AI in Indian Collections: What Actually Works on the Ground
The gap between AI collection vendor promises and what NBFC field teams actually experience. Based on operator conversations and Findrive research.
The Indian collections technology market is full of vendors promising AI-powered recovery optimisation. The pitch is compelling: predictive models that tell you which borrowers to call, when to call them, and what to say. The reality on the ground is more nuanced.
Through Findrive's research and conversations with NBFC operations teams, a pattern emerges. Most AI collection tools work well in two specific scenarios: early-stage delinquency (0-30 DPD) where borrowers are reachable and responsive, and large-portfolio segmentation where statistical models can identify high-probability recovery buckets. Outside these scenarios, the technology hits fundamental limits.
The core problem: field collection agents visit borrowers with minimal context. No payment history, no communication trail, no disposition data from previous visits. The AI model may have scored the borrower correctly, but the agent executing the visit has none of that context. The intelligence exists in the system but doesn't reach the point of execution.
This observation drove the "NBFC collection agents lack real-time borrower context" problem on the ColdStart Labs problem board. The hypothesis: real-time context delivery to field agents — not better scoring models — is the highest-leverage intervention in Indian collections.
Related experiments
Newsletter for fintech operators covering AI in collections, NBFC infrastructure, and the India lending stack.
Related journal entries