By Malcolm Gomes, Chief Operating Officer, IDfy, Head of Privy by IDfy
Every AI system a bank runs stands on data. Not on models, not on compute, not on talent, though all of those matter. In financial services that data is personal, regulated, and consequential, which makes one point unavoidable: privacy and data governance is not a control beside the AI programme; it is the foundation the programme stands on. Treated as a compliance layer, privacy is the thing that slows AI down, the tax on ambition. Treated as a foundation, it decides whether the AI above it holds or falls. In banking, where every worthwhile AI use case sits on regulated data, that is not a constraint on AI. It is the precondition for it.
The direction is already set. The RBI’s FREE-AI committee estimated AI could improve banking efficiency by up to 46%, and AI is already deployed across customer service, risk management and KYC in Indian banks. The ambition is not in question. What runs underneath it is.
Here is the number that should give every board pause. Only 37% of Indian organisations have implemented AI access controls, and close to 60% either have no AI governance policy or are still drafting one. The technology is racing ahead. The foundation, in most cases, is not yet built, and that gap accumulates risk in a particular way. It compounds.
It compounds because the same customer data now flows across credit, fraud, personalisation and analytics. Data collected for one purpose becomes an input to a model serving another, and a weakness at the base does not stay at the base. An unclear consent becomes a model trained on data it should not have used, becomes a decision no one can explain, and becomes a liability carrying penalties of up to Rs 250 crore under DPDP, separate from whatever the breach itself costs.
The collision shows up in three specific places. Data minimisation, where DPDP requires collecting only what a purpose needs while a machine learning team accumulates by instinct, and the fix is to know what is going into a model before training, not after. Purpose limitation, the sharpest conflict, where data gathered to service a loan quietly becomes training data for a cross-sell model the customer never agreed to. This is the most common way a well-intentioned AI programme drifts into non-compliance, because the data was already in the building, and reusing it felt free. And erasure, which almost nobody has solved: a deleted customer’s records may already have trained a fraud model, their influence now dissolved into its weights. Deleting the row does not delete the learning, and regulators will eventually ask about it.
There is a human cost the compliance language hides. Every training dataset is made of real people: the borrower whose repayment history teaches a risk model and the applicant an automated system rejected for reasons no one can explain. Under DPDP, that person can ask how a decision about them was made. A bank that cannot trace which data shaped which model cannot answer, and “the algorithm decided” is not an answer the law accepts.
But the same intersection that accumulates liability is where advantage is built, and it compounds for the same reason. A strong foundation propagates upward too.
Start with speed. When a bank knows what data it holds, what each dataset was collected for, and which consents govern it, the questions that normally stall an AI project – is this usable? Can we defend it? Will legal sign-off be answered before they are asked? Governance stops being where AI initiatives wait and becomes the reason they ship.
Then supply, which banks underestimate. A large share of enterprise data sits unused, not because it lacks value but because no one can establish that using it is lawful. A governed foundation thaws it. AI is only ever as good as the data it is allowed to learn from, and most banks allow it far less than they could.
The RBI did not build FREE-AI to slow banks down. It opened with public trust as its first principle, on the logic that trust is the asset AI must sustain rather than spend. Read alongside DPDP, the message is consistent: govern the data well and you get to do more with it, not less. So the paradox is not the real shape of the problem. The instinct is to picture a dial, privacy at one end and AI ambition at the other, and argue about where to set it. But privacy is not the opposite end of that dial. It is the surface the dial is bolted to. Govern it badly and both fail at once. Govern it well, and both improve together.
The institutions that understand this are not using less AI. They built the ability to see their own data first, and everything else followed. They can trace which dataset fed which model and under what consent and show an auditor not just that a model works but what it was trained on and why that was lawful. That capability is unglamorous and slow to build, and it is the difference between the banks that out-innovate the field and the ones that spend the next decade explaining themselves. A bank that can prove its foundation is sound never has to stop and defend it when the regulator or the customer asks.