Not a dashboard query. Not a pre-built report. The kind of question a board member asks twenty minutes before a meeting. What is your NPA rate by relationship manager for Q4, excluding restructured accounts?
Go ahead and time it.
If the answer comes back in under an hour, your organisation has solved a problem most Indian banks have not. If it comes back in a day or two, you are not behind on technology. Your business rules live in people's heads, not in your systems.
This is worth sitting with. Indian banking is in better shape than it has been in a generation. Gross NPAs across scheduled commercial banks fell to 2.1%. The problems that consumed credit risk teams through the NPA crisis like provisioning, resolution and regulatory firefighting have receded. What remains is the question of how well these institutions can actually think.
The answer, in most cases, is slower than it should be.
The two-day portfolio query is not a data problem. Every bank in India has more data than it can use. It is not a technology problem either. IT spend in Indian banking projected to grow 13.4% in 2026 (IBA Technology Survey 2024-25) is mostly going to cloud infrastructure, fraud detection, and origination speed. PSB Loans in 59 Minutes has compressed MSME in-principle approval to under 60 minutes and disbursal from 45 days to under 10. Origination is fast. Portfolio intelligence is not.
The bottleneck is path dependency. Every time a new question arrives - a different cut of NPA, a new delinquency segment, a cohort the regulator wants- someone has to locate the person who knows what the columns mean. What counts as a restructured account in your system? Which exclusions apply to relationship manager attribution? What is the approved definition of early delinquency? These are not arcane questions. They are the basic grammar of credit risk. But in most institutions, that grammar is oral, not written. It exists in the heads of three or four people who have been there long enough to remember why the database was built the way it was.
Therefore, every query starts from zero. The analyst is not slow. The system is not broken. The organisation simply never encoded what it knows.
Now consider what changes when AI enters this picture. The standard response has been to automate the existing workflow: get the analyst's answer faster, reduce the number of steps, and cut the turnaround from two days to four hours. This is productivity improvement and it is not unimportant. But it is the wrong ambition, because it optimizes an architecture designed for a constraint that no longer exists, or should exist.
The human was in the loop at every step of the query chain not because human judgment was needed at every step, but because only a human could carry the business rules in their head and apply them consistently. That was the constraint. An analyst who had been at the bank for seven years knew that "restructured accounts" meant one thing in the retail book and something slightly different in the SME book, and that the two definitions had never been reconciled in the database. So every query passed through them. This was the only available architecture given what the systems could not hold.
Now AI can hold the rules. This changes the architecture, not just the speed. Business logic is encoded into a layer that sits between the data and the people. Where every metric has an approved formula, every term has a precise definition, and every query draws from the same source, the human's role shifts. In technical terms we call it a semantic layer, or, an ontology. They stop retrieving answers and start interrogating them. The analyst is no longer the bottleneck; they are the judgment layer. That is a fundamentally different operating model, and it is not achieved by making the old workflow faster.
The institutions spending on cloud and AI are not wrong to do so. But most of that 13.4% will not reach this layer automatically, because the portfolio intelligence layer does not show up clearly in a vendor roadmap or a regulatory mandate. It shows up in how long it takes to answer a question your board just asked. The RBI's Comprehensive Credit Risk Management Framework, introduced in January 2026, now requires banks to integrate Risk Appetite Statements directly into automated decision engines and run real-time Early Warning Systems on borrower health. These mandates assume an institution can answer portfolio questions quickly, consistently, and with the same definition every time. That most cannot (yet) is my bet.
The institutions that close this gap will not just move faster. They will think better, because the knowledge will be in the system rather than leaving with every analyst who resigns. Portfolio calls get made on current data, not on a spreadsheet built last Tuesday.
So: how long does it take your team to answer a portfolio question?
If you had to think about it, you already know whether you are automating the old architecture or building a new one.