The term RoDI — Return on Data Invested — does not yet appear in any RBI circular on NBFC board report. But it should. India's NBFC sector has crossed Rs 40 lakh crore in consolidated assets, and technology-enabled lenders now account for over three-quarters of loan sanction volumes by count — yet measuring what the underlying data investment actually returns remains almost entirely absent from boardroom conversations.
The TechFin sector has grown enamoured with model sophistication. Gradient boosting, deep neural networks, self-learning systems — the vocabulary of advanced machine learning has migrated from research papers into board presentations with remarkable speed. The visible costs are understood: infrastructure, compute, data science talent. What is rarely quantified is the structural cost complexity introduces over time.
Consider three. First, explainability: a complex ensemble model cannot, in most cases, produce a simple, auditable explanation for why a borrower was declined. From a regulatory standpoint, the RBI's evolving guidance is moving clearly toward explainable, auditable decisions. Operationally, a team that cannot explain a decline cannot learn from its errors. And institutionally, a model only three people truly understand is a concentration risk — when those three leave, it becomes a black box the organisation is simultaneously dependent on and unable to govern. Second, the feedback loop: self-learning models improve on clean, voluminous outcome data — in a startup with two or three years of portfolio history, the self-learning narrative is often aspirational rather than operational. Third, talent dependency: the ongoing cost of maintaining specialised model infrastructure rarely appears honestly in any build-versus-simplify analysis.
Where Simple Wins
There is a well-established but underappreciated finding in applied machine learning: beyond a threshold of data quality and volume, the marginal performance gain from increasing model complexity diminishes sharply. A well-specified scorecard, trained on clean, relevant features, will frequently match the performance of a poorly-specified neural network — at a fraction of the cost to build, maintain, explain, and govern. In Indian credit, the most predictive features are well-known: bureau score trajectory, banking conduct, GST filing consistency, business vintage, and segment-specific platform data. They require clean pipelines, disciplined feature engineering, and honest validation — not sophisticated models. Complexity should be introduced when there is a demonstrable performance gap simpler approaches cannot close, not as the default because it signals capability.
The Human-In-The-Loop Case
Somewhere in the TechFin conversation, full automation became conflated with progress. A Human-in-the-Loop model — where a system-generated assessment is reviewed by a credit officer with domain knowledge — is not a compromise; it is often the architecturally superior choice. The system handles what machines do best: processing structured data at scale, cross-referencing sources without fatigue, applying rules without bias. The human handles what humans do best: context, judgment, and accountability. Critically, this does not require a veteran analyst — it requires a credit officer who can read a system-generated summary intelligently and know when to escalate. A trainable profile that democratises decision-making without sacrificing governance.
Four Questions Every TechFin CRO Should Answer
— Can your risk team explain, to a regulator or a declined borrower, why your model made a specific decision — in plain language, without a data scientist in the room?
— What is the fully-loaded annual cost of your model infrastructure — talent, compute, maintenance, explainability gaps — and what measurable performance delta justifies it over a simpler alternative?
— Does your model have sufficient, clean outcome data to actually self-improve — or is it iterating on a dataset too thin to generate a reliable learning signal?
— If your lead model architect resigned tomorrow, how long would it take your organisation to understand, validate, and rebuild what they built?
What Deliberate Architecture Delivers
At CapitalXB, we confronted these questions directly when designing our credit assessment infrastructure for export factoring — a segment where relevant data signals are rich but non-standard, spanning marketplace performance, trade documentation, foreign currency receivables, and multi-framework compliance records. We built a structured, multi-dimensional assessment engine with interpretable scoring logic, binary gate controls for absolute risk thresholds, and a Human-in-the-Loop review layer. We did not build a self-learning neural network. We built a system our credit team can interrogate, our compliance function can audit, and our board can understand.
The measured outcome:
90% reduction in manual underwriting effort — standard cases require minimal analyst intervention.
4x AUM growth — with no incremental headcount in the credit function.
2 Hours credit assessment turnaround — down from the better part of a working week, for standard factoring assessments.
100% decisions with documented rationale — every approval and decline carries a full, human-readable explanation for the borrower, credit team, and regulator.
Nil NPA portfolio credit quality — maintained through disciplined origination and continuous monitoring.
The Bill Always Arrives
The TechFin sector has told compelling stories about what its models can do. The more honest conversation — the one that will define the sector's next phase of maturity — is about what those models cost, whether simpler alternatives were considered, and whether the humans alongside these systems are equipped to govern them or merely to operate them.
Data, like capital, compounds when put to work intelligently. It depreciates when it feeds a model nobody fully understands, producing decisions nobody can account for. RODI is the discipline of treating data investment with the same rigour applied to any other allocation of resource. Getting this right does not require building less — it requires building more deliberately. And in an industry that has celebrated speed above almost all else, deliberateness may be the most underrated competitive advantage of all. After all, the most sophisticated thing a model can do is know its own limits — and so far, that particular capability remains stubbornly human.
