AI Risks in Financial Services
AI Risks in Financial Services
A synthesis of reports from ESMA, IOSCO, ECB, FCA, FSB and related authorities — April 2026
Executive Summary
The deployment of artificial intelligence across financial services has accelerated dramatically. As of 2024–2025, over 75% of UK financial services firms are already using AI, and global regulators have moved rapidly to identify, categorise, and address the risks it introduces. This page synthesises findings from the most authoritative regulatory and institutional sources to provide a structured overview of the principal risk categories that AI poses to the financial system.
The picture that emerges is one of significant promise accompanied by equally significant risk. Regulators are broadly pro-innovation, but consistently emphasise that AI adoption must be matched by proportionate governance, transparency, and accountability. The risks are not hypothetical: several are already materialising in credit decisions, trading systems, fraud detection, and customer advisory tools.
Six Risk Categories Dominate the Regulatory Agenda
- Model & Data Risks: bias, opacity, hallucination, and data quality failures
- Operational Risks: resilience failures, third-party concentration, and vendor lock-in
- Cybersecurity & Financial Crime: AI-enhanced attacks, deepfakes, and market manipulation
- Systemic & Financial Stability Risks: herding behaviour, feedback loops, and AI infrastructure concentration
- Consumer Protection & Conduct Risks: discrimination, reduced consumer agency, and erosion of trust
- Governance & Accountability Risks: accountability gaps, skills shortages, and regulatory fragmentation
Key Findings by Source
ESMA — European Securities and Markets Authority
ESMA's May 2024 statement on AI in investment services established the EU-level benchmark for how firms must apply MiFID II obligations when using AI tools. ESMA identified risks including algorithmic bias, opacity, reliance on third-party AI providers, and 'AI washing' in fund documentation. Firms are expected to maintain comprehensive records of AI deployment, decision-making logic, and data sources. ESMA's February 2025 TRV article documented specific instances of AI washing in EU mutual fund documentation.
IOSCO — International Organization of Securities Commissions
IOSCO's March 2025 Consultation Report (CR/01/2025) is the most comprehensive global survey of AI risks in capital markets. Based on surveys of 24 member jurisdictions, IOSCO identified four primary risk clusters: (1) malicious uses of AI, (2) model and data considerations, (3) concentration and third-party dependency, and (4) systemic interactions and financial stability. A key quantitative finding: 40% of machine learning models in capital markets institutions are implemented through vendor tools and cloud services.
ECB — European Central Bank
The ECB's May 2024 Financial Stability Review article identified three interconnected systemic risks: operational risk amplification if AI suppliers are concentrated; increased herding behaviour and market correlation; and 'too-big-to-fail' externalities if AI infrastructure becomes dominated by a small number of providers. The ECB stressed that the overall impact on financial stability will depend critically on how data quality, model development, and deployment challenges are addressed.
FCA — Financial Conduct Authority
The FCA has the most extensive and evolving AI risk framework among national regulators. Its January 2025 research note documented bias risks in NLP and credit scoring. The January 2026 Mills Review—a long-term review of AI's impact on retail financial services—identified bias, opacity, reduced consumer agency, AI-enabled fraud (including deepfakes), and accountability gaps as the defining challenges. The UK Treasury Committee (2025) found that over 75% of UK firms use AI and warned that regulators are not doing enough to manage the risks.
FSB — Financial Stability Board
The FSB's October 2025 report on monitoring AI vulnerabilities identified ongoing challenges including third-party dependencies, market correlations, cyber risks, and model risk governance. The FSB also highlighted specific attack vectors including data and model poisoning and prompt injection as emerging threats to financial AI systems.
Risk Register: 27 AI Risks in Financial Services
The table below provides a structured register of the principal AI risks identified across regulatory reports, organised by category. Sources are cited for each risk.
| Risk | Category | Description | Source |
|---|---|---|---|
| Algorithmic Bias | Model & Data Risk | AI models trained on historical data can embed and amplify biases, leading to discriminatory outcomes in credit scoring, insurance pricing, and lending. Disproportionately harms protected or vulnerable groups. | FCA (Jan 2025); ESMA (May 2024); IOSCO (Mar 2025) |
| Model Opacity / Lack of Explainability | Model & Data Risk | Many AI models are 'black boxes' whose decision logic cannot be explained to regulators, consumers, or the firms deploying them, undermining accountability. | FCA Mills Review (Jan 2026); ESMA (May 2024); IOSCO (Mar 2025); ECB FSR (May 2024) |
| Data Quality & Poisoning | Model & Data Risk | AI systems are vulnerable to poor-quality or manipulated training data. Attackers can corrupt training data or model weights to cause systematic errors or create backdoor vulnerabilities. | IOSCO (Mar 2025); FSB (Oct 2025); ECB FSR (May 2024) |
| Hallucination & Model Errors | Model & Data Risk | Generative AI models can produce plausible but factually incorrect outputs ('hallucinations'), leading to erroneous financial advice or flawed risk assessments if not overseen. | FCA Mills Review (Jan 2026); IOSCO (Mar 2025) |
| AI Washing / Misleading Claims | Model & Data Risk | Firms may overstate AI capabilities of their products or investment funds, misleading investors. ESMA has documented instances of 'AI washing' in EU fund documentation. | ESMA (Feb 2025, TRV Article) |
| Operational Resilience Failures | Operational Risk | Heavy reliance on AI creates new single points of failure. AI-driven outages can disrupt critical financial services at scale, with cascading effects across institutions. | ECB FSR (May 2024); Bank of England FPC (Apr 2025) |
| Concentration Risk – Third-Party AI Providers | Operational Risk | A small number of cloud and AI providers supply tools to most of the financial sector. Disruption at one provider could simultaneously affect thousands of firms. | IOSCO (Mar 2025); FCA (2025); FSB (2024); ECB FSR (May 2024) |
| Nth-Party / Supply Chain Risk | Operational Risk | AI vendors depend on sub-contractors, creating layered dependency structures that firms cannot fully observe. 40% of ML models in capital markets are implemented via vendor tools. | IOSCO (Mar 2025, citing BoE/FCA survey) |
| Vendor Lock-In | Operational Risk | Over-reliance on a single AI or cloud provider can make migration prohibitively costly, reducing firms' ability to switch suppliers and increasing systemic dependency. | FCA (2025 AI Update); IOSCO (Mar 2025) |
| AI-Enhanced Cyber Attacks | Cybersecurity & Financial Crime | Malicious actors use AI to automate and sophisticate cyberattacks including phishing, social engineering, and vulnerability exploitation, expanding the attack surface for financial institutions. | ECB FSR (May 2024); BIS Project Raven (Apr 2024); FSB (Oct 2025) |
| Deepfakes & Synthetic Identity Fraud | Cybersecurity & Financial Crime | AI-generated deepfakes and synthetic identities can defeat customer onboarding controls, enable impersonation fraud, and facilitate money laundering via automated criminal ecosystems. | FCA Mills Review (Jan 2026); IOSCO (Mar 2025) |
| Prompt Injection Attacks | Cybersecurity & Financial Crime | Attackers can manipulate inputs to GenAI/LLM tools to extract confidential data, override safety controls, or trigger unintended actions—posing specific risks for customer-facing applications. | FSB (Oct 2025); BIS (Dec 2024) |
| AI-Enabled Market Manipulation | Cybersecurity & Financial Crime | AI tools could facilitate sophisticated manipulation including cross-market strategies and spoofing that are harder to detect than traditional methods. | FCA (2025 AI Update, Market Abuse TechSprint); IOSCO (Mar 2025) |
| Herding Behaviour & Market Correlation | Systemic & Financial Stability Risk | If multiple institutions use similar AI models, their behaviour may become highly correlated, amplifying market movements and creating synchronised booms and busts. | ECB FSR (May 2024); FSB (2024); IOSCO (Mar 2025) |
| Too-Big-To-Fail AI Infrastructure | Systemic & Financial Stability Risk | Concentration among AI providers could create 'too-big-to-fail' dynamics in technology infrastructure, with systemic externalities if a dominant platform fails. | ECB FSR (May 2024); FSB (Oct 2025) |
| Market Feedback Loops & Flash Events | Systemic & Financial Stability Risk | Automated AI systems interacting in real time could generate self-reinforcing feedback loops, increasing the risk of flash crashes and extreme market volatility episodes. | IOSCO (Mar 2025); FSB (2024); Bank of England FPC (Apr 2025) |
| AI Collusion Risk | Systemic & Financial Stability Risk | Multiple AI agents optimising independently could develop emergent coordination or 'scheming' behaviour without explicit programming, raising competition and stability concerns. | IOSCO (Mar 2025); CFA Institute (Apr 2025) |
| Discriminatory Outcomes & Consumer Exclusion | Consumer Protection & Conduct Risk | AI-driven decisions in credit, insurance, and advice may systematically disadvantage consumer groups, particularly those from protected characteristics or with limited data histories. | FCA (Jan 2025; Mills Review Jan 2026); ESMA (May 2024); EU AI Act (2024) |
| Reduced Consumer Agency & Autonomy | Consumer Protection & Conduct Risk | As consumers delegate more financial decisions to AI agents, they may lose understanding of and control over their financial lives, with risks of unconscious manipulation. | FCA Mills Review (Jan 2026); FCA (2025 AI Update) |
| AI-Driven Mis-selling & Unsuitable Advice | Consumer Protection & Conduct Risk | AI systems optimising for proxies rather than genuine outcomes may recommend unsuitable products or misleading advice, particularly in autonomous advisory applications. | ESMA (May 2024, MiFID II); FCA Mills Review (Jan 2026); IOSCO (Mar 2025) |
| Erosion of Consumer Trust | Consumer Protection & Conduct Risk | Opaque AI decisions and high-profile failures risk undermining consumer confidence in financial services, with broader implications for market participation. | FCA (Mills Review Jan 2026); UK Treasury Committee (2025) |
| Accountability Gaps | Governance & Accountability Risk | Diffuse AI supply chains make it difficult to assign clear responsibility when AI-driven decisions cause harm. Existing frameworks (e.g. SM&CR) may not adequately capture AI accountability. | FCA (Mills Review Jan 2026; 2025 AI Update); UK Treasury Committee (2025) |
| Inadequate Model Governance | Governance & Accountability Risk | Many firms lack robust processes for validating, monitoring, and auditing AI models in production. Models may update continuously in ways that are difficult to audit. | FCA (2025 AI Update); ESMA (May 2024); IOSCO (Mar 2025) |
| Regulatory Arbitrage & Fragmentation | Governance & Accountability Risk | Divergent AI regulatory frameworks across jurisdictions create inconsistent standards and arbitrage opportunities, complicating compliance for cross-border firms. | IOSCO (Mar 2025); European Parliament ECON (Nov 2025); BIS (Dec 2024) |
| Skills Gap | Governance & Accountability Risk | Institutions and supervisors face a shortage of professionals combining financial and AI/ML expertise, limiting their ability to safely develop, deploy, and oversee AI systems. | CFA Institute (Apr 2025); BIS (Apr 2025); FCA (2025 AI Update) |
| Data Privacy & GDPR Violations | Data Privacy & Ethics Risk | AI systems processing large volumes of personal financial data create heightened risks of data breaches, unlawful processing, and non-compliance with data protection regulations including UK GDPR. | FCA (2025 AI Update); ESMA (May 2024) |
| Environmental & ESG Risks | Data Privacy & Ethics Risk | Training and running large AI models requires significant energy consumption, contributing to environmental risks and raising ESG concerns about AI-intensive institutions. | IOSCO (Mar 2025); ICMA Response to IOSCO (Apr 2025) |
Conclusions & Regulatory Direction
The regulatory consensus across ESMA, IOSCO, ECB, FCA, and the FSB points to a consistent set of priorities for responsible AI deployment in financial services:
- Explainability and transparency: firms must be able to articulate how AI models reach decisions, particularly for high-stakes applications in lending, insurance, and fraud detection.
- Bias and fairness testing: bias audits must be embedded into model governance, with particular attention to consumer-facing applications that could disadvantage protected or vulnerable groups.
- Third-party oversight: firms must strengthen vendor due diligence, contractual safeguards, and exit planning to manage concentration and supply chain risks.
- Accountability under existing frameworks: SM&CR, Consumer Duty, MiFID II, and DORA collectively provide the principal accountability architecture.
- Systemic monitoring: regulators are developing new frameworks for monitoring AI-driven market correlations, concentration in AI infrastructure, and feedback loop dynamics.
The direction of travel is clear: AI adoption in financial services will continue at pace, but the regulatory bar for governance, explainability, and consumer protection is rising in parallel. Firms that treat AI governance as a strategic capability—rather than a compliance burden—will be best positioned to benefit from AI's potential while managing its risks.
Sources: ESMA (May 2024, Feb 2025), IOSCO CR/01/2025 (Mar 2025), ECB Financial Stability Review (May 2024), FCA AI Update (2025), FCA Mills Review (Jan 2026), FSB AI Reports (2024, Oct 2025), Bank of England FPC (Apr 2025), UK Treasury Committee (2025), CFA Institute (Apr 2025), BIS (Dec 2024, Apr 2025).