Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Ai Financial Inclusion topic
No spam. Unsubscribe anytime.
Fed governor and panelists say AI can broaden financial inclusion but needs testing and guardrails
Summary
A Federal Reserve governor and technology, nonprofit and academic leaders told a Fed convening that AI tools can expand access to financial guidance for low- and moderate-income consumers but require rigorous, empirically based testing, human-in-the-loop designs and policy-level guardrails to prevent harms.
Governor Cook, a Federal Reserve governor and FOMC member, moderated a panel on artificial intelligence and financial inclusion and said the technology could demystify financial tools but raised urgent questions about accountability, bias and consumer data protection. ‘‘What happens when AI tools fail? Who is accountable?’’ Cook asked, framing the discussion around the Fed’s interest in equitable access and its dual mandate.
Laura Blattner, head of Impact at the Bike Shop Applied AI Lab at MIT, told the panel that AI products often fail for design—not technology—reasons and recommended designing systems to augment trusted human counselors rather than replacing them. Blattner said teams should introduce AI assistance once a counselor has established trust with a client, and use ‘‘applied evaluations’’ (or ‘‘evals’’) made from realistic, difficult cases to test whether a tool efficiently gathers relevant information and reasons correctly compared with experienced advisers.
Sarah of SaverLife, a nonprofit fintech, described the organization’s AI "navigator" and said early deployment reached a large roll‑out and produced promising engagement and outcome signals: members who received AI-generated recommendations were reported by the panel as about 10 times more likely to act on those recommendations, and the panel cited a roughly $200 median increase in savings among engaged users. She said SaverLife initially limited recommendations to trusted sources and avoided referral partnerships that would create conflicts of interest.
Sanjay Subramanian of PwC said modern models are enabling firms to serve ‘‘edge cases’’ that traditional systems excluded by synthesizing alternative data and coaching users, but stressed that model providers must be transparent about data sources, run rigorous testing and be prepared to monitor performance in production. He noted both an efficiency play—making narrow‑margin products cheaper to operate—and an access play driven by new data and models.
Panelists emphasized an iterative rollout approach: start with small, controlled experiments; use humans to review targeted cases; build automated monitors; and expand only after confirming reliability. Blattner warned that synthetic data can help but must be generated from varied, representative seed cases to avoid simply copying limited conversations.
On policy and coordination, Cook said the Fed has used roundtables to bring stakeholders together and recommended more cross‑sector dialogue. Blattner suggested regulation could play a role similar to how model risk management became codified, while panelists urged sandboxes, shared evaluation frameworks and continued human oversight.
The session closed with Cook thanking the guests and reiterating the Fed’s convening role. No formal actions or votes were recorded.
