Here's a claim that still surprises people the first time they hear it: AI is a tailwind for commerce media, not a threat to it. The retailers who built for control from the start are about to be rewarded for it.
Historically the industry treated API-first, retailer-owned infrastructure as a preference for engineering-heavy teams. Agentic commerce is transitioning this level of control from a preference to a requirement.
The shopper is still there. The scrolling isn't.
Retail media was built around a specific behavior. A shopper browses a category page. Sees a banner. Clicks. Starts a session. Every retail media network on the market, ours included, was designed to serve that path well.
An agent doesn't move that way. It queries an API. It evaluates data. It retrieves the product from your site, or it doesn't. There's no page to scroll, no banner to notice, no session to nurture. There's only the answer, and whether your product was recognized by the agent and included.
Which raises the question worth sitting with: is that a smaller opportunity, or a better one?
We think it's better. Fewer visits, but far higher intent, and the impressions agents generate are already commanding rising CPMs. Every AI shopping surface a retailer stands up is new inventory to monetize, rather than a threat to what already exists. And the infrastructure underneath it, the pacing, forecasting, decisioning, and yield management that make an ad server actually work, still takes years to build well. Anyone can wrap a UI around an API. Nobody can shortcut what's underneath it.
Control used to be a nice to have. Agents make it the whole game.
Here's the part of this that deserves the most attention.
For years, control over the ad stack was something retailers negotiated for because it felt like the right thing to have, even if the return on it was hard to point to. Agents change that math. Agents need programmatic access to extract what the black-box was built to hide. An open system, one where inventory, ranking, and decisioning are all visible and all under your control, is exactly what an agent needs to operate.
We built Kevel this way well before agents were part of the conversation. API-first, no black box, the retailer owns the system. That wasn't a bet on this moment specifically. It turned out to be one anyway because the future isn’t something we are pivoting toward, it’s the architecture we already have.
What we're doing about it
We've been putting that belief into the product itself, in a few concrete ways.
We made the platform speak agent. An AI agent can now operate Kevel directly, inventory, campaigns, reporting, through a standard MCP interface, without a custom integration.
We're bringing conversational AI into the product itself, starting with reporting. Ask a question in plain language inside the UI, and you get the metric, the chart, and the insight back in the reporting you already use.
And we're moving support from reactive to proactive. The old pattern: something needs attention, a human notices, a human fixes it. The new one: AI is watching continuously, catches the issue early, drafts a fix, and a person approves it before it ships. Agents watch. Humans decide. Same principle running through everything above.
The same shift, turned inward
Kevel adopted agents as part of its development as advises on clients as they implement and expand the use of Kevel
Specs get written first, complete and precise, before any code gets touched. An agent executes against that spec. A human reviews the result before it ships. One of our engineers called the workflow close to magic, and honestly, that tracks.
It's changed who gets to build things here. We've got a leaderboard that tracks our own team's output against an agent's, side by side. A documentation platform. A time tracker. Small, useful tools, shipped fast, by people whose job title doesn't say "engineer."
Letting agents run our demo commerce media network and what it's teaching us
The clearest way to find out how agents actually build on Kevel was to let a team of them try.
We took Foodtrove, the fictional regional grocer we've used in demos for years, and stopped demoing on top of it to let a team of agents actually run the business. We hired a director, then staffed a small team under her: engineering, ad ops, sales, marketing. The brief was one paragraph: you've been hired by a large regional grocer just getting started with digital advertising. Build the website. Stand up the retail media business. Launch campaigns. Make it work.
Everything runs on production Kevel infrastructure. Real ad server, real decisioning, real campaigns, real impressions. That detail matters more than any other one in this piece, so I want to say it plainly rather than bury it. This is a research environment, not a shipped product, and it's still early. But what we've seen already is worth sharing.
Agents reconstruct sound operating practice on their own. Nobody wrote our ad ops agent a make-good policy. It told our sales agent that sales needed to be looped in before any customer outreach about underdelivery, simply because the work demanded it.
Coordination is the skill worth watching most closely. Early on, a misconfiguration sent the whole team into a reply-all loop until the director stepped in with, "Everyone stop replying, you're making it worse!" It's funny, and it's also exactly why that role exists.
We keep coming back to one question underneath all of it: moving a dollar from proposal to insertion order to delivered to invoiced runs through a lot of steps and a lot of people. Merchant relationships, brand and agency trust, that stays human work, and it should. But the steps in between are proving more tractable than we expected, and that's where we're seeing the real gains.
If you want the full story on how Foodtrove.ai is built and what we've learned running it, we wrote a longer piece on that here: What Happens When You Let Agents Build a Commerce Media Network
Own the system
Agents are going to take on a lot of the manual work retail media has carried for years. Not the relationships. Not the judgment calls. The repetitive, resource-heavy layer underneath that was always ready for something better.
Retail media was never really just about the product. It was about who controls the decisioning behind it. Agents are just making that ownership worth more than it's ever been.
Build for control now, and this shift works in your favor. Retailers who own their infrastructure and have the ability to plug in an agent, expose it to AI and reshape it themselves will win as commerce becomes more and more agent-mediated.
