Creator control begins with the offer, the product information they can maintain, the merchant systems they operate and the customer relationship they are permitted to continue. Agent-platform control begins with the conversation that discovers the need, interprets intent and decides which products enter the frame.

Agentic commerce can make a purchase feel as if the entire transaction happened inside one conversation. The shopper asks, compares, chooses and buys without moving through the familiar sequence of search page, product page and checkout page.

But a smoother surface does not erase the handoffs underneath it.

What the current protocols document

OpenAI's current Agentic Commerce documentation separates the conversational surface from the merchant system. Merchants provide structured product feeds containing identifiers, descriptions, pricing, inventory, media and fulfillment information so products can be surfaced accurately.

The checkout can be rendered inside ChatGPT, but the merchant creates and updates the authoritative cart state, validates the order, calculates tax and fulfillment options, processes payment through its existing provider, accepts or declines the order and handles fulfillment and support.

Google's Agent Payments Protocol addresses a different boundary: proof that an agent was authorized to transact. AP2 uses signed mandates to capture a user's intent and, when applicable, the exact cart and price. For delegated tasks, the intent can include conditions such as merchant, timing and spending limits before the agent acts.

These systems are not identical. One describes how a conversational platform and merchant exchange catalog, checkout and order data. The other creates evidence of user authorization across an agent-led payment flow. Both make the hidden handoffs more visible.

Discovery can move without merchant-of-record control moving

OpenAI says participating merchants keep orders, payment processing, fulfillment and customer support on their own systems. That is meaningful practical control. It distinguishes an embedded checkout from a marketplace that becomes the seller.

At the same time, the conversational platform can shape the discovery layer. It sees the shopper's question, the comparisons requested and the path from a broad need toward a particular product. The merchant may receive the information required to complete the order without receiving the whole discovery conversation.

DucimusAI's inference is not that one side secretly owns everything. It is that the economic loop is being divided differently: the agent platform can hold early intent and ranking context while the merchant holds the authoritative cart, payment, fulfillment and post-purchase obligations.

That division matters for creators who sell products, courses, memberships or licensed work. A merchant can remain merchant of record while losing visibility into why the buyer began searching, which alternatives were considered or which language moved the decision.

DucimusAI's perspective: the new storefront has two ledgers

The merchant ledger records what was offered, ordered, paid, fulfilled, refunded and supported. The conversation ledger records what the shopper asked, what the agent inferred, which products were surfaced and how the choice narrowed.

A transaction can cross both ledgers without either side holding the complete story. Signed intent or cart mandates can prove authorization without automatically transferring the discovery history. A product feed can make an offer eligible for discovery without guaranteeing placement, attention or conversion.

There is a reasonable alternative view. Shoppers may benefit when an agent reduces search friction, applies constraints consistently and keeps payment details from spreading across unfamiliar sites. Merchants may gain qualified demand while preserving their existing checkout and fulfillment systems.

The creator question is not whether conversational commerce is good or bad. It is which evidence, relationship and return path the business still controls after the sale.

The Conversation-to-Customer Map

Map one real offer across four handoffs.

Discovery: Where did the shopper express the need, and which system decided that your offer was relevant?

Intent evidence: What preferences, limits or authorization records exist, and which parties can access them?

Cart and payment authority: Which system defines the authoritative price, inventory, tax, fulfillment options, payment acceptance and order status?

Relationship and return: Who receives the customer information needed for fulfillment and support? What permission exists for future contact? Where will the shopper begin the next search?

Mark each answer as creator-controlled, merchant-controlled, platform-controlled, shared or unknown. Do not use the location of the buy button as a shortcut for the whole architecture.

A specific creator decision

Choose one offer you want an AI agent to understand. Make the public product facts accurate and machine-readable: title, description, price, availability, variants, images, fulfillment terms and canonical URL.

Then protect the layers the feed does not solve. Keep the product page and policy pages under your control. Preserve order and support records in the merchant system. Give customers a clear, voluntary way to return directly when they want updates, without treating a completed order as permission for unrelated outreach.

Finally, measure the two ledgers separately. Track what the merchant can verify from order data and what remains unknown about the upstream discovery conversation. That gap is not proof of wrongdoing. It is the boundary to design around.

The checkout can move into the conversation while the merchant still controls the transaction. The larger shift is that the conversation may become the new front door—and the business may not own the path that led the customer there.