Skip to content
Banking & Finance

Banking Under Pressure: AI, Fraud and the Challenge of Connected Resilience

A global perspective on the connected challenges facing banking: AI-enabled cyber threats, fraud, supplier concentration, payment integrity, financial resilience and quantum preparedness.

Banking · Global regulatory perspectives · Risk and resilience

AI-enabled threats, sophisticated fraud, concentrated technology suppliers and financial uncertainty are reshaping banking risk. Signals from Australia, the UK, Singapore, Hong Kong, the euro area, the United States and India show why resilience needs to be designed across the whole banking service.

A payment can fail in several ways before anyone sees a failed transaction. A customer may be impersonated, an automated assistant may exceed its authority, a supplier may become unavailable, or the bank may lose a reliable view of the payment’s outcome. Each event starts differently. Each can end with a customer unable to trust what happened to their money.

Recent official publications bring these connections into focus. Australia’s payments discussion addresses agentic commerce and merchant choice. UK authorities examine cyber resilience, AI and quantum preparedness. Singapore sets out AI governance expectations, Hong Kong highlights fraudulent payment support, and supervisors elsewhere connect operational weaknesses with wider financial risks. The instruments differ, but together they offer a useful view of the pressures banking leaders must manage.

Our perspective: banking modernisation succeeds when institutions can control automated decisions, maintain essential services, establish trustworthy transaction outcomes and protect customers during disruption. These capabilities depend on coordinated governance across business, technology, security, operations and risk.

What authorities are saying—and what kind of publication it is

This perspective draws on selected official publications available as at 20 November 2026. Their status matters: a consultation proposes an approach, a survey reports evidence, an alert describes a threat, and directions or guidelines apply within their stated scope and timetable. The comparison below preserves those distinctions.

Jurisdiction and publicationStatus and timingRelevant signal
Australia — RBA Payments System Board update [1]19 November 2026; policy and oversight update. Wider review priorities due by end-2026.Agentic commerce, merchant routing choice, payment costs and resilience remain connected policy questions.
UK — AI survey and joint cyber publications [2,9–11]Survey: 26 October 2026. Consultation and separate non-binding AI/quantum practices: 19 November. Consultation closes 26 February 2027.Lifecycle oversight, cyber preparedness, bounded automation and planning for cryptographic change.
Singapore — MAS AI Risk Management Guidelines [3]Issued 7 October 2026. Sections 3–4 expected from 7 October 2027; Sections 5–6 by 7 October 2028.Risk-proportionate governance, lifecycle controls and accountability for third-party AI.
Euro area — ECB supervisory priorities 2027–29 [4]Published 19 November 2026; forward-looking supervisory priorities.Financial resilience to geopolitical/macroeconomic shocks, operational resilience, ICT and risk reporting.
United States — Federal Reserve Financial Stability Report [5]13 November 2026; financial-system assessment, reflecting information through 28 October.Financial vulnerabilities and cyber dependencies can transmit stress across firms and markets.
India — RBI technology/cyber directions and resilience speech [6,7]Directions issued 31 July 2026, updated 1 October; speech published 25 September. Distinct instruments.Board oversight, technology controls and resilience sit alongside governance of AI-enabled services.
Hong Kong — HKICL scam alert carried by HKMA [12]20 November 2026; specific public warning about a fraudulent website.Fraud can imitate refunds, payment-dispute reporting and customer support.

1. Financial resilience still matters in a technology-led discussion

The ECB’s priorities explicitly pair financial resilience with operational resilience. They focus on prudent credit standards, capital planning, exposures to non-bank financial institutions and reliable risk reporting, alongside technology risks. A global banking article therefore cannot reduce today’s challenge to AI adoption or cyber defence alone. [4]

The Federal Reserve’s November report similarly assesses valuations, household and business borrowing, financial-sector leverage and funding vulnerabilities. It describes funding risks as moderate and notes that most domestic banks maintained substantial liquid assets. That is a reason for a measured assessment, not a claim that banking systems are uniformly close to crisis. [5]

The management implication is to test plausible combinations. What happens if a sector’s borrowers face weaker cash flows while the bank’s funding becomes more expensive? Can treasury see usable liquidity when a key reporting dependency is unavailable? Which collateral, counterparties and income streams move together under the same shock?

These are proposed scenario questions, not forecasts. They help connect underwriting, concentration limits, funding plans and service continuity without treating a good system-wide average as evidence that every institution has the same capacity to absorb stress.

2. AI adoption is moving faster than some oversight practices

The UK survey reports AI use by 86% of respondents to the full survey. It also finds that pre-deployment controls were more widely reported than ongoing monitoring, leakage testing and periodic review. The report cautions that its findings describe participating firms rather than the entire financial sector; 105 respondents completed the full questionnaire. [2]

This is a useful distinction for a bank’s AI inventory. Recording a system at approval is only the beginning. A vendor can change a model, an internal team can broaden its use, and a helpful drafting tool can acquire access to business data or execution tools. The original approval may no longer describe what the service actually does.

Singapore’s MAS guidelines provide a related supervisory response: proportionate governance across the AI lifecycle, clear responsibility and assurance over third-party use. The phased 2027–28 implementation dates matter. Issued guidance should not be described as if every provision is already an immediate obligation in November 2026. [3]

A practical control is to record the system’s purpose, data access, permitted actions, materiality and accountable owner, then revisit those details when capability or usage changes. Human review should mean an informed person can challenge or stop an outcome, with sufficient time and evidence to do so.

The UK regulators’ 19 November AI cyber practices add a security dimension to that lifecycle. The paper is non-binding and separate from formal supervisory expectations. It discusses restricting access and actions, maintaining human oversight and ensuring that automation can be stopped or handed over safely. For a banking assistant, the design question becomes concrete: can it explain a payment, prepare one, approve one or execute one—and where does each permission end? [10]

3. Fraud is testing identity, process and judgement together

The US FinCEN deepfake alert is older—13 November 2024—and is included as background, not presented as a new 2026 announcement. It describes misuse of generated media and fraudulent identity documents to bypass verification, provides red flags and discusses suspicious-activity reporting. Its practical relevance persists as banks expand digital onboarding and remote interactions. [8]

The RBI’s September 2026 speech likewise discusses deepfakes, voice cloning, social engineering and the risks around identities, APIs and connected services. It argues for governance and monitoring alongside AI adoption. A speech signals supervisory thinking; it does not itself establish a new set of binding technical rules. [7]

For the bank, a convincing face, voice or document should be assessed within a broader decision process. An unusual beneficiary change, unexpected device, altered communication channel and urgent request can be more informative together than any single signal. The response should be proportionate investigation and secure verification—not automatic exclusion of every customer with a technical difficulty.

The same principle applies to employee fraud and payment approvals. A familiar-looking message should not bypass a defined approval path, and an escalation should be checked through a trusted route. Training, access controls, monitoring and customer redress need to operate as one service.

Hong Kong’s 20 November alert shows that the deception can continue into customer support. HKICL warned of an unaffiliated website presenting itself as FPS refund, unauthorised-transaction reporting and support services, then directing users to an impersonated customer-service representative. HKICL clarified that it does not directly provide FPS services to individual members of the public. This is a scam warning rather than a change to payment rules. [12]

The practical lesson is to treat help-seeking as part of the protected customer journey. A person worried about a missing payment may be particularly receptive to a convincing offer of assistance. Banks can make verified support routes prominent in authenticated channels, explain how suspicious contact should be reported and ensure that genuine staff follow a consistent verification process.

4. Third-party dependency is becoming a concentration question

An institution can diversify its applications while still relying on the same cloud platform, identity provider or model service underneath. Different vendor names do not necessarily mean independent failure paths.

The UK survey reports concentration in hosting relationships and uncertainty over oversight of third-party AI. The ECB’s priorities separately focus on ICT suppliers, cloud dependencies and remediation of previously identified weaknesses. These sources support a dependency question, rather than a claim that outsourcing is inherently unsuitable. [2] [4]

A useful dependency map follows the critical customer service through the application, infrastructure, authentication, connectivity, data and support chain. It identifies subcontractors and common providers, then asks how the bank operates if one becomes unavailable or unusable.

An exit clause is not the same as an executable exit. The institution needs to understand data portability, credentials, configuration, replacement capacity, transition time and the financial implications. A tested fallback can be more useful than an unverified claim that another supplier is available.

5. Payments innovation must preserve choice and transaction integrity

The RBA’s 19 November update reiterates its expectation that dual-network debit functionality remain available across form factors, including those used in agentic commerce, to support merchant choice. It also reviews payment costs and infrastructure resilience. This is not a comprehensive new authorisation or liability rulebook for AI agents. [1]

The transferable design issue is where decisions move when a customer delegates a task. An assistant may search, select a merchant and initiate payment. Those permissions should be separated and bounded. A token or integration should not quietly remove an eligible choice that the intended service is supposed to preserve.

A payment timeout adds another challenge. Missing confirmation does not prove that the financial instruction failed. An automated retry through another provider can create a duplicate if the original attempt succeeded. Stable references, authoritative status checks, controlled retries and reconciliation protect the customer as well as the institution.

Australia’s network arrangements do not automatically apply elsewhere. Indian banks and providers must interpret their own applicable rules and scheme arrangements. The common architecture questions concern authority, choice, evidence and recovery.

6. A recovered application is not necessarily a recovered banking service

Operational recovery needs more than a server becoming available. The bank must establish which payments completed, which accounts are accurate, which obligations remain and how customers will be informed. Decision-makers also need a trustworthy view of exposures and liquidity while that work proceeds.

India’s RBI cybersecurity and technology directions address governance, access, third-party arrangements, surveillance, continuity and assurance. The particular Commercial Banks directions cited here exclude Small Finance Banks, Payments Banks and Local Area Banks from that definition. Each institution must check its own applicable instrument; the scope cannot be generalised from the title. [6]

The operational lesson is to test business outcomes under realistic dependency failures. Can the bank authenticate staff, query payment outcomes, reconcile balances, communicate clearly and prioritise critical obligations? Can it demonstrate that a previous finding is closed through evidence, rather than the completion of a project milestone?

The joint UK consultation CP17/26 and GC26/9, issued on 19 November, proposes principles-based cyber-risk and resilience expectations. It clarifies how firms should meet existing requirements and does not propose new rules. Responses close on 26 February 2027; the proposals should be assessed within each firm’s applicable supervisory scope. [9]

The separate AI cyber practices emphasise continuous exposure monitoring, prioritisation and the speed of remediation. They also bring leadership decision-making into view. In a bank, a team may know how to deploy a fix yet still wait for authority to interrupt a customer service. A useful rehearsal measures the time needed to recognise the exposure, decide on containment, coordinate suppliers, implement the change and verify recovery. [10]

7. Cryptographic change is becoming a resilience planning issue

Quantum preparedness adds a longer planning horizon to the same dependency problem. The UK’s 19 November effective-practices paper states that no cryptographically relevant quantum computer is currently known to exist. Preparation nevertheless takes time, and information intercepted today may remain sensitive long enough to face a future decryption risk. The paper is non-binding, sits outside formal supervisory expectations and establishes no universal migration deadline. [11]

Its suggested practices include identifying cryptographic assets and dependencies, assessing supplier readiness, prioritising migration and testing changes without weakening operational resilience. A bank needs to understand where encryption, authentication and digital signatures protect its services before it can plan how those protections will change. [11]

Our architectural implication is to include cryptographic dependencies in service maps and procurement discussions. A core application, identity service, API gateway and external payment connection may change on different schedules. Planning should expose those constraints early and connect technical migration with customer-service continuity.

An illustrative stress event across several teams

Consider a fictional bank whose AI assistant helps a small business prepare supplier payments. During market volatility, a shared vendor becomes intermittent. Payment confirmations are uncertain, treasury needs the outgoing obligations total, and a convincing caller requests an urgent beneficiary change. The customer also finds a purported refund-support page while searching for help.

This is a fictional scenario, not a report of an incident. It brings the challenges together: delegated authority, identity checks, third-party dependency, transaction status, customer communication and liquidity information. Separate teams can follow their own procedures and still leave gaps at the handovers.

The response needs one agreed set of facts: preserve payment references, verify outcomes, independently validate the beneficiary request, restrict uncertain automation and direct the customer to verified support. Treasury needs a reconciled obligations view. If containment requires an interruption, a named decision-maker must weigh and authorise it promptly. Recovery is complete when the bank can demonstrate trustworthy outcomes and support affected customers.

A practical agenda for banking leadership

ChallengeEvidence to ask for
Credit, capital and funding under stressLinked scenarios showing how borrower deterioration, market moves and operational disruption affect the institution.
AI and delegated decisionsCurrent inventories, action limits, validation, stop and handover controls, monitoring and named accountable owners.
Fraud and identityEvidence that onboarding, beneficiary changes and high-risk requests are assessed across trusted independent signals, with verified customer-support routes.
Supplier concentrationA critical-service dependency map and a demonstrated fallback or exit plan with realistic constraints.
Payment and service recoveryAuthoritative status checks, duplicate prevention, reconciliation and tested customer-support procedures.
Risk data and management reportingTimely, traceable figures with consistent definitions and controlled explanations for unresolved differences.
Cyber response decisionsA timed exercise showing who authorises containment, how suppliers coordinate and how recovery is verified.
Cryptographic preparednessAn inventory of critical cryptographic dependencies, supplier readiness and a prioritised transition plan.

These are proposed management questions, not a single global compliance checklist. Requirements differ by jurisdiction and institution. Their value is to connect risk ownership to the evidence needed to keep a bank functioning.

The shared message is disciplined modernisation: understand where automation acts, where dependencies concentrate, how financial stress travels and what customers experience when a service fails. The quality of those connections will influence how well a bank can turn innovation into sustained trust.

Learning companion: several alerts, one banking service

Practise the decisions in our interactive Decision Lab: The systems are back. Is the banking service safe? Explore the concepts, test your response and earn the Connected Banking Resilience Steward badge.

Decision exercise. An AI-assisted payment remains unresolved during a supplier outage. An impersonated customer requests a beneficiary change, a fake refund-support page appears, and treasury needs an accurate obligations total. Containment may briefly interrupt a service. Which response best protects the banking service?

  1. Retry every payment through another channel and accept the caller’s instructions to clear the queue.
  2. Leave the incident entirely with the technology vendor because the bank’s core is running.
  3. Coordinate operations, fraud and treasury: verify payment status, independently validate the request, limit uncertain automation and reconcile the obligations view.
Reveal the preferred response

Option 3. This is an end-to-end banking service problem. A vendor outage does not remove the bank’s responsibility to understand payment outcomes, protect customer instructions and maintain reliable financial information.

Reflection: Who can authorise containment, how will customers find verified help, and what evidence confirms safe recovery? Identify one critical service and bring its operational, customer and financial owners together to answer these questions.

Sources and interpretation

Sources reviewed as at 20 November 2026. This article brings together policy updates, supervisory priorities, a consultation, non-binding practices, a survey, a financial-stability assessment, guidelines, directions, a speech and a scam alert. The FinCEN document is historical background from 2024. The scenarios and management recommendations are the author’s interpretation. Applicability depends on jurisdiction and institution.

  1. RBA — Payments System Board Update, 19 November 2026
  2. Bank of England/FCA — Artificial intelligence in UK financial services, 26 October 2026; methodology and results
  3. MAS — Responsible AI supervisory expectations and implementation dates, 7 October 2026
  4. ECB Banking Supervision — Supervisory priorities 2027–29, November 2026
  5. Federal Reserve — Financial Stability Report, 13 November 2026
  6. RBI — Commercial Banks Cybersecurity Technology Risk Resilience and Assurance Framework Directions, 2026; updated 1 October 2026
  7. RBI — From Digital Banking to Resilient Banking, speech published 25 September 2026
  8. FinCEN — Alert on Deepfake Media Targeting Financial Institutions, 13 November 2024
  9. UK PRA/Bank of England/FCA — CP17/26 and GC26/9, 19 November 2026; consultation closes 26 February 2027
  10. UK regulators — Effective practices: AI-related cyber risk and resilience, 19 November 2026; non-binding
  11. UK regulators — Effective practices: quantum-related cyber risk and resilience, 19 November 2026; non-binding
  12. HKICL alert carried by HKMA — Fraudulent website imitating FPS support, 20 November 2026
Learn the topic

Put this perspective into practice.

Explore the concepts, make a decision and test what changes when the situation changes.

Leave a Reply

Your email address will not be published. Required fields are marked *

An invitation to contribute

Everyone has a perspective worth hearing.

A lesson from your profession. An experience that changed your mind. A question you think deserves a better conversation.

Share your perspective