Artificial intelligence was the recurring theme this year’s A-Team Group RegTech Summit in New York. Across conversations on AI governance, agentic workflows, crypto compliance, surveillance, AML transformation and regulatory reporting, a single theme cut through: AI is becoming embedded in the regulatory fabric of financial services, but its adoption must remain grounded, explainable, and anchored in strong governance frameworks.
A regulatory keynote opened the summit by underscoring that supervisors are already treating AI as a present-day concern. Far from advocating new rulebooks, the emphasis was on applying existing laws and expectations to evolving AI use cases. As the speaker noted, “Many of the use cases for AI do come under our existing laws.” What supervisors now expect is clarity around model testing, governance, and transparency – particularly in areas touching consumer outcomes or critical business functions.The regulatory discussion also highlighted how supervisors themselves are modernising – investing in data lakes, formal AI-use policies, staff training, and analytics tied to digital assets and blockchain. The message for industry was clear – if supervisors are upgrading their internal architectures, large institutions will need to keep pace.
Navigating the frontier of agentic AI in RegTech
During this Saifr and Theta Lake co-sponsored panel, the discussion noted a marked shift in institutional behaviour. One panellist observed that “there’s been some level of FOMO… everyone is afraid of being left behind,” yet most firms remain in early deployment stages. What has changed is that use cases are no longer theoretical. Participants described practical implementations ranging from digital employees managing horizon scanning and regulatory mapping, to agents assisting in KYC reviews, to real-time risk assessments combining structured and unstructured data.
These early deployments were developed with a clear recognition of data dependency. As one panellist noted, “The biggest problem with data is trust.” Agentic architectures can only perform as well as the data environments around them, and the panel returned repeatedly to the importance of provenance, calibration to risk appetite, and establishing clear lines of accountability. A panellist summed up the implementation challenge with stark candour: “95% of POCs don’t make it into production.” The arbitration point, they said, is rarely model performance – it is explainability.
That theme carried into a detailed fireside discussion on the architectural requirements for enterprise-scale deployment. A technology strategist argued that the long-term success of agentic AI depends on auditability rather than raw model capability. “The chain of reasoning must be reconstructable – that’s the biggest challenge,” they noted, calling for the creation of “audit APIs” capable of storing inputs, context, model versions, and intermediate reasoning steps. Hallucinations diminish significantly when models are “constrained by strict contextual boundaries.” The greater risk lies in integration and identity management – particularly the permissions granted to autonomous agents inside complex estates.
Growing scrutiny around digital assets and stablecoins brought AI into focus from a different angle. A regulatory contributor involved in crypto oversight described the first federal crypto statute as a watershed moment, framing stablecoins as instruments that could become invisible to consumers but embedded deeply within payments and treasury operations. With traditional institutions “playing catch-up after four years of lost time,” the speaker argued that AI will be essential for analysing blockchain-scale data flows, detecting illicit patterns, and maintaining the level of transaction transparency regulators now expect.
The discussion on financial crime and AML drew a similar conclusion: AI is delivering immediate value, but mostly through workflow acceleration rather than autonomous decision-making. Institutions highlighted the growing use of intelligent document processing, automated extraction from KYC files, negative media classification, and early work on AI-assisted onboarding. One panellist asked why firms should “go through our client profiles every five years when we can do continuous CDD?” Yet the room remained realistic about boundaries. As a challenger-bank representative commented, “SAR filings will never be fully AI-based… humans have skin in the game.” The panel emphasised that critical guardrails need “real people, not models,” to avoid inappropriate prompts or misinterpretation of outputs.
The end of manual compliance? How to build, govern and trust an AI driven regulatory change programme
The theme of governance over glamour became even more pronounced in a spirited conversation on regulatory change management (RCM), co-sponsored by RegAlytics and COMPLY. Participants agreed that generative AI performs well in classification, paraphrasing and accelerating document review but remains fragile when faced with cross-jurisdictional nuance or complex obligations. As one panellist cautioned, “Everyone wants to put AI in front of all their data – that’s a fool’s errand.” The effectiveness of an AI-led RCM plan is heavily dependent on metadata quality, drift monitoring, and the ability to establish clear audit trails for control mappings. A recurring point was that the final decision-maker for interpreting regulatory obligations “cannot be a model.”
The surveillance and data-management panels explained the importance of solid data foundations for AI success. Digital communications (e-Comms), trade events, mobile channels, and emerging digital assets are combining into a multimodal surveillance landscape. Panellists stressed that without unified taxonomies, strong metadata, and canonical data models, surveillance teams remain exposed. One practitioner captured the sentiment succinctly: “If you don’t know what data goes into your surveillance, you’re flying blind.” Another speaker explained how LLMs are already proving effective at filtering out newsletters, research and other “instant-skip” messages, but emphasized that human oversight remains essential for “judgment-based escalations.”
Best practices for streamlined and agile regulatory reporting
The final panel sponsored by 1Global, concluded with a practical debate on regulatory reporting, where panellists highlighted the importance of data standardisation and lineage. The goal of no-touch reporting was determined to be achievable, but only for highly structured, real-time regimes (e.g. OTC Derivatives), at least for the near future. For most reporting obligations, explainability and controls continue to require human oversight. As one practitioner observed, “You can automate submissions, but not the governance and controls.”
Across the entire event, the message was clear – AI is moving rapidly into the operational core of compliance. But continued adoption will depend on the maturity of a firm’s data, governance, auditability, and oversight frameworks. Where these exist, AI offers transformational potential. Where they do not, AI risks magnifying underlying weaknesses.
A regulatory contributor captured this balance with a final take-away, “Ultimately, accountability rests with the entity.”
Subscribe to our newsletter
