Hero FinCorp rebuilt lending around AI context — credit costs fell 15% in six months
The lender's Project Dhruv Tara reads 100+ micro-market signals per borrower; company-reported credit costs fell 15% and GNPA improved about 20% in H2 FY26.
What was changed
Hero FinCorp, the Hero group's non-bank lender serving more than 13 million customers across over 96% of India's PIN codes, built Project Dhruv Tara, an AI-led lending architecture spanning the whole credit lifecycle: customer sourcing, underwriting, onboarding, portfolio monitoring, servicing and collections. The problem it targeted: lending decisions built on bureau scores, repayment histories and documents tell only part of the story, especially for customers with thin formal credit histories whose files understate their actual risk.
The architecture replaces the single-score view with context. A cloud data lake combines bureau and repayment data with banking signals, GST data, account aggregator inputs, demographics, asset characteristics and geography — more than 100 micro-market data points and 40-plus APIs that let the lender read patterns at pin-code level, from local income behaviour to asset norms. Fifteen dynamic risk and collections models score repayment capacity, stability and intent, and drive predictive collections that intervene earlier and more appropriately instead of treating all borrowers alike.
The company reports the operational results: credit costs declined 15% in H2 FY26 versus H1 FY26, and gross NPA improved by around 20% over the same period. AI also moved into the front line — chatbots, IVR, automated workflows and AI-enabled email now handle nearly 60% of customer queries, saving more than 4 lakh human minutes every quarter, and one onboarding journey shrank from 90 fields to three.
Chief operating officer Priya Kashyap frames the discipline: the real value of AI in financial services is not speed alone but better decisioning — improved risk selection, early stress detection, less customer friction. Human teams keep contextual judgement, governance and accountability while the models process patterns at scale.
Why it worked
Thin-file borrowers were priced on absence of data; micro-market context recovered signal the bureau file never had.
Uniform collections treated every delinquent borrower the same; predictive models time and target interventions by behaviour.
The build put intelligence into the decision path itself rather than adding AI tools beside the old process.
What can be applied
When everyone underwrites on the same bureau file, the edge moves to context — where the borrower lives, how local income behaves — not to another look at the same score.
Aftermath
The figures are company-reported; the material gives no independent audit. The stated next phase is deeper context across the lending lifecycle: richer signals for earlier risk identification, lower customer friction and personalised engagement, with human judgement retained for decisions where accountability matters.