Underwriting assistant for the personal discussion

The leak

For borrowers without payslips, the lender's credit decision rests on the personal discussion: a field agent visits the customer, picks the category they fall into, a grocery shop, a restaurant, a trade, and works through a questionnaire of around a hundred questions about the business, its income, its expenses and its risks. The agent interviews, listens, and fills the form. In practice, filling the form competes with listening. Answers get summarised, sections get skipped, and the gaps show up later at credit review, when the agent is three visits away and the customer is waiting.

After the form comes the calculation: how much can be lent, is it within policy, what are the risk flags. That was a second manual step with its own delays and its own inconsistencies between underwriters.

The constraint

The interview happens in the field, often with poor connectivity, in whatever language the customer speaks. The recording is personal financial information and had to stay on infrastructure the company controls. The transcription had to separate the agent from the customer and keep up with the conversation, because the point is to fill the form while the agent is still in the room, not afterwards. And the system is decision support: it fills, flags and calculates, and the credit decision stays with a person and a policy.

The system

The agent selects the customer's category and the questionnaire for that category appears, sectioned. As the interview is recorded, the system transcribes and diarises it live and analyses it section by section, filling in every answer it can hear and leaving visible what it cannot. The agent corrects and confirms rather than types. When a section is thin, the system says so while the customer is still there, so the agent asks the missing question instead of discovering the gap a week later.

With the form complete, the system runs the numbers: the income assessment, the loan amount the policy allows, the affordability check, and the risk flags, each with the policy rule it applied, so an underwriter sees the reasoning and not only the result. The transcript, the form and the calculation are stored together as the record of the discussion.

The next version brings the proofs into the same analysis: bank statements, photographs of the customer's own account books, bills and receipts, read through the document intelligence pipeline and reconciled against what the customer said.

The number

The form filled section by section from the interview itself, gaps flagged in the room, and the eligibility calculation produced from the completed form with its rules visible. In pilot with credit, at the request of the risk function.

What I would do differently

Pilot in one branch with one category before generalising the questionnaire. The system handles every category, but the first branch's grocery shops taught more about how agents actually ask questions than the design phase did. Starting narrow would have reached that learning sooner.