The announcement sounds like a milestone: FIS, the banking technology group that by its own account processes transactions worth nearly 12 per cent of global economic output ("powering nearly 12% of the global economy"), is developing a Financial Crimes AI Agent together with Anthropic. Claude serves as the reasoning engine; according to FIS, the agent is meant to compress anti-money laundering (AML) investigations from days to minutes, with Canada's BMO and the US-based Amalgamated Bank involved in the development. Read the announcement soberly, however, and one thing stands out: the product does not yet exist as available software – general availability (GA) is planned only for the second half of 2026. The real story is therefore less about a breakthrough and more about a shift: agentic AI is moving into highly regulated compliance workflows while supervisors in the US and in Europe have no rules for exactly that. This piece is accordingly an analysis of the announcement, not a test report on a working product.
What: On 4 May 2026, FIS and Anthropic announced a Financial Crimes AI Agent – Claude as the reasoning engine, embedded in FIS infrastructure; general availability planned for the second half of 2026
Promise: AML investigations from days to minutes, according to FIS; BMO and Amalgamated Bank are involved in the development according to FIS, with Amalgamated Bank confirming this in its own press release
Context: FIS is a follower, not a first mover – Oracle has offered comparable agentic AML capabilities globally since March 2025, with NICE Actimize and Hawk AI also in the market
Governance: The Federal Reserve's SR 26-2 (17 April 2026) replaces SR 11-7 and SR 21-8 but explicitly excludes generative and agentic AI; the Anti-Money Laundering Authority (AMLA) has issued no requirements on AI agents in money laundering prevention
Open questions: Responsibility of the money laundering reporting officer (Section 7 of the German Anti-Money Laundering Act, GwG), failure to file a suspicious activity report (Section 43 GwG), explainability towards BaFin and the AMLA, no independent benchmark for the promised time savings
What FIS promises – and what of it is substantiated
Functionally, FIS describes the agent as follows: it automatically assembles evidence from the bank's core systems, assesses cases against known money laundering typologies, prioritises the highest-risk matters and prepares the narratives for suspicious activity reports (SARs) in the US. Decision-making authority is meant to remain with the human investigator, every conclusion the agent reaches is meant to be traceable to its source data and auditable, and customer data remains in FIS-controlled infrastructure. Anthropic embedded its Applied AI team directly inside FIS for this purpose; a knowledge transfer has also been agreed so that FIS can build and scale additional agents independently in future – the roadmap names credit decisioning, deposit retention, onboarding and fraud prevention.
So far the vendor's account – and that should be taken literally, because almost all the core claims come from the FIS press release of 4 May 2026. The central performance promise of compressing investigations "from days to minutes" is a vendor promise without an independent benchmark; elsewhere, FIS material itself speaks of "hours to minutes", which points more towards marketing language than towards a robust measurement. BMO's involvement is documented only via FIS; no statement of the bank's own exists. Amalgamated Bank is different: it confirmed the collaboration on 7 May 2026 in its own press release. Its Chief Information and Operations Officer Sean Searby emphasises the co-design character: the bank's own teams are contributing to the design of the agent rather than just deploying a technology. President and CEO Priscilla Sims Brown frames the pilot as part of shaping how advanced tools are introduced into regulated workflows "with governance and controls".
That the problem is real is beyond question. The United Nations estimates illicit funds in the global system at around two trillion US dollars annually; FIS puts the AML operating costs of US financial institutions alone at 35 to 40 billion US dollars per year, without citing a source of its own for that figure. And industry estimates, including from PwC, suggest that up to 95 per cent of all AML alerts are false positives – although no independent primary study of this figure exists. It nevertheless aptly describes why large banks employ hundreds to thousands of analysts whose capacity is mostly tied up in cases that turn out to be harmless. The 2024 enforcement action against TD Bank showed the flip side: unworked alerts caused by false-positive overload were an explicit point of criticism by the US regulators there.
A follower with distribution power: the competitive landscape
The timing also belongs to the assessment. The announcement fell on the same day as Anthropic's 1.5 billion US dollar joint venture with Blackstone, Hellman & Friedman and Goldman Sachs – 4 May 2026 was visibly staged as Anthropic's financial services day. That points to a co-marketing logic: FIS gets the AI narrative, Anthropic gets the proof that Claude is arriving in regulated workflows. It is also notable that the press release is not listed on Anthropic's own website; the driving party was FIS.
In substance, FIS is a follower, not a first mover. Oracle has had comparable agentic AML capabilities globally available in its Investigation Hub since March 2025 – Jason Somrak, Head of Financial Crime Products at Oracle, described at the time an agent that queries data sources, collects evidence, follows an investigation plan, recommends a decision and writes the complete narrative; Oracle estimates that at least 80 per cent of that work could be automated at large banks. NICE Actimize has offered AI agents in its Xceed platform since 2025, Hawk AI positions an agentic AML overlay, and HSBC has worked with Google Cloud for years on machine-learning-based risk scoring – with, by its own account, 60 per cent fewer false positives at two to four times the hit rate, albeit without LLM agents. What actually differentiates FIS comes down to three points: access to the core banking data of thousands of institutions, the co-design model with embedded Anthropic engineers including knowledge transfer, and data residency within its own infrastructure. That is a distribution and architecture differentiation, not a technological lead.
The double governance vacuum: SR 26-2
The most regulatorily charged part of the story is not in the FIS press release. On 17 April 2026, the Federal Reserve recast its model risk guidance with SR 26-2, superseding two predecessor documents: SR 11-7, the standard from 2011, and SR 21-8 – the Interagency Statement on model risk management for systems under the Bank Secrecy Act (BSA), in other words exactly the guardrail that had applied to AML systems until now. SR 26-2 is risk-based and principles-based, but it now applies only to banks with more than 30 billion US dollars in total assets – and it explicitly excludes generative AI and agentic AI. A separate consultation on AI has been announced but not issued.
That produces a double vacuum: smaller institutions fall out of the model risk guidance altogether, and for the technology FIS and Anthropic want to roll out, no binding framework exists even at the large institutions. This becomes pointed in combination with the proposed rule of the Financial Crimes Enforcement Network (FinCEN) of April 2026, which will assess AML programmes by effectiveness rather than by process compliance: the pressure to deploy more capable tools is rising at exactly the moment the governance requirements for them are missing. The FIS design, with auditability, human-in-the-loop and sovereign data residency, anticipates a regulation that does not yet exist.
The DACH dimension: Section 25h KWG, Section 7 GwG and the AMLA
In Europe the picture is structurally similar, with one decisive difference: here, responsibility is codified more sharply. Section 25h of the German Banking Act (Kreditwesengesetz, KWG) obliges German institutions to operate adequate data processing systems for money laundering prevention – whether an AI agent whose conclusions rest on a language model meets that criterion is a question the Federal Financial Supervisory Authority (BaFin) has not yet answered. Section 7 of the German Anti-Money Laundering Act (Geldwäschegesetz, GwG) assigns the money laundering reporting officer responsibility for compliance with anti-money laundering obligations. From this follows the question that appears in no press release: who is liable when the agent wrongly closes a genuine suspicious case and the suspicious activity report under Section 43 GwG is therefore never filed? The officer cannot delegate that responsibility to a piece of software – he or she must be able to explain the agent's decision logic to BaFin, including the cases the agent has filtered out.
Added to this is the new European supervisory layer. The Anti-Money Laundering Authority (AMLA) began operations in Frankfurt am Main on 1 July 2025, took over all AML mandates of the European Banking Authority (EBA) on 1 January 2026 and will, from 2028, directly supervise around 40 cross-border high-risk financial institutions; Bruna Szego chairs the authority. The AMLA has so far issued no specific requirements on agentic AI in money laundering prevention. For DACH institutions this means: anyone piloting an AML agent today is doing so in a grey area – and should document their own governance so that it withstands a future AMLA examination whose standards nobody yet knows.
What compliance leaders should do now
Four work packages follow from the analysis. None of them presupposes buying the FIS agent – all of them presuppose being prepared before the first business unit arrives with a pilot.
Immediately: The FIS agent will be available in the second half of 2026 at the earliest; Oracle has been in the market since March 2025, as have NICE Actimize and Hawk AI. A vendor evaluation should consistently treat vendor promises such as "days to minutes" as unsubstantiated and apply verifiable criteria instead: traceability of agent decisions, source linkage, data residency, and the integration effort into the existing transaction monitoring landscape.
By Q4 2026: Neither SR 26-2 nor the AMLA nor BaFin currently specifies how agentic AI in AML systems is to be validated. Precisely for that reason, the institution should set its own standards: validation cycles for non-deterministic systems, thresholds for human escalation, and logging of every agent decision including its source data. Those who document this now can feed it into consultations later instead of chasing interpretation letters.
By Q1 2027: Before any agentic AML pilot, it should be settled in writing which case classes the agent may pre-prioritise on its own, where a human must decide, and how the money laundering reporting officer discharges his or her responsibility under Section 7 GwG when the agent filters cases out. The critical path is the omitted suspicious activity report under Section 43 GwG – that scenario must be addressed explicitly in the control design, not implicitly.
2027: From 2028, the AMLA will directly supervise the largest cross-border institutions; BaFin already requires adequate systems today via Section 25h KWG. Every procurement should therefore contractually require the provider to document the agent's decision paths in a supervision-proof manner – including the question of which model and which version a decision was made with. What cannot be explained to the supervisor does not belong in the suspicious activity reporting process.
Risks and open questions
Three caveats belong to an honest assessment. First, the evidence base: around four fifths of the story hangs on a single press release. There is no independent benchmark for the time savings, no statement of BMO's own and no confirmation of the model version used; the product is not generally available. Second, the governance risk: an agent that prioritises cases and prepares narratives shifts decision-making weight in practice, even when a human formally signs off – the automation bias of overloaded investigation teams is well documented, and it is exactly this constellation for which the framework is missing on both sides of the Atlantic. Third, the incentive structure: the FinCEN effectiveness logic and the cost pressure of an estimated more than 60 billion US dollars in annual AML compliance costs worldwide are pushing institutions towards deployment before supervisors have set standards. Those who scale too early carry the interpretation risk alone.
The strategic conclusion: the FIS-Anthropic announcement is neither the breakthrough it is marketed as nor a footnote. It marks the moment at which agentic AI turns from an experiment into a procurement decision in money laundering prevention – with genuine distribution power behind it. Institutions that define their own governance, their liability questions and their explainability requirements now will negotiate from a position of strength later. The others will adopt their vendor's standards.
Christian Schablitzki
Strategy & Management Consultant · Agentic AI expert for financial institutions
Over 20 years in investment banking and derivatives trading, followed by more than 10 years advising financial institutions. Currently a Partner at Infosys Consulting in Germany. Certified in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.
LinkedIn profile →Keep reading – every 14 days in your inbox.
Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free.
GDPR-compliant. Unsubscribe at any time.