While the U.S. banking industry continues debating AI in mostly theoretical terms, other markets are getting an uncomfortable preview of what happens when automation meets real-world infrastructure. In an illustration of AI working at capacities that stretch the limits of the physical world, Alibaba’s chatbot Qwen recently offered users an AI-driven promotional coupon for Boba tea that immediately broke China’s gig-worker food delivery system. Ten million orders materialized faster than businesses could process them, forcing Alibaba to halt the entire promotion.
This wasn’t a sophisticated attack or deliberate sabotage; it was simply scale meeting automation, and the results were immediate chaos. Banks should pay attention. The same dynamic that overwhelmed Chinese logistics networks will hit the financial infrastructure, except the consequences won’t be delayed packages and tepid tea. They will be fraudulent transfers, system failures, and customer panic at speeds we’ve never experienced.
When Velocity Becomes Vulnerability
Silicon Valley Bank’s collapse taught us that the infrastructure we built to serve customers faster can also destroy institutions faster. SVB didn’t fail over weeks like historical bank runs; it lost US$42 billion in deposits in a single day. Social media transformed uncertainty into panic while instant payment rails executed the exodus before management could even realize what was happening.
Now add agentic AI to this equation. We are moving from AI as a tool humans use to AI as an agent that acts on our behalf. This creates the potential for a “double-sided run.” Imagine AI agents that can simultaneously generate a panic signal AND execute the bank run. Commercial treasury systems with automated liquidity management will react to algorithmic triggers. Retail customers will use AI assistants that monitor account safety and move funds instantly based on social media sentiment analysis.
The technology banks celebrate—real-time payments, API banking, and instant settlement—becomes the transmission mechanism for AI-amplified contagion. Unlike SVB, where human panic had some natural friction, AI-driven runs will operate at machine speed across thousands of institutions simultaneously. This cascade could occur due to malevolent actors or simply through overactive systems creating a spiral of unintended consequences.
The Epistemological Crisis is Already Here
Velocity of funds is only half the crisis. The deeper issue is that customers are rapidly losing the ability to verify what is real. Epistemology—how we understand and verify truth—is shifting at such a pace that most institutions are still operating as “business as usual.” Sources of truth are becoming corrupted as consumers and businesses learn to adapt to a world of deepfakes and rumors evolving at unparalleled speed.
For retail banking customers, every trust signal is now compromised. AI-generated phishing emails are indistinguishable from legitimate bank communications, and voice cloning defeats phone verification. When your customer service rep or your fraud alert could be an AI-generated fake, what channel remains?
The commercial side is arguably worse. CFOs already receive dozens of fraudulent invoices weekly. Now, those invoices are perfect replicas, down to vendor terminology and approval workflows. Treasury managers face payment authorization requests using cloned executive voices that capture speech patterns and background noise. The entire verification infrastructure of commercial banking assumes humans can eventually validate authenticity through multiple channels. That assumption is breaking.
Infrastructure Wasn’t Built for This
The Alibaba incident showed what happens when AI automation hits systems designed for human-paced activity. Banking faces an asymmetric infrastructure problem: our front-end (AI agents and instant rails) moves at machine speed, but our back-end (fraud reconciliation and legal recourse) still moves at human speed.
Payment rails built for predictable volume will face AI agents executing thousands of micro-decisions per second. Commercial card programs could see transaction volume that break existing reconciliation processes. Even nontransaction banking becomes impossible; try booking a meeting when AI scheduling assistants are competing for every available slot. Much the same way that scalpers weaponized bots to corner ticketed events, agentic AI has the potential to overwhelm the banking system.
Smart Latency Can Help Financial Institutions Adapt
Banks that treat AI as a capability to add rather than a systemic risk to address are making a grave mistake. Financial institutions need to redesign authentication beyond any channel AI can replicate.
Most critically, banks need “smart latency,” circuit breakers for velocity-based risk. The lesson from SVB wasn’t just that runs happen faster, it’s that modern infrastructure provides no natural stopping points. Just as high-frequency trading uses “speed bumps” to prevent flash crashes, banking needs automatic mechanisms that create friction without requiring human intervention. We must distinguish legitimate automation from attacks without breaking the customer experience, creating “intentional friction” for high-velocity, agent-initiated transfers.
Trust Infrastructure IS Bank Infrastructure
The Alibaba boba tea incident was a small promotional offer that revealed fundamental fragility. Banking’s equivalent won’t be delayed packages—it’ll be fraudulent payments executed at scale and customer panic amplified by synthetic signals.
The real challenge isn’t whether to adopt AI; it’s whether banking systems can maintain trustworthiness when customers can no longer verify reality and capital moves at algorithmic speed. Some institutions will solve this; others will discover too late that trust infrastructure is banking infrastructure, and you can’t rebuild it during a crisis. The epistemological crisis is coming. The question is which banks will still be trusted when it arrives.
For more thoughts and insights on the use of AI in commercial banking, please reach me at [email protected].
