Don’t build the agent. Build what the agent buys.

Media has been late to every tech shift. This time, being early doesn't mean what you think.

It is a Tuesday morning in 2027. A media buyer at a mid-sized agency opens a chat window and types a brief: reach EV-intenders aged 30 to 50 in West Sweden, brand-safe news environments, under a set CPM, live by Friday.

There is no RFP. No phone call to your sales team. No media kit attached to an email.

The buyer's agent goes out and asks the market a question. Seller agents answer. Within minutes, budgets are allocated to the publishers who could answer clearly, with data the agent could read, verify and price.

Your inventory was either one of those answers, or it was invisible.

Over the past editions of the Samhub Signal, I have written about first-party data as the publisher's revenue moat, what personalization actually means, the revenue stack, and why pageviews are lying to you. The thread running through all of it has been the same: know your audience, own your data.

This edition is about who is going to read that data next. Increasingly, it will not be people. It will be machines.

We have seen this film before

Media has a pattern, and most of us in the industry know it well. We are rarely first to a technology shift. We are often the ones who pay for it.

When programmatic arrived, publishers treated it as a channel for remnant inventory. A decade later, a large share of every advertising euro was being taken by the intermediaries in between, and the relationship with the buyer had moved to someone else's platform.

When social distribution arrived, many newsrooms reorganised around Facebook's reach. Some of them pivoted to video because the platform said so. When the platform changed its mind, the audience went with it.

When search became the front door to news, publishers optimised for an algorithm they did not control. Now that answer engines summarise the article instead of linking to it, that traffic is quietly draining away.

Here is what I think is the real lesson. In each of these cases, being late to the technology was not what hurt. What hurt was not owning the asset the technology ran on: the audience relationship, the data, the direct line to the buyer.

The publishers who came through those shifts in good shape were not necessarily the fastest adopters. They were the ones who kept control of what made them valuable while everyone else rushed to the new thing.

Agentic advertising is the next shift. And the question of what to own matters more than the question of what to build.

What is actually being built, and why nobody knows the end state

I want to be honest about the uncertainty here, because I think the trade press swings between "agents will replace media buying next year" and "this is just programmatic with a new name." Neither is useful if you are the one deciding where to put next year's budget.

What is true: the plumbing for agents to buy and sell advertising is being built right now, in public, and fast.

In October 2025, a group including Yahoo, PubMatic, Scope3 and Optable launched the Ad Context Protocol, AdCP. It is built for AI agents from the ground up and covers the whole campaign lifecycle, from discovering inventory to buying, creative and reporting. It is now governed by an industry association, AgenticAdvertising.org, and is reported to be in production at companies such as Vox Media and Netflix.

In January 2026, IAB Tech Lab launched its own framework, AAMP. Rather than starting fresh, it layers agents on top of the rails the industry already runs on: OpenRTB, OpenDirect, VAST. Version 2.0 arrived in June with buyer and seller agent toolkits.

What is also true: the two are not coordinated. IAB Tech Lab has positioned AAMP as an alternative to AdCP rather than a complement. The large agency groups are running their own agent-to-agent buys on their own terms. And every SSP and DSP is shipping its own "agentic" layer.

So if you are a media house asking which standard will win, which vendor to bet on, or what the final architecture will look like, the honest answer is: nobody knows yet. Anyone who tells you otherwise is selling something.

That is exactly why this is a hard investment decision. Build too early and you may build on the wrong standard. Wait too long and the market gets shaped without you, just like it did with programmatic.

Wait or move? That is the wrong question

When I talk to commercial directors about agentic, the conversation almost always lands on the same binary: do we invest now, or do we wait until the dust settles?

I think that framing is the problem. "Agentic" is not one investment. It is at least three, and they carry very different levels of risk.

Layer 1: The data layer. No regret. Invest now. 
Whatever protocol wins, an agent will need to understand who your audience is, what context they are in, and what a placement is likely to deliver. That requires consented, structured, first-party audience data. This holds under AdCP, under AAMP, and under whatever comes after both. It is the one part of the agentic future that does not depend on guessing the winner.

Layer 2: The access layer. Partner, do not build. 
Your inventory needs to be discoverable by buyer agents. For most media houses, this will come through the SSPs, sales houses and platforms you already work with, many of whom are shipping seller agents right now. Your job is not to write that code. It is to make sure your partners expose your inventory and your data properly, and on your terms.

Layer 3: Your own agents. Experiment, do not commit. 
Building your own seller agent, or an agent-driven sales desk, can make sense for the largest groups. For most regional and national publishers, it is a bet on a standard that has not settled. Run small pilots, learn, and keep the investment proportional until the market chooses its rails.

Put simply: waiting is a perfectly reasonable strategy for layer 3. It is an expensive one for layer 1.

The data layer takes the longest to build. Consent flows, identity, audience definitions and clean taxonomies cannot be switched on in a quarter. And the data improves with every month you have been collecting it. 

A publisher who starts in 2027 does not catch up with one who started in 2025. They stay two years behind.

Data is the currency agents trade in

Here is the shift I think most media houses underestimate. Agents do not buy banners. They buy answers.

A human buyer can be persuaded by a strong brand, a good lunch and a well-designed media kit. An agent cannot. It asks structured questions: who is this audience, in what context, at what price, with what expected outcome, and can I verify it? If your data cannot answer those questions in a form a machine can read, your inventory does not get considered. It does not even get rejected. It simply never shows up.

That changes what "premium" means. In an agentic market, premium is not the size of your brand. It is the clarity of your signal.

Look at where the largest players are putting their money. Omnicom now owns Acxiom. WPP bought InfoSum. Publicis agreed this year to buy LiveRamp, and framed the deal explicitly as fuel for smarter agents. The holding companies are not mainly buying agent builders. They are buying data. They understand that the agent is the interface, but the data is the asset.

Publishers should draw the same conclusion. The agent will increasingly be a commodity, offered by your SSP, your ad server or your sales house. Your audience understanding will not be.

Remember the COPE example from INMA Berlin that I wrote about in edition #6: advertisers moved performance budgets from Google to COPE because COPE could tell them something about their audience that the black box could not. In an agentic market, that advantage gets stronger, not weaker. The agent will find the publisher with the clearest answer, every time.

There is also a European angle here that I think works in our favour. Agents will need to verify not just who an audience is, but whether the data behind it was collected lawfully. Consent that is GDPR-native, documented and machine-readable stops being a compliance cost and becomes a trust signal. Nordic and DACH publishers who have done the work on consent are, perhaps for the first time, ahead of the curve rather than behind it.

Three things to do before you build a single agent

Unify your first-party data. 
Most media houses I meet have their audience spread across five systems: the subscription platform, the CMP, the ad server, the newsletter tool and the analytics stack. None of them agree on who the reader is. An agent cannot read five half-answers. Bring them together into one consented view of your audience before anything else.

Make your audiences machine-readable. 
Define your segments in standard taxonomies, describe them precisely, and attach the signals that prove them: reach, recency, consent status, context. If a human planner needs a phone call to understand what you are selling, an agent will not understand it at all. The test is simple: could a machine buy this segment without asking you a question?

Write your rules of engagement. 
Decide now what agents are allowed to buy, at what floor price, in which environments, and with what data attached. At INMA Berlin this year, I heard media leaders say about AI crawlers: do not let them in the door until you are sure there is something in it for you. The same principle applies to buyer agents. Agentic should not become programmatic all over again, with your value captured by whoever sits in the middle.

The common thread

Agents will change how media is bought. That part is close to certain. Which protocol wins, which vendors survive, and what the final architecture looks like is still wide open.

But whatever happens, the agent will need something to buy. And the publishers who get bought will be the ones whose data gives the clearest answer.

Media has been late to almost every technology shift. This time, being early does not mean building the agent. It means building what the agent buys.

Until the next edition.