AI identification, inventory and risk materiality in finance

Cross-cutting synthesis

Three moves show up in almost every framework, and they run in order. First you identify where AI is used across the firm, including the embedded and vendor tools the business may not think of as AI. Then you record what you find in an inventory. Then you rate how much each entry matters, and govern in proportion.

• The model-risk tradition already requires the inventory and the rating, and the AI frameworks inherit them. SR 11-7, SS1/23 and OSFI E-23 have long required a central inventory (purpose, owners, inputs, status, validation dates) and a risk rating to prioritise work. MAS, RBI, NIST and FINMA adopt the same controls and extend the inventory to AI-specific attributes: model type including generative, third-party dependencies, and data sources. How much autonomy a system has feeds the materiality rating.
• The materiality rating runs on the same few dimensions. MAS is the clearest: impact on customers and the firm, model complexity, and how much the firm relies on the model or lets it act. These map onto OSFI's vulnerability-and-materiality and SR 11-7's impact-and-consequence test. The rating then sets how deep the downstream controls go.
• The EU AI Act fixes materiality in law. Instead of leaving the judgement to the firm, it sorts systems into tiers - prohibited, high-risk, limited, minimal - and makes the tier the materiality call, with high-risk systems registered in an EU database.

Where they differ is scope and granularity. OSFI triages: only models with non-negligible risk enter the inventory and attract full governance, which keeps the register proportionate but leans on a defensible triage. The EU sets scope by use-case category rather than firm judgement. RBI goes furthest on transparency, with semi-annual updates, supervisory access, and a proposed sector-wide repository, echoing the FSB's point that supervisors lack data on where AI is actually deployed. One tension runs through all of them. Generative and agentic systems, embedded vendor AI, and tools the business does not think of as "models" are easy to miss, so the OCC and MAS both say the inventory has to capture AI by what it does and what it risks, not by whether someone called it a model.

What each framework says

MAS AIRG - identify AI across the firm, inventory it, and rate its materiality.

• A control function identifies AI use consistently across all business and functional areas and acts as the final arbiter of what counts.
• An accurate, up-to-date AI inventory records purpose, scope of use, model type, data, dependencies, lifecycle status, risk-materiality rating and key roles, linked to the firm's other inventories and reviewed so it keeps pace with newer and third-party AI.
• Materiality is rated on three dimensions - impact (the harm if the AI fails, to the firm, customers and stakeholders), complexity (how novel the technology and the data are), and reliance (how much autonomy the system has, and whether there is an alternative).
• That rating drives everything downstream, setting how much lifecycle control each system attracts.

Guidelines on AI Risk Management (MAS, 2025)

United States, SR 11-7 - a firm-wide model inventory, with materiality setting the rigour. Banks keep a comprehensive record of models in use, under development or recently retired, and a single party owns the firm-wide inventory. It describes purpose, products, use and restrictions, inputs and their sources, outputs, whether the model works, last-update and validation dates, and responsible people. Where a model's failure would be particularly harmful, the model risk management framework has to be more extensive and rigorous. SR 11-7 / OCC 2011-12 (Federal Reserve / OCC)

United Kingdom, PRA SS1/23 - a comprehensive inventory plus risk-based tiering. The inventory identifies sources of model risk, feeds management reporting, and shows how models depend on each other, covering models that are live, in development or decommissioned. Risk-based tiering then prioritises validation and other controls, and a formal model definition sets what falls inside the framework. SS1/23: Model risk management principles for banks (Bank of England / PRA)

Canada, OSFI E-23 - identify and triage every model, then rate it. Institutions track all models in use or recently decommissioned, including vendor and third-party models, through periodic surveys and triage. Models with non-negligible risk go into an enterprise inventory that is comprehensive, evergreen and controlled. Each gets a model risk rating on clear, measurable criteria reflecting inherent vulnerability and the materiality of impact, reviewed when a trigger event occurs. Guideline E-23: Model Risk Management (OSFI)

European Union, AI Act - materiality set in law by risk tier. The Act sorts systems into tiers: some practices are prohibited, high-risk systems carry the full obligations, and a listed high-risk system can be treated as not high-risk only where it poses no significant risk of harm. Providers of high-risk systems register themselves and the system in an EU database. The tier is, in effect, a materiality judgement held in a public register. Regulation (EU) 2024/1689, the Artificial Intelligence Act

NIST AI RMF - an AI system inventory as the base for portfolio risk. Under the Govern function, NIST expects mechanisms to inventory AI systems, resourced by the organisation's risk priorities, supporting accountability, context mapping, and prioritising risk-management effort across the portfolio. AI Risk Management Framework 1.0 (NIST)

ISO/IEC 42001 - know and record what is in scope. The organisation has to determine the scope of its AI management system, which means knowing and recording which activities and AI systems it covers, then applying control in proportion to each use case. ISO/IEC 42001:2023, AI management system

India, RBI FREE-AI - a transparent inventory, refreshed and visible to supervisors. Firms keep a comprehensive inventory covering model type including generative AI, purpose and function, use cases, dependencies on third parties, cloud and data, a High, Medium or Low risk category set by board-approved policy, and a record of grievances. It is updated at least twice a year and available to supervisors, and the report proposes a sector-wide AI repository so supervisors can see adoption in aggregate. FREE-AI Committee Report (Reserve Bank of India, 2025)

Switzerland, FINMA - a detailed inventory, with effort scaled to materiality. FINMA expects a detailed inventory of AI systems and risk-management effort that tracks the significance of each application. Guidance 08/2024 on Governance and Risk Management when using AI (FINMA)

IAIS - repositories of deployed models for high-risk AI. For high-risk applications, IAIS recommends keeping repositories of all deployed models to support oversight and risk assessment. Application Paper on the Supervision of Artificial Intelligence (IAIS)

FSB - limited data on AI use makes the risks hard to monitor. The FSB treats the lack of data on where AI is used, at firm and system level, as a gap that leaves AI-related vulnerabilities hard to monitor. It urges authorities to close those data gaps and strengthen monitoring of AI adoption across the sector, which is the supervisory case for firm inventories and sector-level monitoring. Monitoring Adoption of AI and Related Vulnerabilities in the Financial Sector (FSB, 2025)

IOSCO and HKMA are lighter here. Both expect risk assessment and proportionate controls, but neither specifies a formal inventory as a named control, treating it as implied by governance and record-keeping.

Comparison

Framework	Position on identification, inventory and materiality
MAS AIRG	Firm-wide AI inventory with approved scope; materiality rated on impact, complexity and reliance (including autonomy)
US SR 11-7 / SR 26-2	Firm-wide model inventory with rich attributes; materiality drives the rigour of the framework
UK PRA SS1/23	Comprehensive inventory (including under-development and decommissioned); risk-based tiering prioritises controls
OSFI E-23	Identify and triage all models; enterprise inventory of non-negligible-risk models; model risk rating on measurable criteria
EU AI Act	Materiality fixed in law via risk tiers; high-risk systems registered in an EU database
NIST AI RMF	Mechanisms to inventory AI systems, resourced by risk priority; supports portfolio prioritisation
ISO/IEC 42001	Determine and record the scope of the AI management system; risk-based control per use case
RBI FREE-AI	Comprehensive AI inventory (type, use, dependencies, risk tier, grievances); semi-annual updates; sector-wide repository
FINMA	Detailed AI inventory; risk effort scaled to materiality
IAIS	Repositories of all deployed models recommended for high-risk AI
FSB	Firm and sector data gaps obscure AI-related vulnerabilities; FSB urges closing them and monitoring adoption
IOSCO / HKMA	Risk assessment and proportionate controls, but inventory is implied rather than a named requirement

FAQ

Is one AI inventory enough, or do we need a separate AI register? Most frameworks favour one authoritative, firm-wide register linked to related inventories rather than a parallel AI-only list. SR 11-7 and OSFI want a single enterprise record, and MAS links the AI inventory to the firm's other inventories. The point is that AI systems, including embedded vendor and generative tools, are not missed.

What attributes should the inventory capture? The frameworks converge on purpose and use case, approved scope and restrictions, model type including generative AI, inputs and data sources, third-party dependencies, owner and responsible people, risk rating, validation and update dates, and status (live, in development, decommissioned). RBI adds a record of grievances; SR 11-7 adds the model's expected valid lifetime.

What dimensions drive the materiality rating? MAS names impact on customers and the firm, model complexity, and degree of reliance or autonomy. OSFI frames it as inherent vulnerability and the materiality of impact, and SR 11-7 as the impact on business decisions and the harm if the model fails. These are compatible; most firms combine an impact axis with a complexity or autonomy axis.

Does everything the firm uses go in the inventory? Not necessarily at full governance. OSFI triages, so only non-negligible-risk models attract full lifecycle governance, which keeps the register proportionate but needs a documented, defensible triage. The EU AI Act instead fixes which systems are in scope by use-case category.

How often must the inventory be refreshed? It should be evergreen and updated on trigger events, and refreshed at least periodically. RBI specifies updates at least twice a year and availability to supervisors; the UK expects an annual self-assessment cycle that keeps the inventory current.

Read the sources

Every framework quoted above is in the full AI risk management in finance resource list. For the Singapore picture, see the AIRG explained and the AIRG compared to the EU AI Act, NIST and ISO 42001.

Work with me

I train banks, insurers, and supervisors on building an AI inventory and a risk-materiality rating that actually drives controls, grounded in the AIRG and the frameworks above. See the courses and workshops, or get in touch.