Explain It in 30 Seconds
A buyer approves a purchase or invoice payment.
Software can help execute it, but the merchant needs to recognize the agent and the payment system needs an appropriate credential.
The buyer also needs the original limits to survive the automated interaction.
Visa supplies capabilities around those requirements.
A recognized agent, a valid payment credential and permission for a particular purchase answer three different questions. A useful system needs all three to fit together. [S21][S22][S95]
Where It Sits in Visa
Intelligent Commerce connects agent-facing controls to existing credential, acceptance and payment infrastructure within Visa.
Issuers retain the accounts; merchants retain commercial relationships. Tokenization and agent recognition perform different jobs beneath the automated interaction. [S29][S95]
The Problem It Solved
Online merchants have spent years distinguishing legitimate customers from unwanted automated activity.
An authorized shopping agent creates a different situation.
The software is automated, but the buyer may genuinely want it to act.
Simply treating every bot as a customer would be unsafe. Treating every automated interaction as hostile can also block a permitted transaction.
The business problem is recognizable delegation.
The merchant needs information about the agent and the person or organization behind it. The buyer needs control over what has actually been authorized.
Visa's Trusted Agent Protocol addresses that recognition problem using signed information about intent, consumer recognition and optional payment information. [S95]
How It Works — Simple Version
A simplified authorized-agent payment has several stages.
1. The buyer supplies a permitted instruction. 2. The agent operates within that instruction. 3. The merchant or portal recognizes the supported agent interaction. 4. An eligible payment credential is presented. 5. The transaction follows the supported acceptance and processing path. 6. Both sides receive information needed to understand the outcome.
The sequence describes the operating logic, not a universal protocol implementation.
An agent can be recognized without being permitted to make every purchase.
Likewise, a payment credential can be valid while a proposed amount or recipient falls outside the buyer's instruction.
Three Controls With Different Jobs
Agent recognition
Identify the interacting software through the supported framework.
Payment credential
Connect the payment to an eligible account relationship.
Transaction scope
Preserve the buyer's permitted amount, recipient and action.
These controls are complementary. Recognized software with a valid credential can still propose an action outside its authority.
Intelligent Commerce Connect
Visa introduced Connect in April 2026 as a selected-partner pilot through Visa Acceptance Platform.
It supported Visa and non-Visa payment APIs and different protocol/token-vault arrangements. [S21]
The integration reduces separate connection work. It does not supply buyer consent, enroll every merchant or establish universal production availability.
Connect is a capability within the wider portfolio, not the whole offering.
A Live Business-Payment Example
The October 20, 2026 BILL and Billtrust announcement describes a live pilot with participating buyers and suppliers.
The registered BILL Payment Agent uses tokenized credentials and controls aligned with approved payment instructions.
Billtrust's supplier side recognizes the agent and accepts the payment according to the supplier's policy. [S22]
The agent does not decide which obligations the buyer should pay.
It executes eligible payments already approved by the buyer.
That makes the example commercially specific: it automates a remaining execution step across buyer and supplier systems.
It does not establish that every BILL customer or supplier invoice is already covered.
Who Uses It?
Buyers
They want supported work completed without losing control over the payment instruction.
Agent builders
They need a way to connect their software with recognized payment and merchant capabilities.
Merchants and supplier platforms
They need to distinguish authorized activity from abuse while preserving their own acceptance policy.
Payment providers
They connect credentials, controls and processing to the agent's execution.
The visible agent can belong to one company while the underlying account and payment services come from others.
That division of work is central to the product.
How the Economics Work
The opportunity is less repetitive entry and a traceable execution process, weighed against integration, maintenance and exceptions.
Invoice principal is not Visa's revenue. The announcements disclose no universal agentic-payment fee or standalone margin.
Completing a pilot payment proves a bounded workflow, not the cost of supporting every customer and failure mode.
What It Depends On
A valid instruction
A request for recommendations is not necessarily permission to buy.
The buyer's approved scope must remain understandable.
Merchant participation
The merchant or supplier platform must support the interaction and retain its own controls.
Agent recognition is not permission to ignore those controls.
Eligible credentials and accounts
The underlying payment relationship still exists.
An agent does not create its own unrestricted source of funds.
Exceptions and human handoffs
An unavailable item, changed invoice or uncertain result can require another decision.
A permitted implementation needs a way to stop or escalate rather than invent new authority.
Those handoffs are part of making automation usable.
What Depends on Visa Intelligent Commerce
Participating platforms can use selected recognition, credential and control capabilities.
That supports the payment interaction without certifying the AI model's wider recommendations or resolving every purchasing decision and commercial dispute.
Why It Matters to Visa
The portfolio seeks to preserve Visa's infrastructure role as software acts between buyer and merchant.
It relies on recognizable authority and payment connections, not ownership of every agent.
Protocols and pilots establish implementation steps; broad commercial adoption still requires participant cooperation.
Before / After
Manual portal execution
An employee may need to log into a supplier portal and enter the payment information after an invoice has already been approved.
Supported agent execution
The recognized agent can complete the eligible approved step while the supplier retains its acceptance policy and both sides retain visibility. [S22]
The changed constraint is repetitive execution.
The business has not delegated unrestricted financial judgment to the software.
What Most People Misunderstand
“The AI chooses every bill to pay.”
The documented business pilot executes eligible buyer-approved payments.
“Recognizing an agent proves a purchase is authorized.”
Identity and transaction authority are separate questions.
“Connect is the whole Intelligent Commerce portfolio.”
It is an integration capability within the broader offering.
“A live pilot means worldwide general availability.”
The announcements identify particular participants and readiness stages.
“The merchant loses its own acceptance controls.”
The supplier policy remains part of the documented workflow. [S22]
Key Milestones
October 2025 — Trusted Agent Protocol
Visa described an ecosystem framework for recognizing legitimate agent interactions, with phased scope. [S95]
April 2026 — Connect pilot
The selected-partner integration pilot broadened the implementation proposition. [S21]
October 2026 — BILL and Billtrust live pilot
The announcement established an actual bounded business-payment workflow rather than only a simulated demonstration. [S22]
Evidence
- [S95] Agent recognition, intent and the protocol's original phased scope.
- [S21] Connect's integration design and selected pilot stage.
- [S22] The live business-payment sequence and explicit buyer-approval boundary.
- [S29], [S40] Distinct credential and checkout layers.
The control and operating-economics discussion connects those facts. It does not certify an AI model's judgment or every participant's implementation.