AI agents stuck in pilot mode as banks demand trust, auditability, and on-chain primitives

Banks are keeping AI agents in “pilot mode” because they still do not trust delegated automation in high-stakes finance. Citi is the latest example: on April 22, Citi launched “Citi Sky,” an AI wealth assistant built with Google Cloud and Google DeepMind. Citi’s wealth tech head Dipendra Malhotra said the key limitation is “memory,” asking how long a client conversation can continue before hallucinations appear—an operational risk for advisory and portfolio execution. Industry adoption is uneven. A Deloitte poll of 3,300+ finance professionals found 80.5% expect agentic AI (AI agents and GenAI chatbots) to become standard within five years, but only 13.5% say their organizations are already using AI agents. McKinsey estimates 50% to 60% of bank operations are within AI-agent reach, yet experts warn of “pilot purgatory,” where proofs of concept run without changing operating models. Ethereum is pushing possible infrastructure. Draft standards ERC-8004 and ERC-8183 aim to add on-chain primitives for agent identity, reputation, validation, and escrow/evaluator attestation. The goal is to support verifiable delegation—who the agent is, what it did, and how jobs are funded and completed. Key unresolved questions remain: who is responsible for losses caused by AI agents, whether reputation can be trusted (agents can inflate signals at machine speed), who has control at scale, and what regulatory framework applies when an agent acts outside its scope. The article frames these issues as the central barrier to scaling AI agents in finance.