Visa · Flagship
The Visa Model of Find & Buy Agentics
The claim
A Find & Buy Agent is not a chatbot with a card. It is a principal’s mandate carried by software. Visa’s model is the first network design that treats that sentence as a payments problem: prove who the agent is, replace the card number with a bounded token, and let the network decline anything the mandate did not allow.
Find & Buy Agentics is the name for that split. Finding can be open, probabilistic, and curious. Buying cannot. The moment capital moves, the agent stops being a researcher and becomes a fiduciary. Visa Intelligent Commerce, the Trusted Agent Protocol, and the mandate object are how a card network draws that line without asking the agent to impersonate a person.
Why agentics is a payments problem
Most of the public conversation about shopping agents is about the find. Which model reads the page. Which one compares ten retailers. Which one understands a photo of a lamp. That work matters, and it is unfinished. It is also not the part that breaks commercial law.
The break happens at buy. A card network spent fifty years assuming a human was present, or at least that a human had typed a number into a form the merchant could challenge. An autonomous agent violates that assumption on purpose. If the network keeps treating the agent as either a human or an attack, every honest agent payment looks like fraud, and every fraud looks like a helpful assistant.
Visa’s chief executive, Ryan McInerney, put the obstacle in plain language in late October 2026: trust is proving an agent’s identity, linking it to a customer, and confirming the transaction stays inside the customer’s instructions. That is not a model-quality problem. It is a credential problem. Visa said it was already working with nearly fifty AI partners in production, and that more than half of its 5.2 billion credentials were tokenized, with tokenized credentials on more than 60 percent of e-commerce transactions. The rails were ready before the agents were trustworthy.
The split Visa already enforces
Find & Buy Agentics borrows a distinction the network has started to encode. In the find state, the agent may search, rank, and propose. No token is requested. No merchant is charged. A wrong guess costs compute, not money. In the buy state, the agent may act only inside a mandate the principal can read: what may be bought, a ceiling, a clock, and sometimes a merchant or a category.
Visa’s own developer example is almost a Find & Buy sentence: find and purchase these supplies, and keep the total under $100. Initiating that instruction lets Visa Acceptance Platform check that the agent is still inside the customer’s intent. The mandate is not a prompt. The fields that have been described publicly are a mandate identifier, a human-readable description, an effective-until time, and a decline threshold in an amount and a currency. A model can misunderstand a jacket size. It cannot quietly raise the decline threshold. The network holds that number.
| Find | Buy | |
|---|---|---|
| Risk | No capital moves | A real authorization |
| Engine | Open reasoning over catalogs, links, and photos | Deterministic checks against a mandate |
| Credential | None | A network token, not a 16-digit number |
| Who can say no | The principal, before approval | The issuer, the merchant, and Visa, at authorization |
| Failure mode | A bad shortlist | An out-of-bounds charge, which should decline |
Three checks, and the one that is still yours
A useful way to read the stack is as three checks the network can perform, and one it should not pretend to perform.
- The merchant verifies the agent. The Trusted Agent Protocol, introduced with Cloudflare on 14 October 2025 and built on HTTP message signatures, lets a merchant accept a signed agent request instead of a shared password or a scraped session.
- The issuer verifies the token. The agent never holds the primary account number. It requests a token scoped to the purchase. In June 2026 Visa described richer token data — transaction type, where the token is used, and who is paying — plus an assurance signal drawn from how the token was provisioned and how it has behaved.
- Visa applies the consumer’s limits. The decline threshold, the expiry, and the instruction the principal actually signed are network facts. If the authorization does not fit, it does not clear.
The fourth check is the principal’s, and no network should steal it. Did this purchase match the purpose, not only the ceiling? “Under $100” does not mean “any $99 object.” A Find & Buy Agent on this site already refuses to buy until a person approves. That approval is the consumer-side half of the mandate. Visa can enforce the ceiling. The agent’s operator still has to show the person what will be bought.
What the October 2026 pilot actually proved
On 20 October 2026 Visa, BILL, and Billtrust launched a live pilot in which a trusted agent initiates and completes approved invoice payments through supplier portals. The BILL Payment Agent is registered with Visa’s Trusted Agent Protocol and an agentic directory. It uses tokenized credentials and controls aligned to the buyer’s approved instructions. Billtrust then recognizes that agent and lets it into the supplier portal, instead of demanding a human login that made automation impossible.
Read the pilot narrowly, which is how it deserves to be read. It is business invoice payment, not a consumer hunting for a cheaper television. The “find” has already happened: an invoice exists, a buyer has approved it, a supplier expects the money. The agent’s job is execution under an instruction, with both sides able to see who acted. Darren Parslow, Visa’s global head of commercial solutions, described the surrounding shift as businesses automating the work around payments — identifying supplier opportunities, approving invoices, and executing — while keeping control. That is Find & Buy Agentics in a procurement office. It is the proof that the buy engine can be live. It is not proof that open-web discovery is solved.
How a Find & Buy Agent should sit on the model
The interface can stay human and small. Type the need, paste a link, or hand over a photo. That is find. The agent returns a shortlist and a price. Nothing is bought. That pause is not a missing feature. It is the moment a mandate is born.
Approval should compile into something a network can hold: a description a person can read, a ceiling, an expiry, and enough merchant detail that a later dispute can compare what was authorized with what was charged. The payment credential, when this is a card purchase, should be a Visa token created for that mandate, not a number stored in the agent’s context. If the agent is paying a metered API rather than a merchant of record, the buy engine may be x402 instead. The find engine does not get to choose the rail by improvising. The principal’s instruction does.
What this model is not
- It is not a claim that FindBuyAgent.com is a Visa product, a partner integration, or an issuer. The model is the standard. The product is the ask-first agent.
- It is not permission for an agent to wander a card number through prompts. The number stays at the issuer.
- It is not the end of shopping. It is the end of using the same interface for a chore and for a pleasure. Replenishment can be mandated. A Saturday in a bookshop should not be.
- It is not settled law. The October pilot shows a path for approved B2B payments. Consumer retail, returns, and disputes when a model buys the wrong size are still being written, and they should be written against the mandate, not against a screenshot of a chat.
The award, if there is one, is not for a cleverer search box. It is for refusing to let the cleverness touch the money until the instruction is narrow enough for a network to enforce. That is the Visa model of Find & Buy Agentics. Find in the open. Buy inside a mandate. Decline the rest.
Published by FindBuyAgent.com. Canonical URL findbuyagent.com/blog/visa-model-of-find-and-buy-agentics.