Sponsored by

The Accountability Record

Hey {{first_name | there}},

If Microsoft's outage gave you an unexpected head start on the weekend, spare a thought for travellers in the US. While Outlook, Teams and Copilot were struggling to stay online, a separate FAA systems failure grounded flights across America on Friday.

Two unrelated incidents, but a timely reminder of just how much we depend on systems we don't fully control.

And speaking of control, this Friday, 20 November, new EU consumer credit rules take effect. Consumers whose creditworthiness is assessed using automated processing will have the right to request human intervention and a review of the assessment and lending decision.

The question for compliance leaders is easy to ask and harder to prove: can your human reviewer genuinely overturn the model's decision, and can you demonstrate that in your records?

Welcome to this week's Accountability Record.

In this week's Accountability Record:

  • 🛠️ Practitioner's corner: test whether your human reviewers can really overturn the model

  • 🇪🇺 🔴 Act now: Automated credit decisions now need a human who can overturn them

  • 🇪🇺 🟠 Review this quarter: The ECB is now grading banks' frontier-AI action plans

  • 🇦🇺 🟢 Monitor: APRA wants firms to demonstrate AI control, not describe it

🛠️ Practitioner's corner: is your human review real?

  1. Map the decision points. Ask the head of credit risk to list every point where automated processing feeds a consumer credit decision, including vendor scorecards.

  2. Run a live sample. Have the operations lead send 20 recent automated declines to the named reviewer as if each customer had asked for human intervention.

  3. Record what the reviewer saw. The reviewer notes which variables they could see, time spent, and whether they had authority to reverse the outcome.

  4. Test the pattern. The second-line control reviewer checks override rates and review times. Zero overrides at under a minute each is a rubber stamp.

  5. Sign the register. The accountable executive signs a human review register: decision point, reviewer role, authority to overturn, explanation template, test date and result.

Usage-Based Pricing Is Here. Is Your Finance Team Ready?

More B2B companies are moving to usage-based and hybrid pricing — and finance teams are feeling it. Revenue recognition gets messier, forecasting gets harder, and the manual work compounds fast.

Tabs and PwC teamed up to break it down. In this on-demand session, Rebecca Schwartz and Amit Dhir share how leading finance teams are handling the operational reality of dynamic pricing models — and where AI fits in.

Watch the recording for concrete examples, a practical rev rec framework, and a clear-eyed look at what it takes to scale without the overhead.

If your team is navigating this shift, this session is worth your time.

🇪🇺 EU | Consumer credit | Final rule | 🔴 Act now

Automated credit decisions now need a human who can overturn them

  • What happened: The revised Consumer Credit Directive (EU) 2023/2225 applies from 20 November, giving consumers a right to human intervention whenever a creditworthiness assessment involves automated processing.

  • The details:

    • Article 18: consumers can obtain human intervention, a clear and comprehensible explanation of the assessment, give their view and request a review of the decision.

    • Scope widens to third-party buy now, pay later. Assessments must not use special category data, such as health, or social network information.

    • National laws set the detail and penalties.

  • The decision: Decide which automated credit decisions a person can genuinely reverse, who that person is, and what they must see to do it.

  • The owner: the management body member accountable for retail lending, with the head of credit risk owning the model logic and compliance testing that reviews are genuine.

  • The evidence: A human-intervention log for every request (date, named reviewer, information seen, outcome and reason), plus the approved consumer notice and explanation templates.

  • Across regimes: UK GDPR Article 22 gives a comparable right for solely automated decisions; US ECOA and Regulation B require specific adverse action reasons however complex the model. CCD2 is EU-only and triggers on any automated processing.

  • Also for: Head of model risk: can the model produce the main variables behind each decline? COO: is the review queue staffed for national response times?

🇪🇺 EU | Banking | Speech | 🟠 Review this quarter

The ECB is now grading banks' frontier-AI action plans

  • What happened: On 9 November, ECB supervisory chair Claudia Buch told the Eurogroup the ECB is assessing banks' frontier-AI action plans.

  • The details:

    • Plans were due by end-October, following the 7 July Dear CEO letter on AI-enabled cyber threats. The ECB will "identify areas for improvement" and share good practice.

    • Frontier models make it "quicker and cheaper" to find and exploit IT vulnerabilities. More than 90% of supervised banks use AI tools.

  • The decision: Decide whether the plan you submitted is being delivered on its own dates, before feedback arrives.

  • The owner: the management body member responsible for ICT risk, with the CISO delivering and internal audit testing.

  • The evidence: An action plan tracker (measure, owner, due date, status) and the board minutes that reviewed it.

  • Across regimes: DORA's ICT risk framework applies EU-wide; UK operational resilience rules and APRA CPS 230 are the nearest equivalents. The action plan request covers only ECB-supervised banks.

🇦🇺 Australia | Banking, insurance, superannuation | Research | 🟢 Monitor

APRA wants firms to demonstrate AI control, not describe it

Working Work Out GIF by Northern Ireland

Gif by northernireland on Giphy

  • What happened: APRA's System Risk Outlook, published 12 November, puts AI at the centre of its risk assessment and promises tougher supervision.

  • The details:

    • Four AI risk channels: use within firms, use against them, shared dependencies, economy-wide effects. Supervision will test whether firms "can demonstrate" risks are managed.

    • It cites a June incident in which an OpenAI agent circumvented controls to access an Australian Government website.

    • 80 of the top 100 material service providers are based overseas. The 2027 stress test will model AI-driven disruption.

  • The decision: Decide which agents could act outside approved scope, and what stops them first.

  • The owner: the accountable executive under the Financial Accountability Regime for the business line using the agent, with the CRO owning the risk framework.

  • The evidence: An agent permission register listing each agent's authorised actions, the preventive control, and when it last blocked an out-of-scope test.

  • Across regimes: FAR's named accountability mirrors UK SM&CR; CPS 230 maps to DORA and UK operational resilience. The AI stress test is APRA's own.

🇬🇧 UK Desk

AI, risk and accountability across UK financial services this week.

🟢 FPC member: give every AI agent least privilege and its own identity

In his final FPC speech on 11 November, Jon Hall said high-stakes agents should be sandboxed under least privilege, each with a unique identity enabling an audit trail and revocation. A Bank and FCA survey found only 2% of agents have broad autonomy. Bank of England

Evidence to have ready: an agent identity register mapping each agent ID to its permissions, its accountable owner and any revocation.

🟢 Bailey: public trust depends on assured human intervention

On 9 November the Governor called for more extensive model testing before and after release, and better AI incident reporting, monitoring and follow up. Bank of England

Evidence to have ready: an AI incident log recording who intervened, how quickly, and what changed afterwards.

Lender lens: motor finance "not owed" letters due by 18 November

No AI lender item qualified. Lenders should tell customers who complained by 30 June about post-April 2014 agreements, and are not owed compensation, by 18 November. Final hearing dates for the legal challenges are not yet fixed. FCA

Evidence to have ready: if automated rules screened eligibility, a person-checked sample of "not owed" outcomes and the rule version.

Coming up

  • 20 November: ICO agentic AI call for evidence closes ICO

  • 18 March 2027: FCA operational incident and third party reporting rules come into force FCA

⚡ Quick hits

Americas

  • 🇺🇸 [Banking] 🔴 Comments on the interagency third-party risk management proposal are due today. Federal Reserve

Asia-Pacific & Middle East

  • 🇸🇬 [Payments] 🟢 The Singapore FinTech Festival (18 to 20 November) puts agentic commerce and AI-enabled payments on its agenda. MAS

  • 🇰🇷 [Fintech] 🟢 Korea's FSC has set its 25 to 27 November Fintech Week around AI agents and digital assets. FSC

  • 🇦🇺 [Insurance] 🟢 APRA says data-centre concentration is stretching insurers' capacity as AI enters underwriting and claims. APRA

🧰 RegTech & security watch

  • Wells Fargo's updated online access terms take effect 24 November, saying outside AI tools are "not operated by Wells Fargo" and access may be stopped. Consequence: banks allowing customer agents need a documented position on who bears losses. Wells Fargo

  • Feedzai expanded its Digital Trust product on 12 November to flag whether a human or an AI agent is behind a session. Consequence: once agents are visible, allowing them becomes a permission decision with a named owner. PR Newswire

📅 Calendar

  • 20 November | EU Consumer Credit Directive (EU) 2023/2225 applies | 🔴

  • 30 November | Korea FSC checks on external access points at financial firms due | 🟢

📊 Pulse check

When a person reviews an automated decision at your firm, can they overturn it?

Login or Subscribe to participate

Results in next week's issue.

Facing this decision at your firm? Hit reply and tell me how you are handling it.

From the editor's day job

Regulatory updates and consultation papers can have a direct impact on the authority delegated to AI agents.

BuildMoat connects regulatory developments to an organisation's internal policies, controls and affected AI agents, identifying where delegated authority may need to change. Through a simple, auditable workflow, accountable Senior Managers (SMFs) can review the impact, understand the rationale, approve changes to an agent's authority and demonstrate reasonable steps and effective oversight. BuildMoat connects regulatory change to enforceable controls at runtime, turning accountability from a paper exercise into demonstrable action.

Learn more: buildmoat.co.uk

Regulatory commentary is for information only and is not legal, compliance or financial advice. Any product mention is disclosed.

That’s it for the week!

Spotted something wrong, or something we missed? Hit reply. I read every one.

Know someone accountable for AI at their firm? Forward this to them.
Was this forwarded to you? Subscribe here. It’s free, every Monday.

Emeka Ogbonnaya

Editor, The Accountability Record
Former Bank of England · Founder, BuildMoat
LinkedIn · X