Agentic AI is rapidly becoming one of the most important topics in financial services, and few rooms feel that more acutely than the Point Zero Forum, which gathered some 2,000 central bankers, regulators, policymakers and technologists in Zurich from 23–25 June for three days of dialogue. The forum is co-organised by the Global Finance and Technology Network (GFTN) and Switzerland’s State Secretariat for International Finance. GFTN is a Singapore non-profit, launched by the Monetary Authority of Singapore in 2024, and the same body that runs the Singapore FinTech Festival.
Of everything on the forum’s three-day agenda, the MRKT 3.0 editorial team had circled one session in advance: “The Agentic Leap: Deploying AI at Scale in Financial Services.” The move from AI demonstrations to live, regulated deployment sits at the heart of what we cover, so we made a point of being in the room for it.
Moderating was Arjun Vir Singh, Global Head of Fintech and Digital Assets at the consultancy Arthur D. Little. With him sat Giuliano Benjamin Clark of Amazon, Ian Rogers of the digital asset security firm Ledger, Dr. Jochen Papenbrock of NVIDIA, Nicolas de Skowronski of the Swiss private bank Julius Baer, and Tin Pei Ling of the Singapore payments company MetaComp.
Why Most Agentic AI Projects Never Reach Production
Singh set the tone by emptying his pockets of statistics he was careful not to source. Around 95% of generative AI pilots, he said, deliver no measurable financial return. Some 88% of AI agent proofs of concept never reach production. One large research house expects 40% of agentic AI projects to be cancelled by 2027. He called it the production gap, and underneath it he saw something less glamorous than model quality: a data and governance problem wearing a technology costume.
Then he handed the microphone to the only banker on the stage and asked him to name something agentic that was actually in production.
Agentic AI Use Cases Already Running in Banking
Nicolas de Skowronski, who leads digital business transformation at Bank Julius Baer and has spent nearly 25 years there, did not flinch. His first move was to gently correct the consultants. There is, he said, a vast underestimation of the complexity of the business banks actually run, and of the regulatory framework they operate inside. Plenty of people have promised to solve his software in a week. It is, he noted, a little more complex than that.
Julius Baer took a different route from most competitors. It chose to run everything on premise, to keep full control of its own data, and to work with open source models rather than expose itself to runaway token prices if usage took off. On that foundation, de Skowronski rattled off live cases rather than slideware. Relationship managers query an agent that sits on top of millions of research and product documents, so a banker can pull the house investment opinion and the right product in seconds. In risk and compliance, a name and media screening agent has cut false positives by 50%, clearing millions of hits that would otherwise have landed on a human desk. Payment agents process payments automatically.
His verdict on the production gap was measured. The full journey, from use case to development to regulatory compliance to process redesign to change management, is genuinely harder than people expect. But the idea that 90% never go live, he said, simply does not match his experience.
Failure, Reframed as a Feature
The technologists on the panel were happy to let the failure rate stand, because they read it differently. A high pilot mortality rate, the counter-argument ran, is actually a good sign. It means the technology has unlocked enough agility that teams can test ideas they would never previously have prioritised. The old cycle of spending months scoping a product, then months more building it, only to discover it was the wrong product, has collapsed into something far shorter and far cheaper to get wrong.
The wider mood in the room was that none of this should surprise anyone who has lived through a technological shift before. The industry is at the early, over-excited stage where experiments are plentiful and most of them are meant to fail. The job, on this view, is to keep teams free to experiment and free to fail until the genuinely valuable use cases reveal themselves.
The Full Stack Argument
If the bottleneck is not the model, what is it? Dr. Jochen Papenbrock, NVIDIA’s EMEA Head of Financial Technology, answered with infrastructure. Scaling agentic AI is a full stack problem. He described an AI factory of layered components, where compute, the AI software stack, and lifecycle management are co-designed so that open source models are aware of the silicon beneath them. He kept returning to one image: the token factory. Long-running agents with complex tasks consume more and more tokens, so the platform underneath has to convert energy and capital into tokens as efficiently as possible.
Governance, in his telling, starts with ownership. There is a difference between renting a frontier model and owning one. Download a foundational model, tune it on your own data, and you hold the full lifecycle and the full control. He pointed to transaction foundational models built with partners including Mastercard, Adyen and Revolut, where a company’s own transactional history is turned into a unified intelligence layer that also informs its agents, all of it kept in house.
Building Trust in Autonomous Financial Agents
On the demand side, the industry’s early assumption was wrong. Everyone expected fully autonomous agents to take off fast. Giuliano Benjamin Clark, who heads product for agentic payments at Amazon, said customers told him the opposite. They do not trust it yet. They want to see what the agent is doing, and they want to approve it. So the market sits at the left end of a spectrum that runs from a buyer checking everything to a fridge that reorders milk on its own. For now, the autonomous fridge is a long way off, and because the current use cases are e-commerce adjacent, existing consumer and data regulation already serves them. His advice on new legislation was to wait. This is not the moment to reinvent the wheel.
Moving along that spectrum is a question of trust, built the way trust is always built, through repeated positive interactions until you stop questioning. Start with product discovery, check the agent’s work, watch it be right more often, and gradually hand it more leeway. The foundation, Clark insisted, is accuracy and the right controls.
Ian Rogers, Chief Human Agency Officer at the Paris-based digital asset security firm Ledger, gave that intuition a sharper edge. Borrowing a line he attributed to Balaji Srinivasan, the tech entrepreneur, investor and former Coinbase Chief Technology Officer, he said AI is middle to middle and humans are end to end. The future of enterprise work is many agents per human, with people orchestrating and verifying at both ends. The danger is that we do one end and skip the other, accepting whatever answer arrives. Fine for a poem. Not fine for a transaction.
His security point was blunt. Approving a multi-million dollar transaction on a phone or a laptop is simply not good enough. The fix is a clean division of labour. Let the probabilistic model run wild, because it holds no keys. Compliance lives in a deterministic policy engine, not in the model. Before any transaction executes, the policy engine checks it and the human signs it. Without transaction-time security, he warned, traceability is just a record of what a probabilistic creature happened to do.
Rogers also made a quieter prediction about commodities. Switching costs have never been lower. Whether you type one model’s name or another’s at the command line is, in his words, irrelevant. The durable value sits not in the model but in the trust an Amazon or a Ledger has earned, and in the enforcement layer between the model and the secrets.
Governance Frameworks for Agentic AI
Tin Pei Ling, Co-President of the Singapore payments firm MetaComp, made the case that governance can accelerate rather than obstruct. An interoperable framework that jurisdictions and large players agree on gives everyone a baseline they can trust, which is exactly what lowers the inertia of the first step. Singapore runs governance first and legislation later, introducing a model agentic AI governance framework in January and updating it in May, alongside changes to its data protection law. She pointed to a wave of parallel effort, a recent set of rules from the FSB and work at the EU Commission, the IMF and IOSCO, as evidence that international coordination is moving from forum talk into practice.
For finance specifically, she laid out a sequence: give every agent a verifiable identity tied to a human, score it and set permissions, then monitor its behaviour for drift with a dynamic risk rating, much like a client risk rating that updates over time rather than freezing at onboarding. The hardest frontier, she said, is governing agent-to-agent and ecosystem-to-ecosystem interaction, where standards and interoperability finally have to deliver. Her closing principle was the one the room kept circling: human agency must never be undermined.
Clark turned the same logic into a payment rails wish list. Customers and merchants stay in control. Existing consumer and data protections are preserved, never weakened. Governance is proportionate, because wiring a million dollars is not the same risk as saving five euros on a cable. And the system is interoperable on open standards, because the whole promise of agentic commerce collapses the moment any agent cannot transact with any merchant. He reached for the shipping container as his analogy. Multiple competing standards came first, then convergence under ISO, then a decade of retrofitting. Technology will compress the timeline, but the order will hold.
The Swiss Moat
The most sobering moment came back from Julius Baer. De Skowronski expects the model itself to become a commodity, which is precisely why private banking will differentiate on accountability, privacy and judgment. He offered two unsettling anecdotes. Clients now record conversations with their relationship managers and challenge them because an AI assistant told them the bank was wrong. Others arrive convinced overnight that they have discovered a risk-free options strategy. Garbage in, garbage out, he said, and the judgment to catch it has to be human. His sharpest worry was privacy, watching people feed account screenshots and the intimate details of their financial lives into whatever model is nearest, for free.
On auditability, Papenbrock was optimistic, and he leaned on the open source community to keep him honest. He cited NVIDIA’s open Nemotron model family, OpenShell, a gatekeeper runtime that lets an agent act until a policy blocks the action, and NeMo Guardrails for checking complex policies at runtime without slowing execution. Building safe, auditable agentic systems, he argued, takes serious engineering and good tooling, which is why he wants the community building the frameworks rather than any single vendor.
The panel ran four minutes over and ended, as the moderator noted, with a table full of optimists. Seen from Switzerland, the optimism has a particular shape. The technologists want room to experiment and fail. The banker wants control of his data and his judgment. The governance advocate on the panel wants a baseline everyone can trust. None of them think the model is the hard part. All of them think the human, sitting at the end of the chain with a key and a final yes, is the part worth protecting.
Author: Klaudia Archimowicz
See Also:
Slash’s AI Banker Can Now Move Money Without You. What Could Go Wrong?
