NBFCs Can Automate Most of Lending. The Credit Call Stays Human
NBFC technology heads at the ETBFSI NBFC Leaders' Retreat 2026 see 50 to 80% of lending operations automated, with loan rejections, restructuring and write-offs kept under human judgment.
Most of an NBFC's lending operation can be handed to AI. The decisions that hurt people when they go wrong can't. That was roughly where among technology heads at the ETBFSI NBFC Leaders' Retreat 2026, landed, as reported on 12 October.
How far automation could go
Nobody agreed on a number. Dharmvir Singh, CTO of Sammaan Capital, put it at 50 to 60% automation across business functions over the next two to three years. Premraj Avasthi, CIO of GIC Housing Finance, went further: 75 to 80% autonomy, with humans still accountable for consequential decisions.
Sudesh Puthran, CTO of CreditAccess Grameen, drew the line by task. Document extraction, identity verification and data entry can be automated heavily. Loan rejections, restructuring, write-offs and legal proceedings should stay with people. Himanshu Gupta, co-founder and Chief AI Officer at WeRize, reduced it to one rule: automation should be inversely proportional to risk.
What is already working
- Long documents. Sammaan Capital uses AI to read legal and technical reports that run to around 100 pages, classify documents and summarise what an evaluator needs.
- Collections. Wonder Home Finance uses AI to spot missing documents and flag incomplete applications. CTO Binod Roy said collections through digital channels are up about 30%.
- Build time. A loan origination journey on Salesforce took Sammaan five to six months. A comparable self-service journey on an open-source stack took about three.
- Rural income. CreditAccess Grameen handles around 1.8 lakh loan applications a day. It has finished a proof of concept that estimates dairy income from photos of livestock taken by field officers, with the officer validating the result. It is now moving to implementation.
Where it gets hard
Data is the first problem. Avasthi said legacy systems and fragmented data are the main obstacles, and GIC Housing Finance started its transformation with a three-layer bronze, silver and gold lakehouse. Roy argued for replacing legacy platforms in phases rather than all at once.
Governance is the second. Puthran said consent, purpose limits and retention under India's DPDP framework have to be designed into the architecture from day one. Gupta said AI agents should be treated like digital employees, with access controls and accountability, and warned that models trained on historical lending data can keep excluding the borrowers who were excluded before. His advice was to start with small exposures.
Avasthi also raised cost. He flagged token-based LLM pricing as a constraint on scale and asked providers for more predictable commercial models.
My read
The percentages matter less than the risk-based split. Every NBFC CTO can list which steps in their loan journey are clerical and which are judgment calls. That list is the automation roadmap, and it is also the list a regulator will ask for when something goes wrong. Lenders who write it down now, with an owner against each automated step, will have an easier time than those who automate first and document later.