An expiry date does not establish that an authorised verification exception stopped affecting new decisions—or that earlier AI-assisted decisions remain reconstructable.

Methodology note: This article examines publicly available Reserve Bank of India (RBI) material and develops an illustrative, synthetic examination question. No banking system, customer record, provider telemetry or operational test result was examined.

Executive takeaways

  1. Expiry and cessation are different. An approved end time establishes intended authority; operating records are needed to assess what individual services actually enforced.
  2. Cessation and history are different. A bank must separately retain lawful, usable links between earlier decisions and the rule, model context and authority under which they were made.
  3. An unresolved record is not a successful closure. Management should distinguish demonstrated behaviour, evidenced failure and insufficient evidence, then decide remediation or bounded reliance according to materiality.

An expired approval does not close the evidence question

Consider a bank that temporarily authorises an alternative customer-verification route during a servicing disruption, with an approved end time of 18:00. A configuration change may be recorded as complete while a cached rule, queued case or provider-side session still uses the earlier route. Equally, an otherwise correct withdrawal may leave an earlier decision difficult to explain if its rule version or authority reference was overwritten.

The management decision is whether the exception can be closed on evidence that relevant new decisions ceased using the old authority and that affected earlier decisions remain reconstructable. These are two separate examinations; neither is guaranteed by an expired approval.

The example assumes an alternative route is legally permissible and authorised for the specified use. It does not recommend bypassing identity-verification or customer due diligence duties. In an actual bank, applicable know-your-customer, outsourcing, records and privacy requirements must be assessed before such an exception is allowed. [5]

Why existing change control may not answer it

This is not a failure mode exclusive to artificial intelligence (AI). Conventional change-management records can establish what configuration was approved and when a deployment occurred. They may not establish which version was consumed by a particular decision: a channel may use cached state, a queued job may retain a previous rule identifier, or the provider’s enforcement service may update later than the bank’s change register.

Where a model contributes a score, classification or recommendation, the decision record may additionally need the model version, relevant input references, threshold or policy-rule version, and any human intervention. The rule service and the AI model may change independently; preserving one version identifier does not necessarily preserve the other.

Reconstruction means explaining the decision’s basis and path from lawful retained records. It is not necessarily a bit-for-bit replay: some model outputs and upstream data dependencies may change between runs. An acceptable means of reconstruction must be specified for the actual workflow, rather than selected after a disputed outcome.

Specify the expiry boundary before testing it

Illustrative scenario — synthetic, not based on an examined system. A case opens at 17:59:59, is queued for retry and reaches a decision service at 18:00:01. A separate new case arrives at 18:00:01. Whether either may use the old exception depends on the approved authority attachment rule—for example, whether authority is checked at case initiation, at each consequential decision or at completion. None is a universal default; the institution must determine which is appropriate and lawful.

At minimum, the responsible business and technology functions should agree:

  • Authority and scope: which actions, channels and systems the exception authorises, and when that permission attaches to a case
  • Time and continuity: the authoritative clock, acceptable differences between clocks, and treatment of retries, sessions and queued work after expiry
  • Operational dependency: which rule services, caches, providers and model or policy versions may affect the decision, and who can revoke or restrict the exception
  • Evidence ownership: what decision-time links should exist, how long they may lawfully be retained, and who can retrieve them for an authorised review

A stateful session may retain an earlier authorisation, while another service rechecks authority for each decision. Neither implementation alone establishes acceptable behaviour; the observed result must be compared with the bank’s approved rule and applicable obligations.

Regulatory context is relevant but does not prescribe this test

Two Indian regulatory sources provide context without mandating the specific examination above. The FREE-AI Committee Report, released by RBI on 13 August 2025, contains committee recommendations on responsible AI governance; those recommendations are not, by themselves, a binding rule for exception expiry. [1][2]

On 24 June 2026, RBI released draft Guidance on Regulatory Principles for Model Risk Management, with comments invited through 24 July 2026. It addresses model approvals, exception governance, validation, monitoring, change management, third-party models and business continuity. A model-approval exception is not the same as a temporary identity-verification routing exception. The draft’s paragraphs 23–24 contemplate retention of decommissioned model inventory entries and related model documentation; they do not establish an equivalent retention period for every individual customer decision. [3][4]

RBI’s 2025–26 Annual Report described model-risk directions as approaching completion, but the official sources reviewed for this article did not establish a subsequently issued final instrument. The draft’s current regulatory status must therefore be independently confirmed before anyone relies on it as a regulatory instrument. [6] Separate, operative commercial-bank customer due diligence requirements remain relevant to the illustrative verification use case. [5]

What management should inspect

Question Evidence to seek What it can establish within scope What remains unresolved
What authority was granted? Approved scope, end time, amendments, decision and rule ownership Documented intent and approved treatment of boundary cases What an operational service actually applied
Did new use cease? Post-expiry decision/routing records and service-state observations Observed enforcement for the examined population or sample and period Unobserved routes, later changes or incomplete logging
What happened to in-flight work? Correlated case events, retry/queue treatment, clock references Behaviour of examined cases against the approved time rule Races and dependencies not covered by the review
Can earlier decisions be reconstructed? Decision references linked to applicable rule, model context, source records and human action Basis and path of selected historical decisions Population completeness or identical model replay
What contradicts, or is missing from, the record? Failed or denied events, alerts, completeness checks and reconciliations Adverse observations and evidenced gaps against explicit criteria Unlogged events and criteria not yet defined

Evidence provenance means establishing where a record originated, which system and version produced it, and whether its integrity and chain of custody can be assessed. Evidence must be appropriate for the question: a search across all available post-expiry decision records may reveal more than a small sample, but its completeness still needs verification; selective reconstruction can then examine the basis of individual decisions.

If reliable evidence establishes that a required record or control is absent against an applicable criterion, an examiner may identify a Gap. If record existence, coverage or the criterion itself cannot be established, the correct result is Unknown, not an invented pass or failure. A successful sample supports a bounded conclusion about that sample and the conditions examined.

Prioritise exceptions by consequence and uncertainty

A bank with numerous exceptions should not assume that every case requires identical examination depth. ODA3 Institute’s proposed sequencing starts with exceptions that change customer eligibility or verification, because their effects can alter customer outcomes; span several channels or providers, because the same approval may be enforced differently; cross queue or retry boundaries, because authority can persist between processing steps; or lack decision-time links, because reconstructing affected cases may be difficult. These are judgement factors, not a validated risk scale.

Management decision Accountable function Decision and revisit condition
Define and approve permissible authority, including in-flight treatment Workflow business owner, supported by risk, legal and engineering Revisit after a material workflow, rule, service or applicable-duty change
Reconcile observed enforcement with approval Platform/service owner; business owner accepts closure Investigate anomalous post-expiry use; assess affected decisions before closure
Preserve lawful records and access rights Data/records governance with privacy and provider-contract owners Revisit if retention duties, provider arrangements or evidence access change
Independently challenge material claims Competent second-line risk/validation function or independent examiner; internal audit as third line where applicable Scope, criteria, negative observations and limitations must be documented
Decide on unresolved material cases Risk leadership; Board committee receives material or recurring issues Remediate, restrict reliance or document authorised residual-risk treatment; reassess as evidence changes

A closure decision may be supported within a specified scope when both operational cessation and historical reconstruction are demonstrated to the agreed extent. A verified failure requires remediation and assessment of affected cases. Unknown means the closure claim remains unsubstantiated within the examination scope; it does not automatically establish a breach. Board oversight should focus on material unresolved matters rather than routine exception sign-offs.

Beyond control review: a limited insurance question

If a particular incident or customer claim arose, the same authorisation, operating-state and decision-history records might be relevant to an insurer’s information request about the specified exposure. No insurer practice or policy response for this scenario was verified. Technical failure does not prove loss; loss does not prove liability; liability does not prove coverage. Those questions require their own facts, contractual wording and applicable law.

Notably Absent

  • No bank or provider system, customer case, operational exception or test result was examined.
  • No evidence establishes how common post-expiry exception persistence or decision-record loss is in banking.
  • No cited RBI source prescribes this precise two-part operational examination, and the draft model-risk document is not an individual-decision retention rule.
  • No verified insurer request, claim, pricing result or coverage determination supports an insurance consequence.
  • No sample result, however successful, establishes complete enforcement across every channel, version and later deployment.

The absence statements are limited to the reviewed material; they do not establish that no such records or events exist elsewhere.

The management question

An exception should not be treated as operationally resolved merely because its approved period has ended. Before sign-off, management can ask: Which records demonstrate that the old authority ceased governing relevant new decisions—and which records still explain the decisions it governed before expiry, including a case crossing the boundary?

This article provides general research analysis of an illustrative examination question. It is not legal, regulatory, compliance or insurance advice, and it reports no finding about a deployed institution.

References

[1] RBI, Report of the Committee to Develop a Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) in the Financial Sector, press release, 13 August 2025. https://www.rbi.org.in/Scripts/BS_PressReleaseDisplay.aspx?prid=61018 (accessed 15 November 2026).

[2] RBI, FREE-AI Committee Report, 13 August 2025. https://rbi.org.in/scripts/PublicationReportDetails.aspx?ID=1306 (accessed 15 November 2026).

[3] RBI, Draft Guidance on Regulatory Principles for Model Risk Management, press release, 24 June 2026. https://www.rbi.org.in/Scripts/BS_PressReleaseDisplay.aspx?prid=63006 (accessed 15 November 2026).

[4] RBI, Guidance on Regulatory Principles for Model Risk Management, 2026, consultation draft, including paragraphs 15 and 23–24. https://www.rbi.org.in/Scripts/bs_viewcontent.aspx?Id=5089 (accessed 15 November 2026).

[5] RBI, Reserve Bank of India (Commercial Banks – Know Your Customer) Directions, 2025, consolidated text updated 1 October 2026. https://rbi.org.in/scripts/BS_ViewMasDirections.aspx?id=13141 (accessed 15 November 2026).

[6] RBI, Annual Report 2025–26, section VI.18. https://www.rbi.org.in/Scripts/AnnualReportPublications.aspx?Id=1466 (accessed 15 November 2026).

Methodology note

Source-bounded desk review with a synthetic 17:59:59/18:00:01 boundary case; separate proof obligations for operational cessation and historical reconstruction; document authority attachment rule, clock, queues, caches, versions and lawful evidence retention; distinguish bounded Gap from Unknown.

Notably Absent

No bank/provider system, customer case, operational exception, field test, prevalence estimate, customer loss, insurance pricing/claim or coverage determination examined. No RBI source prescribes the two-part test or a universal individual-decision retention term. No sampled success proves all channels or versions.