Agentic AI is gaining attention across finance, but the industry’s biggest obstacle is no longer whether the models are powerful enough. The harder problem is whether banks, asset managers, and treasury desks have the infrastructure to delegate financial tasks to autonomous systems without losing control of money, accountability, or compliance.
A Deloitte poll of more than 3,300 finance and accounting professionals showed the gap clearly: 80.5% said AI-powered tools such as agents and GenAI chatbots could become standard within five years, but only 13.5% said their organizations were already using agentic AI.
Citi Sky showed why the infrastructure debate matters
Citi launched Citi Sky, an AI-powered wealth assistant built with Google Cloud and Google DeepMind technologies, on April 22. The tool was developed using Google’s Gemini Enterprise Agent Platform and is set for a phased rollout to Citigold clients in the U.S. this summer.
The launch gave the agentic AI debate a live banking example. Citi wealth technology head Dipendra Malhotra pointed to memory as a central constraint for high-stakes advisory AI, asking how long a client can keep a conversation going before the system starts hallucinating.
Most agents rely on retrieval-augmented generation to extend memory through external databases. Context windows still cap how much information an agent can hold at once.
In financial advice, treasury management, or portfolio execution, that memory ceiling becomes more than a technical issue. It becomes an operational risk.
MihnChi Park, co-founder of CoinFello, said the conditions for trustworthy delegation are simple: the agent can only act within user instructions, the user can halt it, and the underlying assets never move to a third party.
Ethereum drafts on-chain primitives for agent identity
Ethereum proposal ERC-8004 introduces systems for agent identity, reputation, and validation. The draft standard sets out three registries: an Identity Registry, a Reputation Registry, and a Validation Registry.
Together, they are meant to help autonomous agents prove who they are, build a record of behavior, and support verification by other market participants.
ERC-8183 takes a narrower route. It proposes a job escrow standard with evaluator attestation, where a client funds a job, a provider submits work, and an evaluator completes or rejects the outcome.
The proposal does not provide arbitration or formal dispute resolution, but it gives agent-based markets a framework for escrowed tasks and verifiable completion.
The arXiv paper “The Agent Economy: A Blockchain-Based Foundation for Autonomous AI Agents” maps a five-layer architecture for this shift, covering physical infrastructure, on-chain identity, cognitive tooling, economic settlement, and collective governance.
The reputation layer still carries a structural vulnerability. Agents can generate activity at a speed and scale humans cannot match, making it possible to inflate trust signals over short periods.
That leaves financial institutions with a difficult question: when an agent has a good record, is that record evidence of reliability or just evidence of repeated automated activity?
McKinsey puts 50% to 60% of bank operations in scope
McKinsey estimates 50% to 60% of bank full-time equivalents are tied to operations. Experts warn of “pilot purgatory,” where institutions run narrow proofs of concept without rewiring the operating model.
As Cryptopolitan reported from the Hong Kong Web3 Festival, McKinsey projected that the agentic AI market would grow from $5.25 billion in 2024 to roughly $200 billion by 2034.
Porter Stowell, CEO of W3.io, said: “Enterprises have no way to see, control, or audit what autonomous systems are doing with their money. Human oversight doesn’t disappear. It just moves up the stack.”
Four questions remain unresolved: who is responsible when an AI agent causes financial loss, whether its reputation can be trusted, who is in control once these systems deploy at scale, and what regulatory framework applies when an agent acts outside its scope.
