Agentic media buying is real now. The wall isn't the AI, it's the account model.

Three things happened in advertising this week, and the trade press filed them as separate stories. NBCUniversal and FreeWheel completed what’s being called a first AI-agent-led programmatic guaranteed deal, where the agent negotiated and executed the buy. A major holding company confirmed it has agents placing live media buys, not piloting them. And OpenAI began rolling out conversion-optimized campaigns on ChatGPT Ads, the first time that surface has had a performance layer rather than just an awareness one.

Filed separately, each is a “look, the agents are here” story. Filed together, they’re the same story, and it’s not the one the headlines are telling. The headline story is that AI can now make the buying decision. That’s true, and it’s the least interesting thing about this week. The buying decision was always going to fall to a model eventually. Bidding into an auction is a closed problem with a clear objective function. It is exactly the kind of thing machines beat humans at, and the capability is commoditizing on a quarterly clock.

I’ve spent the last stretch building an agent that constructs campaigns across networks. Not bidding, building. Standing up the campaign, the ad sets, the targeting, the creative assignment, the conversion wiring, across more than one platform from a single instruction. And the thing I keep running into has nothing to do with whether the model is smart enough. The model is smart enough. The wall is structural, and it’s worth describing precisely, because it’s the part that decides who can actually operate in this era and who just gets to talk about it.

Every network is a different country

Here is what the demos skip. When an agent builds a campaign across networks, it is not working against one system. It is working against three or four or seven systems that share almost nothing. Each network has its own account model, the structure of who owns the ad account and how clients and sub-accounts nest underneath it. Each has its own API surface, with its own auth flow, its own object hierarchy, its own idea of what a “campaign” even contains. And each has its own half-formed notion of how an autonomous agent is supposed to identify itself and act on someone’s behalf.

There is no shared passport. An agent that can fluently operate inside one network’s account structure hits a wall the moment it crosses into the next one, because the next one models the world differently. The targeting primitive that exists on one platform has no clean equivalent on another. The conversion object that one network expects you to define up front, another expects you to attach after the fact. The agent doesn’t fail because it’s confused about marketing. It fails because the territory it’s crossing has no common language, and someone has to do the unglamorous work of translation for every border.

This is exactly the fight playing out one level up, in the standards argument over how agents authenticate and act across services. There are competing proposals for how an agent proves who it is and what it’s allowed to do when it shows up at a platform it doesn’t own. That fight is unresolved, and until it resolves, every cross-network agent is carrying a stack of bespoke translations instead of a passport. The intelligence is free. The plumbing is the entire job.

The account model wasn’t built for you

If you only buy for yourself, the previous section is an annoyance, not a wall. You stand up your accounts once, eat the integration cost once, and the agent operates inside a structure that fits you.

Agencies don’t have that luxury, and this is where the newest surfaces get genuinely awkward. The AI-conversation ad inventory everyone is rushing toward is shaped for the direct advertiser. One brand, one account, one entity that owns the relationship and the billing and the data. That is a perfectly reasonable way to build a young ad product. It is also the exact shape an agency cannot use, because an agency is a single operator holding the accounts of many clients, each of whom owns their own spend, their own data, and their own compliance posture.

So the agency hits a question the press releases never mention. Which account holds the buy? If the surface only really supports one advertiser per account, does each client get their own, and who administers the sprawl? If the agency holds a master account, whose data is commingled inside it, and is that even allowed under the client’s contract? The agent can build the campaign in thirty seconds. The question of which legal and structural container that campaign is allowed to live in is not a thirty-second question, and on the newest surfaces it frequently doesn’t have a clean answer yet.

We’re running the ChatGPT Ads beta, so this isn’t theoretical. The performance capability is real and it’s improving fast. The account structure underneath it is still shaped for a world with one advertiser per account, and the managed-services operator has to improvise the structure that the platform hasn’t built yet. That improvisation, not the bidding, is the work.

What this means for who wins

The instinct in a week like this is to react to the capability. The agents can buy, so we need an agent that can buy. That instinct is correct and also nearly worthless, because everyone is going to have an agent that can buy, the same way everyone eventually had a programmatic seat. Capability that arrives for everyone on the same quarterly cadence is not a moat. It’s table stakes that haven’t finished being set.

The durable advantage is one level down, in the part that doesn’t commoditize. Account structure that’s clean across networks. A cross-network identity story that lets an agent act on behalf of many clients without commingling what shouldn’t be commingled. Conversion wiring that you own, so the agent has a consistent signal to optimize against no matter which network it’s operating in. These are not exciting. They are not demoable. They are exactly the kind of structural plumbing that gets deferred in favor of the shiny capability, and they are exactly what determines whether the shiny capability can run at scale or just in a screenshot.

The operators who win the agentic-buying era are the ones who treated this as an infrastructure question before the agents arrived. They figured out, ahead of the rush, how their accounts nest, how identity flows across platforms they don’t own, and where the deterministic conversion signal lives. When the agent showed up, it had a clean territory to operate in. The operators who treated it as a capability question are standing up the agent now and discovering that the agent’s first task is to navigate a map nobody drew.

The objection worth taking seriously

The strong counter-argument is that the platforms will fix this themselves. The account models are awkward today because the products are young. The standards fight will resolve, a shared passport will emerge, the networks will build agency-shaped account structures because agencies are too much spend to ignore, and in two years the plumbing I’m describing will be a solved layer you don’t think about. Why build around a problem the platforms are racing to eliminate?

Partly true, and it’s why I wouldn’t tell anyone to build a permanent business on today’s specific gaps. But notice the shape of the bet. “The platforms will eventually make cross-network operation clean” is a prediction about a future where every operator gets the same clean substrate at the same time. If that’s the world, then the buying agent really is the whole game, and it’s a commodity, and there’s no advantage anywhere. I don’t think that’s the world, for the same reason the deterministic layers underneath advertising have stayed scarce through every prior wave. The platforms will make their own surfaces cleaner. They have no incentive to make the cross-network seam clean, because the seam between networks is precisely where no single network wants you to have a unified, portable, network-independent view. The passport that works everywhere is the one nobody who owns a border wants to issue.

So the plumbing stays valuable exactly where it’s hardest, which is across the networks rather than inside any one of them. That’s the part to own.

The takeaway

Agentic media buying became real this week in the only way that counts, which is live deals with real money instead of stage demos with hypothetical budgets. The correct operator response is not to be impressed by the buying and not to panic about the capability. It’s to ask the boring question underneath the exciting one. When the agent can build the campaign anywhere in thirty seconds, the constraint moves to the thing the agent can’t manufacture for itself: a clean account structure, a working cross-network identity, and a conversion signal you own. The intelligence came in the box. The territory it has to operate in is something you either built in advance or are scrambling to draw now.

The wall isn’t the AI. The AI is the easy part. The wall is the account model, and the operators who already climbed it are the ones who get to use this week’s news instead of just reading it.

Agentic media buying stopped being a panel topic this week. NBCU and FreeWheel closed a first agent-led programmatic deal, Omnicom confirmed live agentic buys, ChatGPT Ads turned on conversion campaigns.

Having built the agent, the bottleneck isn't the buying decision. It's that every network is a different country and there's no shared passport yet.

https://joelcitron.com/posts/2026-06-12-agentic-media-buying-account-model/

This week agentic media buying stopped being a conference panel and became live deals. An AI agent negotiated and executed a programmatic guaranteed buy between a major broadcaster and its sell-side platform. A holding company confirmed agents are placing live buys. The new AI-conversation ad surface turned on conversion-optimized campaigns.

The reflexive read is "the machines are bidding now." That's the boring part, and it's the part that's commoditizing fastest.

I've spent the last stretch building an agent that constructs campaigns across networks, and the actual bottleneck is unglamorous. Every network is a different country. Its own account model, its own API, its own half-finished idea of how an autonomous agent should authenticate and act. There is no shared passport yet. The standards fight over how agents identify themselves across platforms is exactly this argument, and it is unresolved.

It gets worse for agencies specifically. The surfaces everyone is most excited about are built for the direct advertiser, not for the operator holding dozens of clients' accounts. So "the agent buys the media" runs straight into "which account can even hold the buy, and who owns it."

The operators who win the agentic-buying era won't be the ones with the cleverest bidding agent. They'll be the ones who solved account structure and cross-network identity before the agent showed up, because that's the part that doesn't come in the box.