Reporting sits on top of all this, assembled by hand from several systems on a cycle, with definitions that vary slightly depending on who built which return. Regulatory and internal reporting draw from overlapping data with different rules, and reconciling them is somebody’s recurring week.
The thing that makes this sector distinctive is not the volume; it is that an examiner will ask how a decision was reached and expects a coherent answer. That single requirement rules out any design where a model’s output is the decision, and rules in a design where the model narrows and evidences and a person decides.
