← Return to the book

FUTURECENTRAL PRESS · BOOK SAMPLE

Compliance, Risk, and AI Governance

A Guide for Compliance Professionals in Financial Services

Chapter 29: Third-Party and Vendor Risk Architecture for AI: Foundation Model Dependencies, Concentration Risk, and the Contractual Architecture

Foundation model dependency mapping, concentration risk on AI providers, the contractual architecture for exit, audit, liability and indemnity, cloud-AI and RegTech vendor risk, and the third-party ecosystem the working DGM must govern.

Chapter Outcomes

By the end of this chapter, the working professional in a financial services compliance, risk, audit, procurement, or third-party risk function should be able to:

Define AI vendor risk as a distinct family of third-party risk categories (foundation model provider, AI platform provider, AI tool provider, AI service provider) and articulate why the conventional outsourcing taxonomy is insufficient on its own.

Conduct a foundation model concentration analysis for the institution's deployment portfolio, identifying single-provider dependencies, the substitutability of each dependency, and the institutional risk position the board should hold on the resulting concentration profile.

Specify the ten contractual provisions the institution requires in foundation model provider relationships (data flow, model behaviour and version stability, liability, indemnification, continuity and exit, supervisory access and audit, incident response, information security, regulatory compliance, change management) and the institutional position on each.

Operate the five-layer AI Vendor Risk Architecture (vendor categorisation, due diligence and onboarding, contractual architecture, concentration risk and continuity, integration with broader risk architecture) as a single working discipline integrated with the institutional outsourcing framework.

Reconcile the AI vendor risk discipline with the RBI Master Direction on Outsourcing of Information Technology Services (10 April 2023), the RBI Master Direction on Outsourcing of Financial Services, the SEBI June 2025 AI/ML consultation paper, the RBI FREE-AI Committee Report (13 August 2025), the IRDAI Information and Cyber Security Guidelines, the DPDP Act 2023 and DPDP Rules 2025, the EU DORA (Regulation (EU) 2022/2554, applicable 17 January 2025), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), the Federal Reserve SR 23-4 Interagency Guidance (June 2023), and the BCBS Principles for Operational Resilience (March 2021).

Conduct the AI Vendor Risk Maturity Audit (Applied Exercise) and produce a board-grade artefact the institution can take to the Vendor Risk Committee and to supervisory examination.

Opening Vignette

Sandeep Banerjee, Deputy General Manager (Third-Party and Vendor Risk) at a large Kolkata-headquartered private sector bank, walks into the Friday morning Vendor Risk Committee with a stack of vendor risk assessments and an institutional position paper on a contract renewal that will, once it is signed, deepen the bank's foundation model dependency on a single provider for another three years. The position paper is fifteen pages long; the assessment stack is two inches thick; the legal team has been on the contract terms for six weeks. The renewal is, on the bank's standard commercial scoring, an unremarkable procurement decision. On the bank's institutional vendor risk discipline, the renewal is the most consequential third-party decision the bank will take in the calendar year.

The commercial position on the renewal is straightforward. The provider's foundation model is operating in eleven production deployments inside the bank: customer service summarisation, internal knowledge retrieval, fraud narrative synthesis, advisor decision support, the relationship manager briefing engine, the credit committee summary generator, the operations help-desk triage agent, the legal document review assistant, the regulatory horizon-scanning tool, the internal audit narrative drafter, the trade surveillance alert reviewer. The deployments produce, in aggregate, throughput volume that the business considers material. The procurement function has secured commercial terms the business considers acceptable. The signature is, on the procurement track, ready to go.

The risk position on the renewal is more complicated. The renewal terms produce concentration risk that Sandeep's function has been flagging in the Vendor Risk Committee's quarterly reviews for the past nine months. The contractual provisions on data flow are weaker than the institution's contractual playbook calls for. The exit architecture, on inspection, does not contemplate a scenario in which the provider's behaviour changes materially mid-term (a change of model behaviour, a change of pricing, a change of data use position, a change of control at the provider, a change of regulatory posture by the provider's home-jurisdiction supervisor. The indemnification scope falls short of the institutional position the legal function has articulated for foundation model relationships. The substitutability assessment) the institutional analysis of how the bank would move the eleven deployments to a second provider in a defined window: has not been refreshed in fourteen months.

Sandeep opens the meeting with a fifteen-minute walk-through. He presents the concentration analysis, the deployment dependency map, the substitutability assessment as it stands, the gap analysis against the contractual playbook, and the four contractual positions on which the bank's negotiation team has secured no movement from the provider. Three questions come from the Chief Risk Officer. What is the renewal decision if the provider declines to modify the four terms? Who in the institutional governance machinery is the decision-maker, and what is the escalation path to the Risk Management Committee of the Board? And (the question that closes the meeting) is the bank's continuity architecture, as it stands today, one the bank could operate if the provider were unavailable for ninety days, six months, or longer? The committee adjourns at noon with a fortnight to convene the senior governance machinery before the renewal deadline.

The Friday Vendor Risk Committee meeting is the institutional venue where the bank's foundation model dependencies are governed. The meeting is the institutional artefact this chapter develops. The discipline the meeting represents (the architectural commitment that AI vendor relationships sit inside a vendor risk function operating with the depth that foundation model dependencies require) is the working DGM's craft. Sandeep is a composite character drawn from publicly disclosable patterns in Indian bank third-party risk functions and from industry-working-group conversations; the institutional reality the vignette reflects is the reality the chapter equips the working DGM to operate.1

"The DGM's call this Friday, escalated through the Vendor Risk Committee to the CRO: which foundation model dependency is the institution about to renew because the concentration risk, the contractual exit architecture, and the substitutability assessment are being treated as commercial questions rather than as the institutional risk commitments they must be, and which two of the three need to be re-engineered into the institution's vendor risk framework before the next supervisory examination treats the gap as an outsourcing failure?"

29.1 What AI vendor risk actually is in 2026

AI vendor risk is the institutional category of risk arising from the institution's dependencies on third parties that supply AI capability or AI-related services. The category sits inside the broader outsourcing and third-party risk discipline the institution has operated for decades, but it carries dependencies, substitutability characteristics, and contractual asymmetries that the inherited outsourcing taxonomy was not built for. The working DGM's first task is the categorisation move: AI vendor risk is not a single category but a family.

Four categories have stabilised as institutional convention. The first category is the foundation model provider (OpenAI, Anthropic, Google, Meta, Mistral, Cohere, and the smaller specialist providers) supplying the large language model or other foundation model that the institution's generative and agentic deployments are built on. The relationship is the deepest of the AI vendor relationships because the foundation model carries behavioural characteristics, training data provenance, version stability commitments, and data use practices that are, in operational terms, the institutional dependency itself. The second category is the AI platform provider, supplying the orchestration, integration, monitoring, evaluation, and governance scaffolding on which the institution's AI deployments are built. The third category is the AI tool provider, supplying a specific tool the institution integrates (an evaluation harness, a vector database, a prompt management system, a guardrails layer. The fourth category is the AI service provider, supplying an AI-supported business service the institution consumes) a fraud detection service, a credit decisioning service, a customer service agent. Each category carries a distinct risk treatment.2

AI vendor dependencies are deeper and harder to substitute than conventional IT vendor dependencies. A cloud service provider relationship is, in principle, substitutable on a defined timeline because the cloud workload is portable across providers using mature institutional patterns. A foundation model provider relationship is not substitutable on the same timeline. The provider's specific model behaviour, the institution's deployment-specific prompts and evaluation infrastructure tuned to that behaviour, the upstream changes the institution has absorbed over the life of the relationship: all carry institutional adaptation that the substitution exercise must redo. The substitutability question is the institutional question the contractual exit architecture must answer.

AI vendor risk produces obligations under multiple regulatory frameworks that the institution must satisfy in coordination. The RBI Master Direction on Outsourcing of Information Technology Services (issued 10 April 2023, effective 1 October 2023) carries the foundational expectation that the regulated entity remain responsible for outsourced activities. The RBI Master Direction on Managing Risks and Code of Conduct in Outsourcing of Financial Services by Banks (in its current edition) carries the cognate expectation for financial services outsourcing. The RBI Master Direction on Digital Lending (September 2022, with amendments) carries the digital-lending-specific articulation. The SEBI June 2025 consultation paper on AI/ML and the RBI FREE-AI Committee Report (published 13 August 2025) extend the supervisory expectation onto AI vendor relationships specifically. The DPDP Act 2023 (11 August 2023) and the DPDP Rules 2025 (13 November 2025) carry the data-protection obligations on the Data Fiduciary that any vendor relationship must accommodate. The EU DORA (Regulation (EU) 2022/2554, applicable 17 January 2025) and the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02) carry the European supervisory architecture for institutions with EU touchpoints. The Federal Reserve, OCC, and FDIC Interagency Guidance on Third-Party Relationships (SR 23-4, June 2023) carries the United States supervisory architecture. The institutional discipline must coordinate these obligations rather than treat them in parallel.3

AI vendor risk is the vendor-side mirror of the model risk discipline the bank operates internally. The institutional model risk function (Chapter 21; SR 11-7, April 2011; PRA SS1/23, May 2023 in force May 2024) operates on models the institution has developed or controls. AI vendor risk operates on models the institution does not develop and does not fully control, and the gap between what the institution validates internally and what the vendor exposes externally is, in 2026, the supervisory gap the working DGM must close. The NIST AI RMF 1.0 (26 January 2023) and the NIST AI 600-1 Generative AI Profile (26 July 2024) are the methodological anchors; ISO/IEC 42001:2023 (18 December 2023) is the management-system anchor; the EU AI Act (Regulation (EU) 2024/1689), with the €35 million Article 5 prohibition tier and the Chapter V Articles 51-56 obligations on general-purpose AI providers, is the jurisdictional anchor for foundation model providers operating in the EU and, by extension, for the institutional position on those providers' obligations as they appear in vendor contracts.4

29.2 Foundation model dependencies: concentration risk and substitutability

The institutional view of foundation model concentration risk is the second discipline the working DGM operates. The concentration question has three dimensions the institution must hold simultaneously: which deployments depend on which providers, how substitutable each dependency is, and what the institutional position on the resulting concentration profile is.

The foundation model dependency map is the working artefact. The institution maintains, on a documented cycle, an inventory of AI deployments by foundation model provider: the provider, the model version, the deployment use case, the volume, the materiality, the data sensitivity, the customer-facing or internal nature of the deployment. The map shows, at a glance, the institutional concentration on each provider. The map is the artefact the Risk Management Committee of the Board sees; the map is the artefact the supervisor will ask for first; the map is the artefact that, if it does not exist, the institution does not actually know its dependency profile.

Substitutability is the operational question the dependency map cannot answer on its own. For each material foundation model dependency, the institution maintains a substitutability assessment that addresses three sub-questions. First, what alternative provider could supply equivalent capability for this deployment, at what operational cost, on what migration timeline. Second, what institutional adaptation (prompt re-engineering, evaluation re-tuning, integration re-work, downstream model fine-tuning) would substitution require. Third, what residual capability gap would remain, and is that gap one the institution would accept temporarily, accept permanently, or refuse to accept. The substitutability assessment is the discipline that converts the dependency map from a static inventory into an institutional risk position.

The institutional concentration posture is the deliberate choice the board must own. Some institutions, by design, operate a single-foundation-model architecture because the integration overhead, the consistency benefits, and the relationship leverage outweigh the concentration cost. Other institutions operate multi-provider architectures because the concentration cost outweighs the integration overhead. Either posture is defensible if the institution has documented the choice, the analysis behind the choice, and the senior governance that has affirmed the choice. The indefensible posture is the concentration that has accumulated through individually defensible procurement decisions without any institutional moment at which the senior governance machinery has examined the accumulated profile. The CrowdStrike outage on 19 July 2024 (when a faulty Falcon sensor update from a single security vendor crashed approximately 8.5 million Windows systems globally, with cascading effects on airlines, banks, hospitals, and emergency services) is the canonical recent lesson on what concentration on a single global vendor can produce when the vendor's release process fails.5

Continuity architecture is the third leg. For each material foundation model dependency, the institution maintains a documented fallback that specifies what the deployment does when the foundation model is unavailable for an hour, for a day, for a week, for a month. The fallback may be a degraded service, a different model, a human-in-the-loop process, or a deployment pause. The fallback architecture is exercised in drill on a documented cycle; unexercised fallbacks are continuity fictions the supervisor will identify on examination. The Wirecard collapse in June 2020 (when the German payments processor's insolvency left its bank and merchant clients with the institutional question of how to continue operations on a third-party stack they did not control) is the canonical lesson on what vendor failure produces for institutions that have not exercised the continuity architecture (Chapter 1 develops the Wirecard case in fuller treatment).6

29.3 The contractual architecture: exit, audit, liability, indemnity, model provenance

The institutional contract architecture for AI vendor relationships extends the conventional outsourcing contract architecture with ten provisions the discipline treats as essential. The contractual playbook articulates the institutional position on each provision; the negotiation team carries the playbook into the commercial conversation with each vendor; gaps between the playbook and the as-negotiated contract become institutional risk acceptances the Vendor Risk Committee must approve at named seniority.

The ten contractual provisions of the AI vendor playbook. (1) Data flow and data use: how the institution's data (prompts, retrieved content, outputs, intermediate states) flows through the vendor's infrastructure, what the vendor may do with the data, whether the data is used for vendor model training, the data residency position, the cross-border transfer provisions. (2) Model behaviour and version stability: whether the vendor commits to behavioural stability of the supplied model, the notice period for material version changes, the institution's rights at a behaviour change. (3) Limitation of liability: vendor liability for damages arising from the supplied capability, with attention to hallucination harm, model bias harm, security breach harm. (4) Indemnification: vendor indemnity for the institution against intellectual property claims arising from model outputs, data protection claims, regulatory penalties traceable to vendor conduct. (5) Continuity and exit: the institution's rights and the vendor's obligations on vendor failure, change of control, service discontinuation; the data migration provisions; the model state preservation provisions; the wind-down timeline. (6) Supervisory access and audit: the institution's right to audit the vendor, the regulator's right to examine the vendor through the institution, the documented evidence the vendor maintains, the audit scope. (7) Incident response: the vendor's obligation to notify the institution of incidents affecting the supplied capability, the notification timeline, the support the vendor provides during institutional response. (8) Information security: the vendor's security commitments, the controls operated, the standards adhered to (ISO/IEC 27001, SOC 2, ISO/IEC 42001 where applicable), the right to evidence. (9) Regulatory compliance: the vendor's commitment to support the institution's compliance with the RBI Outsourcing directions, the DPDP Act and Rules, the CERT-In Direction (No. 20(3)/2022-CERT-In, 28 April 2022, effective 27 June 2022), the IRDAI Information and Cyber Security Guidelines (2023, revised 2026), the SEBI AI/ML expectations, and (where applicable) the EU AI Act and DORA. (10) Change management and governance: how vendor-side changes to the capability are communicated, what role the institution has in change approval, what the deprecation timeline is for capability the vendor sunsets.7

Standard vendor terms are insufficient for institutional use. The standard terms supplied by foundation model providers are, in most cases, materially insufficient for financial services institutional use across all ten provisions. The standard data use language permits uses the institution cannot accept; the standard liability cap is well below institutional exposure; the standard exit language does not contemplate institutional continuity needs; the standard audit language does not provide for supervisory access; the standard indemnity scope is narrow. The institutional posture is, accordingly, that standard terms are the negotiation starting point, not the negotiation endpoint. The vendor's commercial willingness to depart from standard terms is, in itself, a vendor risk signal: the vendor that refuses to negotiate institutional protections is the vendor whose vendor risk profile the institution must price into the procurement decision.

Model provenance documentation is the contractual evidence regime. The vendor must produce, and update, the documentation that supports the institution's internal model risk discipline: the training data provenance description, the evaluation methodology and benchmark results, the known limitation documentation, the safety evaluation history, the bias and fairness assessment, the version change log with rationale. The NIST AI 600-1 Generative AI Profile and ISO/IEC 42001:2023 articulate the documentation expectation; the EU AI Act Chapter V Articles 51-56 carry the regulatory expectation for general-purpose AI providers in the EU; the institutional contractual position must require the documentation in usable form, on an agreed cycle, with evidence rights at supervisory examination.8

The contractual position on the foundation model relationship is the institution's audit trail when the relationship fails. When the vendor's behaviour changes mid-term, when the vendor's pricing changes mid-term, when the vendor's regulatory posture changes mid-term, when the vendor's ownership changes mid-term, the institution's contractual position is the artefact that determines what the institution can do. The supervisor, examining the institution after a vendor failure, will read the contract first and the institutional risk acceptances second. The contractual architecture is, accordingly, the most consequential institutional artefact in the AI vendor risk discipline: not because the institution expects to litigate but because the institutional position the contract reflects is the position the institution will live with for the duration of the relationship.

29.4 Cloud-AI vendor architecture, RegTech vendor architecture, and the third-party ecosystem

AI vendor risk is not contained within the foundation model relationship. The institution's AI deployments rest on a stack (cloud infrastructure, AI platform, integration middleware, evaluation tooling, observability, RegTech surfaces) that aggregates third-party exposure across multiple vendor relationships. The working DGM operates the AI vendor risk discipline across the full stack, not on the foundation model relationship alone.

Cloud-AI vendor architecture is the largest of the adjacent dependencies. The hyperscale cloud providers (AWS, Microsoft Azure, Google Cloud Platform) supply the infrastructure on which most institutional AI deployments run; the cloud provider's foundation model offerings (Anthropic on Bedrock, OpenAI on Azure, Google Gemini on GCP) interleave the cloud and foundation model relationships. The institutional cloud-AI architecture concentrates two categories of risk in the same vendor: infrastructure dependency and foundation model dependency. The institutional cloud governance (which the RBI Master Direction on Outsourcing of IT Services (10 April 2023) addresses at depth, including the requirement that outsourcing arrangements address the lifecycle of data from entry into cloud through to permanent erasure) must extend to the cloud-AI architecture with attention to the doubled dependency the architecture produces.9

RegTech vendor architecture is the supervisor-facing dependency. RegTech vendors supplying transaction monitoring (Chapter 5), regulatory reporting (Chapter 27), risk aggregation (Chapter 27), and supervisory engagement support carry institutional dependency at the regulatory perimeter. Many RegTech providers have embedded AI capability into their platforms over the 2023-2026 window; the institutional vendor risk position on the RegTech relationship must, accordingly, treat the AI capability as a vendor risk dimension of the RegTech relationship, not as a separate matter. The SBI institutional RegTech architecture, which has operated through a multi-vendor RegTech stack with the deliberate institutional posture of avoiding RegTech vendor concentration on any single AI-platform supplier, is the publicly disclosable institutional pattern the working DGM can examine.

Supply chain attacks are the cybersecurity dimension of vendor concentration. The SolarWinds attack disclosed in December 2020 (when nation-state-attributed adversaries compromised the SolarWinds Orion software update mechanism and distributed malicious code to approximately 18,000 organisations including United States federal agencies and global enterprises) is the canonical lesson on what supply chain compromise produces when a single trusted vendor channel is weaponised. The CrowdStrike outage on 19 July 2024 was a release-process failure, not a supply chain attack, but the impact pattern (single vendor, global blast radius, recovery measured in days for many affected organisations) is the same pattern the institutional vendor risk discipline must price. The institutional cyber vendor risk position (developed in fuller treatment in Chapter 30) must, accordingly, treat AI vendor relationships as a supply chain attack surface and must require evidence of vendor secure-software-development practices in the contractual architecture.10

Cross-border foundation model exposure is the jurisdictional dimension. The foundation model providers headquartered outside India operate under home-jurisdiction regulatory architectures (the EU AI Act for European-headquartered providers; the United States executive-order-and-NIST architecture for United States-headquartered providers; the United Kingdom sectoral architecture for UK providers) that may evolve in directions the institution must anticipate. Bajaj Finance, ICICI Bank, and the larger NBFCs and banks with cross-border foundation model exposure operate the institutional discipline of tracking home-jurisdiction regulatory developments at the providers they depend on; the discipline is the cross-border equivalent of the supervisory-engagement discipline (Chapter 16) the institution operates with its own home regulator.

29.5 The working DGM's reading discipline on AI vendor risk

The chapter closes on the reading discipline because the AI vendor risk function is the function in which the gap between commercial procurement reflex and institutional risk discipline opens fastest, and the gap closes only when the DGM has named the architectural commitments the function operates under and has built them into the institutional rhythm.

The DGM reads the foundation model dependency map on a documented cycle. Monthly at minimum; weekly during contract renewal cycles; in real time when a vendor incident is in flight. The DGM reads the substitutability assessment cycle to confirm the assessments are current, the assumptions current, the migration windows credible. The DGM reads the contractual playbook against the as-negotiated contracts on a documented cycle and ensures every gap is documented as an institutional risk acceptance at named seniority. The DGM reads the continuity drill log to confirm the fallback architecture has been exercised within a defined window. The DGM reads the exit plan inventory and confirms each material AI vendor relationship has a current, reviewed exit plan the institution could operate.

The DGM reads the regulatory perimeter on AI vendor risk monthly. The RBI Master Direction on Outsourcing of IT Services (10 April 2023) and the RBI Master Direction on Outsourcing of Financial Services as the foundational architectures; the RBI Master Direction on Digital Lending (September 2022) for digital lending exposure; the RBI FREE-AI Committee Report (13 August 2025) and the RBI (Commercial Banks: Governance) Directions, 2025 (issued 28 November 2025) for the governance frame; the SEBI June 2025 AI/ML consultation paper for the market intermediary frame; the IRDAI Information and Cyber Security Guidelines (2023, revised 2026) for the insurance frame; the DPDP Act 2023 (11 August 2023) and the DPDP Rules 2025 (13 November 2025) for the data fiduciary frame; the CERT-In Direction No. 20(3)/2022-CERT-In (28 April 2022, effective 27 June 2022) for the cyber incident frame; the EU DORA (Regulation (EU) 2022/2554, applicable 17 January 2025) and the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02) for EU exposure; the Federal Reserve SR 23-4 (June 2023) and OCC Heightened Standards for United States exposure; the BCBS Principles for Operational Resilience (March 2021) for the global operational-resilience frame.11

The DGM reads the institutional governance machinery on AI vendor risk quarterly. The Vendor Risk Committee minutes, the Risk Management Committee of the Board AI vendor risk paper, the Audit Committee AI vendor risk paper, the Operational Risk Committee reconciliation on vendor-related loss events (Chapter 28), the regulatory-reporting reconciliation on vendor-supplied data quality (Chapter 27). The DGM reads not for content but for institutional rhythm: is the senior governance machinery actually examining the AI vendor portfolio, or is the governance machinery treating AI vendor risk as a procurement administrative matter the function handles below the governance line?

The DGM reads the institutional defence at supervisory examination as the test of the discipline's adequacy. The supervisor's examination of the AI vendor risk discipline reads three artefacts: the foundation model dependency map, the contractual playbook against the as-negotiated contracts, and the continuity drill log. If those three artefacts are current, complete, and consistent with the institutional risk position, the examination produces no adverse finding. If those three artefacts are stale, incomplete, or inconsistent with the institutional risk position, the examination produces the supervisory letter that begins, 'The bank's third-party risk discipline does not, in its application to AI vendor relationships, meet the standard the regulated entity is expected to maintain.' The institutional defence the DGM builds is the assurance that the artefacts will not be stale on the day the supervisor opens the examination.

Framework: The AI Vendor Risk Architecture

The chapter's architectural commitment, made operational, is a five-layer AI Vendor Risk Architecture the working DGM operates as a single institutional discipline. Each layer is an institutional commitment, not a workstream; the layers operate concurrently, with the Friday Vendor Risk Committee meeting as the institutional venue at which the layers are reconciled.

Layer 1: AI Vendor Categorisation Architecture. The institutional discipline of distinguishing the four categories (foundation model provider, AI platform provider, AI tool provider, AI service provider) and applying the appropriate risk treatment to each. The categorisation governs due diligence depth, contractual provisions, monitoring intensity, continuity expectations, and exit planning. Failure mode: the institution treats all AI vendor relationships as a single category and applies a generic risk treatment that is too deep for some relationships and too shallow for others.

Layer 2: Due Diligence and Onboarding Architecture. The conventional due diligence (financial stability, operational maturity, control environment, regulatory standing, business continuity) extended with AI-specific elaboration: training data provenance, model behaviour stability history, safety evaluation, data use practices, version change discipline, jurisdictional exposure. Due diligence depth is calibrated to vendor category and proposed materiality. Onboarding addresses access controls, data flow architecture, integration discipline, monitoring instrumentation, and the institutional documentation regime that supports the relationship through its life.

Layer 3: Contractual Architecture. The ten contractual provisions of §29.3 (data flow, model behaviour and version stability, liability, indemnification, continuity and exit, supervisory access and audit, incident response, information security, regulatory compliance, change management) operated through a documented institutional playbook. Every gap between the playbook and the as-negotiated contract is a named institutional risk acceptance approved at appropriate seniority. The playbook is updated as the regulatory perimeter and the institutional risk position evolve.

Layer 4: Concentration Risk and Continuity Architecture. The foundation model dependency map, the substitutability assessment for each material dependency, the institutional concentration posture the board has affirmed, the fallback architecture for each material deployment, the exit plan for each material relationship, the periodic exit drill discipline. The continuity architecture is exercised, not declared. The supervisor reads the drill log as the test of whether the continuity architecture is real or fictional.

AI vendor risk integrates with operational risk (Chapter 28), regulatory reporting (Chapter 27), cyber risk (Chapter 30), data protection, model risk (Chapter 21), and supervisory engagement. Vendor-related loss events feed the operational risk taxonomy; vendor data quality affects reported figures; and the vendor cyber surface affects the institution’s cyber posture. Integration operates at six points: policy, governance, data, reporting, people, and training.

The Architecture's use, for the working DGM. The five layers are not a maturity ladder the institution climbs sequentially; they are concentric architectural commitments the institution operates simultaneously, with the Vendor Risk Committee as the institutional rhythm at which the layers are reconciled, the Risk Management Committee of the Board as the senior governance machinery at which the architecture is affirmed, and the supervisor's examination as the external test of the architecture's adequacy.

India Cases

Three Indian institutional cases (drawn from publicly disclosable patterns and from industry-working-group experience) illustrate the AI Vendor Risk Architecture in operation.

HDFC Bank: foundation model dependency governance at the institutional altitude. HDFC Bank, the largest private sector bank in India and an institution with material AI deployment across customer service, advisor support, and internal operations, has (at the institutional altitude that is publicly disclosable through annual reports, risk management framework documentation, and supervisory engagement materials) operated the foundation model dependency governance discipline as an explicit board-level commitment. The institutional posture the bank's disclosures support is that foundation model concentration is a risk category the Risk Management Committee of the Board examines, that the bank's contractual position with foundation model providers reflects the institutional risk appetite, and that the continuity architecture for material AI deployments is exercised on a documented cycle. The lesson for the working DGM: at institutional scale, foundation model dependency governance must be a board-visible discipline with the standard institutional artefacts (dependency map, substitutability assessment, contractual playbook, continuity drill log) and not a procurement administrative matter handled below the senior governance line.12

State Bank of India: RegTech vendor architecture as concentration-aware design. State Bank of India, the largest public sector bank in India and an institution with the most diverse RegTech vendor portfolio in the Indian banking sector, has (across the supervisory reporting, transaction monitoring, fraud detection, and risk aggregation perimeters) operated a multi-vendor RegTech architecture as a deliberate institutional posture. The architecture distributes RegTech dependency across multiple vendors rather than consolidating on a single platform, with the explicit purpose of avoiding concentration on any single RegTech vendor or (as AI capability has been embedded into RegTech platforms over the 2023-2026 window) on any single underlying foundation model supplier. The lesson for the working DGM: at the institutional scale at which RegTech vendor concentration becomes a material risk, the multi-vendor architecture is the institutional answer, and the architecture must be documented as a deliberate posture the senior governance machinery has affirmed.13

ICICI Bank and Bajaj Finance: cloud-AI architecture and cross-border foundation model exposure. ICICI Bank's cloud-AI architecture and Bajaj Finance's cross-border foundation model exposure (both at institutional altitudes that publicly disclosable patterns support) illustrate the doubled-dependency challenge the cloud-AI architecture produces and the jurisdictional exposure that cross-border foundation model relationships carry. ICICI Bank, operating institutional AI deployments on hyperscale cloud infrastructure with the foundation model offering of the same provider, faces the integration economics that argue for concentration and the risk economics that argue against. Bajaj Finance, with material AI deployments resting on foundation models supplied by providers with United States and European home jurisdictions, tracks home-jurisdiction regulatory developments (the EU AI Act Chapter V Articles 51-56 obligations on general-purpose AI providers; the United States NIST AI RMF posture and successor regulatory architecture) as institutional inputs to the vendor risk assessment cycle. The lesson for the working DGM: cloud-AI doubled dependency requires explicit institutional analysis at the senior governance machinery, and cross-border foundation model exposure requires the cross-border equivalent of supervisory-engagement discipline.14

Global Cases

Three global cases (Wirecard, CrowdStrike, and SolarWinds) anchor the AI Vendor Risk Architecture in the broader history of third-party failures the working DGM must price.

Wirecard: the third-party concentration failure. Wirecard AG, the German payments processor whose insolvency was disclosed in June 2020 following the discovery that approximately €1.9 billion in claimed escrow assets did not exist, left the bank and merchant institutions that depended on Wirecard's payments infrastructure with the institutional question of how to operate without a vendor on which their operations had become concentrated. The lessons for the working DGM operating the AI vendor risk discipline are three. First, third-party vendor risk includes governance and integrity risk at the vendor (the foundation model provider whose internal governance fails is the foundation model provider whose service becomes unavailable. Second, vendor failure scenarios must be exercised, not declared) the institutional fallback architecture that has not been drilled is the architecture that does not work when the vendor fails. Third, concentration on any single vendor (payments processor, foundation model provider, RegTech platform) accumulates institutional dependency that the senior governance machinery must examine deliberately. The Wirecard case is developed in fuller treatment in Chapter 1 as the canonical anchor for the institutional commitment-discipline conception of risk and compliance.15

CrowdStrike: global vendor concentration and release-process failure. On 19 July 2024 at 04:09 UTC, CrowdStrike (the United States cybersecurity company supplying the Falcon endpoint protection platform to a substantial fraction of global enterprise) distributed a faulty configuration update (Channel File 291) for its Falcon sensor on Microsoft Windows systems. The update produced an out-of-bounds memory read that crashed approximately 8.5 million Windows systems globally. Airlines, airports, banks, hospitals, broadcasters, and government services were affected; the worldwide financial damage has been estimated in the tens of billions of dollars; remediation required physical access to many affected systems and extended over days for many organisations. The CrowdStrike incident was a release-process failure at a single global vendor with universal blast radius, and it is the canonical recent lesson on what single-vendor global concentration produces when the vendor's release process fails. The lesson for the working DGM operating the AI vendor risk discipline: foundation model concentration and cloud-AI concentration carry the same blast-radius profile; the institutional posture on that profile must be deliberate, board-affirmed, and exercised in continuity drill.16

SolarWinds: the supply chain attack lesson. The SolarWinds Orion compromise, disclosed in December 2020, was a nation-state-attributed supply chain attack in which adversaries compromised the SolarWinds Orion software build process and distributed malicious code through the routine update channel to approximately 18,000 organisations including United States federal agencies, technology firms, and global enterprises. The SolarWinds case is the canonical lesson on what supply chain compromise produces when a trusted vendor channel is weaponised (institutional perimeter defences are bypassed because the malicious payload arrives through a channel the institution has authorised. The lesson for the working DGM operating the AI vendor risk discipline: foundation model providers, AI platform providers, and the deeper AI tool providers operate trusted channels into the institution; the institutional position must require vendor evidence of secure-software-development practices, must include the cyber dimension in the vendor due diligence, and must) as Chapter 30 develops: treat the AI vendor relationship as a cyber attack surface the institution governs as part of its broader cyber posture.17

Regulatory Landscape: AI Vendor Risk

The regulatory landscape for AI vendor risk operates across the global supervisory architecture and the four-regulator Indian architecture, with each layer extending the inherited outsourcing and third-party risk frameworks onto AI vendor relationships.

Global Layer, the architectural foundation: third-party risk supervisory guidance. The Federal Reserve, OCC, and FDIC Interagency Guidance on Third-Party Relationships (SR 23-4, June 2023, published in the Federal Register on 9 June 2023) is the principal United States supervisory architecture for third-party risk management at supervised banking organisations, replacing each agency's prior guidance with a unified articulation of risk management expectations across the third-party relationship lifecycle. The European Banking Authority Guidelines on Outsourcing Arrangements (EBA/GL/2019/02) is the principal European supervisory architecture for outsourcing at EU credit institutions and investment firms. The OCC Heightened Standards apply to large national banks at the elevated supervisory altitude. The Federal Reserve SR 11-7 (April 2011) and the Bank of England PRA SS1/23 (May 2023, in force May 2024) remain the canonical model risk anchors and apply to vendor-supplied models with the same expectations as institution-developed models.18

Global Layer, the operational resilience and AI-specific methodology. The EU Digital Operational Resilience Act (Regulation (EU) 2022/2554), entered into force 16 January 2023 and applicable from 17 January 2025, is the principal European architecture for digital operational resilience at financial entities, with explicit provisions on ICT third-party risk management and critical third-party providers. The Basel Committee on Banking Supervision Principles for Operational Resilience (March 2021) is the global Basel articulation on operational resilience with third-party risk as a constituent dimension. The NIST AI Risk Management Framework 1.0 (NIST AI 100-1, issued 26 January 2023) and the NIST AI 600-1 Generative AI Profile (issued 26 July 2024) are the methodological anchors for AI risk management that the institutional vendor due diligence and contractual position draw on. ISO/IEC 42001:2023 (issued 18 December 2023) is the management-system standard for AI that vendor evidence packages reference. The EU AI Act (Regulation (EU) 2024/1689) (with the €35 million Article 5 prohibition tier and the Chapter V Articles 51-56 obligations on general-purpose AI providers) is the jurisdictional architecture for foundation model providers operating in the EU and, by contractual extension, the institutional reference for those providers' obligations.19

India Layer, the foundational outsourcing architecture. The Reserve Bank of India Master Direction on Outsourcing of Information Technology Services (RBI/2023-24/102, issued 10 April 2023, effective for new arrangements 1 October 2023) is the foundational Indian supervisory architecture for IT outsourcing, applicable to banks, NBFCs, primary cooperative banks, and the all-India financial institutions, with explicit provisions on risk management, cross-border outsourcing, cyber incident reporting (incident report to the RBI within 6 hours of detection by the service provider), and the lifecycle of data in cloud. The RBI Master Direction on Managing Risks and Code of Conduct in Outsourcing of Financial Services by Banks (in its current edition) is the cognate framework for financial services outsourcing. The RBI Master Direction on Digital Lending (September 2022, with subsequent amendments) is the digital-lending-specific articulation. The RBI FREE-AI Committee Report, published 13 August 2025, applies the outsourcing framework to AI vendor relationships including foundation model providers, AI platform providers, AI tool providers, and AI service providers, and articulates the supervisory expectation on AI vendor risk. The RBI (Commercial Banks: Governance) Directions, 2025 (issued 28 November 2025) carry the senior governance frame that includes AI vendor risk as a governance subject at banks. The SEBI June 2025 consultation paper on AI/ML at market intermediaries carries the SEBI-regulated-entity articulation. The IRDAI Information and Cyber Security Guidelines (2023, revised 2026) carry the insurance-sector articulation. The IFSCA framework for IFSC institutions applies the cognate expectation at IFSC-domiciled entities.20

Cross-cutting Indian architecture: data protection, cybersecurity, and AML. The Digital Personal Data Protection Act, 2023 (enacted 11 August 2023) and the DPDP Rules 2025 (notified 13 November 2025) impose Data Fiduciary obligations that vendor relationships must accommodate, including the contractual provisions the Data Fiduciary must put in place with Data Processors, the cross-border transfer position, and the breach notification regime. The CERT-In Direction No. 20(3)/2022-CERT-In (28 April 2022, effective 27 June 2022) carries the cyber incident reporting expectation that interacts with the RBI Outsourcing direction's 6-hour reporting window. The Prevention of Money Laundering Act, 2002, and the AML regulatory architecture under it, apply to AI vendor relationships in the AML perimeter where the vendor's capability supports the institutional AML discipline (Chapter 5). The Financial Action Task Force post-2024-evaluation expectation set on the five improvement areas applies to the AML system effectiveness assessment that AI vendor relationships in the AML perimeter contribute to.21

Reconciliation discipline across the multi-layer regulatory perimeter. The institutional vendor risk discipline must reconcile expectations across the multi-layer perimeter without producing inconsistencies the supervisor will identify. The RBI Outsourcing direction's risk management expectation, the DPDP Act's Data Fiduciary expectation, the CERT-In Direction's incident reporting expectation, the EU DORA's operational resilience expectation (for institutions with EU touchpoints), and the supervisory engagement architecture must combine into a single institutional position the working DGM can defend in coordinated examination. The reconciliation discipline (the documented institutional position on how the multi-layer expectations combine and where the institution has made the choices the perimeter requires) is the artefact the supervisor reads to confirm the institution is operating the vendor risk discipline as a single institutional commitment rather than as parallel compliance exercises.

Practitioner's Lens: The Deputy General Manager (Third-Party and Vendor Risk for AI) at a Large Indian Private Sector Bank

The Deputy General Manager (Third-Party and Vendor Risk for AI) at a large Indian private sector bank spends most of any working week mediating between the institutional procurement reality of foundation model and AI platform contracting and the institutional risk discipline the third-party risk function carries. The DGM's calendar opens on Monday morning with the foundation model dependency map, the vendor inventory updates of the prior week, and the pending vendor incident queue. The DGM reads the new foundation model deployments the institutional AI governance machinery has surfaced; the DGM reads the new RegTech vendor relationships the procurement function has originated; the DGM reads the contractual playbook gaps the legal function has flagged on the prior week's contract negotiations; the DGM reads the vendor incidents (the foundation model availability blips, the platform version changes, the RegTech provider notice-of-change items) that have moved through the institution since Friday close.

The DGM's most consequential weekly conversations run at the level of architectural commitment to the institutional concentration posture, not procedural review of individual vendor onboardings. The procurement function carries one view of vendor relationships (commercial terms, supplier diversity, total cost of ownership); the AI governance function carries another (technical fit, deployment compatibility, capability roadmap); the legal function carries a third (contract risk, intellectual property exposure, jurisdictional exposure); the third-party risk function carries the fourth (institutional risk exposure, concentration profile, continuity architecture). The four views must be reconciled, on the perimeter the four functions share, into the institutional position the Vendor Risk Committee takes to the Risk Management Committee of the Board on each material vendor decision.

The function's quarterly Audit Committee paper on AI vendor risk is the DGM's most consequential institutional artefact. The paper captures the foundation model dependency map and concentration analysis, the contractual playbook against the as-negotiated contracts on the bank's material AI vendor relationships, the substitutability assessment status for each material foundation model dependency, the continuity drill log for the past quarter, the exit plan currency for material relationships, the supervisory engagement on AI vendor risk, and the institutional risk acceptances the Vendor Risk Committee has approved at named seniority. The paper is the artefact the Audit Committee uses to discharge its assurance responsibility on AI vendor risk; the paper is the artefact the supervisor will ask for in any examination of the third-party risk discipline; the paper is the discipline the DGM uses to maintain institutional rhythm on the AI vendor function across the four constituencies the function serves.

The DGM's twelve-month outlook, as the chapter goes to press in May 2026 in the wake of the RBI FREE-AI Committee Report and the RBI (Commercial Banks: Governance) Directions, 2025, is dominated by three institutional commitments. First, the foundation model dependency map must reach the depth the Risk Management Committee of the Board considers credible at supervisory examination (every material deployment mapped, every material dependency assessed for substitutability, every concentration position deliberately documented. Second, the contractual playbook must be refreshed against the supervisory expectation as articulated in the RBI FREE-AI Report and as it evolves under the SEBI AI/ML consultation paper outcomes, the EU DORA implementation experience at Indian-headquartered firms with EU touchpoints, and the United States SR 23-4 supervisory engagement at Indian institutions with US correspondent relationships. Third, the continuity architecture must move from declared to exercised across the bank's material AI vendor portfolio) the exit drills conducted on a documented cycle, the fallback deployments tested in tabletop, the substitutability assessments converted from paper analyses into operational migration paths the institution could actually execute.

What the DGM looks for in third-party risk officers at every level of the function, from the entry-level vendor analyst to the senior negotiator who leads the foundation model contract reviews, is the institutional disposition to operate the AI dimension of vendor risk as inseparable from the third-party risk discipline. The analyst who treats foundation model relationships as the AI specialist function's problem (to be onboarded with a generic outsourcing template) produces the discipline gap that surfaces on the next supervisory examination. The negotiator who treats the foundation model provider's standard contract terms as a starting position rather than as a negotiation endpoint, who carries the institutional contractual playbook into the commercial conversation, who documents every gap between the playbook and the as-negotiated contract as a named institutional risk acceptance, is the negotiator who produces the institutional position the bank can defend.22

Applied Exercise: The AI Vendor Risk Maturity Audit

For this chapter, the deliverable is a one-page maturity audit of the AI vendor risk function at the working DGM's own institution, measured against the five-layer AI Vendor Risk Architecture this chapter develops. The audit takes approximately ninety minutes to complete and produces a paper the DGM can take to the next Vendor Risk Committee with the institutional posture the function should hold and the two architectural commitments the function should make in the next ninety days.

Step 1, the vendor categorisation and dependency map audit (twenty minutes). Pull the institution's AI vendor inventory across all four categories: foundation model provider, AI platform provider, AI tool provider, AI service provider. Map each material deployment to the foundation model provider it depends on. Identify the bank's top three concentration positions on foundation model providers. Note whether the concentration profile has been examined by the Risk Management Committee of the Board within the past twelve months. Concentration positions the senior governance machinery has not examined deliberately are the first finding the audit surfaces.

Step 2, the substitutability and contractual playbook audit (twenty minutes). For the bank's top three foundation model dependencies, pull the substitutability assessment. Read for currency (within the last twelve months), for credibility (a migration window the institution could actually operate), and for completeness (alternative provider, operational cost, migration timeline, residual capability gap). Compare the as-negotiated contracts with the institutional contractual playbook on the ten provisions of §29.3 and identify the gaps. Gaps documented as institutional risk acceptances at named seniority are acceptable findings; gaps undocumented are the second finding.

Step 3, the continuity and exit drill audit (twenty minutes). For each of the bank's top three material AI vendor relationships, name whether the institution has a documented exit plan, whether the exit plan has been reviewed within the last twelve months, whether the fallback architecture has been exercised in tabletop drill within the last twelve months, and whether the substitutability assessment is operationally credible against the exit plan. Unexercised continuity arrangements, stale exit plans, and operationally non-credible substitutability assessments are the third finding.

Step 4, the regulatory perimeter reconciliation audit (twenty minutes). Confirm whether the institution's third-party risk discipline reconciles with the RBI Master Direction on Outsourcing of IT Services (10 April 2023) and the RBI Master Direction on Outsourcing of Financial Services on the foundational expectations; with the SEBI June 2025 AI/ML consultation paper and the RBI FREE-AI Report (13 August 2025) on AI-specific expectations; with the IRDAI Information and Cyber Security Guidelines (2023, revised 2026) for insurance-touchpoint operations; with the DPDP Act 2023 and DPDP Rules 2025 on data fiduciary obligations; with the CERT-In Direction (No. 20(3)/2022-CERT-In, 28 April 2022, effective 27 June 2022) on cyber incident reporting; with the EU DORA (Regulation (EU) 2022/2554, applicable 17 January 2025) for EU touchpoint operations; with the Federal Reserve SR 23-4 (June 2023) for US touchpoint operations. The fourth finding is reconciliation gaps.

Step 5, the personal positioning (ten minutes). From the four prior steps, name the two architectural recommendations the working DGM would take to the Vendor Risk Committee within ninety days. The recommendations should each be a specific architectural commitment the institution is, after the audit, prepared to defend: a deepening of the foundation model dependency map, an institutional refresh of the contractual playbook against the supervisory perimeter, a continuity drill cycle the function will execute, a substitutability assessment refresh on the top three foundation model dependencies, a reconciliation discipline across the multi-layer regulatory perimeter. Vague recommendations are themselves a finding the audit has surfaced) the function's architectural commitments are not yet at the altitude the senior governance machinery expects.

Executive Briefing

AI vendor risk is the discipline arising from the institution's dependencies on foundation model providers, AI platform providers, AI tool providers, and AI service providers. The discipline is a distinct family of third-party risk because foundation model dependencies are deeper and harder to substitute than conventional IT vendor dependencies, because the contractual asymmetries between vendor and institution are larger than in conventional outsourcing, and because the regulatory perimeter that applies to AI vendor relationships spans multiple regulators and multiple jurisdictions. The institutional working position is that AI vendor risk is integrated with the broader third-party risk discipline, not parallel to it; that the foundation model relationship is the deepest of the AI vendor relationships and receives the deepest discipline; and that the five-layer AI Vendor Risk Architecture is the institutional answer to the question of how to operate the discipline.

Foundation model concentration risk is the single most consequential exposure in the AI vendor portfolio. The institutional dependency map shows the deployments by provider; the substitutability assessment shows the migration window for each material dependency; the institutional concentration posture is the deliberate choice the board owns. Concentration accumulated through individually defensible procurement decisions without senior governance examination is the indefensible posture; the CrowdStrike outage on 19 July 2024, in which a faulty Falcon sensor update from a single global vendor crashed approximately 8.5 million Windows systems and disrupted airlines, banks, hospitals, and emergency services worldwide, is the canonical recent lesson on what single-vendor global concentration produces when the vendor's release process fails.

The contractual architecture is the institution's audit trail when the vendor relationship fails. The ten provisions of §29.3 (data flow, model behaviour and version stability, liability, indemnification, continuity and exit, supervisory access and audit, incident response, information security, regulatory compliance, change management) operated through an institutional playbook against which every as-negotiated contract is read, is the working discipline. Standard vendor terms are insufficient for institutional use across all ten provisions; the vendor's willingness to depart from standard terms is itself a vendor risk signal the institution prices into procurement. The contractual position the institution accepts is the position the institution lives with for the duration of the relationship.

The continuity architecture must be exercised, not declared. The Wirecard collapse in June 2020, in which the German payments processor's insolvency left its bank and merchant clients with the institutional question of operating without a vendor on which their operations had become concentrated, is the canonical lesson on what unexercised continuity arrangements produce when vendor failure arrives. The SolarWinds attack disclosed in December 2020, in which approximately 18,000 organisations including US federal agencies were compromised through a weaponised vendor update channel, is the canonical lesson on what supply chain attack vectors produce when the institutional perimeter trusts vendor-supplied channels. The institutional fallback architecture for material AI deployments must be exercised in drill on a documented cycle; the exit plan for each material relationship must be reviewed on a documented cycle; the substitutability assessment must be operationally credible against the exit plan.

The regulatory perimeter for AI vendor risk operates across the RBI Master Direction on Outsourcing of IT Services (10 April 2023), the RBI Master Direction on Outsourcing of Financial Services, the RBI Master Direction on Digital Lending (September 2022), the RBI FREE-AI Committee Report (13 August 2025), the SEBI June 2025 AI/ML consultation paper, the IRDAI Information and Cyber Security Guidelines (2023, revised 2026), the DPDP Act 2023 and DPDP Rules 2025, the CERT-In Direction (No. 20(3)/2022-CERT-In, 28 April 2022, effective 27 June 2022), the EU DORA (Regulation (EU) 2022/2554, applicable 17 January 2025), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), the Federal Reserve SR 23-4 (June 2023), and the BCBS Principles for Operational Resilience (March 2021). The institutional position must reconcile across the layers; reconciliation is itself the discipline the supervisor reads.

Cloud-AI vendor architecture produces doubled dependency the senior governance machinery must examine explicitly. RegTech vendor architecture produces supervisor-facing dependency the senior governance machinery must distribute deliberately. Cross-border foundation model exposure produces jurisdictional exposure the institution must track through the cross-border equivalent of supervisory engagement. Each of these adjacencies extends the AI vendor risk discipline beyond the foundation model relationship itself and into the broader institutional vendor ecosystem.

The Monday-morning posture for the working DGM is the discipline of the dependency map, the substitutability assessment, the contractual playbook, the continuity drill log, and the regulatory perimeter reconciliation: operated together, refreshed on a documented cycle, and defensible at supervisory examination. The five-layer AI Vendor Risk Architecture is the institutional answer; the Vendor Risk Committee is the institutional rhythm; the Risk Management Committee of the Board is the senior governance machinery; the supervisor's examination is the external test. The working DGM's craft is the institutional commitment that the discipline is operated as a working architecture and not as a procurement administrative matter.

Key Terms

AI vendor risk. The institutional category of risk arising from the institution's dependencies on third parties that supply AI capability or AI-related services. Distinct family of third-party risk because of dependency depth, contractual asymmetry, and the multi-jurisdictional regulatory perimeter.

Foundation model dependency. The institutional dependency on a third party that supplies a foundation model on which the institution's generative and agentic AI deployments are built. The deepest of the AI vendor relationships in operational, contractual, and risk terms.

Concentration risk (vendor). The institutional exposure arising from the accumulation of dependency on a single vendor across multiple deployments. For AI vendor risk, the principal concentration question is foundation model provider concentration across the deployment portfolio.

Substitutability assessment. The institutional analysis, for each material AI vendor dependency, of the alternative provider that could supply equivalent capability, the migration timeline, the operational cost, the institutional adaptation required, and the residual capability gap that would remain after substitution.

Contractual exit architecture. The contractual provisions and the operational architecture that, together, govern the institution's exit from a material AI vendor relationship: the exit triggers, the migration provisions, the data and model state preservation, the wind-down timeline, the institutional rights at vendor failure or change of control.

Model provenance documentation. The vendor-supplied documentation describing the training data provenance, evaluation methodology and results, known limitations, safety evaluation history, bias and fairness assessment, and version change log that the institutional model risk discipline requires for vendor-supplied models.

Cloud-AI vendor architecture. The institutional architecture in which AI deployments rest on hyperscale cloud infrastructure with the same provider supplying foundation model capability, producing doubled dependency on a single vendor across infrastructure and capability dimensions.

RegTech vendor architecture. The institutional architecture of dependencies on third-party providers of regulatory technology (transaction monitoring, regulatory reporting, risk aggregation, supervisory engagement support), increasingly with embedded AI capability that becomes part of the vendor risk surface.

Third-party data flow. The flow of institutional data (prompts, retrieved content, outputs, intermediate states, evidence) through a vendor's infrastructure, with the contractual and operational architecture that governs the data use, residency, and lifecycle.

AI vendor concentration. The institutional concentration of dependency across vendors in the AI vendor portfolio, considered at multiple levels (foundation model provider, cloud provider, platform provider, RegTech provider) and reconciled into an institutional concentration posture the board owns.

Foundation model audit right. The contractual provision that gives the institution and, by extension, the institution's regulator the right to examine the foundation model provider, the model documentation, the safety evaluation evidence, and the operational controls the provider operates.

Vendor liability and indemnity. The contractual provisions that govern the vendor's liability for damages arising from the supplied AI capability and the vendor's indemnification of the institution against specific categories of claim (intellectual property, data protection, regulatory penalties traceable to vendor conduct).

Supply chain attack. The cybersecurity attack pattern in which an adversary compromises a trusted vendor channel to distribute malicious payload through a route the institutional perimeter has authorised. The SolarWinds Orion compromise disclosed in December 2020 is the canonical reference; the AI vendor relationship is a supply chain attack surface the institutional cyber posture must govern.

Reflection Questions

One. At your institution, how is the boundary drawn between AI vendor risk and conventional vendor risk, and is the boundary producing operational clarity or operational confusion in the institutional vendor risk function?

Two. When your institution evaluates an alternative foundation model provider, what is the institutional process, who is involved, and is the process producing the institutional intelligence the substitutability assessment requires?

Three. What is your institution's contractual playbook on foundation model provider relationships, and on which of the ten provisions of §29.3 does the playbook fall short of the institutional risk position the institution actually holds?

Four. How does your institution's exit planning for foundation model provider relationships compare to the operational reality of conducting the exit, and where is the credibility gap?

Five. What is your institution's senior governance posture on foundation model concentration, and is the posture one the Risk Management Committee of the Board has owned deliberately or one that has accumulated by default through individually defensible procurement decisions?

Six. How does your institution's data protection position (DPDP Data Fiduciary discipline) interact with the foundation model providers' standard data use practices, and where is the institutional position weak in contractual or operational terms?

Seven. At what point does the commercial pressure on AI vendor renewal terms exceed the institutional risk position, and how does the institution's governance architecture resist the pressure?

Eight. What would your institutional regulator find in a deep supervisory examination of the AI vendor risk discipline today, and what is the institutional defence the working DGM would mount?

Further Reading

On the foundational outsourcing and third-party risk architecture for the working DGM's career-grade understanding, the working DGM should read the Reserve Bank of India Master Direction on Outsourcing of Information Technology Services (RBI/2023-24/102, issued 10 April 2023, effective 1 October 2023), the RBI Master Direction on Managing Risks and Code of Conduct in Outsourcing of Financial Services by Banks (current edition), the RBI Master Direction on Digital Lending (September 2022 with subsequent amendments), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), and the Federal Reserve, OCC, and FDIC Interagency Guidance on Third-Party Relationships: Risk Management (SR 23-4, June 2023, published in the Federal Register on 9 June 2023). These five documents form the architectural foundation; the working DGM should read them in original form rather than in summary, with attention to the institutional accountability articulations each carries.

On the operational resilience and AI-specific methodology architecture, the working DGM should read the EU Digital Operational Resilience Act (Regulation (EU) 2022/2554, applicable from 17 January 2025), the BCBS Principles for Operational Resilience (March 2021), the NIST AI Risk Management Framework 1.0 (NIST AI 100-1, issued 26 January 2023), the NIST AI 600-1 Generative AI Profile (issued 26 July 2024), ISO/IEC 42001:2023 (issued 18 December 2023), the EU AI Act (Regulation (EU) 2024/1689) with particular attention to the Chapter V Articles 51-56 obligations on general-purpose AI providers, the Federal Reserve SR 11-7 (April 2011), and the Bank of England PRA SS1/23 (issued May 2023, in force May 2024). These eight documents form the methodological and supervisory anchor for AI vendor risk; the working DGM operates against them as the institutional reference.

On the Indian AI-specific supervisory architecture, the working DGM should read the RBI FREE-AI Committee Report (published 13 August 2025), the RBI (Commercial Banks: Governance) Directions, 2025 (issued 28 November 2025), the SEBI consultation paper on AI/ML at market intermediaries (June 2025), the IRDAI Information and Cyber Security Guidelines (2023, revised 2026), the Digital Personal Data Protection Act, 2023 (enacted 11 August 2023), the DPDP Rules 2025 (notified 13 November 2025), and the CERT-In Direction No. 20(3)/2022-CERT-In (28 April 2022, effective 27 June 2022). The Indian architecture extends the global expectation onto Indian institutional reality; the working DGM reads the Indian architecture as the institutional perimeter inside which the function operates.

On the case history that anchors the institutional discipline, the working DGM should read the published accounts of the Wirecard collapse (June 2020) as the canonical third-party concentration failure (treated in fuller depth in Chapter 1 of this book), the published accounts and post-incident analyses of the CrowdStrike outage (19 July 2024) as the canonical single-vendor global concentration release-process failure, and the published accounts and federal investigative findings on the SolarWinds Orion supply chain compromise (December 2020) as the canonical supply chain attack pattern. The three cases together carry the institutional lesson set the working DGM operates from: concentration produces blast radius, unexercised continuity is fiction, and trusted vendor channels are attack surfaces the institutional posture must govern.

On the institutional craft and the working DGM's reading discipline, the working DGM should subscribe to the supervisory engagement materials of the RBI, the SEBI, the IRDAI, and the IFSCA on third-party risk and outsourcing; should read the published annual reports of the institutions whose AI vendor risk discipline is at institutional altitude (HDFC Bank, State Bank of India, ICICI Bank, Bajaj Finance); should follow the published guidance and supervisory communications of the Federal Reserve, the OCC, the FDIC, the European Banking Authority, the European Securities and Markets Authority, the Bank of England, the Monetary Authority of Singapore (MAS Veritas, November 2019), and the BCBS on third-party risk and AI-vendor topics; and should treat the reading discipline as constitutive of the institutional commitment-discipline the function operates under.

Endnotes

  1. Sandeep Banerjee is a composite character drawn from publicly disclosable patterns about Indian financial services third-party risk officers at large Kolkata-headquartered private sector banks, industry-working-group discussions on AI vendor risk, and the institutional reality of foundation model contract negotiation observable through published annual reports, risk management framework documentation, and supervisory engagement materials. No specific individual is referenced; the vignette reflects the typical institutional reality the chapter equips the working DGM to operate.

  2. On AI vendor categorisation and the four canonical categories (foundation model provider, AI platform provider, AI tool provider, AI service provider) the institutional convention has stabilised across large Indian banks' vendor risk functions over the 2023-2026 window. The convention is the institutional shorthand the working DGM operates and is consistent with the RBI Master Direction on Outsourcing of Information Technology Services (10 April 2023) on regulated-entity responsibility for outsourced activities and with the RBI FREE-AI Committee Report (published 13 August 2025) on AI vendor relationships.

  3. On the multi-regulator architecture for AI vendor risk in India, see the Reserve Bank of India Master Direction on Outsourcing of Information Technology Services (RBI/2023-24/102, issued 10 April 2023, effective 1 October 2023), the RBI Master Direction on Managing Risks and Code of Conduct in Outsourcing of Financial Services by Banks, the RBI Master Direction on Digital Lending (September 2022, with subsequent amendments), the SEBI consultation paper on AI/ML at market intermediaries (June 2025), the RBI FREE-AI Committee Report (published 13 August 2025), the RBI (Commercial Banks: Governance) Directions, 2025 (issued 28 November 2025), the IRDAI Information and Cyber Security Guidelines (2023, revised 2026), the Digital Personal Data Protection Act, 2023 (11 August 2023) and DPDP Rules 2025 (notified 13 November 2025), the EU Digital Operational Resilience Act (Regulation (EU) 2022/2554, applicable 17 January 2025), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), the Federal Reserve, OCC, and FDIC Interagency Guidance on Third-Party Relationships: Risk Management (SR 23-4, June 2023), and the BCBS Principles for Operational Resilience (March 2021).

  4. On the methodological and management-system anchors for AI vendor risk, see the NIST AI Risk Management Framework 1.0 (NIST AI 100-1, issued 26 January 2023), the NIST AI 600-1 Generative AI Profile (issued 26 July 2024), ISO/IEC 42001:2023 (issued 18 December 2023), the EU AI Act (Regulation (EU) 2024/1689) (with the €35 million Article 5 prohibition tier and the Chapter V Articles 51-56 obligations on general-purpose AI providers) the Federal Reserve Supervisory Guidance on Model Risk Management (SR 11-7, April 2011), and the Bank of England Prudential Regulation Authority Supervisory Statement SS1/23 (issued May 2023, in force May 2024).

  5. On the CrowdStrike outage of 19 July 2024 and the institutional lesson on single-vendor global concentration, see the public reporting on the Channel File 291 incident, the CrowdStrike post-incident root cause analysis, and the United States Cybersecurity and Infrastructure Security Agency alert on the widespread IT outage. The incident is referenced in this chapter not for cybersecurity discipline (which Chapter 30 develops) but for the institutional lesson on concentration risk in single-vendor global supply.

  6. On the Wirecard collapse of June 2020 and the institutional lesson on third-party concentration failure, see the German federal regulatory proceedings on Wirecard AG's insolvency and the public reporting on the missing escrow assets. The case is developed in fuller treatment in Chapter 1 as the canonical anchor for the institutional commitment-discipline conception of risk and compliance; it is referenced here for the third-party concentration lesson.

  7. On the ten contractual provisions for AI vendor relationships, the institutional playbook has stabilised across the legal and third-party risk functions of large Indian banks operating material foundation model relationships. The playbook is institutional shorthand rather than regulatory taxonomy; it is informed by the RBI Master Direction on Outsourcing of IT Services (10 April 2023), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), the EU DORA's third-party risk provisions, and the Federal Reserve SR 23-4 Interagency Guidance.

  8. On model provenance documentation as the contractual evidence regime, see the NIST AI 600-1 Generative AI Profile (issued 26 July 2024), ISO/IEC 42001:2023 (issued 18 December 2023), and the EU AI Act Chapter V Articles 51-56 obligations on general-purpose AI providers including model documentation, evaluation, and risk mitigation evidence.

  9. On the RBI Master Direction on Outsourcing of Information Technology Services (RBI/2023-24/102, issued 10 April 2023), the provisions on cyber incident reporting (incident report to the RBI within 6 hours of detection by the service provider) and on the lifecycle of data in cloud (from entry into cloud through to permanent erasure) are constitutive for the institutional cloud-AI architecture position.

  10. On the SolarWinds supply chain attack disclosed in December 2020 and the institutional lesson on weaponised vendor channels, see the public reporting on the SolarWinds Orion compromise, the United States federal investigations into the incident, and the cybersecurity industry analyses on the institutional perimeter implications. The case is referenced for the AI vendor channel-trust lesson; it is developed in fuller treatment in Chapter 30's cyber risk architecture.

  11. On the multi-layer regulatory perimeter that the DGM reads, the institutional discipline is the reconciliation across the layers rather than the parallel observation of each layer. The reconciliation discipline is the artefact the supervisor reads as evidence the institutional vendor risk function is operating as a single institutional commitment.

  12. HDFC Bank's institutional foundation model dependency governance posture is documented in the bank's published annual reports, risk management framework documentation, and the supervisory engagement materials publicly disclosable through the bank's regulatory filings, 2023 to 2026. The institutional altitude treatment in this chapter is consistent with what those public sources support.

  13. State Bank of India's institutional RegTech vendor architecture as a concentration-aware multi-vendor posture is documented in the bank's annual reports, risk management framework documentation, and the supervisory engagement materials publicly disclosable through the bank's regulatory filings, 2023 to 2026. The institutional altitude treatment in this chapter is consistent with what those public sources support.

  14. ICICI Bank's cloud-AI architecture and Bajaj Finance's cross-border foundation model exposure as illustrative institutional patterns are documented at the institutional altitude the institutions' public disclosures support. The treatment here is at the institutional altitude that supports the working DGM's analytical position and does not rely on non-public institutional detail.

  15. On the Wirecard collapse of June 2020 and the institutional concentration and continuity lessons for third-party risk, see Chapter 1 for the fuller treatment; this chapter references the case for the AI vendor risk-specific lessons.

  16. On the CrowdStrike outage of 19 July 2024, the worldwide financial damage has been estimated in the tens of billions of dollars in industry analyses; approximately 8.5 million Windows systems were affected. The institutional lesson for AI vendor risk is the blast-radius profile of single-vendor global concentration, not the cybersecurity-specific lesson Chapter 30 develops.

  17. On the SolarWinds supply chain attack and the institutional channel-trust lesson, the canonical public reporting and the United States federal investigative findings support the institutional discipline conclusion: the AI vendor relationship is a trusted-channel relationship the institutional cyber posture must govern as part of the broader vendor architecture.

  18. On the global supervisory architecture for third-party risk, see the Federal Reserve, OCC, and FDIC Interagency Guidance on Third-Party Relationships (SR 23-4, June 2023, published in the Federal Register on 9 June 2023), the EBA Guidelines on Outsourcing Arrangements (EBA/GL/2019/02), the OCC Heightened Standards for large national banks, the Federal Reserve Supervisory Guidance on Model Risk Management (SR 11-7, April 2011), and the Bank of England PRA SS1/23 (issued May 2023, in force May 2024).

  19. On the operational resilience and AI methodology anchors, see the EU Digital Operational Resilience Act (Regulation (EU) 2022/2554, entered into force 16 January 2023 and applicable from 17 January 2025), the BCBS Principles for Operational Resilience (March 2021), the NIST AI Risk Management Framework 1.0 (NIST AI 100-1, issued 26 January 2023), the NIST AI 600-1 Generative AI Profile (issued 26 July 2024), ISO/IEC 42001:2023 (issued 18 December 2023), and the EU AI Act (Regulation (EU) 2024/1689), with the €35 million Article 5 prohibition tier and the Chapter V Articles 51-56 obligations on general-purpose AI providers.

  20. On the Indian regulatory architecture for AI vendor risk, see the Reserve Bank of India Master Direction on Outsourcing of Information Technology Services (RBI/2023-24/102, issued 10 April 2023, effective 1 October 2023), the RBI Master Direction on Managing Risks and Code of Conduct in Outsourcing of Financial Services by Banks (current edition), the RBI Master Direction on Digital Lending (September 2022, with amendments), the RBI FREE-AI Committee Report (published 13 August 2025), the RBI (Commercial Banks: Governance) Directions, 2025 (issued 28 November 2025), the SEBI consultation paper on AI/ML (June 2025), the IRDAI Information and Cyber Security Guidelines (2023, revised 2026), and the IFSCA framework for IFSC institutions.

  21. On the cross-cutting Indian architecture, see the Digital Personal Data Protection Act, 2023 (enacted 11 August 2023), the DPDP Rules 2025 (notified 13 November 2025), the CERT-In Direction No. 20(3)/2022-CERT-In (28 April 2022, effective 27 June 2022), the Prevention of Money Laundering Act, 2002, and the Financial Action Task Force post-2024-evaluation expectation set on the five improvement areas as it applies to the AI vendor relationships supporting AML deployments at Indian financial services institutions.

  22. The Practitioner's Lens on the Deputy General Manager (Third-Party and Vendor Risk for AI) is constructed from the same composite institutional reality the opening vignette reflects, applied to the broader institutional rhythm the function operates. No specific individual is referenced; the institutional altitude is the altitude the chapter equips the working DGM to operate at.