Omnicom closed its $13.5 billion acquisition of IPG last November, creating a company with pro forma revenue above $25 billion. In the months since, the agencies moving fastest on AI have been mid-size independents. They have engineering talent and development budgets, and they don't have six layers of approval between an idea and a shipped product. 

Speed of execution now decides who wins with AI, and holding companies are built in a way that makes them slow. 

Look at what they are building. Omnicom folded IPG's Acxiom into its Omni platform this year, built on 2.6 billion identity records. Publicis is doing the same with its pending acquisition of LiveRamp, pairing it with Epsilon's data assets. A global brand running hundreds of campaigns across markets wants one data layer instead of a dozen, and these platforms deliver exactly that. 

Assembling them across separately acquired agencies takes years and demands standardized approval chains to stay reliable. Nobody wants a shared system spanning dozens of countries to break because one team pushed something untested. So new AI capabilities route through the platform before reaching a client, and the platform sets the pace for everyone downstream. How teams work with these tools changes every few weeks. A review process guarding a multibillion-dollar system cannot turn around on that clock, and it was never designed to. 

The spending has been drifting for years. The six largest holding companies’ share of U.S. ad spending fell to 29.6% in Q1 2024 from 44.6% in 2019, according Advertiser Perceptions data cited by Emarketer. That slide started well before the current AI wave and has kept going. 

Being small is no guarantee either. AI coding tools now let a two-person shop prompt its way to something that looks finished in a weekend, and looking finished is not the same as working. Anything a client depends on needs staging and production environments, real data warehousing, and monitoring that catches failures before customers do. Vibe coding produces a convincing demo with nothing underneath it. Agentic engineering treats the system as infrastructure and builds accordingly. 

So the agencies ahead right now sit in the middle. They have enough engineering discipline to build something that holds up, and they are small enough to approve it in one meeting. In practice, that looks like a few specific habits. They work in cross-functional pods rather than routing requests between siloed departments. They build a working prototype before writing the brief about it, because a rough version that runs settles an argument faster than a deck. And they test new capabilities on internal work first, so the judgment calls are already made by the time anything carries a client's name. 

None of this erases what the holding companies own. Acxiom's identity graph and LiveRamp's data network are assets no mid-size shop can replicate, and that will hold for a long time. But owning infrastructure is a different thing from moving on it. Scale buys access to a capability. It does not buy the speed to ship that capability while it still matters. 

Omnicom and Publicis will be running the largest networks in this industry five years from now. Scale, owned data, and global reach are not going away. But the work that changes what AI in advertising looks like will come out of smaller rooms, where the build starts the same afternoon as the idea.