Smt. Sitharaman on AI in Banking: The Collateral Blind Spot

An Early Warning Signal is a defined indicator that a loan account may be heading toward fraud. The Hon'ble Finance Minister's remark about AI in banking pointed at two jobs built on those signals, detecting potential fraud and monitoring loan exposure, and Indian lenders are already under a regulatory obligation to do both.

Most banks meet that obligation using data they already hold about the borrower. The security is re-examined too, but usually at a lighter touch than the diligence run at sanction, and against the file on record rather than the registries where a change would actually surface.

A lender holding a first charge will often discover a competing charge only when recovery begins, at the point where the security is being enforced and a second lender appears. The information was available the whole time, in a central registry, searchable in seconds. Since 31 July 2026, how quickly a bank finds it has a number attached to it.

What did the Hon'ble Finance Minister say about AI in banking?

She highlighted the potential of AI in banking for helping financial institutions detect potential fraud and monitor loan exposure, and described the relevance of that work as national in scope. Smt. Sitharaman made the remarks during an interaction with the founder of AdvaRisk at a startup conclave in Rajasthan.

For a credit team the useful part is the pairing. Fraud detection and loan exposure monitoring named together describe a control that runs continuously, not a check performed once at sanction.

The remark is not an endorsement. The published account is reported speech rather than a verbatim transcript, and neither it nor the session on the Hon'ble Finance Minister's channel records approval of any product. Business Standard reported it as describing such technology as nationally relevant for banks across the country, and this article does not treat it as more.

What is an Early Warning Signal, and who requires one?

A signal obliges the bank to examine the account and decide whether to classify it as a Red Flagged Account. The governing instrument behind that duty changed this year. On 31 July 2026 the Reserve Bank issued the Reserve Bank of India (Commercial Banks - Fraud Risk Management) Directions, 2026, numbered RBI/DoS/2026-27/412, superseding the July 2024 Master Directions. A separate set was issued the same day for Regional Rural Banks. The operative change for a credit team is that the sequence now carries clocks.

The structure is still a funnel: signals fire, an account is examined, a Red Flagged Account decision is recorded, and where fraud is established the incident is reported. What the 2026 Directions add is a ceiling on the examination step, so the framework now turns on what feeds the top of the funnel and on whether the evidence can be assembled inside the window.

Why do most EWS frameworks miss the collateral?

Most frameworks are built from data the bank already holds internally, and the collateral record is not internal data. Account conduct, transaction patterns, bureau scores and statutory filings all arrive through existing pipes. Land records, encumbrance histories and litigation filings sit in state systems the bank has no standing connection to.

The result is a monitoring asymmetry. A borrower's cheque return is visible the same day. A second mortgage registered against the same property in a different Sub-Registrar office may not surface for years, because the due diligence run before sanction is rarely repeated with the same rigour after disbursal.

That access gap is narrowing, though less evenly than the coverage suggests. The Reserve Bank Innovation Hub's Unified Lending Interface exposes state land records through a single API, and the RBI's Trend and Progress of Banking in India report records 64 lenders on the platform as of 12 December 2025, drawing on digitised land records from eight states.

Eight is the number to hold on to. ULI is a pipe rather than a diligence product: it returns a record, not a chain of prior deeds, an encumbrance position across a look-back period or a litigation status. Where it reaches, a periodic re-check gets cheaper. Where it does not, the record still has to be pulled from the state system that issued it, in that system's language.

The asymmetry is worse for secured lending than for unsecured, which is counter-intuitive. The security was supposed to be the thing that made the exposure safe, and it is the thing least often re-checked.

Which collateral events should trigger a red flag?

Four events change the bank's actual security position and are verifiable against a source outside the borrower's control. Each is a discrepancy to investigate rather than a finding of fraud.

All four are observable without the borrower's cooperation. An EWS signal that depends on the borrower disclosing something is a weak signal precisely in the cases that matter.

Why does a CERSAI search miss a competing charge?

Because a search returns the position as filed, in the form it was filed, and four things sit between that and the position as it actually stands.

The 30 day filing window. Rule 5 of the SARFAESI (Central Registry) Rules, 2011 requires a security interest to be registered within 30 days of the transaction. A charge created three weeks ago can therefore be entirely legitimate and entirely invisible to a search run today.

Charges predating the registry. CERSAI became operational on 31 March 2011. Security interests created earlier entered the register only if filed during the transitional window that followed, so an older equitable mortgage may never have reached it.

Name-matched searches. A search run on the borrower's name depends on the name matching what was filed. Initials, expanded and abbreviated forms, transliteration differences and a change of entity name all break that match. The charge is registered; the search does not return it.

Asset description. An asset based search depends on the property being identified the same way in the registration as in the bank's own file. Survey numbers, plot numbers and addresses are recorded inconsistently across states and across decades, and a mismatch reads as an absence.

None of this argues for skipping CERSAI. It is why one search, run once, on one spelling of one name, is weaker evidence than it appears. The controls that close the gap are unglamorous: search on the asset as well as the borrower, run the name and address variants deliberately, and re-run on a schedule instead of only at sanction.

What does the RBI expect once an account is red-flagged?

Examination, then classification, then reporting to the Reserve Bank and to law enforcement where fraud is established, against the timelines above.

One provision bears directly on collateral evidence. The 2026 Directions allow banks to write terms into agreements with third party service providers holding them accountable where wilful negligence or malpractice on their part is found to be a causative factor for a fraud, and to report such parties to the Indian Banks' Association for caution listing. The standard applied to outsourced diligence is contractual as well as regulatory.

In a property-secured exposure that evidence is documentary: the title chain, the encumbrance position, the revenue entries and the litigation status, each traceable to the office that issued it. A signal a bank cannot substantiate with a sourced document is one it cannot act on, and now also one it may struggle to examine inside 30 days. The Insolvency and Bankruptcy Board of India applies the same evidentiary logic once a matter reaches the National Company Law Tribunal, where asset tracing and beneficial-interest work takes over from monitoring.

What can collateral monitoring not tell you?

It cannot tell you that a title is sound, that fraud has occurred, or that an asset is recoverable. It tells you only that something on the public record has changed since you last looked. The following sit outside what any monitoring system sees.

• Unregistered instruments. Oral family arrangements, unregistered agreements to sell and undisclosed tenancies appear in neither the registry nor the revenue record.
• Search boundaries. An encumbrance certificate covers the period and the sub-registrar office searched. A transaction registered elsewhere, or outside the window, is simply absent.
• Unindexed litigation. A suit filed last week may not reach a court index for some time, so a clean litigation search is evidence only of a clean index.
• Survey and boundary mismatches. The deed schedule, the revenue extent and the physical boundary can all disagree, and only a survey resolves it.
• Forgery and impersonation. A well-made forged deed can pass a documentary review, which is why documentary diligence and physical verification are different controls.
• Lag. Revenue records and mutation entries update on their own timetable. Absence of a signal today is not proof that nothing happened.

A framework that presents collateral signals as conclusions will produce false positives, annoy relationship teams and eventually be ignored. One that presents them as discrepancies requiring explanation is usable.

Where automated collateral monitoring fits

Automation belongs to the mechanical half of the work: re-running CERSAI and Sub-Registrar searches across many properties in many states, running the name and asset variants rather than a single query, normalising records that arrive in different formats and languages, and surfacing only what changed. Deciding what a change means is a separate job.

The 30 day window is what turns that mechanical half from a convenience into a constraint. If examining an EWS alert on a secured exposure needs a fresh charge search, a sub-registrar check and a litigation status across two or three states, the turnaround on those searches decides whether the examination closes inside the window or runs past it.

AdvaRisk states it was empanelled by the Indian Banks' Association for fraud detection and monitoring large credit. It states experience investigating banking exposure of more than ₹1,50,000 crore since it began operating in 2016, and its published claim for automated collateral monitoring is a reduction in turnaround time of up to 90%. That is a claim about the mechanical step, not about the judgement that follows it.

The judgement stays with the bank. A change in the record is an input to a credit decision and, where red flagging is being considered, to a documented examination the RBI expects to be substantiated. Where a legal opinion is required, an empanelled advocate reviews the evidence and signs it. AI that reads records at scale changes how quickly a lender learns something has changed. It does not change who is accountable for deciding what it means.