Glossary

Words this field keeps mixing up

Financial agentic AI borrowed language from labs, vendors, and risk committees. These are the senses used on this site.

Agentic AI

A system that can plan steps and call tools (cores, CRMs, payment APIs, document stores) to complete a job, rather than only generating text. In finance the tool call is the product: refund, dispute, memo, allocation, KYC update.

Not a synonym for “chatbot,” “Copilot,” or “LLM.” Those can be components of an agent.

Agentic commerce

A purchase path where an AI agent discovers a product, confirms details, and completes payment with the consumer’s authorization. Networks (Mastercard Agent Pay / Agent Connect, Visa assistants, Klarna’s product protocol) are racing to be the rail.

Discovery without authorization is a catalog. Authorization without an audit trail is a risk event.

Multi-agent architecture

A design that splits work across an orchestrator and specialist agents instead of one prompt that “does banking.” Specialists are independently testable; the orchestrator should not also hold irreversible tools.

Capital One Chat Concierge is a public example of this split on the customer side.

Containment rate

Share of intents resolved without a human. Useful, and easy to game by making it hard to reach a person. Pair it with handoff quality, repeat-contact rate, and loss given error.

Klarna’s first-year figures are a containment story. The second year is a quality-of-handoff story.

Maker-checker

Dual control: one party proposes a posting, another disposes. For agents, the maker is often the model and the checker is a human or a second, independently prompted model above a dollar or risk threshold.

The oldest control in banking operations, and still the right one.

Model risk

The risk that a model is wrong, misused, or silently drifts. Agents inherit it. A planner that calls a credit model is in the model inventory, even if a vendor branded it an “agent.”

US banking practice sits on SR 11-7 / OCC 2011-12 lines; the EU adds AI Act and DORA traceability.

MCP (Model Context Protocol)

A standard way for a model to call tools and data sources. Coinbase’s Payments MCP is a payments-shaped instance: the agent gets a typed payments server instead of a browser window.

MCP is plumbing. Policy on that plumbing is the control.

Orchestrator

The planner that breaks a user job into steps and assigns them to specialists. It should route and constrain, not quietly wire money.

If your diagram has no orchestrator, you have a single prompt with extra arrows.

Human-in-the-loop

A required human gate on defined actions (above a limit, on a reason code, on a new merchant, on a SAR). Not a slogan — a workflow state with an owner and an SLA.

Human-on-the-loop (watch and interrupt) is weaker; use it only where unwind is cheap.

Trace / replay

A hop-by-hop record of what the agent saw, which tool it called, and what it posted, so a case can be replayed in an audit. Without replay you cannot explain a payment, a decline, or a credit decision.

This is the operational meaning of “explainability” in a bank.