FAQ

Financial agentic AI, answered

Short answers to the searches this observatory is built for. Longer treatments live in the sector guides and production cases.

What is agentic AI in banking?

Agentic AI in banking is software that can plan a multi-step job and call bank systems to do it — file a dispute, draft a credit memo, change a payment plan — rather than only answering a question. The live deployments are narrow: JPMorgan’s employee LLM Suite and legal workflow, Capital One’s Chat Concierge, CommBank’s payment-dispute tool. A general “AI teller” is still mostly a slide.

How is agentic AI different from a chatbot in finance?

A chatbot generates language. An agent is allowed to take an action with side effects: a refund, a core posting, a KYC update, a portfolio weight. That action needs the same identity, limit, and audit controls as a human in the same seat. If it cannot call a tool, it is a chatbot, whatever the landing page says.

Which banks use agentic AI in production?

Public, named examples as of September 2026 include JPMorgan Chase (LLM Suite at employee scale; Legal Agentic Workflow; research-grade regime agents), Capital One (Chat Concierge), Commonwealth Bank (payment disputes), plus fintech and network rails at Klarna, Mastercard, Visa, Coinbase, and Anthropic’s financial-services agents used in bank partnerships. Survey claims that “70% of banks use agentic AI” mix Copilot seats with production agents — this observatory lists the named ones.

What are the risks of agentic AI in payments?

Unauthorized or poorly authorized payments, weak mandate, missing replay, scheme-rule breaches, and no unwind path. A payments agent must log payer, mandate, amount, merchant, MCC, authorization time, model version, and the human who can reverse it. Networks are building Agent Pay-style authorization specifically because a free-form model must not be the cardholder.

How do you govern financial AI agents?

Put them in the model inventory. Split low-risk and high-risk review. Pin prompts and tool allow-lists. Require maker-checker above thresholds. Log every tool call. Keep a kill switch. EU AI Act and DORA add traceability for in-scope firms; they do not replace existing credit, conduct, or payments rules.

What is a multi-agent architecture in financial services?

A split into channel, orchestrator, specialists, tools, and governance. Specialists (credit, KYC, disputes, legal) are independently testable. The orchestrator routes. Tools talk to cores and networks. Governance can refuse a step. One mega-prompt that “does banking” is not an architecture.

Does agentic AI reduce bank costs?

McKinsey has published 15–20% of cost base as a net AI opportunity, and 30–50% efficiency on workflows that early movers actually rebuilt. Klarna has claimed ~$40 million annual profit impact from its service agent, then later rebalanced toward human quality. Treat firm-level savings as case-specific and usually self-reported.

What is agentic commerce?

A shopping path where an agent finds a product, confirms it, and pays with the consumer’s authorization. Mastercard Agent Connect / Agent Pay, Klarna’s Agentic Product Protocol, and Visa’s embeddable assistant are the 2026 rail-builders. Catalog access is not a completed payment.