Why Domain Matters More in the AI Era

The code was correct

Rohit had been at the NBFC in Jaipur for seven months when he built the loan-repayment calculator. The company offered personal loans to small shopkeepers and self-employed workers across Rajasthan — the kind of loans that a large bank would not bother with. Loan amounts of thirty thousand, fifty thousand, occasionally a lakh. Monthly repayments collected by field agents who visited each borrower in person.

The product team had asked for a new calculator that would show borrowers, at the point of application, what their monthly repayment would be. Rohit built it. The formula was in the product specification. He implemented it precisely, wrote tests, confirmed the output against the spec's own worked examples, deployed it, and moved on.

Three weeks later, a credit-operations veteran named Meena called a meeting. Meena had been in lending for nineteen years, longer than Rohit had been alive. She had pulled the calculator's output for a sample of recent loans and compared it to what borrowers should have been paying under the RBI's Key Facts Statement guidelines. The numbers did not match. They were close enough that nobody had noticed in testing. They were wrong enough that the regulator would have noticed eventually.

The problem was not in Rohit's code. His code was perfect. The problem was in the specification. The spec had used a flat-interest formula — interest calculated on the full original principal for the entire loan term. The RBI's Key Facts Statement requires NBFCs to disclose the reducing-balance rate — interest calculated on the remaining principal after each payment. The difference is not a rounding error. On a fifty-thousand-rupee loan over twelve months, the difference between flat and reducing can be over two thousand rupees. On ten thousand loans, it is over two crore. On the wrong side of the regulator.