TL;DR

• SR 11-7 defines model broadly — any quantitative system transforming inputs into outputs used for business decisions, including collections propensity scoring, self-cure identification, and contact routing logic
• 68% of US banks with AI collections have at least one system not in their SR 11-7 inventory — vendor-hosted propensity models are the most common unregistered gap
• OCC issued formal MRAs to 14 institutions in 2024 for collections AI inventory gaps — remediation averaged 8 months and required dedicated MRM headcount additions
• Vendor-hosted collections AI models remain the bank’s SR 11-7 responsibility — outsourcing deployment does not outsource model governance accountability
• iTuring auto-registers every collections model in the bank’s inventory from deployment day with full documentation — zero retrospective registration effort required

SR 11-7 Model Inventory Requirements for Collections AI: What Banks Must Know in 2026

Most model risk management teams at US banks built their collections governance frameworks years before AI entered the collections workflow. That gap is now visible in examination findings. SR 11-7: Guidance on Model Risk Management, issued jointly by the Federal Reserve Board and the OCC in April 2011, alongside the OCC 2011-12 companion bulletin and the Q3 2024 updated examination guidance on AI and self-learning models, establishes the baseline obligation for every supervised institution. The guidance requires US banks to maintain a comprehensive inventory of all models used in business decisions, validate those models independently, and document their conceptual soundness, limitations, and ongoing performance. AI-driven collections workflows fall squarely within this scope because they produce quantitative outputs: propensity scores, contact priority rankings, self-cure predictions, and treatment assignment decisions that directly affect recovery outcomes and consumer treatment. This article provides a practical reference for the Head of Model Risk responsible for SR 11-7 model risk management across collections AI at a bank. After reading it, you will know exactly which collections AI components qualify as models, what documentation examiners request, where the most common inventory gaps exist, and how to close them before the next examination cycle.

The Three SR 11-7 Obligations US Banks Most Often Miss When Inventorying Collections AI

SR 11-7 / OCC 2011-12 Model Risk Management Guidance is explicit on one point that many collections operations overlook: all quantitative systems transforming inputs into outputs used for business decisions must be in the model inventory. This includes collections propensity scoring, self-cure identification, contact timing optimization, and treatment assignment logic. Each of these constitutes a distinct model under the guidance. Banks must document each model’s purpose, methodology, assumptions, limitations, data inputs, output usage, and the identity of the model owner. Any system not yet registered must either be added to the inventory or formally restricted from production use. The restriction must be documented and enforceable, not merely noted in a policy memo.

The scope question is where most banks stumble. According to the Moody’s Analytics Bank Model Governance Survey 2025, 68% of US banks with AI collections platforms have at least one system not registered in their SR 11-7 model inventory. Vendor-hosted propensity models are the most common unregistered gap. Banks that rely on third-party AI vendors for scoring or segmentation frequently assume the vendor carries the governance burden. SR 11-7 makes no such distinction. The bank retains full accountability for every model used in its decision-making, regardless of where that model is hosted or who built it. OCC examiners reviewing collections-specific model inventories in 2025 have been particularly focused on this vendor gap.

The enforcement record confirms the risk is not theoretical. The OCC issued formal Matters Requiring Attention to 14 institutions in 2024 for collections AI inventory gaps. Remediation averaged 8 months and required dedicated MRM headcount additions at each institution (OCC Matters Requiring Attention Report 2024). The cost of retroactive compliance consistently exceeds the cost of building governance into the deployment process from the start. The implementation checklist later in this article covers the specific gaps most US banks teams need to close before the next examination.

What OCC Examiners Request From Collections AI That US Banks Consistently Cannot Produce

The data layer is the first failure point. SR 11-7: Guidance on Model Risk Management requires banks to document the data inputs feeding each model, including source, transformations, quality checks, and lineage from raw data through to model output. Most collections AI platforms ingest dozens of data streams: bureau attributes, behavioral signals, payment history, contact attempt outcomes, and demographic variables. Standard collections systems rarely maintain a traceable lineage from each input variable back to its source system. When examiners request the data dictionary and lineage documentation for a specific propensity model, banks frequently cannot produce it within the examination window.

The process layer gap compounds the data problem. SR 11-7 examination findings for collections-specific model governance gaps increased 41% year-over-year in 2024 as examiners expanded scope to include agentic and self-learning collections AI systems (Federal Reserve Model Risk Management Examination Trends Report Q4 2024). Examiners now ask for evidence of model change control: who approved the last retraining cycle, what performance metrics triggered it, and whether an independent validation was completed before the updated model entered production. Banks maintaining their model inventory for collections AI under OCC examination standards frequently discover that their change control records exist in email threads and spreadsheets rather than in a governed, auditable system.

The audit trail gap is the most damaging in examination settings. Examiners request on-demand access to historical model versions, prior validation reports, monitoring dashboards showing performance drift, and records of every material change made to a production model. Standard collections operations cannot produce these artifacts because they were never designed to retain them. The pattern is consistent: US banks teams that built their collections AI before the Q3 2024 updated examination guidance on AI and self-learning models are operating governance frameworks that predate the obligation.

How iTuring Satisfies SR 11-7 Model Inventory, Validation, and Documentation Requirements

Automatic model inventory registration: every collections model registered with purpose, data lineage, owner, and version history from deployment day one without manual entry

iTuring ML Governance automatically registers every model at the moment it enters production. The registration record includes the model’s stated purpose, complete data lineage tracing each input variable to its source system, the assigned model owner, version history, and the approval chain that authorized deployment. This eliminates the retrospective registration effort that consumes MRM teams at most banks. From the examiner’s perspective, the inventory is always current. If an OCC examiner requests the inventory for all collections models in production, the Head of Model Risk can generate a complete, timestamped report within minutes. Every entry includes the model’s classification under SR 11-7, its risk tier, and the date of its last independent validation.

SR 11-7 evidence pack auto-generation: conceptual soundness documentation, validation reports, monitoring summaries, and change control records compiled per examination cycle in 30 minutes

The obligation under SR 11-7 is clear: all quantitative systems transforming inputs into outputs used for business decisions must be documented with sufficient detail for independent review. iTuring ML Governance compiles conceptual soundness documentation, validation reports, ongoing monitoring summaries, and change control records into a single evidence pack per model. The compilation takes approximately 30 minutes per examination cycle. Each evidence pack maps directly to the SR 11-7 documentation requirements: model description, intended use, methodology, assumptions, limitations, validation outcomes, and performance monitoring results. A collections propensity model, for example, would include its feature importance rankings, back-testing results against realized roll rates, and any threshold adjustments made during the monitoring period, with platform-native model governance with immutable audit trail and maker-checker approval ensuring every record is tamper-proof.

A collections AI platform that scores accounts, routes contacts, and identifies self-cure candidates contains at minimum three SR 11-7 models. Most bank model inventories list one, or none.

Vendor model governance wrapper: iTuring-hosted models documented under the bank’s inventory with full training data lineage and methodology documentation available on demand

For banks using iTuring-hosted collections models, the governance wrapper ensures each model is documented under the bank’s own inventory with full training data lineage, methodology documentation, and performance monitoring accessible on demand. The Head of Model Risk retains complete visibility into model behavior without needing to request documentation from a third-party vendor during an examination. This directly addresses the SR 11-7 vendor model governance requirement for collections at any bank: the guidance holds the institution accountable for every model used in its decisions, and iTuring’s governance layer ensures that accountability is backed by complete, current documentation. SR 11-7: Guidance on Model Risk Management does not permit banks to delegate governance to vendors. iTuring’s architecture reflects that principle by placing the bank’s MRM team in control of every approval, every version change, and every validation cycle, regardless of where the model was originally built.

Before the Next OCC Examination: The Collections AI Inventory Gaps Most Banks Need to Close

1. Start with a complete census of every AI collections model currently in production. This means identifying not only the primary scoring model but also any secondary models handling contact optimization, self-cure prediction, treatment assignment, or payment channel routing. Under SR 11-7 / OCC 2011-12 Model Risk Management Guidance, each of these qualifies as a separate model requiring its own inventory entry. Most banks discover two to four unregistered models during this exercise.
2. Assemble the documentation package for each identified model. The records must satisfy the SR 11-7 requirement that all quantitative systems transforming inputs into outputs used for business decisions be documented with conceptual soundness memos, data dictionaries, assumption logs, and limitation disclosures. Collections propensity scoring and self-cure identification models require particular attention because their outputs directly influence consumer contact frequency and treatment intensity.
3. Establish a formal approval process for material model changes before any change enters production. SR 11-7 model risk management for collections AI at a bank requires that retraining events, feature additions, threshold adjustments, and methodology changes each pass through an independent review and approval workflow. The approval must be documented with the reviewer’s identity, the date, and the rationale for the decision.
4. Set a monitoring cadence that satisfies SR 11-7: Guidance on Model Risk Management. Quarterly performance monitoring is the minimum most examiners expect for collections models, with out-of-cycle reviews triggered by significant performance drift, changes in portfolio composition, or macroeconomic shifts that alter borrower behavior. Document the triggers explicitly in your model risk policy so examiners can verify that the monitoring cadence is not arbitrary.

Real Results: US Community Bank (assets <$10B)

A US community bank with assets under $10 billion faced an SR 11-7: Guidance on Model Risk Management compliance challenge after deploying AI-driven collections scoring without registering the underlying models in its inventory, a gap identified during internal audit preparation for an upcoming OCC examination. The bank deployed iTuring ML Governance to automatically register all collections models, generate documentation packages, and establish governed change control workflows, achieving full compliance readiness within 60 days of deployment.

Results after deployment:

 SR 11-7 examination passed with zero findings on all collections AI models

Before the Next OCC Examination: Which Collections AI Systems Your Model Inventory Must Show

The single highest-risk gap for most MRM teams is the vendor-hosted model that collections operations treat as a “tool” rather than a model: if it produces a score or a ranking that influences a collections decision, it belongs in the inventory. Examination preparation should begin with a joint session between model risk and collections operations leadership to reconcile what is in production against what is registered. The banks that pass examinations cleanly are the ones that treat model inventory maintenance as a continuous process rather than an annual project.

SR 11-7 examination intensity scales with bank asset size and model risk profile: community banks face proportionate requirements, but the model inventory and validation obligations apply at all asset sizes without exception.

The cost of retroactive SR 11-7 compliance for collections AI is measured in months of MRM team effort, dedicated headcount additions, and the operational disruption of restricting production models mid-examination. Banks that have embedded governance into their collections AI deployment process from day one consistently avoid these costs. The examination record from 2024 and 2025 makes the trajectory clear: OCC examiners are expanding their scope, not contracting it, and collections AI is firmly within that expanding scope.

For MRM teams preparing for their next examination cycle, request a demo to see how iTuring ML Governance registers, documents, and governs every collections model in your inventory from deployment day one. The 60-day compliance readiness timeline that community banks have already achieved is available to institutions at every asset tier.