Goldman Sachs and KYC Automation: Scalable Agentic AI in Action 

Goldman Sachs has been co-developing autonomous AI agents with Anthropic for six months, with the AI company's engineers embedded directly at the bank. The deployment targets two specific functions: accounting for trades and transactions and client vetting and onboarding, both high-volume, process-intensive tasks that combine the need to parse large quantities of data and documents with the application of rules and judgment. 

Marco Argenti, Goldman's Chief Information Officer, told CNBC that the agents function as digital co-workers for roles within the firm that are large-scale, complex, and process-intensive. The approach started with an autonomous coding agent, but Goldman found that the same model's capacity for step-by-step reasoning applied equally well to compliance and accounting. The bank says it was "surprised" at how capable the system was in areas beyond coding. 

CEO David Solomon has framed the broader initiative as a multiyear reorganisation around generative AI, with the explicit goal of constraining headcount growth while volumes increase: a form of asymmetric scaling in which business capacity expands without proportional growth in operational cost. 

Agentic AI vs Chatbots: Redefining Back Office Automation in Banking 

What Goldman is deploying is categorically different from what most financial institutions have introduced so far, namely chatbots. A chatbot answers questions, while an AI agent executes tasks by accessing systems, applying rules, making decisions within a defined scope, and producing outputs that trigger further actions downstream. 

In trade reconciliation, this means the agent does not just surface information: it processes it, identifies discrepancies, and returns outcomes. In KYC onboarding, it does not just retrieve documents: it evaluates them against compliance criteria and progresses a case. The human is removed from the repetitive steps, not from the judgment call at the end. 

This distinction matters because the failure modes are different. A chatbot that gives a wrong answer creates a bad experience. An agent that applies an outdated compliance rule, misreads a document, or acts on stale data creates a regulatory event. The stakes of getting knowledge management wrong scale with the autonomy of the system. 

“An agent that applies an outdated compliance rule does not produce a bad conversation. It produces a regulatory event.” 

Why European Banks Lag Behind in Agentic AI Architecture 

The contrast with European banking is significant. No major European institution has announced an equivalent agentic deployment in compliance or back-office functions. The Bank of Italy's Director General Luigi Federico Signorini, in a speech published in July 2025, noted that Italian banks' investments in AI-based projects have more than quadrupled in two years, with generative AI accounting for the largest share of both investment and new projects. Around three-quarters of larger banks are using AI for credit risk assessment. But the same speech also emphasised the need to "preserve the key role of the critical judgement and experience of human supervisors”.  

This framing reflects where European institutions currently are: accelerating adoption while managing governance carefully. 

The examples most often cited as evidence of agentic AI in European banking reveal the same pattern. Lloyds Banking Group's 2025 deployment of a financial assistant for millions of customers focuses on spending analysis and savings advice through a mobile app: a personalisation tool, not a system operating autonomously on regulated processes. Société Générale created a dedicated AI entity in 2025 with multiple domains in production including onboarding and back office, while Anthropic's January 2026 partnership with Allianz to build custom agents signals growing structural ambition. These are meaningful steps. None of them constitutes an architectural transformation of mission-critical processes comparable to what Goldman Sachs is building. 

The deeper distinction is structural: no European financial institution has engaged foundation model provider engineers to co-develop proprietary systems on-site. European banks remain dependent on external infrastructure rather than building their own. 

EU AI Act and Compliance: Architectural Governance for AI Agents

The EU AI Act is sometimes framed as the reason European banks are moving more slowly. The more accurate framing is that it is the reason European banks need to move more deliberately. For high-risk AI systems, those that support or influence decisions materially affecting individuals, such as creditworthiness assessment, KYC, and compliance screening, the Act requires meaningful human oversight, audit trails, and traceable, correctable outputs. These are not optional additions. They are architectural requirements. 

Steven Maijoor of the DNB, the Dutch regulator, stated in an interview with Bloomberg that the large-scale adoption of AI systems means, intrinsically, greater dependence on American tech giants, concentration of risk, and geopolitical vulnerability. It is a strategic evaluation that explains part of the European caution: not only the weight of compliance, but a deliberate calculation about who controls the infrastructure on which critical financial systems run. 

The EBA's Risk Assessment Questionnaire confirms the structural picture. Most EU banks access generative AI through cloud-based APIs from a small number of dominant model providers. A smaller proportion integrates systems on-premises for greater control, while only a handful are building their own. Generative AI deployment is most advanced in customer-facing support and back-office efficiency, with limited penetration into core risk processes. In areas closer to core banking risk, such as AML/CFT, client profiling, risk modelling, around 10% of EU banks are in active experimentation, not production. The EBA attributes this measured pace to reliability limitations including hallucinations, explainability challenges, ICT concentration risks, and a scarcity of skills required to implement meaningful human-in-the-loop oversight, particularly in smaller institutions. 

The institutions that will navigate this successfully are not the ones that wait for regulatory clarity to be perfect before acting. They are the ones that build governance into the architecture from the start, so that when an agent makes a decision in a regulated context, there is a traceable record of what it retrieved, what it produced, who reviewed it, and what changed as a result.
 

Agentic AI 
An AI system composed of autonomous agents that act independently to achieve defined objectives, accessing systems, applying rules, and executing multi-step tasks without requiring step-by-step human direction. 

Expert-in-the-Loop AI for Regulated Banking: KYC, Compliance, and Approval Workflows. 

For autonomous AI agents operating in high-regulation environments, an Expert-in-the-Loop framework is not a nice-to-have: it is what makes Agentic AI defensible. 

The Syllotips architecture is designed for exactly this context. Every response or action the system produces is automatically scored for reliability. Anything below threshold is routed to the relevant Subject Matter Expert before it reaches a downstream process, not after. The expert's review is logged, digitally signed, and timestamped. That correction enters the shared knowledge base and propagates in real time, so the same gap does not produce the same compliance exposure in the next transaction. 

In highly regulated environments, Syllotips can add a further layer: a structured approval workflow in which an Approver/Controller does not only review individual responses but validate the update to a procedure before it propagates across the organisation. When a compliance rule changes, a KYC criterion is revised, or an exception workflow is updated, the corrected knowledge does not enter the shared knowledge base until an SME with the relevant competence has confirmed that the update is accurate, complete, and consistent with the governing regulatory framework. What reaches agents in every office is not simply the latest available update. It is the one that has been validated by the right expert. 

In banking specifically, this means that the procedural knowledge governing KYC checks, trade exceptions, and onboarding edge cases is not static. It is continuously updated by the people who know when rules change, when exceptions apply, and when a case requires a different path. The audit trail this generates is not compliance overhead; it is the evidence that human oversight was meaningful rather than nominal.