AI in microfinance: the inclusion promise and the over-indebtedness risk

Microfinance is where AI's best and worst instincts collide. The same alternative-data model that brings a first-time borrower into the formal system can also push a vulnerable household into a debt spiral — and the governance scaffolding to tell the two apart is mostly missing.

The upside is real and quantified. The IFC has documented AI models built on mobile-money flows, airtime-purchase patterns and social-graph signals cutting default rates by 20–35% while approving 30–40% more borrowers than traditional microfinance scoring. McKinsey projects AI-driven personalisation could unlock trillions in new credit issuance by 2030, much of it to people previously excluded by thin credit files. Globally, the MFI loan portfolio sits above US$183 billion, women remain a majority of borrowers, and the sector compounds at double digits annually.

But alternative-data scoring carries failure modes that conventional underwriting does not. A model trained on phone and behavioural data can encode proxies for caste, gender, region or religion without anyone choosing to discriminate. It can approve a borrower who is already over-leveraged across three other lenders because the model optimises for repayment probability, not for the borrower's total debt burden. And in a sector serving low-literacy, low-recourse customers, an opaque 'computer said no' — or a too-easy 'computer said yes' — is a consumer-protection failure waiting for a regulator.

India's direction of travel is clear. RBI's FREE-AI framework explicitly elevates fairness, explainability and consumer protection as first-order obligations across all regulated entities, NBFC-MFIs included. The March 2026 NCAER–MFIN study on regulated small-borrowing in India put over-indebtedness and borrower protection squarely back on the policy agenda. And RBI's new draft Model Risk Management guidance pulls AI/ML scoring models into a formal lifecycle-governance expectation.

What an MFI should build now: a documented model inventory with risk ratings; bias testing across protected and proxy attributes before deployment and on every material change; a hard rule that high-impact credit decisions carry human-in-the-loop review and a borrower-facing explanation; affordability and total-indebtedness checks that the model cannot override; and a grievance-redressal path that actually reaches a person. Inclusion without governance is just faster harm.