Carat and agency peers put human oversight at center of AI media buying

Carat, NBCUniversal and other major agency groups are testing a more automated way to plan and buy media, but the central question is no longer whether AI can recommend inventory. It is who can authorize a decision when that recommendation affects an advertiser's money.

At Advertising Week, agency leaders described an emerging operating model that pairs AI agents with explicit human oversight. The new systems promise faster planning and optimization, but they also force marketers to define where automation ends and accountability begins.

Table of contents

Jump to each section:

Carat brings agents into upfront buying

The agency approaches are diverging

Why human approval is becoming a product feature

What marketers should know about agentic media

Carat brings agents into upfront buying

Carat, NBCUniversal, FreeWheel and Newton have developed an agentic planning and optimization solution for upfront advertising commitments. Their [official announcement](https://az-euw1-p-dentsu-web-01.azurewebsites.net/us/en/media-and-investors/carat-and-nbcuniversal-introduce-bespoke-agentic-capabilities) describes a system combining agency audience intelligence, NBCUniversal identity data and FreeWheel infrastructure to support buying recommendations. An unnamed luxury retail advertiser is the first client.

The distinction between an AI recommendation and an autonomous transaction matters. A media plan can be analyzed quickly without necessarily granting software the authority to commit budget. For marketers, those are different levels of operational risk.

The launch also builds on an earlier NBCUniversal, FreeWheel and Newton initiative with agency RPA involving agent-assisted video buying. The latest collaboration extends the proposition to Carat's client-specific upfront workflows.

The agency approaches are diverging

Other agencies are opening different parts of their operations to AI. Tinuiti introduced a Bliss Point Model Context Protocol server that connects its measurement layer with external agents. Stagwell, meanwhile, positioned Machine OS as part of its relaunched Stagwell Media organization.

These are not interchangeable products. Carat's system addresses media investment planning and optimization. Tinuiti's approach makes measurement intelligence available to agents. Stagwell is presenting agentic technology as an agency operating layer.

The emerging competition is as much about workflow access as algorithm quality. An agent can only make useful decisions when it has access to relevant data, permission to act and a clear definition of the outcome it is meant to optimize.

Why human approval is becoming a product feature

Dentsu platforms executive Rebekah Shalit described oversight as a sliding scale, with repeatable reporting tasks requiring a different degree of supervision from strategic planning. FreeWheel product chief David Dworin similarly pointed to approval steps and specific controls as ways to prevent agents from acting beyond their remit.

WPP Media's Lauren Wetzel made a further distinction between probabilistic AI recommendations and deterministic execution. For decisions with financial or activation consequences, she described approval thresholds and accountable human oversight.

That distinction is becoming a competitive feature, not merely a legal precaution. The more consequential the action, the more important it is for advertisers to understand who can stop, reverse or explain it.

An agent that buys faster is not automatically an agent that buys better. Its value depends on whether the marketer can verify the inputs, challenge the recommendation and measure the outcome independently.

What marketers should know about agentic media

Brands evaluating agency AI systems should examine the decision rights behind the interface, not just the promise of speed.

Separate analysis from activation. Ask whether the agent only recommends media choices or can actually place orders and shift budget.

Specify approval thresholds. Define which spend levels, inventory categories and strategic changes require a named human decision-maker.

Inspect data access. Understand which audience, identity and performance signals shape recommendations and whether the system can compare options across media owners.

Require explainability. Ask for records showing what the agent recommended, what was approved and how the final decision performed.

The strategic challenge is not to remove people from the media-buying process. It is to place human judgment where it changes the quality and accountability of decisions.

As agentic tools move from demonstrations into agency workflows, governance will increasingly be judged by what the software actually permits, rather than by the assurances attached to its launch.