What does human oversight of AI mean in practice for a bank?
RBI Governor Sanjay Malhotra described human oversight as the ability to explain, intervene and override an AI decision. In a bank's systems that means three things: a record for every decision, a place where a person can step in, and a switch that stops the model without stopping the business.
Banking & payments2 min read
Human oversight of AI in a bank means a person can explain any decision the model made, step in before or after it takes effect, and override or stop it. RBI Governor Sanjay Malhotra described it in those terms at FIBAC on August 11, 2026, and said it must be a design principle. It is built into the system, not added through a policy document.
He also said responsibility for a decision stays with the bank and cannot pass to an algorithm or a vendor.
Explain: keep the record at the moment of decision
To explain a decision later, the system must store what the model saw and said at the time: the inputs, the model version, the score, the threshold and the reasons. A model retrained next quarter cannot tell you why it declined a loan last month. Logging this belongs in the integration with the core system, built in the first release.
Intervene: choose the oversight model per decision
Not every decision needs a person to approve it. The choice depends on what a wrong answer costs and whether it can be reversed.
- A person approves each action before it happens, for decisions that move money or affect a customer's standing.
- A person supervises and can stop or reverse actions, for high-volume decisions that are cheap to undo.
- A person reviews samples and trends, for low-risk assistance such as drafting or search.
Whichever pattern applies, the record matters. On a metropolitan transit network's voice line in 11 languages, the system resolves 70% of calls without a person and logs every call for audit in the network's core systems.
Override: a switch that does not stop the bank
Every model needs a way to turn it off and a fallback for when it is off: the previous rules, a manual queue, or a more conservative threshold. Test the switch before launch. A model that cannot be stopped without stopping lending or payments is a model nobody will stop, even when they should.
Someone with authority
Oversight only works if the person reviewing has the information, the time and the authority to disagree with the model. If approvers are measured on throughput alone, they will approve everything. Measure how often reviewers change or reject a recommendation; a rate near zero is a warning, not a success.
Questions
- What does RBI mean by human oversight of AI?
- In the RBI Governor's words at FIBAC 2026, the ability to explain AI decisions, intervene in them and override them, built in as a design principle.
- Does every AI decision in a bank need human approval?
- No. Decisions that move money or affect a customer's standing usually need approval first. Cheap, reversible decisions can run under supervision, and low-risk assistance can be reviewed by sampling.
- How do you explain an AI decision months later?
- Store the inputs, model version, score, threshold and reasons at the moment of decision. A retrained model cannot reconstruct why an earlier version decided as it did.
Sources
- RBI Governor Sanjay Malhotra Sets Out AI Playbook for Indian Banks, The Kanal, August 13, 2026
More on banking & payments
All writingNovember 2026
RBI's API advisory: the controls banks need by December 3
RBI has reportedly told banks to tighten API-based payments as AI-assisted attacks were reported: a full API client inventory, audits of API clients, payout safeguards and transaction limits, with key deadlines on December 3, 2026. What each control means in the systems.
October 2026
Fraud scoring inside a core banking system, step by step
Fraud scoring pays off most when the score reaches the transaction before it is approved. Here is how to get it there. When we put real-time scoring and alerts inside a major financial institution's banking systems, fraudulent transactions fell by 25%.
August 2026
What RBI's FIBAC speech asks of Indian banks' AI programs
RBI Governor Sanjay Malhotra told banks on August 11 that responsibility for an AI decision stays with the bank. He named four priorities: an inventory of AI systems, a board-approved policy, explainable customer decisions and red-teaming. Here is what each one means in the systems.
Tell us which system the AI has to run inside
If that system is not ready for AI yet, tell us that too. An engineer will reply and say plainly whether we are the right firm for it.