Banks act, Brussels negotiates

Highlight

Banks act. Brussels negotiates.

Five days after this newsletter lands, Parliament, Council and Commission will wrestle in Brussels with the most consequential lever of the European Union Artificial Intelligence Act (EU AI Act) since it entered into force: will the high-risk deadline of 2 August 2026 be postponed to December 2027? Political agreement in the trilogue is scheduled for 28 April – until then, the original deadline remains operationally in force.

While Europe still debates whether it can meet its own deadlines, banks worldwide have already moved: Goldman Sachs is building autonomous agents with Anthropic for trade accounting, Know Your Customer (KYC) and trading surveillance. Piraeus and Accenture have launched an Anthropic-powered AI Hub for Greece's largest bank, CIMB Niaga has put Google Cloud agents into production in Indonesia – and HSBC has appointed David Rice as the first Chief Artificial Intelligence Officer (CAIO) in its history.

The asymmetry is the real theme of this issue: regulators are negotiating timelines, operational units are already deploying. Both eventually converge in the same compliance organisation – and the question is not whether they collide, but when.

LinkedIn Featured

Goldman Sachs AI Agents: When Machines Audit the Books

Goldman Sachs embedded Anthropic engineers into its own technology teams for six months to develop autonomous agents for trade accounting, KYC/AML and trading surveillance. It is the most ambitious attempt by a tier-1 bank to replace rules-based back-office processes with AI agents – and a test case for the high-risk requirements of the EU AI Act taking effect in August 2026. The targets Goldman reports are not incremental: –80 % exception queues in trade accounting, –30 % onboarding time in client onboarding, +20 % developer productivity.

Full article: Goldman Sachs AI Agents →

Agentic AI

Frontier Models, Platforms & Governance

Anthropic News: Claude Opus 4.7 hits general availability – with new cyber safeguards

Anthropic released Claude Opus 4.7 to general availability on 16 April. The model outperforms Opus 4.6 on software engineering tasks and long autonomous workflows at identical pricing: USD 5 per million input tokens and USD 25 per million output tokens. The more interesting point is the positioning: Opus 4.7 serves as Anthropic's testing ground for differentiated cyber safeguards that will later be transferred to Mythos-class models. Digital-banking platform Q2 Holdings is already using the model for Claude-Code-assisted development of regulation-compliant banking software.

Platform Release: Oracle and Google Cloud ship the platform toolkit – two banking deployments on the same day

On 14 April the platform competition thickened: Oracle Financial Services extended its banking platform with seven specialised agents for corporate banking – including credit validation, trade finance review and supply chain finance. The same day, CIMB Niaga, Google Cloud and Artefact launched production agents for relationship managers and contact centres at Indonesia's second-largest private bank. Taken together, both announcements signal the same thing: the agent stack is being commoditised – tier-1 banks increasingly configure rather than build.

Governance: Microsoft Agent 365 – control plane for autonomous agents, GA on 1 May

Microsoft announced the Microsoft 365 E7 "Frontier Suite" back on 9 March; the partner briefing of 21 April now fills in the details: general availability from 1 May 2026 at USD 99 per user per month. The bundle includes Microsoft 365 Copilot, Entra Suite and, for the first time, Agent 365 as a central control plane for the governance of autonomous agents. For banks operating under the Digital Operational Resilience Act (DORA), the BaFin MaRisk framework and the EU AI Act, this is the regulatorily most relevant part of the package: audit trails, permission structures and real-time monitoring for agents are exactly the requirements that have so far blocked productive agent rollouts inside banks.

Banking & Regulation

What actually matters now

EU AI Act Digital Omnibus: trilogue agreement targeted for 28 April

Since 26 March, Commission, Parliament and Council have been negotiating the Digital Omnibus on Artificial Intelligence. Both institutions share the core position: the high-risk deadline of 2 August 2026 is to be replaced with fixed alternative dates – 2 December 2027 for standalone high-risk systems (Annex III AI Act) and 2 August 2028 for systems embedded in products (Annex I). The political decision will fall five days after this newsletter appears. For banks this means: until an Official Journal publication lands, the August deadline remains operationally binding – anyone running credit scoring, anti-money-laundering or risk models with AI must be compliant by then.

BaFin's 9th MaRisk amendment: structural break in supervisory law, consultation closes 8 May

Germany's Federal Financial Supervisory Authority (BaFin) published the 9th amendment of the Minimum Requirements for Risk Management (MaRisk) for consultation at the beginning of April. The draft shrinks from 122 to 82 pages, introduces a three-tier classification of institutions by balance sheet size and integrates DORA as the exclusive framework for information and communications technology (ICT) risk. Significant Institutions fall out of scope – all Less Significant Institutions (LSI) and Small and Non-Complex Institutions (SNCI) must review their implementation status. Comments are due by 8 May 2026, with the final version expected in Q3 2026.

Signal & Noise

What your time is worth

• HSBC appoints David Rice as its first Chief AI Officer – Banking Dive
  The symbolism travels beyond HSBC: when one of the world's largest banks elevates a dedicated CAIO into the C-suite, others follow. 85 % of HSBC staff already have access to generative AI tools – the organisation now gets a leadership architecture to match.
• OpenAI acquires fintech startup Hiro Finance – TechCrunch
  Second financial sector acquisition within six months after the Roi deal. Ethan Bloch's team moves to OpenAI in full, the product is shut down. The direction is unmistakable: AI companies are pushing into the financial sector, not the other way round – and, as MIT professor Andrew Lo warns, buying into fiduciary-duty gaps as they go.
• Meta launches Muse Spark – its first proprietary frontier model – Meta AI Blog
  Meta leaves the open-source rail. The newly formed Meta Superintelligence Labs ship a natively multimodal model as a private preview. For banks that have relied on Llama as a cost-effective open-weight alternative, a strategic turning point.
• EBA consults on a 50 % reduction of supervisory reporting – European Banking Authority (EBA)
  The biggest structural simplification in years – while simultaneously integrating IFRS 18, ESG and the Fundamental Review of the Trading Book (FRTB). Consultation until 10 July 2026, application from September 2027. Reporting teams should shape the consultation now rather than be surprised later.
• Piero Cipollone: Tokenisation and the Role of Central Banks – European Central Bank (ECB)
  ECB Executive Board member Piero Cipollone outlines the European stance: tokenisation as a transformation of the financial system, anchored by central bank money. Anyone who has followed Appia and Pontes will find the frame both projects belong to.

"We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run."