Customer Case Study · Credit Operations · GenAI Document Intelligence
A home-loan file arrives as a stack of scanned identity, income, bank and property documents. HMC built a GenAI assistant that sorts every page, extracts every field, cross-checks the documents against each other, answers questions with citations and drafts the credit note, so underwriters spend their time on judgement, not reading.
Figures indicative; validated metrics available on request.
Each home-loan file carries a dozen or more documents in different formats. Before any credit judgement, someone had to sort them, type the key numbers into the system and check that they agree with each other.
Identity proofs, salary slips, Form 16, bank statements, sale agreements, title deeds and plans, all as scans.
Names, incomes, addresses and values typed in by hand, with typos that surface later.
A name variant, a salary that doesn't match the bank credit or a different tower number on the title, found late or not at all.
Most of the underwriter's time went into reading, leaving less for the actual credit decision.
GenAI does the reading and the reconciling; underwriters keep every decision.
Select a step to see what happened.
Mapped how a file is read today.
Deliverable: Document taxonomy and field list.
Turned tacit knowledge into rules.
Deliverable: Extraction schema and check library.
Four capabilities on one pipeline.
Deliverable: Working assistant and review screen.
Human-in-the-loop from day one.
Deliverable: Pilot playbook and review metrics.
The same pipeline grows.
Deliverable: Roll-out roadmap.
One synthetic loan file runs through all five modules. Pick a scenario in module 02 and the questions, answers and credit note in later modules change with it.
Eleven scanned pages arrive in random order. Press Process: each page is classified into a document type, then each document is read and its fields extracted as structured data with a confidence score.
Synthetic applicant and documents; identity numbers masked.
Choose a scenario and run the checks. The assistant compares fields across documents, applies policy rules and explains every result with its sources.
Pass = consistent. Review = explainable difference for a human to confirm. Fail = must be resolved before sanction.
Ask in plain English. The assistant retrieves the most relevant passages from this file, answers only from them and cites each source. Try asking something the file can't answer.
Passages used for the answer appear here, ranked by similarity.
Generate the draft note for the current scenario. Then set the confidence threshold: fields below it go to a human reviewer. Higher thresholds mean fewer errors slip through, but less straight-through processing.
Press "Draft credit note" to generate the underwriter's summary for this file.
Threshold curve from a simulated validation set of 4,000 fields; illustrative.
Files arrive through the day. In the top lane, people do data entry, checks and underwriting. In the bottom lane, the assistant reads and checks, people review flagged fields, and underwriters decide. Watch the queues.
Simulation with indicative stage times; not measured data.
Reading and re-keying shrink; the minutes left are spent on the credit decision.
Each dot is one loan file, minutes from intake to underwriter-ready
Per file, by type of work
Figures indicative; validated metrics available on request.
Structured fields with confidence scores are what make checks, routing and audit possible.
Underwriters trust an answer they can trace to a page; guardrails keep the assistant inside the file.
Let GenAI read and reconcile; let thresholds route doubt to people, and leave credit judgement with them.