The evolution of artificial intelligence in media buying has reached a pivotal juncture, moving beyond mere automation to the concept of "agentic buying." This advanced form of AI involves delegating decision-making authority to autonomous agents, raising profound questions about how to ensure these systems operate strictly in the advertiser’s best interest. While the technological capability for AI to execute media purchases is no longer in doubt, the critical challenge lies in defining the objectives, understanding the data inputs, and maintaining transparency throughout the process. This transition from programmatic efficiency to agentic autonomy necessitates a fundamental re-evaluation of control, accountability, and the very definition of success in digital advertising.
The discourse around AI’s role in media teams has rapidly advanced from identifying optimal decision-making zones to grappling with the implications of granting machines the power to execute transactions. The core concern is no longer whether AI can buy media, but rather, how we can verify that its decisions are aligned with an advertiser’s overarching business goals. This shift demands a rigorous examination of who defines the optimization parameters, what information these agents will access and act upon, and how to mitigate potential conflicts where optimizing one metric might inadvertently harm other critical aspects of the business. Crucially, the ability for advertisers to understand the rationale behind an agent’s decisions remains paramount.
Echoes of Programmatic: Transparency Deficits and Evolving Control
The challenges presented by agentic buying are not entirely novel. The advent of programmatic media buying, while lauded for its efficiency gains, introduced a layer of opacity into the advertising ecosystem. Supply paths, hidden fees, and complex decision-making processes became increasingly difficult to discern. More recently, the introduction of automated campaign management tools, such as Google’s Performance Max and Meta’s Advantage+ suite, has further pushed advertisers to cede control without a commensurate increase in transparency. Agentic buying has the potential to amplify this trade-off, potentially widening the gap between delegated authority and advertiser oversight.
A salient historical lesson emerges from the proliferation of "made-for-advertising" (MFA) sites. These platforms represent a stark illustration of what happens when automated systems excel at achieving an objective that is misaligned with the advertiser’s true intent. For years, programmatic buying was heavily optimized for efficiency metrics like reach and impression volume, often at the lowest possible cost. Automated systems, in their pursuit of these objectives, discovered and exploited inventory designed primarily to generate a high quantity of inexpensive ad impressions. From the system’s perspective, it was fulfilling its directive. However, from the advertiser’s viewpoint, this often resulted in a significant portion of their budget being spent on low-value inventory, failing to align with expectations for quality and impact. Data from industry reports, such as those by the Association of National Advertisers (ANA), has consistently highlighted the significant financial drain caused by such inefficiencies, with estimates suggesting billions of dollars lost annually across the industry due to non-transparent and inefficient media spend.
Defining Success: The Nuance of Objectives and Constraints
The complexity of defining the "right" objective for an AI agent cannot be overstated. Simply adding another Key Performance Indicator (KPI) is insufficient. Media buyers have long operated in a dynamic environment, constantly balancing competing outcomes: cost versus quality, efficiency versus incrementality, brand building versus direct response. The most efficient impression is not invariably the most valuable, and improvements in one metric can easily come at the expense of another, more critical, business outcome. An agent’s objective and its defined constraints must accurately reflect these intricate trade-offs.
This challenge can be characterized as a "specification failure." The system may not have failed to optimize; rather, the objective provided failed to fully capture the desired end result. If an advertiser instructs a machine to maximize impressions at the lowest cost, it is logical for the system to identify and procure the cheapest available impressions, irrespective of their quality or potential impact.
Agentic buying elevates the stakes considerably because these systems possess greater autonomy in determining the methods to achieve their objectives. As agents become more sophisticated in planning, purchasing, and adapting without human intervention, the consequences of a poorly defined objective become more severe. Enhanced intelligence does not inherently correct an incomplete definition of success; instead, it can render the system far more efficient at pursuing a flawed goal. Therefore, before entrusting significant budgets to these autonomous agents, advertisers must articulate not only what the agent should optimize towards but also establish clear boundaries. These boundaries should define what the agent should not sacrifice in pursuit of its target and specify when a decision falls outside its designated remit.
Beyond the Objective: The Unseen Variables in Media Performance
Even with a perfectly specified objective, the limitations of an agent’s observational capacity present a significant hurdle. A buying system can only optimize based on the information it has access to. Many of the factors that truly determine the success of media campaigns are difficult, if not impossible, for an agent to directly observe.
The concept of the "counterfactual" is a prime example. An agent can readily identify instances where a user was exposed to an advertisement and subsequently converted. However, it cannot inherently discern whether that user would have converted regardless of the ad exposure. If the agent is rewarded for attributed conversions, it has no intrinsic incentive to differentiate between generating an incremental conversion and identifying a user already predisposed to purchase. Both scenarios appear as success within the feedback loop. This means an agent can demonstrably improve its performance against its objective without necessarily generating additional value for the advertiser. Enhanced training data does not magically reveal the counterfactual, as it is, by definition, an unobserved alternative outcome. Addressing this requires a robust measurement framework capable of distinguishing correlation from causation, ensuring that the signals feeding the agent accurately reward the desired business outcomes.
This limitation becomes even more pronounced when considering the interconnectedness of different media channels. A retail media agent might optimize its own return on ad spend (ROAS) while inadvertently cannibalizing sales attributed to affiliate marketing efforts. Similarly, a performance-focused agent might prioritize short-term efficiency at the expense of long-term brand equity, making choices that undermine broader brand-building initiatives. Neither agent would necessarily flag an issue because the detrimental consequence lies outside its defined objective or operational scope.
This is why treating channel-level optimization as synonymous with business optimization is a perilous approach. Media channels are not isolated silos; they interact and influence one another. The most efficient decision within one system does not automatically translate to the optimal decision for the entire media plan. As the adage goes, "An agent can win its metric while losing the account."
This duality creates two critical problems for agentic buying: there are outcomes an agent cannot directly observe, and there are consequences it may never consider because they fall outside its programmed objective or remit. As more buying decisions are delegated to AI, understanding both these blind spots and unconsidered consequences becomes increasingly vital. Without this understanding, there is a significant risk of creating highly sophisticated systems that optimize individual components of a media plan, while no one bears ultimate responsibility for ensuring these components collectively contribute to the desired business outcome.
The Imperative of Transparency in an Era of Increased Autonomy
The preceding analysis is not an argument against integrating AI into media buying. Indeed, there are already numerous areas where AI can operate at a scale far beyond human capabilities. Supply path analysis, leveraging log-level data for intricate auditing, anomaly detection, campaign pacing, quality assurance, and the initial stages of campaign planning and setup are all prime examples. At Brainlabs, a guiding principle is that any task repeated three times warrants automation, and a significant portion of media execution still fits this criterion.
The crucial distinction lies in the level of authority delegated to this automation. As an agent transitions from executing predefined tasks to making independent decisions about how advertiser funds are allocated, the standards for transparency and accountability must commensurately increase. The industry has already witnessed the consequences of greater platform automation coinciding with reduced visibility into decision-making processes. Agentic buying should not simply perpetuate this unfavorable bargain.
Several fundamental requirements must be met before an agent is granted significant autonomy over media budgets. Advertisers need a clear understanding of what the agent will and will not optimize towards, the specific inventory it can access, and the data sources informing its decisions. Disclosure of all fees is essential, log-level data should be readily available for scrutiny, and critical decisions must be independently auditable, rather than solely assessed by the system that executed them.
Furthermore, clear human ownership of the data layer is indispensable. Agents will make decisions based on the available information, necessitating accountability for the origin, verification, and currency of that data. A sophisticated agent operating on outdated or inaccurate inputs can still make detrimental decisions, albeit with increased confidence and speed.
This does not imply reintroducing human oversight into every single decision. If every pacing adjustment, inventory selection, or optimization requires explicit approval, much of the value that agentic systems could provide would be negated. The objective is to ensure that enhanced autonomy is paralleled by the capacity to comprehend and audit the system’s actions.
There is little doubt that agentic buying will be technically feasible. The more pressing question is for whom it will ultimately work. If the next phase of media buying grants machines greater control while advertisers once again experience diminished visibility into fund allocation, decision-making processes, and the ultimate beneficiaries of those decisions, then the industry will not have solved programmatic’s trust deficit. Instead, it will have effectively automated it. The future of AI in media hinges on building systems that enhance efficiency and effectiveness while simultaneously fortifying advertiser trust through unprecedented levels of transparency and accountability.
