How does AI credit scoring work in India, and is it fair?

What is AI credit scoring?

A credit score is a prediction: how likely is this borrower to repay? Traditional scoring relies mostly on a credit bureau record, meaning past loans, cards and repayment behaviour. That works well for people and firms that have borrowed before. It works badly for those who have not, the "thin-file" or "new-to-credit" borrowers, because there is little to score.

AI credit scoring uses machine-learning models that can combine many more signals and find patterns a scorecard would not. Credit underwriting is one of the main uses of AI in Indian finance. When RBI’s FREE-AI committee surveyed regulated entities, it accounted for 13.70% of the AI applications reported, second only to customer support. The committee also noted that AI-driven lending for small-ticket loans could "onboard first-time borrowers and underserved communities".

What data do these models use?

The most significant change in India is consented access to cash-flow data. Through the Account Aggregator framework, a borrower approves a specific request and their bank sends statement data directly to the lender, instead of the borrower downloading and uploading PDFs. Sahamati, the industry body for the network, counted 352.19 million accounts linked and 595.11 million consents fulfilled as of 30 September 2026. A model reading twelve months of actual inflows can judge a small business that has never borrowed. Our Account Aggregator guide explains how consent works, and the Unified Lending Interface explains the wider lending data rails.

What do RBI’s Digital Lending Directions require?

On 8 May 2025 RBI issued the Reserve Bank of India (Digital Lending) Directions, 2025, consolidating its earlier digital lending guidelines. They apply to banks, co-operative banks, NBFCs (including housing finance companies) and all-India financial institutions when they lend digitally, and to the lending service providers (LSPs) they use. The provisions that matter most for data-driven scoring:

• Know the borrower’s profile. Before lending, the lender must capture the borrower’s economic profile, at a minimum age, occupation and income, and keep it on record for audit.
• Need-based data, with consent. Data collection by a digital lending app must be need-based and with the borrower’s prior and explicit consent, with an audit trail. Apps must not access phone resources such as files and media, contact lists or call logs. Camera, microphone or location may be accessed once, for onboarding or KYC, with consent.
• A Key Fact Statement setting out the loan’s cost and terms before the contract, sent automatically to the borrower’s verified email or SMS after signing.
• Fair comparison across lenders. From 1 November 2025, an LSP that works with several lenders must show a digital view of all matching loan offers, name lenders that did not match, and follow "a consistent approach for similarly placed borrowers". The content must be "unbiased, objective".
• A public directory of digital lending apps, which RBI publishes from data the lenders report.

The Directions do not regulate algorithms as such. What they do is limit what goes into the model, the data, and make the outcome comparable and documented.

Why is fairness the hard part?

A model that predicts repayment well on average can still treat some groups unfairly. It may learn from history in which some borrowers were under-served, or use proxies: a PIN code, a phone model or a pattern of spending can stand in for income, caste, gender or religion without the model ever seeing those fields. RBI’s FREE-AI committee said directly that "AI models used for credit, loan, or insurance decisions may exhibit bias against specific demographic groups", and that decisions "should be sufficiently explainable to ensure that no discrimination, either direct or indirect, occurs". In its annex on digital lending it called for fairness audits to detect and mitigate algorithmic bias.

Can a lender say "the model said no"?

Not if it is a bank. Under RBI’s Fair Practices Code for Lenders, banks and financial institutions have been required since 2007 to convey in writing the main reason or reasons for rejecting a loan application, for all categories of loans irrespective of amount, including credit cards. A model does not change that duty. It means the lender has to be able to translate the model’s output into reasons a borrower can understand. The FREE-AI report goes further, saying customers should have "the means to challenge and seek clarification on AI decisions" and that human oversight and override should be available for critical decisions.

How is this regulated outside India?

The EU’s AI Act treats AI used "to evaluate the creditworthiness of natural persons or establish their credit score" as high-risk, with duties on risk management, data governance, documentation and human oversight. After the 2026 Digital Omnibus, those high-risk rules apply from 2 December 2027. An Indian lender scoring EU residents would be in scope; most Indian lending is not. Our post on AI regulation in financial services covers the detail.

What should a small business borrower know?

• Your bank statements are your credit file. Regular receipts, few bounced payments and clean separation of business and personal money all score well when cash-flow data is read.
• Consent is specific. An AA consent names the purpose, the data and the period. Read it, and revoke consents you no longer need.
• Ask for reasons. If a bank declines you, it should tell you the main reasons in writing. That tells you what to fix.
• Be wary of apps that ask for your contacts or gallery. Under the 2025 Directions, regulated lenders’ apps should not need them. Check the lender against RBI’s directory of digital lending apps.