The news
The Capital One agentic AI strategy puts data before agents, Rashmi Shetty, a vice president of enterprise AI at the bank, told CIO Dive in remarks reported by Banking Dive on September 28, 2026. That day, Plaid announced a partnership with AI support-agent maker Decagon.
Agentic AI refers to systems that can take actions, such as moving money or changing records, rather than only answering questions. "Providing agent context becomes that much easier with a very strong data foundation," Shetty said. She argued that agentic AI has to be treated as an end-to-end system, and warned that layering governance and security onto a patchwork of agents after the fact is much harder than preparing a common platform up front.
Shetty laid out the controls she sees as necessary. A harness, a common software layer around agents, lets an enterprise standardize controls such as tool permissions, which set what an agent is allowed to do. She said enterprises should validate agents through rule-based checks, sandbox simulations and evaluations, known as evals, and plan where humans step in on high-risk actions. At Capital One, she said, observability tools track agent trajectories, tool accuracy and end-to-end latency.
Capital One already runs agents with customers. Banking Dive reported that Chat Concierge, which helps car shoppers connect with dealers and request test drives, was one of the bank's first multiagent applications and has been in production for more than two years. On Capital One's second-quarter earnings call in July, Chief Executive Richard Fairbank said the bank continues to invest in foundational capabilities, AI infrastructure and specific AI experiences.
Plaid's deal targets the same problem from the infrastructure side. Under the partnership, Decagon's AI customer support agents can connect to consumer-permissioned bank accounts through Plaid without sending customers out of the conversation, according to a September 28 Plaid blog post. Agents can verify accounts, check a transfer's status, complete payments, look up balances and keep context across sessions for tasks such as insurance claims, loan applications and disputes.
Consumers grant access through Plaid Link and can review or disconnect connections through Plaid Portal, Plaid said. The post named no customers and gave no launch timeline. AI agent company Sierra announced a similar partnership with Plaid on August 3, 2026, covering tasks such as loan refinancing, payments and fraud resolution.
The numbers
- Time Capital One's Chat Concierge has been in production
- More than two years
- Plaid-Decagon partnership announced
- September 28, 2026
- Plaid-Sierra partnership announced
- August 3, 2026
Why CEOs should care
For bank CIOs, CTOs and chief data officers, Capital One's message is about sequencing. Before funding agent projects in individual business lines, map what data each agent needs, who owns it and whether it is clean enough to act on. Then build one shared platform for permissions, logging and testing, because Shetty warns that adding those controls to scattered agents after the fact is much harder.
For CISOs, chief risk officers and compliance heads, the Plaid-Decagon setup means a third-party agent can complete a payment inside a chat. Treat agent vendors like any other critical third party: ask which actions each agent can take, which require human approval, how consent is recorded, and who is liable when an agent moves money in error. Capital One's use of trajectories and tool-accuracy tracking is a useful benchmark for the audit trail you should expect from any vendor.
For CFOs and boards, neither Capital One nor Plaid disclosed cost, adoption or accuracy figures. That makes pilot metrics essential: resolution rates, error and escalation rates, and time saved per case. Chat Concierge's two-year run suggests starting with a narrow, customer-facing task that has clear boundaries before handing agents higher-risk actions.
The bigger picture
Bank connectivity for AI agents is turning into its own infrastructure layer. With Sierra in August and Decagon in September, Plaid now offers in-conversation account access to at least two agent platforms. Sierra, in its post on the Plaid deal, cast conversations with AI agents as the newest front door to financial services, succeeding physical and digital channels.
That sets up a strategic choice for banks. Capital One is building its own agents on its own data platform. Other institutions may meet customers through third-party agents that reach accounts via data networks like Plaid. Either way, the controls Shetty described, from tool permissions to human review, become the terms banks will need to set or accept.
What’s next
Watch for Plaid and Decagon to name launch customers and availability dates, and for Capital One to describe new agentic uses on its next earnings call. Also watch whether bank supervisors issue expectations for AI agents that initiate payments or change account records, which would turn current voluntary guardrails into requirements.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
