Power Finance Corporation managed a ₹40,000 Cr infrastructure loan portfolio. Important early-warning signals were buried in PDFs, audit notes, and fragmented reviews that no person could inspect systematically at the required cadence.
The obvious automation was not the whole problem.
Extracting entities was only one part of the job. Sensitive data, role-specific access, and operator trust defined whether the intelligence could become part of the credit workflow.
In regulated systems, trust is an architectural constraint. A useful signal must also be explainable, appropriately visible, and connected to an accountable decision process.
Build the operating loop, not only the intelligent step.
- 01
Built named-entity extraction around concrete risk events such as liens, disputes, and construction delays.
- 02
Structured the output for repeatable review rather than producing open-ended model summaries.
- 03
Designed role-based dashboards with automated PII masking and data minimisation.
- 04
Worked with loan officers to move the tool from technical output to a trusted review habit.
Early-warning detection shifted from quarterly review toward a near-weekly rhythm, giving officers a structured way to inspect developing risk before it reached watchlist status.
What stayed after shipping.
The hardest enterprise AI problem is often adoption, not extraction. A model becomes a product only when people understand its boundaries and trust the way it enters their decisions.