.svg "Group 2238 (2)"){kind=small}
.svg)
.svg)
%20(1).svg)
%20(1).svg)
.svg)
Share Article
A wealth management client recently challenged her own banker with three words: "Claude told me..." That collision between AI-fueled overconfidence and cold statistics defined this panel, where one study claims 95% of generative AI pilots deliver no measurable return, while a banker on stage insists that's not what he's seeing in production.
Are 90% of agentic AI pilots actually failing?
Moderator Arjun Vir Singh cited a batch of eye-catching stats making the rounds in banking circles: one study claiming 95% of generative AI pilots deliver no measurable financial return, another putting the failure rate of AI agent proofs-of-concept at 88%, and a forecast that 40% of agentic AI projects would be cancelled by 2027. Nicolas de Skowronski pushed back directly, pointing to Julius Baer's own live production deployments: agents that query the bank's entire research and product universe to answer relationship-manager questions instantly, a negative-media screening tool that had already cut false-positive alerts by 50%, and automated payment-processing agents. He credited the bank's deliberate choice to run everything on-premises with open-source models, prioritizing full data control over convenience. Giuliano Benjamin Clark reframed the failure statistics as a feature rather than a bug: fast, cheap experimentation was replacing the old six-to-nine-month product cycle, letting teams test and kill ideas in weeks instead of committing to them for months.
Ledger's rule for agentic finance: AI suggests, humans sign
Ian Rogers argued the industry was overcomplicating agentic governance by trying to make AI models themselves trustworthy, when the real fix was structural: AI models are inherently probabilistic, so a deterministic layer, a human, or in Ledger's case, hardware wallet infrastructure, has to hold the actual keys and sign off on every transaction. In his framing, the model can suggest a trade or a transfer, but it never has the ability to execute one, which removes most of the governance questions people were asking about the model itself. He added a related, counterintuitive point: crypto rails already solve the interoperable, low-cost, decentralized machine-to-machine payment problem the industry has been trying to design from scratch, arguing the market had been underestimated because it was built with human users, not autonomous agents, in mind.
Clients are already taking financial advice from chatbots — and private bankers are worried
Nicolas de Skowronski shared two anecdotes that stuck with the room: relationship managers now field clients who record their conversations and push back with "Claude told me this isn't correct," and others who arrive convinced overnight that a chatbot helped them discover a "risk-free" options strategy that, on closer inspection, doesn't hold up. As enthusiastic as he was about AI overall, he said his biggest emerging worry wasn't automation, it was privacy: clients routinely feed screenshots of their accounts and details about their health and personal finances into public AI tools without understanding what happens to that data afterward.
Speakers:
Dr. Jochen Papenbrock, EMEA Head of Financial Technology, NVIDIA
Giuliano Benjamin Clark, Head of Product, Agentic Payments, Amazon
Ian Rogers, Chief Human Agency Officer, Ledger
Nicolas de Skowronski, Head Digital Business Transformation, Bank Julius Baer & Co. Ltd.
Tin Pei Ling, Co-President, MetaComp
Host:
Arjun Vir Singh, Partner, Global Head of Fintech & Digital Assets, Arthur D. Little
Point Zero Forum 2026 | Zurich, Switzerland
