Social Rocket
AI agents are increasingly taking over larger parts of planning, buying and optimization in advertising with real budgets. The systems are still far from fully autonomous media buying, but current projects show that the industry is already making the step from assistance to execution.
First real budgets run through agents
ADWEEK reports that Apostra handled around 1,2 million US dollars in automated media buys for major brands over the past three months. The agents can take over tasks such as discovery, negotiation, approval and execution across multiple channels.
A concrete performance example comes from PubMatic. According to ADWEEK, the wellness brand Rouge Care invested 25.000 US dollars in a CTV campaign and generated more than 125.000 US dollars in attributable revenue. The purchase ran through PubMatic's AgenticOS.
This makes the development more concrete than the IAB Europe findings on agentic media buying we had already seen. There, 58 percent of respondents expected agentic systems to be in operational use or at greater scale within a year.
Monks lets agents work autonomously for 36 to 48 hours
Agency work is changing too. S4 Capital's Monks is letting AI agents work on internal software tasks for 36 to 48 hours without ongoing human supervision, according to Chief AI and Revenue Officer Wesley ter Haar.
These long autonomous runs currently mainly concern coding. Monks is, however, already applying the principle to social and performance marketing. There, agents are meant to react to signals such as trends, comment volume or declining posts.
For early 2027, Monks plans, according to Digiday, a first client workflow that is to run from insights through to community management on this agentic architecture.
Token costs become part of the talent budget
Ter Haar now describes compute and model costs as part of the talent budget. Teams therefore have to decide which tasks need an expensive top-tier model and where cheaper models are sufficient.
This changes the traditional calculation of agency work. Alongside working hours, there are ongoing costs for agents, models and tool use. At the same time, the time required for certain tasks is falling. Monks describes internal work cycles that have shrunk from two weeks to two days.
The question of control also lies behind our overview of AI in media buying and platform automation.
Standards for Agentic Media Buying
In September, the IAB Tech Lab presented AAMP 3.0. The OpenProposal specification is meant to turn the still heavily manual process from RFP to purchase commitment into an agent-readable format.
This allows buyer and seller agents to exchange and compare offers in a structured way. At the same time, governance and interoperability remain central issues, because several protocols and platforms are emerging in parallel.
Apostra is an example of this new coordination layer. The company, formerly Scope3, now aligns its business model specifically with agentic media buying. Learn more in our background on Scope3 and Apostra.
What Marketers Should Keep in Mind Now
Agentic media buying is not yet a self-runner. ADWEEK cites Gartner data showing that highly autonomous applications 2025 accounted for less than five percent of marketing use cases. For 2028, 15 percent is expected.
For companies, three points are therefore more important than maximum autonomy: clear budget limits, traceable decisions, and a defined human escalation model. The more an agent is allowed to reallocate, buy or publish on its own, the more important the question of who is responsible for wrong decisions becomes.