Campaign optimization can now operate at impression level through real-time agent decisioning, representing peak efficiency for advertising technology that traditionally relied on periodic human intervention across broader campaign segments. This isn’t just an incremental update; it’s being framed as the logical endpoint of automation, a significant architectural shift for how digital ad campaigns are managed and fine-tuned.
“Can you decision an impression with an agent as opposed to the way it was done in the past?” Joseph Hirsch, CEO of Swivel, asked at the recent Beet.TV/Horizon Media AI Media Summit. “I think the depth is what I’m seeing as the future, all the way down to the impression level or all the way down to the millisecond level.” This moves beyond the broader campaign segments of yesteryear, where human eyes would scan performance metrics and make adjustments perhaps once a week, or even less frequently. Now, the idea is to have AI agents making micro-decisions on every single ad impression.
What does this even look like in practice? Think about tasks traditionally handled by humans that required wading through lists. Hirsch points to the painstaking process of analyzing hundreds or thousands of app names and bundles to identify top performers and then manually cull underperformers. This new paradigm suggests agents can do that continuously. It’s a move from reactive, batch processing to proactive, continuous micro-management. The implication is that incremental value, often lost in the gaps between human interventions, can now be captured.
The Depth Versus Width Debate
This push for impression-level optimization signals a broader trend in AI development within ad tech: a shift from AI’s trying to do a little bit of everything (width) to AI’s mastering specific, deep workflows (depth). For years, the promise was AI could assist with almost any task. Now, the focus is on AI agents that can execute an entire workflow from start to finish, rather than just a partial assist.
Hirsch articulated this well: “There’s a lot of width. Can AI touch everything? Now we’re starting to see more depth where instead of doing 50% or 75% of the workflow, maybe you’re doing 100% of the workflow.” This means agents aren’t just suggesting improvements; they’re executing them, continuously, at a granular level that humans simply couldn’t manage. The promise is efficiency that scales through sheer frequency of operation.
The Natural Language Layer
Swivel’s approach also includes a natural language interface, allowing users to interact with these sophisticated automations using everyday language. This isn’t just about creating automations; it’s about communicating with data, optimizing campaigns, and even generating publisher yield across multiple ad platforms, rather than being confined to single systems. The ambition is a unified interaction layer for a fragmented ad tech ecosystem.
“If you want to use natural language to create an automation in your business to do any task that humans have done in the past, you can do that in Swivel,” Hirsch explained. “If you’re a seller using one, three, five, seven, 10 platforms, now you can use a singular platform to interact with these ad platforms via agent.” This aims to cut through the complexity of managing disparate platform UIs and API integrations.
Is This Actually ‘Peak Efficiency’?
The term ‘peak efficiency’ is a bold claim. Historically, ad tech has chased efficiency through various means – better targeting, more efficient bidding, faster load times. Impression-level AI decisioning is a significant evolution, pushing automation to its extreme. However, my journalistic skepticism kicks in. The history of ad tech is littered with technological promises that failed to live up to the hype, often due to implementation challenges or unforeseen consequences.
For example, while agents can theoretically perform tasks continuously, are the models sophisticated enough to handle the nuances of every single impression without introducing bias or unintended negative consequences? And how much of this advanced functionality remains untapped because users simply don’t know how to prompt the AI effectively? Hirsch alluded to this, comparing it to unexplored functions in large language models. This implies that the potential for peak efficiency might exist, but the realization depends heavily on user expertise and AI maturity.
“We have tools inside of our platform that have never been touched because there are things that AI platforms can do that people don’t know because they’ve never prompted them to do,” Hirsch said. “People will become more expert driven in how they prompt, and they will discover new things that they can do.”
This suggests that the ‘peak efficiency’ isn’t just about the tech itself, but also about the human element in wielding it. It’s a two-way street: AI needs to be powerful, and users need to be adept at unlocking that power. Without both, even the most granular decisioning might just be a very fast, very complex way of doing things incorrectly.
What does this mean for the average media buyer or planner? It likely means a steeper learning curve, a necessity to become more of a prompt engineer, and a reliance on AI agents that are not just intelligent but demonstrably trustworthy. The transition from periodic human oversight to continuous AI micro-management is a profound architectural shift, one that promises unprecedented optimization but demands a new level of user sophistication and confidence in automated systems. It’s less about replacing humans entirely and more about fundamentally changing the nature of their roles—evolving from direct operators to orchestrators and supervisors of highly specialized AI agents.
This isn’t just about doing things faster; it’s about fundamentally reshaping the operational fabric of digital advertising, pushing automation to its most granular limit. Whether this truly represents ‘peak efficiency’ or another ambitious step on a long road remains to be seen, but the architecture of campaign management is undoubtedly being redrawn at an impression-by-impression level.
