What do RBI's Co-Lending Arrangements Directions 2025 change and when did they take effect?

RBI issued the Co-Lending Arrangements Directions, 2025 (Notification No. RBI/DOR/2025-26/139, August 6, 2025), effective January 1, 2026. The Directions replace the earlier Co-Lending Model circular (RBI/2020-21/63, November 5, 2020) and significantly expand the scope: co-lending is no longer restricted to priority sector loans — any bank-NBFC pair (or AIFI-NBFC-HFC pair) can now co-lend on any loan product. The Directions establish that each Regulated Entity (RE) must hold its portion of the co-originated loan on its own balance sheet, must conduct independent credit assessment before co-originating, must follow its own NPA classification criteria, and — critically for AI governance — if one RE classifies a loan as NPA, the co-lending partner must follow within a specified period. This NPA cascade provision means AI-driven credit decisions and NPA-prediction models in one RE can have balance sheet consequences for the other RE without that RE's own model being involved in the determination.

Under MRMF, which Regulated Entity is responsible for validating the AI credit model in a co-lending arrangement?

RBI MRMF June 2026 assigns model risk governance to the board of the Regulated Entity that owns the model. In a co-lending arrangement, the typical structure is that the NBFC partner uses an AI model for front-end origination and credit scoring, and the bank partner conducts a back-book validation or relies on the NBFC's credit assessment. MRMF Section 7 (Third-Party and Vendor Models) applies when a bank relies on a model it did not develop — the bank cannot treat the NBFC's credit AI as exempt from its MRMF obligations. The bank must: (a) ensure the NBFC's AI credit model is in the bank's model inventory as a third-party model; (b) ensure independent validation has been conducted on the model (not just accepting the NBFC's own validation); and (c) establish ongoing performance monitoring obligations. If the co-lending arrangement uses a single blended-rate AI model that both REs feed and rely on, MRMF Section 5 (Model Risk Governance) requires that the board of one RE own the model risk — but the Directions do not specify which RE. The co-lending MOU must resolve this, designating a lead RE for model ownership and validation with defined access rights for the co-lender.

What is the NPA cascade risk from AI model error in co-lending and how should it be governed?

The Co-Lending Arrangements Directions 2025 require that if one RE classifies a co-originated loan as NPA, the co-lending partner must follow suit within a defined period. This creates an AI model error cascade: if one RE's AI NPA-prediction model generates a false positive — incorrectly predicting that a performing borrower is likely to default, leading to early NPA classification — the co-lending partner's book is affected by that AI error even if the partner's own model would not have flagged the loan. Governing this risk requires: (a) the co-lending MOU should specify that AI-driven NPA classification triggers must be disclosed to the co-lending partner before the classification takes effect, giving the partner an opportunity to apply its own model as a check; (b) the RE whose AI model triggers the NPA classification should maintain documented evidence of the model's prediction and the analyst review that confirmed it; (c) where AI model error results in erroneous NPA classification that cascades to the co-lending partner, there should be a defined inter-RE dispute resolution process; and (d) both REs should conduct joint backtesting to assess historical NPA false-positive rates for the co-originated portfolio, as a baseline for monitoring the cascade risk.

How does FREE-AI Pillar 4 (Explainability) apply to blended-rate credit decisions in co-lending?

FREE-AI Pillar 4 requires that AI-driven credit decisions be explainable to the borrower on request — the factors that influenced the decision must be communicated in terms the borrower can understand. In a co-lending arrangement with a blended interest rate, the credit decision has two contributing AI model assessments: the NBFC's front-end scoring model and the bank's back-book validation. When the borrower asks why their credit was denied or why their interest rate is what it is, the explanation obligation falls on the RE with the customer-facing relationship — typically the NBFC originator. But if the bank's back-book model was the operative assessment (the bank exercised its independent credit assessment right and declined), the NBFC may not have access to the bank's model's reasoning. The co-lending MOU must establish: (a) which RE is responsible for borrower-facing explanation delivery; (b) what explanation data the bank is required to share with the NBFC when the bank's assessment is adverse; and (c) how the combined explanation — derived from two models' assessments — is constructed and delivered in a coherent way that satisfies FREE-AI Pillar 4 without disclosing either RE's proprietary model logic.

AI Governance · RBI Co-Lending · MRMF · NBFC-Bank · India BFSI

RBI Co-Lending Arrangements Directions 2025 AI Governance: Resolving the Joint Credit Model Accountability Gap Under MRMF

RBI's Co-Lending Arrangements Directions 2025 (effective January 1, 2026) create a governance gap that no regulatory guidance resolves: when bank and NBFC co-originate using AI credit models, MRMF assigns model risk to individual RE boards — but the co-lending structure creates dual-board model ownership ambiguity. NPA cascade provisions mean one RE's AI model error propagates to the other's balance sheet. No global AI governance vendor has mapped any of this.

Zytra Research Team · January 2026 · 13 min read

📖 Part of the India BFSI AI Governance Complete Guide 2026 — RBI, SEBI, IRDAI, NPCI & DPDP in one place.

Co-lending between banks and NBFCs has been a structural feature of India's credit market for years. The original 2020 circular established the priority sector co-lending model. RBI's Co-Lending Arrangements Directions 2025 (Notification No. RBI/DOR/2025-26/139, August 6, 2025, effective January 1, 2026) expanded co-lending beyond priority sector to any loan product — creating a much larger canvas for bank-NBFC joint origination.

The AI governance implications of this expansion are unaddressed in the Directions, in MRMF, and in any published guidance. The gap is not theoretical: most NBFC-led co-lending arrangements use AI models in their origination workflow. The NBFC's credit scoring model evaluates the borrower; the bank either relies on that assessment or runs its own back-book validation. MRMF assigns model risk governance to individual RE boards. Co-lending creates a joint decision structure where that assignment is ambiguous.

This post maps the specific AI governance gaps created by the Directions and proposes the governance architecture that co-lending agreements must include to resolve them.

What the Co-Lending Arrangements Directions 2025 Changed

The 2025 Directions preserved the structural mechanics of co-lending — balance sheet sharing, blended pricing, joint origination — while expanding the eligible loan universe from priority sector to all products. For AI governance, the most consequential provision is Direction 15's NPA cascade requirement: if one RE classifies a co-originated loan as non-performing, the co-lending partner must follow the NPA classification within a specified timeframe.

This cascade provision, combined with AI-driven NPA prediction models, creates a specific risk: AI model error in one RE's NPA-prediction system can force the co-lending partner's book into NPA classification through no fault of the partner's own credit assessment. The partner RE cannot use its own model's assessment to override the cascade obligation — if the other RE classifies NPA, the cascade follows.

The MRMF Joint Ownership Gap

MRMF June 2026 Section 5 establishes model risk governance with board-level accountability at the RE that owns the model. The model inventory lives with one RE. Independent validation is that RE's obligation. Ongoing performance monitoring is that RE's responsibility.

Co-lending creates a scenario MRMF's drafters did not explicitly address: what happens when two REs jointly originate credit using AI models, and no single RE is the clear model owner?

Three common co-lending AI architectures each have different governance implications:

Architecture 1: NBFC-owned origination model, bank independent assessment. NBFC uses its own AI credit scoring model to generate an approval recommendation. Bank reviews the NBFC's assessment and conducts its own independent credit evaluation (which may or may not be AI-driven). MRMF assigns model ownership to the NBFC for its scoring model. The bank's independent assessment is the bank's own model governance obligation. Governance gap: the bank's independent credit assessment relies on the NBFC's scoring output as an input — is the bank a model user under MRMF Section 7 (Third-Party Models) for the NBFC's model? If so, the bank has independent validation obligations for the NBFC's AI model, in addition to its own assessment.

Architecture 2: Single shared model, joint data. Bank and NBFC co-develop or co-procure an AI credit model that uses both institutions' data in its training. Both REs rely on its outputs. MRMF Section 5 requires one RE's board to own the model risk — but shared development and joint reliance mean no natural single owner emerges. The co-lending MOU must designate a lead RE for MRMF ownership purposes, with defined access and audit rights for the co-lending partner.

Architecture 3: Separate models, blended credit decision. Each RE uses its own AI credit model, and the final credit decision reflects both models' outputs (scored separately, then combined or adjudicated). MRMF is clear that each RE governs its own model. The governance gap is at the output layer: when the blended decision is adverse to the borrower, which model's reasoning is the basis for the FREE-AI Pillar 4 explanation? If both models contributed, the explanation must be constructed from both, which requires data sharing between the REs that may not have been anticipated in the co-lending MOU.

Independent Validation in Co-Lending: MRMF Section 7

MRMF Section 7 (Third-Party and Vendor Models) applies when an RE relies on a model it did not develop. In a co-lending arrangement where the bank relies on the NBFC's AI credit scoring, the bank is in a third-party model relationship with the NBFC's model. The bank's MRMF obligations under Section 7 require:

The NBFC's AI model must appear in the bank's model inventory as a third-party model
Independent validation of the NBFC's model — the bank cannot accept the NBFC's own validation as satisfying its MRMF obligation; a bank-commissioned or jointly-commissioned independent validator must review the NBFC's model
Ongoing performance monitoring — the bank must track the NBFC's model's performance on the co-originated portfolio, not just at the NBFC's reported aggregate level
Change notification — the co-lending MOU must require the NBFC to notify the bank before any material change to the AI credit model, giving the bank time to assess the change's impact on its co-originated book

This creates a significant due diligence obligation that most current co-lending agreements do not include. An NBFC partner that has positioned its AI credit model as proprietary and refused model documentation access to its bank co-lending partners is effectively preventing the bank from meeting its MRMF obligations.

NPA Cascade Risk Governance

Direction 15's NPA cascade means AI model error propagation across REs. Governing this requires provisions that current co-lending MOUs typically do not contain:

Pre-cascade disclosure: The RE whose AI model triggers NPA classification should notify the co-lending partner before the classification takes effect, providing the model output, any analyst override or confirmation, and the opportunity for the partner to apply its own assessment as a check.

Joint backtesting baseline: Both REs should conduct annual joint backtesting to measure historical NPA false-positive rates for the co-originated portfolio. A rate above an agreed threshold triggers a governance review of the triggering RE's AI model.

Dispute resolution: When the co-lending partner disputes an NPA classification that cascades from the other RE's AI model, an expedited inter-RE dispute resolution process should determine whether the triggering RE's AI model produced an error, and what remediation is available to the co-lending partner's book.

Financial remedy: If AI model error in one RE's NPA prediction causes the co-lending partner to incur provision costs for a performing loan incorrectly classified as NPA, the co-lending MOU should specify whether the triggering RE has indemnity exposure. This is a commercial negotiation, but without a provision in the MOU, the co-lending partner has no remedy.

Co-Lending AI Governance Checklist

☐ Co-lending AI architecture mapped: which model(s) are used, who owns each, which RE has the customer relationship
☐ Lead RE for MRMF ownership designated in co-lending MOU: model inventory assignment, board accountability, validation responsibility
☐ Bank's MRMF Section 7 obligations satisfied: NBFC AI model in bank's model inventory as third-party model; independent validation conducted or commissioned
☐ Model documentation access: NBFC must provide sufficient model documentation to the bank for MRMF independent validation — this provision in co-lending MOU
☐ Change notification: MOU requires NBFC to notify bank before material AI model changes; re-validation timeline agreed
☐ Performance monitoring: bank tracks NBFC model performance on co-originated portfolio; PSI/CSI monitoring covers co-origination book
☐ NPA cascade governance: pre-cascade disclosure provision in MOU; joint backtesting baseline; dispute resolution process; indemnity provisions for AI model error cascade
☐ FREE-AI Pillar 4 explainability: responsibility assigned for borrower-facing explanation; explanation data sharing between REs specified in MOU; combined explanation construction process documented
☐ DPDP Act: borrower data sharing between REs for model training or inference has lawful basis; data processing agreement between co-lending REs in place
☐ RBI examination readiness: joint documentation package — both REs' model governance documentation covers co-originated book consistently