AI Safety in Banking - IIDA Perspective

Executive summary

Regulators are not converging on a single AI rulebook for banks. They are converging on a shared set of expectations (governance, inventories, lifecycle testing, third-party oversight, human control of autonomous systems) delivered through very different legal instruments.

Legal form ranges from binding horizontal law (EU) through supervisory guidelines (Singapore, China) and voluntary frameworks (US Treasury, India, FSB) to existing principles-based rules applied to AI (UK). Generative and agentic AI is the common pressure point.

Five things to take away

Five things to take away

  1. Legal status varies widely. Only the EU AI Act is horizontal binding AI law; most other initiatives are guidance, frameworks, or recommendations. Treat each separately.
  2. Generative and agentic AI is treated differently. US model-risk guidance expressly excludes it, while supervisors are asking pointed questions about agent authority, oversight, auditability and shutdown.
  3. “No AI-specific rule” does not mean “no obligation.” UK and US banks remain subject to existing conduct, accountability, model-risk and consumer rules.
  4. Dates are moving. The EU deferred its high-risk deadline to December 2027; Singapore’s expectations phase in through October 2028.
  5. A common core is emerging, so multinational banks can build one control baseline and map it to each regime.

How to read the labels in this briefing

Publications differ in legal force, and treating every one as a rule is a common error. Each initiative carries one of these color-coded labels.

Label

What it means

Example

 BINDING LAW 

Enacted law with enforceable obligations and penalties.

EU AI Act (as amended)

 SUPERVISORY GUIDANCE 

Regulator guidelines assessed in examinations; not always formal rules.

MAS, NFRA, US model-risk guidance

 VOLUNTARY / RECOMMENDED 

Frameworks and sound practices institutions are encouraged to adopt.

US Treasury FS AI RMF, RBI FREE-AI, FSB

 PRINCIPLES-BASED (EXISTING RULES) 

No AI-specific rules; existing technology-neutral rules apply.

UK FCA and PRA

 PROGRAM / PILOT 

Regulator-run testing, sandbox or pilot activity.

FCA AI Live Testing; MAS MindForge

Recommended practice, reported supervisory inquiries and binding obligations are different things. The absence of AI-specific guidance is not permission to deploy without controls.

1. Landscape at a glance

Figure 1. Regulatory approach by jurisdiction.

Region

Principal instrument

Legal character

GenAI / agents

Key dates

United States

Treasury FS AI RMF and AI Lexicon; revised Fed, OCC, FDIC model-risk guidance

 VOLUNTARY / RECOMMENDED 

Excluded from model-risk guidance; supervisory attention growing

Feb 19, 2026; Apr 17, 2026

United Kingdom

Existing rules (Consumer Duty, SM&CR, PRA SS1/23); FCA AI Live Testing

 PRINCIPLES-BASED (EXISTING RULES) 

Technology-neutral; agentic AI in core decisions expected to grow

Testing ends end-2026; report Q1 2027

European Union

EU AI Act as amended by the Digital Omnibus on AI

 BINDING LAW 

Credit scoring is high-risk; chatbot transparency duties

Annex III: Dec 2, 2027; Annex I: Aug 2, 2028

India

RBI FREE-AI: 7 Sutras, 26 recommendations, 6 pillars

 VOLUNTARY / RECOMMENDED 

Lifecycle governance, board policies, incident reporting

Aug 13, 2025

China

NFRA Guiding Opinions: 32 measures, 7 areas

 SUPERVISORY GUIDANCE 

Cautious exploration of financial AI agents

Jun 18, 2026

Singapore

MAS AI Risk Management Guidelines and MindForge toolkit

 SUPERVISORY GUIDANCE 

Covers all AI; toolkit spans traditional, generative, agentic

Oct 7, 2026; phased to Oct 7, 2028

Global (FSB)

12 Sound Practices for Responsible AI Adoption

 VOLUNTARY / RECOMMENDED 

Governance, cyber, data quality, human oversight

Consulted Jun 2026; final reported Oct 15, 2026

2. United States, United Kingdom and European Union

United States

** VOLUNTARY / RECOMMENDED **  SUPERVISORY GUIDANCE 

  • Treasury FS AI RMF and AI Lexicon (Feb 19, 2026). A voluntary framework adapting NIST’s AI RMF to financial services, plus shared definitions. Treasury’s AI oversight group coordinates further work on identity, fraud, explainability and data practices; not every deliverable is complete.
  • Revised Fed, OCC, FDIC model-risk guidance (Apr 17, 2026). Replaces SR 11-7 and covers validation, monitoring, governance and third-party models. It sets no enforceable standards and is most relevant above $30 billion in assets.
  • GenAI and agentic AI are explicitly out of scope, but not exempt from risk management.
  • Existing law still applies. Treasury asks firms to review AI use cases against current laws before deployment and periodically; state laws and fair-lending duties also apply.

So what for banks

Do not stretch traditional model validation to LLMs and agents. Run a separate control path for them, and review legal compliance separately from technical testing.

United Kingdom

** PRINCIPLES-BASED (EXISTING RULES) **  PROGRAM / PILOT 

  • No AI-specific rules planned. The FCA relies on the Consumer Duty and SM&CR; the PRA on model-risk principles in SS1/23. Bank of England roundtables showed broad support for this approach. <sup>[7][8][9]</sup>
  • Pressure for clarity. The Treasury Committee (Jan 20, 2026) called the stance “wait-and-see” and recommended FCA guidance and AI stress tests, and critical-third-party designation of major AI and cloud providers, by end-2026. It has no binding powers; the BoE and PRA disputed the label on Apr 1. <sup>[8][10][9]</sup>
  • FCA AI Live Testing. Eight firms joined the second cohort in April 2026 (including Barclays, Experian, Scottish Widows, UBS) to test AI with live monitoring. Testing ends in 2026; evaluation report in Q1 2027. <sup>[11]</sup>

So what for banks

Show that AI use meets named-manager accountability (SM&CR), good customer outcomes (Consumer Duty) and sound model risk management (SS1/23). Prepare for possible FCA guidance.

European Union

 BINDING LAW 

  • The only horizontal binding AI law. Assessing individuals’ creditworthiness or credit scores is high-risk, as are life and health insurance pricing and employment uses. AI that detects financial fraud is excluded from the credit-scoring category. <sup>[12]</sup>
  • Deadline deferred, not cancelled. The Digital Omnibus (Reg. (EU) 2026/1744, in force Jul 27, 2026) moved stand-alone high-risk obligations from Aug 2, 2026 to Dec 2, 2027, and embedded products (Annex I) to Aug 2, 2028. Some obligations were not deferred. <sup>[13][14][15]</sup>
  • Operational points. A bank building its own scoring model is both provider and deployer; one licensing a vendor platform is a deployer. Affected individuals should be able to understand the factors behind a rejection. <sup>[12]</sup>

So what for banks

Classify use cases now, identify provider or deployer roles, map AI Act duties to existing model-risk programs, and use the extra time to build evidence rather than pause. Non-EU banks serving EU customers should confirm applicability.

3. India, China, Singapore and global standard-setters

India

China

Singapore

 VOLUNTARY / RECOMMENDED 
■    RBI FREE-AI report (Aug 13, 2025): 7 Sutras, 26 recommendations, 6 pillars. [17]
■    Recommends board-approved AI policies, lifecycle governance, extending product approval, consumer protection and audits to AI, stronger cyber and incident reporting, an innovation sandbox, and shared data and compute. [18]
■    Caution: these are committee recommendations, not binding RBI requirements. [17]
So what: Governance paired with inclusion, so responsible adoption reaches beyond the largest banks.

 SUPERVISORY GUIDANCE 
■    NFRA Guiding Opinions (Jun 18, 2026): 32 measures across 7 areas, from governance to supervision. Institutions remain responsible for their AI. [19]
■    Full-lifecycle management; risk-tiered controls; approval and human intervention for high-risk uses; data protection including poisoning defenses; explainability; outsourcing and resilience. [19][20]
■    Calls for cautious exploration of financial AI agents. [19]
So what: Controls the whole system around AI, not only the model. Check obligations against the underlying text.

 SUPERVISORY GUIDANCE    PROGRAM / PILOT 
■    MAS Guidelines (final Oct 7, 2026): all financial institutions, all forms of AI, proportionate to risk. [21][22]
■    Phasing: effective Oct 7, 2027, with oversight and core risk management (Sections 3–4) from then; lifecycle controls and capabilities (Sections 5–6) by Oct 7, 2028. [22]
■    MindForge toolkit (Mar 20, 2026): built with 24 firms; covers traditional, generative and agentic AI. [23] MAS plans to consult on agentic AI in 2027. [24]
So what: Guidelines set expected outcomes; MindForge helps implement them.

Global standard-setters

VOLUNTARY / RECOMMENDED 

  • Financial Stability Board. On Jun 10, 2026 the FSB consulted on 12 non-binding sound practices for responsible AI adoption (comments to Jul 22), aimed at boards and senior management: governance, cybersecurity, data quality and human oversight.A final version is reported for Oct 15, 2026; confirm on fsb.org.
  • Basel Committee, IMF and BIS. Basel published an ICT risk report with an AI example (Jun 2) and discussed AI at its Sept 28–29 meeting. IMF and BIS published analysis of AI-enabled cyber risk and long-term impacts.

So what for banks

Global standards are non-binding but influential. Use the FSB practices as a common governance baseline across jurisdictions and expect national supervisors to align over time.

Dates to watch

4. Where risk arises, what regimes share, and what to do

AI safety scenarios across lines of business

Line of business

Illustrative safety scenario

Key control

Retail and consumer lending

Credit model systematically declines a protected group; customers get no explanation.

Fairness testing, adverse-action reasons, human review

Cards and payments

Drift causes false declines, or an agent initiates payments beyond its authority.

Drift monitoring, limits, approval gates, kill switch

Corporate and commercial

LLM summary misses a loan covenant, or client data leaks via prompts.

Human sign-off, data classification, output validation

Markets and treasury

Similar models herd in stress; an opaque model fails outside its training range.

Stress testing, position limits, challenger models

Wealth and advice

Chatbot gives unsuitable or invented advice to a retail investor.

Guardrails, suitability checks, supervision

Financial crime and compliance

Model misses a new laundering pattern, or deepfakes defeat identity checks.

Independent validation, red-teaming, layered identity checks

Customer service and operations

Assistant reveals another customer’s data or misstates fees.

Retrieval limits, PII redaction, human escalation

Technology, cyber and vendors

Security agent acts beyond authority, or a vendor model change breaks a critical service.

Least privilege, audit logs, exit plans, concentration monitoring

Where the regimes overlap

Theme

US

UK

EU

India

China

SG

FSB

Board and senior governance

●

●

◐

●

●

●

●

Lifecycle validation and monitoring

●

●

◐

●

●

●

◐

AI inventory and risk tiering

●

◐

●

○

●

●

◐

Third-party and vendor risk

●

●

◐

○

●

◐

◐

Human oversight of agents

●

◐

◐

○

●

●

◐

Data security, privacy, cyber

◐

◐

◐

●

●

○

◐

Consumer protection and explainability

●

●

●

●

◐

○

○

Testing sandboxes and pilots

○

●

○

●

◐

●

○

● explicitly addressed ◐ partly or indirectly ○ not a focus of the material summarized. Indicative, based on the cited sources; not a legal analysis.

Action plan: one control baseline, mapped to each regime

#

Action

What good looks like

Draws on

1

Build a tiered AI inventory

Conventional models, generative assistants and action-taking agents recorded separately, with materiality ratings.

US, SG, CN, EU

2

Assign accountability

Board-approved AI policy; a named senior owner per material use case.

IN, UK, SG

3

Classify and map to regimes

Each use case tagged to applicable laws and guidance (e.g., EU credit scoring as high-risk).

EU, US

4

Apply lifecycle controls

Independent validation, pre-deployment testing and live monitoring after launch.

US, UK, CN

5

Separate path for GenAI and agents

Documented permitted actions, approval boundaries, activity logs and tested shutdown.

US, CN, SG

6

Govern vendors and data

Vendor evidence, concentration monitoring, exit plans; pipeline security and incident reporting.

US, CN, UK, IN

7

Assess legal compliance and track status

Fairness and explainability reviews apart from technical testing; a register of binding, supervisory and voluntary items.

US, UK, EU

5. Conclusion and references

Conclusion

The evidence points to convergence in substance but not in form. Governance, inventories, lifecycle testing, third-party oversight and human control recur in every regime, while legal force ranges from EU law to voluntary frameworks. Generative and agentic AI is the open frontier, the calendar (EU Dec 2027, Singapore Oct 2028) is now the constraint, and existing consumer, fair-lending and accountability law applies regardless.

Now
Stand up the tiered inventory, name owners, classify use cases against applicable regimes.

Next 12 months
Build the GenAI and agent control path; strengthen vendor evidence; track the FCA report, FSB final text and any US information request.

2027 to 2028
Evidence compliance with Singapore and EU high-risk expectations; embed live monitoring and incident reporting.

Bottom line

Treat AI safety as an operating capability, not a compliance project: show what AI you run, who owns it, how it is tested and monitored, and how it can be stopped.

References

  1. ICBA: Treasury releases two AI resources
  2. Orrick: Agencies Overhaul Model Risk Management Guidance for Banks: Here’s What Changed
  3. Federal Reserve Board: SR 26-2: Supervisory Guidance on Model Risk Management
  4. PwC: Our take: updated model risk management guidance
  5. Global Banking & Finance Review (Reuters syndication): Exclusive: Banking watchdogs ramp up AI scrutiny as stability fears mount
  6. U.S. Department of the Treasury: Treasury Releases Report on the Uses, Opportunities, and Risks of AI in Financial Services
  7. Covington: UK financial services regulators’ approach to artificial intelligence in 2026
  8. Linklaters: FCA set deadline for giving UK firms more AI guidance
  9. Bank of England: Response to Treasury Committee inquiry report on AI in financial services
  10. HFW: Treasury Committee report on AI in financial services
  11. Financial Conduct Authority: FCA announces second cohort for AI Live Testing
  12. Banking Vision: High-risk AI in banking
  13. Gibson Dunn: EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes
  14. Cloud Security Alliance: EU AI Act’s High-Risk Deadline: Deferred, Not Cancelled
  15. Mondaq: EU Digital Omnibus on AI enters into force
  16. AI2Work: EU AI Act High-Risk Obligations Hit Banking: What Firms Must Do Now
  17. Reserve Bank of India: Report of the FREE-AI Committee: press release
  18. Reserve Bank of India: FREE-AI Committee Report
  19. AML Network: China AI Finance Rules: NFRA’s 2026 Bank and Insurer Framework
  20. HLC: Navigating China’s regulatory landscape for AI applications in the financial industry
  21. Monetary Authority of Singapore: Guidelines on AI Risk Management for Financial Institutions
  22. Monetary Authority of Singapore: Guidelines on Artificial Intelligence Risk Management
  23. Monetary Authority of Singapore: MAS partners with industry to develop AI risk management toolkit
  24. Allen & Gledhill: MAS publishes AI risk management guidelines
  25. Financial Stability Board: FSB consults on sound practices for the responsible adoption of AI
  26. ABA Banking Journal: Financial Stability Board releases sound practices for AI adoption
  27. Resultsense: FSB final 12 sound practices on AI adoption
  28. Regulation Tomorrow (Hogan Lovells): Basel Committee meeting discusses AI
  29. GFMA: GFMA AI newsletter

Disclosure: This briefing is an informational summary of public sources and is not legal advice. Regulatory positions change; confirm current requirements against primary texts and with qualified counsel before acting. Several sources are commentary or news; regulator publications should be preferred where they differ.

AI use disclosure: Author wrote the primary text of this article. They used Sonnet 5.5 and Astra 6 to help reason through my work, brainstorm ideas, build visualizations and elaborate it.