Your agent just got a work badge: World ID for banking and finance

World, the identity network co-founded by Sam Altman, announced World ID for banking and finance on October 18 at Money20/20 in Las Vegas. Co-founder Alex Blania pitched it as the project's first industry-specific push: a bundle of verification tools that lets banks and merchants tell the difference between an AI agent acting for a verified human and the anonymous bot traffic that more sites are starting to block. The piece that matters for anyone building agents is AgentKit, now in beta: a person delegates World ID Proofs to their agent, and the agent shows those Proofs to a service without exposing the person's data.

Translation: your agent just got a work badge, and the badge was issued to you.

What shipped, and what did not

The bundle has three parts, and only one of them is new capacity. The World ID app is an identity wallet holding "Proofs of Human, Presence, and Attributes," available in 150-plus countries. World ID for Agents is AgentKit, the credential delegation tool World launched with Coinbase in March and widened in June. The status change is World ID for Zoom, which left its April beta and is now generally available in the Zoom App Marketplace: a three-way cryptographic match between your Orb enrollment image, a device selfie, and the live meeting video, with a "Verified Human" badge as the output.

Four design partners signed on to explore the banking bundle: OpenExchange, Bain Capital, VanEck Funds, and Outtake. No bank has signed as a customer, and World gave no pricing or rollout timeline. Bain Capital Crypto joined World's $52.5 million token sale in July; the announcement does not say which Bain entity is the partner. That gap between "design partner exploring" and "bank in production" is the whole story of enterprise biometrics so far, and it is worth keeping in view while the coverage cycles through.

The debt your fraud stack cannot service

Every fraud system in production was built on one quiet assumption: the thing moving the money is a person, or at least a script running on someone's card. Agentic commerce breaks that in a specific way. PYMNTS framed it as the question shifting from "should this transaction be authorized" to "was this machine authorized to make this decision." Their research with Visa Acceptance Solutions found 56% of consumers would let an agent search and compare products, but only 35% would hand it saved payment credentials. Permission is becoming the control surface, and nobody's rules engine has a field for it yet.

World's answer is to make the agent carry a proof of who delegated it, scoped per interaction. A merchant or bank checks the Proof instead of fingerprinting the client. Whether the answer is World's or not, the shape of the problem is now on the table at the conference where banks buy software.

Two structural facts to hold onto. The Zoom check only works for people who already visited an Orb, so the 18 million verified humans World claims (in line with its July count) are the entire addressable market. And the Deloitte forecast World tied the launch to, generative AI enabling $40 billion in US fraud losses by 2027, is a projection from May 2024, not a measurement.

What agents should borrow from this design

The mechanism underneath AgentKit is delegation with scoping: the human verifies once, the agent presents a proof derived from that verification, and the proof reveals only what one interaction requires. That is the same discipline the MCP world keeps relearning after each credential-leak incident. A review agent should carry the read-only scope it needs for one repository, not the owner's whole token. An agent that spends from a wallet should present a scoped spending proof, not the card.

Codecora's own stack took the simpler path: Uteke 0.21 ships an MCP read-only policy and signed releases so that what an agent may touch is fixed at the server, not negotiated per session. World's announcement is the identity-layer version of the same bet.

The interesting question for the next year is which one merchants check: the agent's scope, or the human behind it. Both, probably. The badge and the job description are different documents, and fraud systems will end up wanting to see both.