What “agentic” means on a bank floor
A chatbot answers. An agent is allowed to do something irreversible: open a dispute, change a payment plan, pull a credit decision, or hand a file to a human with a recommended action. That permission — a tool call against a core, a card system, or a CRM — is the whole difference.
McKinsey’s December 2025 operations paper on AI in Asian banking estimates that AI could cut certain cost categories by as much as 70%, and 15–20% of a bank’s entire cost base once technology spend is netted. Those are category estimates, not a promise that any one agent will land there. Early movers, McKinsey says, are seeing 30–50% efficiency gains on the workflows they actually rebuilt.
A 2026 MIT Technology Review Insights survey of 250 banking leaders, cited by Open Banking Tracker, found 70% of executives saying their institution is already using agentic AI. “Using” still ranges from a Copilot seat to a production dispute bot. This observatory prefers named tools with a public architecture over survey self-report.
What is actually in production
JPMorgan’s LLM Suite is the scale story on the employee side: ~200,000 people, in a year, with a registered-model control plane. Capital One’s Chat Concierge is the rare customer-facing multi-agent: a car-buying flow, not a general assistant. CommBank’s payment-dispute agent is the copyable operations pattern — high volume, written rules, predictable failure.
Underwriting and collections get talked about constantly and disclosed almost never. Treat vendor decks about “autonomous credit” as proposals until a named book of business, a hold-out sample, and a model-risk opinion exist.
Where these systems break
Identity, mandate, and “who authorized this transfer” do not go away because the channel is an agent. A service agent that can issue a refund is a payments agent. A concierge that can start a loan is a credit agent. Both inherit the same controls as a human in that seat: maker-checker, limit, audit, and a path to unwind.
The useful design is five layers: channel, orchestrator, specialists, tools, governance. Skip governance and you have a demo. Skip specialists and you have a chatbot with extra latency.