Quo Vadis AdTech: Targeting and Data in an Agentic World

Agentic AI has quickly become one of the most discussed topics in the AdTech industry. The idea is simple: instead of systems that follow predefined rules, we move toward agents that can understand a goal and take actions to achieve it. In practice, this means applying AI not just to optimize parts of a campaign, but to operate across the entire workflow - from setup, through targeting decisions, to ongoing optimization. This is why the topic is gaining traction. Rather than another incremental improvement, it represents a shift toward applying AI at scale to how the programmatic advertising industry works as a whole. 

In this article, we focus on one specific area where this shift becomes most visible - targeting and data usage.

‍

From manual setup to agent-driven execution

‍

In today’s setup, campaign execution is still largely manual. Advertisers spend significant time selecting audience segments, configuring line items, defining targeting, and optimizing performance over time.

In an agentic model, many of these operational tasks can be delegated to AI agents. In an agentic model, the advertiser's role shifts to defining campaign parameters, such as goals, budget constraints, CPM targets, or brand safety requirements. The agent then takes over campaign execution by selecting audiences, configuring the campaign, adjusting targeting and bidding, and communicating with other agents, such as DSP-side, SSP-side, or data provider agents.

This introduces a new layer: agent-to-agent communication. Instead of a human selecting data and activating it, agents request, evaluate, and combine inputs from other agents to achieve a goal. This sounds efficient, but it raises practical questions about how targeting decisions will be made and how data usage is attributed.

‍

The two pathways of agentic targeting

‍

To understand how agentic AI affects targeting, it is important to look at how decisions are actually made. Agents do not operate in isolation. They rely on available data, prior knowledge, and interactions with other systems. This means that targeting is no longer just about choosing an audience, but how that choice is informed. It is tempting to assume that agentic systems can replace the daunting task of selecting the best segments from those available in data marketplaces with something entirely new.

That is not the case. There are two realistic ways agents can approach targeting today.

‍

1. Selecting from predefined segments.

‍

An agent can browse and evaluate existing off-the-shelf audience segments available within data marketplaces, DSPs, or SSPs. To do this effectively, it needs knowledge of how those segments perform, and this is where two distinct approaches emerge:

‍

•  Using external historical data: ‍

An agent may rely on performance data collected across multiple advertisers and campaigns, for example, aggregated insights held by a DSP or another platform. This means the agent is not learning on its own. It is using knowledge generated by others. This raises a practical question: If performance data from many advertisers is used to guide targeting decisions for a single campaign, is that a fair default for how the ecosystem should work? Data providers may prefer a different model in which advertisers invest in their own testing and build their own performance knowledge, rather than relying on aggregated insights collected elsewhere. Not only because it benefits them financially, but also because the advertisers also collect the context needed to understand how those results were achieved.

• Testing and comparison (self-collected data):

‍Alternatively, an agent can test multiple segments, compare outcomes, and optimize based on its own results. This is effectively a faster version of what humans already do: testing, comparing, and iterating. It is still a form of brute-force exploration, but at a different speed and scale.

‍

In both cases, the agent is not creating something new. It is trying to select the best targeting settings from those that already exist.

‍

2. Requesting custom audiences on demand.

‍

An alternative approach is for agents to request a new audience instead of selecting an existing one. In this model, an agent does not need to manually craft a prompt. It already operates based on a defined goal or campaign brief, which it can pass forward to another system or agent. That input is then used to generate a custom audience tailored to the campaign objective. This eliminates the need to guess which predefined segment fits best or test multiple options blindly. Instead, the audience is created specifically for the desired outcome.

This is where prompt-based audience creation becomes relevant in practice. At PrimeAudience, this approach is already used to generate audiences directly from a campaign brief, enabling faster and more precise targeting without relying solely on predefined segments.

‍

As a result, agents can:

‍

• Generate audiences dynamically.
• Align targeting more closely with campaign goals.
• Reduce the need for manual selection and iteration.

‍

The first gap: data usage and attribution

‍

So far, both approaches - selecting predefined segments and generating custom audiences - assume that agents can access and use data freely to make decisions. This is where the first major issue appears. As agentic systems evolve, they do not just consume data. They start to accumulate knowledge about audiences, performance, and user behavior across multiple campaigns and sources. This brings us back to the earlier scenario.

If signals or performance insights are aggregated across advertisers, it becomes possible to build a centralized knowledge layer about how data performs. That knowledge can then be exposed to agents to guide targeting decisions. While this may improve efficiency, it introduces a new layer of complexity. To understand this challenge, it is worth introducing a concept that is increasingly discussed in the context of agentic systems: Agentic Audiences. It is a relatively new proposal for reshaping the targeting landscape, originally introduced by LiveRamp. But how does it work in practice?

Instead of passing plain segment membership status, systems exchange signals, such as probabilities, scores, or representations of user intent. Rather than saying, "This user belongs to segment X," the information becomes something closer to: "I am 75% sure this person behaves like someone who would be interested in topic X."

‍

In this model:

‍

• Multiple data sources may contribute signals.
• Those signals are combined and interpreted.
• An agent makes the final targeting decision.

‍

At that point, attribution becomes significantly more complex, especially if you are using signals from more than one provider:

‍

• Which signal influenced the outcome?
• Which data source actually mattered?
• How should value be assigned?
• Is the performance of the campaign correlated with how “sure” the signal is?

‍

If several providers contribute signals, but only some of them have a meaningful impact on performance, determining which contribution actually matters is not straightforward. This creates both measurement and data governance challenges and introduces risks that are not yet fully understood. Once data is aggregated, transformed into signals, and reused across multiple decision-making processes, it becomes difficult to predict how it will be used, how it will evolve, and who ultimately benefits from it. This is not just a technical problem but a structural challenge for the entire ecosystem.

‍

Without clear attribution:

‍

• Usage cannot be verified reliably.
• Value becomes harder to assign.
• The system becomes less transparent.

‍

This problem does not exist with simple segment membership because the logic is binary. A user or URL either belongs to a segment or it does not.

‍

The second gap: explainability and trust

‍

Agentic systems make decisions at a scale and speed that humans cannot follow in detail. This creates a simple challenge: If we cannot understand why a decision was made, how do we trust it? In practice, trust comes from two things: results and understanding.

‍

If results are strong, a lack of explanation is often acceptable. But when results are weak, questions appear quickly:

‍

• Why was this audience selected?
• Why did performance drop?
• What should be changed?

‍

Without answers, it becomes difficult to debug issues, improve performance, and ultimately rely on the system. In an ideal scenario, we have both strong results and clear reasoning behind decisions. In reality, however, trade-offs are likely. It may be possible to operate without full explainability, but not without trust and performance. If those two cannot be consistently delivered, the transition to agentic systems is unlikely to succeed.

‍

Where does this leave targeting and data?

‍

Agentic AI introduces a new way of thinking about targeting, but it does not replace existing approaches. Instead, it adds automated decision-making on top of them and introduces new ways to create and use audiences. At the same time, it exposes gaps that need to be addressed, particularly around how data is used, how value is attributed, and how decisions are understood.

Before Agentic AI can be adopted at scale in AdTech, these questions need clear answers. The technology is moving quickly, but trust, transparency, and accountability will determine how far it can go.

‍

‍