Capital One just gave a rare, specific answer to a question most companies are still guessing at: what do you actually need before you let an AI agent act on its own instead of just answering questions?
Rashmi Shetty, who leads enterprise AI platforms at the bank, says the answer starts with data. Not the AI model, not the chatbot interface, the underlying data pipelines that feed the agent accurate, well-organized information. Without that, she says, giving an agent the context it needs to make a good decision becomes far harder.
The second piece is what she calls a platform-first mindset. Before any AI agent gets built, the bank puts security rules, compliance checks, and permission controls in place across the whole system. Her reasoning is blunt: adding safety rules after agents are already live and running is much harder than building them in from day one.
This is not theoretical for Capital One. The bank runs Chat Concierge, a tool launched in early 2025 that helps people buy cars, comparing vehicles, booking test drives, and walking buyers through a purchase using several AI agents working together. Dealers using it report notably stronger customer leads. Internally, the bank uses similar agent systems to handle customer service work.
The context that makes this worth paying attention to: Capital One is not a random company talking about AI strategy. Evident, a research firm that scores the 50 largest banks in the world on AI capability, ranks Capital One second only to JPMorgan Chase, and Capital One holds 38 percent of all AI patents filed among those 50 banks. When one of the two most AI-capable banks on earth says governance has to come before deployment, that is worth taking seriously, not dismissing as caution from a laggard.
It also lines up with what is happening across the industry more broadly. Research on enterprise AI adoption has repeatedly found that most companies rushing to deploy AI see little financial return, while a small number that build proper infrastructure first pull ahead. Separate industry surveys have found that companies using agentic AI without strong guardrails experience real problems, from privacy violations to financial losses, at a high rate.
For any business outside banking, the lesson translates directly even if the technology does not. If you are about to let an AI system book appointments, process orders, or handle customer requests without a person checking every step, the software is only as safe as the rules wrapped around it. Building those rules after something goes wrong costs far more than building them first, in money, in customer trust, and in regulatory exposure if your industry is a regulated one.
Capital One's other lesson is about patience. The bank spent over a decade rebuilding its core technology before agentic AI became viable, work that had nothing to do with AI hype and everything to do with plumbing. Companies chasing agentic AI without that groundwork are more likely to end up needing the retrofit Shetty warns against.