Brussels Delays, the Supervisors Inspect, the Agents Move In

Highlight

A Deferral With Strings Attached

Three movements shaped the past two weeks, and they pull in opposite directions. Brussels is deferring the EU AI Act's high-risk obligations for credit scoring to December 2027. In the same breath, the supervisors are getting concrete: Frank Elderson (ECB) spoke on 3 June about strengthening operational resilience "for the age of AI", and BaFin is announcing targeted on-site inspections in response to substantial AI risks.

And while regulators push back deadlines and sharpen their inspection plans, the agents are moving into the core system. Morgan Stanley is opening a 1.2-trillion-dollar wealth platform to external AI agents – not for humans, for machines. OpenAI is bringing Codex into every role; Microsoft is reducing its dependence on OpenAI with its own models. The market is not waiting for the next deadline.

That is the through-line of this issue: the deferral is not a free pass. It moves a deadline, not the responsibility. Read the 16 months as a pause and you lose them. Use them as orderly preparation and you walk into the obligations ready – precisely when the technology is moving to the centre of the bank.

LinkedIn Featured

EU AI Act: What the Deferral to December 2027 Really Means

The high-risk obligations for AI in creditworthiness assessment are being deferred by 16 months. That gives banks and insurers breathing room – but the deferral presupposes the stocktaking, it does not suspend it. Three obligations apply unchanged from August 2026, and the 7 May agreement is not even legally binding yet. Anyone who reads the provisional understanding as a "do nothing for now" is optimising for the wrong variable.

Read the full article: What the Deferral to December 2027 Really Means →

Agentic AI

tools, skills & what's trending

OpenAI News: Codex in Every Role – and on AWS

OpenAI has expanded Codex from a coding tool into a productivity platform for knowledge workers: new role-specific plugins for analysts, investors and sales, plus "Sites" for interactive workspaces. At the same time, the frontier models and Codex are now generally available via Amazon Web Services (AWS) – with the procurement and compliance controls banks need for internal approval. The explicit investment-banking plugin signals that OpenAI has identified financial institutions as a primary target group.

Microsoft Build: Own Models Against the OpenAI Dependence

At Build 2026, Microsoft announced its own generative models (the MAI family) as a lower-cost alternative to OpenAI – flanked by an agent strategy built around the open NLWeb protocol, which makes structured websites directly queryable for AI agents (conceptually an HTTP for the agent era). Even OpenAI's largest investor is diversifying. For banks, the signal is clear: enterprise procurement must be designed for model portability and vendor diversification.

Claude Update: Anthropic Opus 4.8 With Dynamic Workflows

Claude Opus 4.8 introduces "Dynamic Workflows": orchestrated multi-agent systems with up to 1,000 subagents per run, plus a lower-cost Fast Mode. Still a research preview, but a foundation for scalable agentic orchestration – for example in compliance analysis or trade surveillance, where parallel subagents process different data sources simultaneously. Anthropic's first native step towards near-production orchestration.

Banking & Regulation

what actually matters now

Agents Move Into the Core System: Morgan Stanley Opens Its Wealth Platform

Morgan Stanley is connecting external AI agents from corporate clients directly to its equity administration platforms ShareWorks and Equity Edge – without a human login; 3,400 administration clients are to follow by the end of 2026. CNBC notes that JPMorgan and Goldman Sachs have so far used agents only internally, for instance in code development. The move shifts the interface from human to machine – and with it the question of who owns Know-Your-Customer (KYC), Anti-Money-Laundering (AML) and liability when no one logs in any more.

The Supervisors Get Concrete on AI Resilience

Frank Elderson (ECB) spoke on 3 June about strengthening operational resilience "for the age of AI"; Isabel Schnabel drew lessons from money market funds to stablecoins on 1 June. In Germany, BaFin President Mark Branson announced targeted on-site inspections in response to "substantial" AI risks and published a tightened consultation draft of the minimum risk-management requirements for investment firms (WpI-MaRisk). The tone has turned: from guidance to inspection practice. Anyone using AI in scoring, surveillance or reporting should document their model governance in an audit-proof way now.

Signal & Noise

what's worth your time

• Gradient Labs raises 26M USD for banking AI agents – FinTech Global
  CommerzVentures is investing in regulated agentic infrastructure (lending, disputes and KYB agents with FCA and AI Act guardrails). 900 per cent revenue growth in a year – the niche is becoming a market.
• AlphaSense: 350M USD at a 7.5B valuation – AlphaSense
  J.P. Morgan Asset Management and Accenture Ventures as lead investors, Goldman Sachs Alternatives as an existing investor – leading banks are buying into agentic market intelligence, not just using it.
• Basel Committee publishes report on ICT risk management – Bank for International Settlements (BIS)
  A range-of-practices report on information and communication technology risk as a core component of operational risk. The bridge between DORA logic and Basel supervision is becoming visible.
• BoE and FCA set out a shared vision for tokenisation in UK wholesale markets – Bank of England
  A Call for Input (18 May) for DLT-based settlement – the British counterpart to the EU DLT pilot regime, with more legal certainty for tokenised assets.
• Michael Burry says neither SpaceX nor Anthropic is worth 1 trillion USD – Business Insider
  The "Big Short" investor warns of the AI valuation bubble. A useful corrective amid all the up-rounds in this issue – scepticism is part of due diligence.

"The real question is not whether machines think but whether men do."