Fraud, risk and customer intelligence at a scale where a fraction of a percentage point moves billions.
Convening inside the industry The council has held its working sessions at Visa facilities in San Francisco and Foster City, bringing payments, banking, engineering and academic members together on the questions institutions cannot answer alone.
Financial services runs AI under three constraints most industries never face at once: a regulator who can ask how a decision was made, an adversary actively working to defeat the model, and a customer who leaves after one bad experience.
Those constraints are why so much of what works elsewhere does not transfer. A model that cannot be explained is not deployable when a declined application has to be justified. A fraud model that is accurate today degrades as soon as the people it catches adapt to it. And the data that would make everything easier is the data institutions are least able to share.
This council works on the problems that sit in that intersection, with people who have carried responsibility for systems operating at national and global scale.
The sector's most valuable signals sit inside institutions that cannot legally or commercially pool them. Synthetic data and privacy-preserving methods are how a fraud pattern learned at one bank becomes useful at another, and the council treats that as its central technical problem.
Fraud is the one AI application where the environment fights back. Models that are not retrained against adaptive behaviour decay in production rather than at the point of testing, and the council works on how institutions detect that decay before losses reveal it.
An explanation that convinces a data scientist is not the same as one that satisfies a supervisor or a customer contesting a decision. The council works on what evidence has to exist at the point a model declines someone.
Combining transaction history, correspondence and service contact allows advice tailored to a specific person rather than a cohort. The line between useful and intrusive is the part worth getting right, and it is a design decision rather than a technical one.
Agents that execute transactions rather than recommend them raise questions about authority, audit and reversal that the sector has not settled. Working through those before deployment is cheaper than after.
Competitors face identical threats and identical supervisors. Sessions run at member facilities under conditions that let people describe what actually happened rather than what their communications team would prefer.
Generative methods producing secure datasets that retain statistical validity, so risk models can be tested rigorously without exposing real customer records.
Models monitoring billions of daily transactions for unusual patterns, with the aim of stopping fraud as it happens rather than reporting it afterwards.
Multimodal analysis across transactions, written enquiries and call records to personalise advice and anticipate churn before it happens.
Causal methods relating economic indicators, news sentiment and global events, rather than pattern matching on price history alone.
Systems that combine real-time financial news with a client's stated preferences to produce recommendations, with the disclosure and suitability obligations that implies.
Computer vision applied to validating insurance documentation and assessing damage, where the accuracy question and the fairness question are the same question.
Two closed sessions hosted inside Visa facilities, which is a level of access that says something about how the council operates.
Signals to Outcomes: AI at Mission Rock, October 2025, at Visa's global headquarters in San Francisco. An invitation-only working session on agentic AI and hyperpersonalisation, structured around a keynote, a panel on moving proof of concept into production, a regulatory briefing, and roundtables that ended with participants holding thirty, sixty and ninety day pilot plans rather than notes.
Applied AI Research Roundtable, December 2024, at Visa's headquarters in Foster City. Sessions on AI in banking, cybersecurity, humanitarian impact and financial systems, with parallel workshops on smart cities and advanced manufacturing. Speakers came from Visa, Bank of America and the University of Ottawa.
Both were built to produce commitments. A day that ends with a dated plan is worth more than one that ends with agreement.
The council is convening technology and financial services executives from across the Bay Area for its next closed session. The agenda is being set with members now, and the working themes are agentic systems where money moves, privacy-preserving collaboration between institutions, and what explainability has to look like when a supervisor asks.
Dates to be confirmed. Attendance is by invitation; council members are invited automatically.
Payments
Banking
Research university
Research university
Sam Hamilton, former Senior Vice President at Visa, with Konrad Konarski.
With Visa's Head of Consumer Credit and Konrad Konarski.
Payments, banking and academic members. The council is growing and actively recruiting, particularly in risk and regulatory functions.
Senior Vice President (Former)
Vice President
Vice President (Former)
BFSI Consulting Partner · formerly Stevens Institute of Technology
Membership suits people accountable for fraud, risk, data or AI inside a bank, payments network, insurer or asset manager, and researchers working on privacy-preserving methods who want their work tested against production constraints.
Members join quarterly working sessions and one hosted symposium, and take part in the council's research and policy work. In practice this comes to about a day each quarter.
You get peers at competing institutions facing identical threats and identical supervisors, and sessions run so that people can speak plainly.
We ask members to describe what happened accurately, including losses and failed deployments. That is the only reason anyone learns anything here.
[CONFIRM — state dues plainly, or “no fee” if that is the model.]
