This episode offers an exceptional look into the practical applications of agentic AI within the banking sector. Drew Sievers details how AI is currently being deployed in "unglamorous but critical" areas like fraud detection, AML, and customer support, providing clear ROI. He also casts a vision for future banking where AI orchestrates permissions and identity, serving autonomous software agents, ultimately allowing banks to operate with a fraction of current headcount.
The conversation covers three main areas: Drew's path from advertising to mFoundry to a stint in investing and back to operating. Drew traces his route from Ogilvy and Saatchi & Saatchi through five years in Japan to founding mFoundry in 2003, before the iPhone existed, to solve mobile banking's two big problems: handset fragmentation and carrier control. The turning point was landing Citi as a client, which pushed him to drop every other vertical (astrology apps, wine guides, Yellow Pages) and bet the company entirely on banking. Within a few years mFoundry had a third of the top 20 US banks as clients before selling to FIS. He's candid that the investing chapter he tried next was his biggest career regret; he now wishes he'd taken months off before jumping back in, which eventually led him back to operating, most recently at Drift. Where agentic AI is actually landing in banking: Drew splits this into near-term and long-term. Near-term, the real activity is in unglamorous, controlled environments: fraud, AML, compliance, reconciliation, onboarding, and customer support, where ROI is provable and regulators are comfortable with augmentation. He cites research showing 88% of finance leaders would allow some form of agentic AI in banking workflows and 71% say AI-initiated connectivity now factors into their choice of bank. Longer-term, he argues the real battleground is orchestration. Whoever controls permissions, identity, and auditability wins. And that banks will eventually serve autonomous software agents, not just people in apps. He also reframes the "AI has to be perfect" objection: human-run banking workflows already run 10-15% error rates, so an AI system that halves costs while improving to 8% error is an easy call for any bank. Could a bank run with a fraction of the headcount, and is AI overhyped: Dan and Drew debate whether a $10B depository institution could be built and run with roughly 50 employees. Drew's answer is yes: building a modern core is n