Book a demo
Topics · REGULATORY

Payments AI Model Risk Controls | Moring

Payments model risk management is the discipline that keeps the scoring, screening, and fraud models behind money movement accountable for their decisions. Supervisory expectations such as SR 26-2 and OSFI Guideline E-23 extend enterprise-wide model risk duties to generative and agentic payment AI, so living inventories, validation, runtime monitoring, and audit-ready evidence now matter more than periodic sign-off.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading

Short answer

Payments model risk management is the discipline that keeps the scoring, screening, and fraud models behind money movement accountable for their decisions. Supervisory expectations such as SR 26-2 and OSFI Guideline E-23 extend enterprise-wide model risk duties to generative and agentic payment AI, so living inventories, validation, runtime monitoring, and audit-ready evidence now matter more than periodic sign-off.

How it works

Model risk management in payments is the discipline that holds the models behind money movement accountable for their decisions. Payment institutions now apply machine learning across transaction fraud scoring, sanctions and watchlist screening, merchant and consumer credit underwriting, and Bank Secrecy Act and sanctions monitoring, and the complexity of these technologies has raised the stakes for model risk management. Improper or insufficient model risk management carries consequences that land directly on a payment provider: erosion of regulator trust, formal or informal regulatory action, expensive look-backs and remediation, regulatory fines, damaged reputation, financial loss, and internal inefficiency. Those outcomes are not hypothetical in an industry where a flawed model can block, release, or misroute funds at machine speed. Sources: https://kaufmanrossin.com/blog/managing-ai-model-risk-in-financial-institutions-best-practices-for-compliance-and-governance/

Supervisory expectations have moved to match the technology. SR 26-2 reaffirms model governance and validation and explicitly carves out generative and agentic AI, which means an agentic payments assistant needs its own governance track rather than inheriting the sign-off of a legacy scoring model. The EU AI Act adds a second clock: high-risk financial services uses such as credit scoring and fraud detection pick up their obligations from December 2027, and the classification work does not wait for the deadline. A payments operator that delays mapping its estate will end up classifying systems under examiner timelines rather than its own. Sources: https://www.yields.io/use-cases/ai-model-risk-management

The perimeter of what counts as a model has widened accordingly. OSFI's finalized Guideline E-23 treats any quantitative tool or algorithm that uses data to generate an output as a model, explicitly pulling AI-powered systems and black-box models that previously sat outside rigorous frameworks into scope. The guideline, which comes into full effect in May 2027, establishes a principles-based, risk-proportional approach and mandates that financial institutions manage model risk across the entire enterprise, regardless of a model's complexity or its specific application. For a payments business, that means a fraud heuristic, a transaction-scoring ensemble, and a customer-facing language model all belong on one inventory under one accountability structure. Sources: https://www.moodys.com/web/en/us/insights/regulatory-news/-published-just-now-chanpreet-mehta-edit--share---from-complianc.html https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027

AI intensifies the underlying risk rather than replacing it. Payment models introduce opacity that complicates validation, high-dimensional and nonlinear relationships that resist explanation, model drift as payment behavior shifts, bias embedded in training data, frequent retraining that multiplies oversight events, and data dependency that ties model quality directly to feed quality. OSFI warns that increased model risk can expose institutions to financial loss from flawed decision making, operational losses, legal implications, or reputational damage, and that models characterized by dynamic self-learning and autonomous decision-making may become prevalent. A payments estate that cannot see drift in its fraud models cannot defend its approval rates to a supervisor. Sources: https://www.yields.io/use-cases/ai-model-risk-management https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027

Third-party dependence sharpens the problem. Payment institutions frequently consume vendor models for screening, scoring, and customer communication, yet regulators hold institutions responsible for mitigating risk and ensuring the conceptual soundness of any model or algorithm their systems use. The April 9, 2021 Interagency Statement on Model Risk Management for Bank Systems Supporting Bank Secrecy Act compliance states plainly that banks remain ultimately responsible for complying even when they choose third-party models. The practical response is to test every model before implementation, evaluate the results, document them thoroughly, and focus controls on preventing inappropriate or incomplete data inputs, biased models, insecure or error-prone models, and privacy breaches. Sources: https://kaufmanrossin.com/blog/managing-ai-model-risk-in-financial-institutions-best-practices-for-compliance-and-governance/

The discipline has also become a market signal. The global market for AI-focused model risk management solutions is projected to reach roughly USD 6.4 billion in 2025 and to grow at more than 12 percent annually through the end of the decade, reflecting rising demand for oversight of predictive systems across banking and payments. Model risk management is the structured process of identifying, assessing, and managing the risks of predictive models so they remain accurate, transparent, and compliant, and it matters to payment operators because it mitigates financial losses, supports regulatory compliance, improves decision quality, and builds stakeholder trust behind every payment approval, decline, and routing decision a model touches. Sources: https://www.metricstream.com/learn/what-is-model-risk-management.html

What converts the policy into protection is a control set that observes, enforces, and records. The foundational control is a living AI inventory, a continuously refreshed bill of materials covering internal models, vendor-embedded AI, agents, and their datasets, because an inventory built from self-reporting undercounts. Risk classification should follow the use case rather than the model, so the same system is treated differently when it summarizes internal notes and when it scores payment fraud. Pre-deployment adversarial testing against prompt injection, jailbreaks, data extraction, and evasion, followed by runtime monitoring with retention of prompts, responses, and tool calls, gives a payments compliance team evidence that behaves like the systems it defends. Sources: https://cranium.ai/8-ai-governance-controls-that-reduce-model-risk/ - federalreserve.gov: https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm - ithandbook.ffiec.gov: https://ithandbook.ffiec.gov/ - artificialintelligenceact.eu: https://artificialintelligenceact.eu/high-level-summary/ - eur-lex.europa.eu: https://eur-lex.europa.eu/eli/reg/2022/2554/oj

Frequently asked questions

What counts as a model under payments model risk management?

Under the expanded supervisory definition OSFI finalized in Guideline E-23, a model is any quantitative tool or algorithm that uses data to generate an output, which reaches well beyond capital and pricing mathematics. For a payments operator that means transaction fraud scoring ensembles, sanctions screening engines, merchant credit underwriting models, routing optimizers, and customer-facing generative or agentic assistants all sit on the same inventory. The practical test is whether a system uses data to produce an output that influences a payment decision, not whether it resembles a traditional statistical model.

Which regulatory frameworks drive payments model risk expectations today?

Three threads dominate. SR 26-2, the revised United States interagency guidance, reaffirms model governance and validation and explicitly carves out generative and agentic AI as systems that need their own governance track. OSFI Guideline E-23 comes into full effect in May 2027 and sets a principles-based, risk-proportional, enterprise-wide mandate for model risk. The EU AI Act assigns high-risk obligations to financial services uses such as credit scoring and fraud detection from December 2027, which means classification of the payments estate cannot wait for the enforcement date.

Why is third-party model risk especially acute for payment providers?

Payment institutions often consume vendor-built models for screening, scoring, and customer communication, and those third-party systems are harder to understand and explain than models developed in-house. Supervisors nonetheless hold the institution responsible for mitigating risk and ensuring the conceptual soundness of any model or algorithm its systems use, an expectation stated as early as the April 2021 interagency statement on bank systems supporting Bank Secrecy Act compliance. The mitigation combines contract and control: test before implementation, document results, and govern the data inputs the vendor model consumes.

What does continuous model risk control look like in a payments estate?

It looks like controls that observe, enforce, and record rather than documents that describe. A continuously refreshed inventory of internal models, vendor-embedded AI, agents, and datasets scopes everything else. Risk classification tied to the use case keeps oversight proportionate across payment workflows. Pre-deployment adversarial testing covers prompt injection, jailbreaks, data extraction, and evasion before a system touches production traffic. Runtime monitoring with retention of prompts, responses, and tool calls then captures model risk as a behavior, because a fraud or routing model shows what it really is only in production under live inputs.