I keep coming back to the same product lesson from an unlikely place: Netflix, the company that trained a generation to believe on-demand was the end state, is reportedly exploring live channels and bundles to lift engagement. The most on-demand product on earth is rediscovering that habit needs a channel. Agentic commerce is running into the same physics, and most teams are still shipping the chat.

The consensus right now says the interface war for AI-mediated commerce will be won by the most fluent conversation. Better prompts, better memory, better tone. My read is that the conversation is the least defensible layer. What holds shoppers, and what turns an agent from a novelty into infrastructure, is a channel: an always-on entry point with persistent state, a defined promise, and a next step that survives the session closing.

That is not what most teams are building. They are building a smarter waiter who forgets your order the moment you walk out.

What a channel actually is in commerce

Strip the word of its media baggage. In an agentic commerce context a channel is four things stacked together.

An entry point that is embedded where the shopper already is. State that persists across discovery, checkout, and post-purchase without asking the buyer to reintroduce themselves. A set of next-best-action rules that route intent to the correct commitment, whether that is a price hold, a shipping cutoff, a reorder, or an escalation. And a disciplined scope: a clear list of what the channel will promise and what it refuses to promise until the underlying systems can carry the weight.

A conversation optimizes for expressiveness. A channel optimizes for commitment. Commerce is a sequence of commitments. That is the mismatch.

The Salesforce evidence is doing more than most people noticed

This week Salesforce moved Shopper, Buyer, and Merchant agents into general availability with native integration into ChatGPT, Google Search AI Mode, and Gemini arriving this summer. The headlines called it another agent launch. The interesting part is buried in the product page.

Agentforce Commerce is pitched as taking a shopper from discovery to checkout and post-purchase with a single conversation, grounded in inventory checks, shipping cutoffs, pricing and tax correctness, and real fulfillment logic. That is not a chatbot description. That is the spec for a stateful channel. The conversation is the surface. The commitments are the product.

Read it that way and the release stops looking like a UI story. It looks like a bid to standardize the channel layer underneath every AI-mediated storefront, before anyone else defines what a durable one should do.

Increase conversion and basket sizes with agent-led guided shopping experiences on your storefronts that take shoppers from discovery to checkout and post-purchase with a single conversation.

That sentence sounds like marketing copy. It is actually a system requirement. A single conversation across three stages only works if the state underneath it survives every interruption a real shopper introduces.

The Netflix mirror, quickly

Netflix has the deepest recommendation graph in consumer software and one of the best on-demand interfaces ever built. It is still adding channels. The reason is simple. Even excellent recommendation systems ask the user to make a choice every session. Channels remove that friction by running whether you show up or not, and by putting habit ahead of preference. The shopper equivalent is a channel that already knows what you are trying to reorder, what is back in stock, and what the merchant is willing to commit to today, without a new prompt.

On-demand is a capability. Channels are a habit. The habit is the moat.

What this makes possible

Once you frame the channel as the primitive, the design work becomes clearer and, in some ways, more constrained in a useful way.

Commerce teams can start shipping agentic surfaces that hold buyer intent across sessions instead of resetting every visit. Merchandising becomes about routing rules rather than page layouts. Post-purchase stops being a support ticket and becomes the next channel state. Loyalty stops being a points program and becomes a persistent context the agent already carries. This is what the adaptive storefront argument was pointing toward, and it is what the handoff-as-product case has been circling. The channel is the object that carries continuity forward.

It also changes what a partnership with an assistant platform means. Native integration with ChatGPT or Google's AI Mode is not distribution in the old sense. It is a channel handoff between your commitments and theirs. If your side of the handoff cannot hold state or honor a promise, the assistant will route around you the second a competitor can.

Why conversations keep losing

A fluent conversation is impressive in a demo and fragile in production. Users close tabs. They come back on a different device. They switch from Google to ChatGPT mid-decision. They abandon and return three days later expecting the price they were quoted. Conversational memory alone does not solve this, because the memory is scoped to the session or the account, not to the commitment.

A channel is scoped to the commitment. That is the difference. And it is why so many agentic commerce pilots stall at the seam between what the model said and what the business can actually deliver. That seam is where integration debt shows up as lost revenue, and it is exactly what the channel layer is supposed to close.

This is also where the discipline shows up. The teams doing this well are ruthless about what their channel will and will not promise this quarter. They would rather ship a narrow channel that always honors its commitments than a broad one that occasionally embarrasses the brand. The scope is the product. Expanding it too fast is the failure mode.

Build the channel first, then the conversation

If I were sequencing an agentic commerce roadmap right now, the order would be inverted from where most teams start.

First, define the channel. Pick a real buyer motion, reorder, restock alerts, price holds, guided configuration, whatever the data says already has habit energy. Wire the state. Define exactly what the agent is allowed to commit to and what it must escalate. Instrument the commitments, not the conversation turns.

Then, and only then, add conversational flexibility on top. Language is the easy part. The models are already good enough. The channel underneath is what almost no one has finished building.

The teams that get this order right will look, from the outside, like they have quieter launches. Fewer viral demos. Fewer screenshots of clever agent replies. What they will have instead is repeat behavior, cleaner attribution across assistant platforms, and a growing library of commitments the agent can honor without escalation. That is the compounding surface.

The closing turn

The question I would ask any team pitching an agentic commerce strategy this quarter is not how good the agent sounds. It is what the channel promises, what it refuses, and what state it carries when the shopper closes the tab.

If the answer is a conversation, you have a feature. If the answer is a channel, you have infrastructure. Netflix figured out that even the best on-demand experience needs a channel layer to build habit. Commerce is about to learn the same lesson, faster, and with money on the line every session.

Build the channel. The conversation will follow.

Sources and further reading