Bajaj Finance disbursed roughly ₹1,600 crore in loans through AI-powered call centers in a single quarter — about 10% of its Q3 FY26 total (Outlook Business, February 2026). Read that fast and it sounds like an AI model is approving loans. Read the underlying numbers and a different picture shows up.
That distinction matters more than it sounds, and it's the part most "AI is transforming lending" coverage skips entirely.
What the AI Was Actually Doing
The same reporting breaks down what those systems are doing under the hood: converting roughly 2 crore customer calls from voice to text, running 46 million face-match checks for KYC validation at 95-96% accuracy, and surfacing 100,000 new loan offers by extracting signal from customer data that was previously sitting unused in unstructured form.
None of that is a credit decision. It's perception and extraction — turning a phone call into searchable text, turning a selfie into a verified identity match, turning scattered transaction history into structured features a scoring model can actually use. The "AI approved your loan" headline is really "AI made the inputs to an existing decision process arrive faster and cleaner."
That's still a meaningful engineering win. Tata Capital has reportedly moved from isolated AI pilots to enterprise-wide deployment across marketing, sales, credit, and collections, and Poonawalla Fincorp has taken an "AI-first" posture on underwriting workflows. But "AI-first workflow" and "AI makes the credit call" are different claims, and conflating them is where a lot of the public conversation goes wrong.
Why the Actual Decision Stays Out of the Black Box
Here's the constraint that shapes every production system in this space: India's RBI Digital Lending Guidelines require that an automated decision affecting someone's credit eligibility be explainable, not just accurate. If a borrower is rejected, a lender needs to be able to point to why, in terms a regulator and an ombudsman can both follow. An opaque model that says "rejected, 73% confidence" with no decomposable reasoning doesn't satisfy that.
This is why almost nobody builds "feed the bureau report to a large language model and accept whatever it says" as the actual underwriting step, however tempting that sounds from a pure capability standpoint. Instead, production pipelines tend to layer it:
- Deterministic hard filters — income floor, age, city, product eligibility. Fully auditable, no model involved.
- A scored model — typically gradient-boosted trees or a logistic model with documented feature importance, not a generative model, because regulators and internal risk teams need to trace which inputs drove which score.
- LLM-assisted extraction and triage — this is where the generative AI actually lives: pulling structured data out of voice calls, documents, and chat transcripts, and routing ambiguous cases to a human underwriter with a pre-summarized case file.
- Human review for edge cases — not eliminated, just applied to a much smaller, pre-filtered set of applications.
The speed gain everyone's reporting comes overwhelmingly from steps 1 and 3 getting faster, not from step 2 being replaced by something less explainable.
The Turnaround-Time Numbers Make More Sense This Way
CredAble's COO Ashutosh Taparia described the shift in MSME lending bluntly: "What was once a 14-day manual process can become a 14-minute autonomous workflow" (Outlook Business, February 2026). A 14-day process isn't 14 days of a human deliberating over credit risk — it's mostly document collection, manual data entry, follow-up calls, and waiting for someone to get to your file. Automating extraction and triage collapses almost all of that dead time. The actual risk decision, when you isolate it, was probably never the slow part.
| Underwriting Step | What Got Automated | What Didn't |
|---|---|---|
| Document/voice intake | Extraction (NLP, voice-to-text) | The credit scoring logic itself |
| Identity verification | Face-match at scale | Manual review of the ~4-5% mismatch tail |
| Offer generation | Surfacing eligible offers from data | Final approval authority |
| Edge-case applications | Pre-summarized case routing | The human underwriter's judgment call |
Where This Honestly Breaks
A 95-96% face-match accuracy rate sounds excellent until you run it across 46 million checks — that's somewhere between 1.8 and 2.3 million mismatches that need a fallback path, not an edge case you can ignore in design. Any system built on "AI handles the bulk, humans handle the tail" lives or dies on how well that fallback path is engineered, not on the headline accuracy number.
The honest framing for anyone building in this space: generative AI is currently doing extraction, summarization, and routing extremely well, and credit decisioning extremely rarely, on purpose. EY's research puts 42% of Indian financial institutions as active investors in AI/GenAI initiatives right now, and that investment is concentrated almost entirely on the parts of the pipeline regulators are comfortable letting move fast.
Where a Marketplace Fits Into This
Every lender behind this kind of pipeline still has its own appetite, its own thresholds, and its own idea of what a good borrower looks like — faster underwriting doesn't make one lender's criteria match another's. A marketplace like SwipeLoan sits on top of that variation, comparing a borrower's profile against offers from 100+ RBI-registered lenders rather than betting everything on whichever single NBFC's pipeline you happened to apply to first. The underwriting speed is the lender's engineering problem; finding the right lender in the first place is a different one entirely.
If you've built any part of an explainability layer for a credit-decisioning system, I'd like to hear how you're handling the audit trail requirement in practice — that's usually the part that doesn't make it into the case studies. Drop it in the comments.
Top comments (0)