6+ agentic ad buying tests reveal per-client AI governance is now mandatory

Agentic media buying, the use of AI agents to autonomously execute programmatic ad purchases, is transitioning from broad experimentation to a fundamentally client-specific configuration challenge. Independent media agency Butler/Till, which has conducted over half a dozen tests since May 2026 across CTV, online video, display, and streaming audio, reports that the governance frameworks for these AI systems must be rebuilt for each client, each large language model (LLM) platform, and each advertising channel. Scott Ensign, chief strategy officer at Butler/Till, emphasized in a Digiday Podcast interview that the future requires an enterprise AI layer where 'those environments are going to have to be controlled on a client by client basis.' This insight reveals a deeper tension: while agentic AI promises efficiency and scale, the reality is that brands' risk tolerances, regulatory environments, and performance expectations vary so dramatically that a one-size-fits-all agent is not viable. Butler/Till’s client roster—spanning pharmaceuticals, medical devices, financial services, agriculture, and alcohol—each demands 'completely error-free work,' reflecting high stakes in compliance and brand safety. Currently, only a low single-digit percentage of total media spend flows through these AI agents, indicating cautious adoption. However, Ensign predicts this will reach double digits as guardrails mature. The agency is actively building these safeguards, but the process is iterative and labor-intensive, involving not just technical setup but deep partnership with clients to define which actions the AI can take autonomously versus those requiring human approval. This stands in contrast to the initial hype around fully autonomous 'set it and forget it' media buying. The need for client-level customization introduces a new layer of complexity for agencies, requiring them to become AI configurators rather than merely adopters. It also raises questions about scalability: if each client demands a bespoke agent setup, the overhead could limit the cost-saving benefits that agentic AI was supposed to provide. Moreover, the reliance on different LLM platforms (the specific ones were not disclosed but implied by ‘LLM platform or channel expansion’) means agencies must navigate varying API capabilities, cost structures, and model behaviors, further fragmenting the landscape. This fragmentation is likely to slow the consolidation of adtech AI into a few dominant vendors, instead fostering a multi-platform ecosystem where integration services become as critical as the models themselves. Ensign’s call for industry-wide collaboration—sharing best practices rather than 'racing to the bottom'—hints at an emerging need for standards bodies or open frameworks for agentic governance. Without such cooperation, each agency will incur duplicative costs, and the collective learning curve will be steep. For publishers and platforms, this means they must offer highly configurable API endpoints and detailed audit logs to support client-specific rules. The implications extend to the workforce: media buyers will evolve into AI supervisors, requiring skills in prompt engineering, risk calibration, and real-time system monitoring. In the long term, the successful deployment of agentic media buying may hinge less on the AI's raw intelligence and more on the sophistication of the control layer wrapping it. The industry is moving from 'can AI buy ads?' to 'can we trust AI to buy ads for this specific brand in this specific context?' That shift, while more mundane than sci-fi visions, is the real path to adoption.