When you apply for a loan on an app and get a decision in ninety seconds, no human read your application. A model scored it. Understanding roughly how that model works tells you a good deal about why you were approved or declined — and what you can actually do about it.
From Scorecards to Models
Traditional credit assessment used a scorecard — a set of weighted factors a credit analyst designed. Income contributes so many points, existing debt subtracts so many, employment tenure adds some. Transparent, stable, and limited by what its designers thought to include.
Machine learning models work differently. They are shown large volumes of past loans with known outcomes and infer which patterns predict default, including combinations no analyst specified.
| Traditional scorecard | ML model | |
|---|---|---|
| Rules come from | Analyst judgement | Patterns in historical data |
| Variables used | Tens | Hundreds or more |
| Interactions between variables | Rarely captured | Captured automatically |
| Explainability | Complete | Requires deliberate effort |
| Speed to update | Slow — redesign required | Faster — retrain |
| Degrades over time | Slowly | Yes — needs monitoring |
What the Model Actually Predicts
A credit model does not output “approve” or “decline”. It outputs a probability of default — an estimate that this applicant will fail to repay within a defined window, typically ninety days past due within twelve months.
The lender then applies a cut-off. Above a certain probability, decline. Below it, approve. Between, perhaps approve at a higher rate or lower amount.
This distinction matters because the cut-off is a business decision, not a model output. Two lenders using the identical model can reach opposite conclusions on the same applicant because one has a higher risk appetite. Being declined by one lender and approved by another is entirely normal and does not mean either is wrong.
What Goes Into the Model
| Category | Examples |
|---|---|
| Bureau data | Credit history, existing loans, repayment record, enquiries |
| Application data | Income, employment, age, location |
| Banking data | Salary credits, balance patterns, bounced payments |
| Existing relationship | Behaviour on the lender’s other products |
| Alternative data | Utility payments, telecom, device and app signals, transaction patterns |
The bottom row is where AI lending genuinely differs from what came before, and it deserves its own treatment.
Alternative Data and the Credit-Invisible
India has a large population with no meaningful credit history — first-time borrowers, informal-sector workers, small traders. A traditional scorecard has almost nothing to work with. They are not high risk; they are unmeasured.
Alternative data attempts to fill that gap using signals correlated with reliability: consistent utility payments, stable telecom usage, regular bank credits, transaction consistency.
The financial inclusion argument is genuine, and it aligns with inclusivity being one of the principles in RBI’s FREE-AI framework. But three cautions apply:
- Correlation is not creditworthiness. A signal that statistically predicts repayment may have no causal connection to it, and may stop predicting when circumstances change.
- Some signals act as proxies for protected characteristics. Location, device type and app usage can correlate with religion, caste or gender, producing discriminatory outcomes without any protected variable being used directly. This is Explainable AI in Lending.
- Data collection has limits. India’s data protection framework constrains what can be collected and for what purpose. Consent obtained for one use does not extend automatically to model training.
The Structural Blind Spot
This is a subtle problem worth understanding, because it affects every credit model ever built.
A model is trained on applicants the lender approved, because only they have repayment outcomes. Rejected applicants have no outcome — you cannot observe how someone would have repaid a loan they never received.
So the model learns from a filtered population, selected by whatever criteria previous models or scorecards applied. If the old process wrongly rejected a category of applicant, the new model never learns they would have repaid. The error perpetuates.
Lenders address this partially through reject inference techniques and by occasionally approving a small random sample below cut-off to generate genuine outcome data. Neither fully solves it.
Why Explainability Is Mandatory Here
A borrower declined credit is entitled to understand why. “The model returned 0.31” is not an explanation.
Explainability is one of the seven principles in RBI’s framework, and it constrains model choice in practice. A highly complex model that cannot produce a defensible reason code may be unusable for lending regardless of its accuracy. Techniques exist to attribute a decision to contributing factors — covered in explainable AI in lending.
What This Means If You Are Borrowing
- Ask for the reason if declined. Institutions are expected to be able to explain adverse decisions.
- Check your bureau report. Errors in bureau data flow straight into the model. Correcting them is the highest-return action available to you.
- Multiple applications hurt. Each enquiry is recorded and clustered applications signal distress to the model.
- A decline is lender-specific. Different cut-offs mean another lender may approve you.
- Consistency matters more than size. Regular, predictable banking behaviour reads better than a large but erratic balance.
- Be careful what permissions you grant. App-based lenders requesting broad device access are collecting alternative data. You are entitled to know what is collected and why.
Key Takeaways
- ML credit models infer default patterns from data rather than following analyst-designed rules
- The model outputs a probability of default; the cut-off is a business decision
- Being declined by one lender and approved by another is normal
- Alternative data can score the credit-invisible — a genuine inclusion gain
- But some alternative signals act as proxies for protected characteristics
- Models train only on approved applicants — a structural blind spot that perpetuates past errors
- Explainability is a regulatory requirement, not an optional feature
- Correcting bureau errors is the highest-return action a borrower can take
Frequently Asked Questions (FAQ)
Q: How does AI credit scoring work?
A machine learning model is trained on large volumes of past loans with known outcomes, learning which patterns predict default. For a new application it outputs a probability of default, and the lender applies a cut-off to decide whether to approve, decline, or approve on modified terms.
Q: Why was I rejected by one lender but approved by another?
Because the cut-off is a business decision, not a model output. Two lenders can use similar models and reach opposite conclusions on the same applicant simply because one accepts more risk. A decline reflects that lender’s appetite as much as your profile.
Q: What is alternative data in credit scoring?
Data outside traditional credit bureau records — utility and telecom payment patterns, bank transaction behaviour, device and app signals. It allows lenders to assess borrowers with no credit history, which is a genuine financial inclusion benefit but raises privacy and fairness questions.
Q: Can AI credit scoring be biased?
Yes. Models trained on historical decisions learn historical patterns including discriminatory ones. Variables such as location or device type can act as proxies for protected characteristics, producing discriminatory outcomes without any protected variable being used directly.
Q: Do lenders have to tell me why I was declined?
Explainability is one of the principles in RBI’s FREE-AI framework, and institutions are expected to be able to explain decisions affecting customers. Asking for the reason is reasonable, and the answer often points to something correctable such as a bureau error.
Q: Does applying to many lenders hurt my chances?
Yes. Each application generates a bureau enquiry, and multiple enquiries in a short period signal financial distress to the model. It is better to check eligibility where possible and apply selectively.
Q: What is the fastest way to improve my chances?
Check your credit bureau report for errors and get them corrected — inaccurate bureau data flows directly into the model. Beyond that, consistent repayment behaviour and stable banking patterns matter more than any single large deposit.
Related Reading: