C:\>DIR Global ‘Agentic Regulator’ Hackathon
Agentic Financial & High-Stakes AI Advice
Mitigate risks to consumers using autonomous LLMs & agentic advisors for high-stakes decisions within financial services and the wider digital economy, including model biases, hallucinations, lack of explainability and compliance breaches.
Context for the Problem Space
Agentic advice systems produce seemingly fluent, personalised outputs but frequently fail to recognise their own limits or adhere to regulatory or compliance "deny lists".
Key risks include but are not limited to:
Guessing missing client details instead of gathering required facts
Giving premature advice rather than pausing to ask questions
Compliance checks that silently fail during real conversations
Unpredictable behavior triggered by minor changes in user phrasing
Made-up AI reasoning that hides how a decision was actually reached
Broken audit trails making it impossible to prove why advice was given
Potential Solution Areas
Navigating the Regulatory Perimeter
How can authorities monitor high-stakes AI advice when the underlying models operate outside their traditional jurisdictions?
Verifying Authentic Explainability
How can regulators technically validate that an AI's explanation reflects its actual logic, rather than fabricated reasoning?
Testing for Safe Refusals
How can authorities audit AI models to ensure they recognise their limits, gather necessary context, and safely refuse to provide regulated advice?
Illustrative Hackathon Prototypes
Advice Market Radar
An automated scanner that maps where consumers receive AI advice for high-stakes decisions across digital channels, sizing public exposure both inside and outside the regulatory perimeter.
High-Stakes Advice Red-Teaming Engine
Adversarial agents that stress-test advice APIs using synthetic personas to expose hallucinations, policy breaches, and a lack of safe refusal behaviours.
Cross-Market Agentic Advice Benchmark
An open evaluation framework to test both regulated and unregulated AI advice systems against a shared scenario library, measuring suitability, bias, and safety.
Consumer Explainability Translator
An Explainable AI (XAI) tool that translates complex "black box" decisions into plain language and intuitive visuals, providing consumers with clear reasons for outcomes like credit rejections or fraud alerts.
Supervisory XAI Toolkit
A privacy-preserving platform (inspired by BIS Project Noor) that equips regulators to independently audit AI models, translating complex logic into practical metrics to actively assess fairness and robustness without relying solely on a firm's own disclosures.
Reference & Resources
FCA (2026), Review into the long-term impact of AI on retail financial services (the Mills Review) - the UK's flagship inquiry into AI and retail advice
EY (2026) https://www.ey.com/en_gl/newsroom/2026/04/nearly-half-of-global-consumers-now-use-ai-to-guide-savings-and-investment-decisions - 49% of 18,000 consumers ;23 countries
BIS, Project Noor https://www.bis.org/about/bisih/topics/suptech_regtech/noor.htm
OpenAI, A new personal finance experience in ChatGPT - 200 million people per month ask ChatGPT with questions about their financial matters.
Lending Standards Board (2025), research on flaws in financial services chatbots - consumer outcomes and FCA bias research
FCA Handbook, COBS 9A.2 (Assessing suitability: the obligations) and COBS 9.2 — the duty to obtain information "necessary … to understand the essential facts about the client," and the rule that a firm "must not recommend … where none of the services or instruments are suitable." Implements Art. 25(2) MiFID II. https://www.handbook.fca.org.uk/handbook/COBS/9A/2.html
FCA Handbook, COBS 9.4 (Suitability reports), esp. COBS 9.4.7R — the report must "explain why the firm has concluded that the recommended transaction is suitable for the client." FCA supervisory commentary stresses that "explain" requires demonstrating the connection between the client's circumstances and the recommendation, not a conclusory statement. https://www.handbook.fca.org.uk/handbook/COBS/9/4.html
FCA, Consumer Duty (PRIN 2A / Principle 12, in force 31 July 2023) — consumer‑understanding outcome and the obligation to evidence good outcomes.
FCA, Senior Managers & Certification Regime — named individual accountability (e.g. SMF16, Compliance Oversight) for the adequacy of the suitability and oversight process. Both regimes presuppose an inspectable, attributable reasoning process.
Sclar, Choi, Tsvetkov & Suhr (2024), Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design, ICLR 2024 — performance swings up to 76 points from meaning‑preserving prompt‑format changes; persists with scale, more examples and instruction tuning. arXiv:2310.11324. https://arxiv.org/abs/2310.11324
Turpin, Michael, Perez & Bowman (2023), Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain‑of‑Thought Prompting, NeurIPS 2023 — CoT explanations systematically misrepresent the true cause of a model's output. arXiv:2305.04388. https://arxiv.org/abs/2305.04388
Chen et al. / Anthropic (2025), Reasoning Models Don't Always Say What They Think — reasoning models verbalise a decisive hint only ~25% (Claude 3.7 Sonnet) / ~39% (DeepSeek R1) of the time; faithfulness plateaus at ~20–28% under outcome‑based RL. https://www.anthropic.com/research/reasoning-models-dont-say-think
Lanham et al. (2023), Measuring Faithfulness in Chain‑of‑Thought Reasoning, Anthropic — methodological companion on CoT faithfulness.
Contact
For additional questions, please contact: contact@cdir.global