Cambridge counts, Brussels stalls, Washington bows out

Highlight

Cambridge counts, Brussels stalls, Washington bows out.

On 28 April 2026, the Cambridge Centre for Alternative Finance (CCAF) – together with the Bank for International Settlements (BIS), the International Monetary Fund (IMF), the World Economic Forum (WEF), the Inter-American Development Bank (IDB), the Consultative Group to Assist the Poor (CGAP) and the Arab Monetary Fund (AMF) – published the most comprehensive stocktake of artificial intelligence in financial services to date. 628 responses from 151 jurisdictions deliver an unambiguous diagnosis: 81 per cent of institutions are using AI, 40 per cent of those at an advanced level, and agentic AI is in adoption at 52 per cent. The pilot phase is officially over.

On the same day in Brussels, the second political trilogue on the EU AI Act Omnibus collapsed after roughly twelve hours of negotiations. The 2 August 2026 deadline for high-risk obligations therefore remains legally in force. Eleven days earlier, the Federal Reserve, the Office of the Comptroller of the Currency (OCC) and the Federal Deposit Insurance Corporation (FDIC) drew the opposite line with their revised Inter-Agency Model Risk Management Guidance (SR 26-2): generative and agentic AI are explicitly excluded from the model risk regime.

Three reactions to the same reality: Cambridge counts, Brussels postpones nothing and fails at it, Washington steps out of the game. For German institutions caught between the Federal Financial Supervisory Authority (BaFin), the European Central Bank (ECB) and US group-parent requirements, this means planning parallel compliance tracks without being able to bet on a single line. Anyone still waiting for harmonised AI supervision in 2026 should stop.

LinkedIn Featured

Adoption as a Façade: What the 2026 Cambridge Study Reveals about AI in Financial Services

81 per cent of financial institutions are deploying artificial intelligence. Only 14 per cent consider it strategically transformative. The April 2026 study from the Cambridge Centre for Alternative Finance paints the most sober picture yet of an industry caught between widespread piloting and unrealised value creation – and a supervisory architecture that is structurally lagging reality.

Read the full article: Adoption as a Façade →

Agentic AI

tools, skills & what's trending

Anthropic News: Finance Agents, Microsoft 365 and Moody's – going vertical on Wall Street

On 5 May 2026, Anthropic released ten pre-configured agent templates for banking, asset management and insurance – from Pitch Builder, Earnings Reviewer and Valuation Reviewer to KYC Screener and Month-End Closer. Claude operates as a single agent across Excel, PowerPoint, Word and Outlook. Moody's Corporation has embedded its full platform as a native Model Context Protocol (MCP) app inside Claude – direct access to credit ratings and data on more than 600 million companies within the interface. The market reaction was unambiguous: FactSet's share price dropped by up to 8.1 per cent, Morningstar lost more than 3 per cent.

Platform: Microsoft Agent 365 generally available – and multi-cloud

Microsoft Agent 365 has been generally available since 1 May 2026 at 15 US dollars per user per month. The Microsoft 365 E7 bundle (E5 + Entra Suite + Copilot + Agent 365), released the same day, is priced at 99 USD per user. The structurally more important news is the registry sync with Amazon Web Services (AWS) Bedrock and Google Cloud Gemini Enterprise – public preview from the same date. That makes Agent 365 the first broadly available multi-cloud governance plane for AI agents – and, for Microsoft 365-centric banks across the DACH region, the audit and identity layer that becomes inescapable in time.

OpenAI News: GPT-5.5 GA with 5.5 Pro – and BNY tests it alongside Anthropic

OpenAI released GPT-5.5 and GPT-5.5 Pro on 23 April 2026, with API availability since 24 April. The Bank of New York Mellon (BNY) is testing the model in parallel to Anthropic offerings – no single-vendor commitment. The pattern is the one European tier-1 banks have long adopted: no single-vendor bet. Anyone still selling a pure Microsoft- or OpenAI-only strategy in DACH today is ignoring market reality.

Open Source: Mistral Medium 3.5 and Codestral 2 under Apache 2.0 – the EU sovereignty card

On 29 April 2026, Mistral released a dense 128-billion-parameter model with agentic features and a "Work Mode" for multi-step autonomous tasks. Codestral 2 ships under Apache 2.0 – a licence loosening compared with the predecessor, beating GPT-4o on HumanEval and MBPP. Voxtral TTS brings open-source speech generation in nine languages, including German. For banks with EU hosting or sovereignty requirements, Mistral remains the only relevant European foundation-model option.

Banking & Regulation

what actually matters now

EU AI Act Omnibus: Trilogue collapsed on 28 April, August deadline remains operative

The second political trilogue between the European Parliament, the Council of the European Union and the European Commission ended without agreement after roughly twelve hours of negotiations. The sticking point was not the high-risk postponement, but the conformity-assessment architecture for AI systems embedded in products under sector-specific EU safety regulation (Annex I of the AI Act). The 2 August 2026 deadline for high-risk obligations therefore remains legally in force – anyone running credit-scoring, anti-money-laundering or risk models with AI must be compliant. The next trilogue is scheduled for 13 May 2026.

BaFin 9th MaRisk Amendment – consultation closes tomorrow

The Federal Financial Supervisory Authority (BaFin) closes the consultation on the 9th amendment to the Minimum Requirements for Risk Management (MaRisk) on 8 May 2026. The draft is markedly more principles-based, integrates ESG risks and the Digital Operational Resilience Act (DORA) as the exclusive framework for information and communication technology risks (ICT), and broadens the opening clauses for Small and Non-Complex Institutions (SNCI). Submissions to konsultation-02-26@bafin.de and marisk@bundesbank.de – final version expected Q3 2026, implementation by end of 2026.

Signal & Noise

what your time is worth

• FIS and Anthropic launch Financial Crimes AI Agent – FIS press release
  Bank of Montreal (BMO) and Amalgamated Bank as lead customers. The first time a core-banking provider markets agentic AI as a native layer – the distribution channel into US tier-2 and tier-3 banks, and the benchmark Temenos, Avaloq and Finastra in Europe will be measured against.
• Fed Vice Chair Michelle W. Bowman names Mythos explicitly, SR 26-2 takes agentic AI out of model-risk scope – Federal Reserve
  A central banker naming a frontier model on the record – notable. More structurally important: Federal Reserve, OCC and FDIC are pulling generative and agentic AI out of the model-risk regime, while the EU/BaFin/ECB are pulling the same content in via the AI Act and MaRisk.
• Deutsche Bank and Google build AI agents for trading surveillance – PYMNTS
  Frankfurt's tier-1 bank is planning Gemini-based agents in the compliance discipline with the highest audit requirements (Market Abuse Regulation, Markets in Financial Instruments Directive II). If the model works here, the bar for other use cases drops.
• DeepSeek V4 Preview – open weights on par with Opus 4.7 at one-seventh of the price – DeepSeek
  1.6 trillion parameters, MIT licence, statistically tied with Claude Opus 4.7 on SWE-bench Verified. Geopolitics make productive deployment in DACH unrealistic – but the pricing pressure on Western providers is real, the closed-source premium is melting.
• Piero Cipollone: tokenised central bank money is "indispensable" – European Central Bank
  The ECB Executive Board member frames tokenisation as a general-purpose technology. Anyone running treasury or capital markets needs roadmaps for DLT-based settlement in 2026 – not 2027.

Personal Note

currently in IT-Finanzmagazin

Agent Registry, Decision Logs, Permission Tiers – wo die Banken-IT jetzt nachlegen muss

An AI agent is not a chatbot. It has system access, plans multi-step actions and makes decisions – yet many banks still monitor it like a single API call. My piece for IT-Finanzmagazin (5 May 2026) lays out three structural gaps that banks' IT architectures need to close now: a central Agent Registry, structured Decision Logs and tiered Permission Tiers along least-privilege lines – the prerequisites for EU AI Act compliance and for keeping operational control over autonomous systems. (Article in German)

Read the article in IT-Finanzmagazin →

"Every bank in the world wants AI that acts, not just assists. The future is about a trusted provider who manages the data, governs the agents, and stands between your customers and the AI."