AI Agents Unsuitable for Real-Time Programmatic Ad Buying - Episode Hero Image

AI Agents Unsuitable for Real-Time Programmatic Ad Buying

The Digiday Podcast · · Listen to Original Episode →
Original Title:

TL;DR

  • AI agents are not suitable for programmatic ad buying due to their propensity for hallucination and inability to handle the complex, real-time variables of bid streams, risking brand reputation.
  • Large language models powering AI agents are too slow for the bid stream's 100-millisecond window, making them impractical for real-time ad bidding compared to deterministic machine learning models.
  • Protocols like Agentic RTB aim to standardize and reduce latency for programmatic systems by co-locating logic, but current LLMs are too large and slow to be containerized effectively within this framework.
  • Major ad tech platforms like Google and Trade Desk are unlikely to support new protocols that enable agents to shift budgets between DSPs, as it undermines their business models.
  • AI excels in programmatic ad buying for tasks like audience ideation and repetitive campaign setup, acting as a tool for human oversight rather than autonomous execution.
  • The lack of widespread adoption of agentic AI in daily media buying suggests the problem it aims to solve is not yet a critical industry crisis, unlike the issues that led to OpenRTB.
  • Companies must train employees on appropriate AI use, emphasizing human oversight and testing purposes, rather than blindly adopting AI for all tasks, to avoid errors and ensure quality.

Deep Dive

The programmatic advertising industry is not yet ready to entrust its core functions to AI agents, as current large language models (LLMs) lack the speed, control, and deterministic accuracy required for real-time bidding and campaign activation. While AI offers significant potential for tasks like ideation, summarization, and repetitive data manipulation, the complexities and risks associated with automated bidding and creative adjustments necessitate a cautious, human-in-the-loop approach for the foreseeable future.

The primary argument against AI agents in programmatic ad buying centers on their fundamental limitations in speed and control. The bid stream, a critical component of programmatic advertising, operates within extremely tight latency windows, often 100 milliseconds or less. Current LLMs, due to their complex parsing, computation, and mapping processes, cannot meet these speed requirements, even with optimizations like containerization or co-location. This speed deficit means that AI agents cannot effectively participate in real-time bidding, a core function where split-second decisions determine ad placement and cost. Furthermore, brands prioritize control over their image and reputation, making them hesitant to delegate critical decisions to AI agents that can "hallucinate" or make unpredictable errors, potentially damaging a brand's legacy. While platforms like Meta's Advantage+ and Google's Performance Max leverage machine learning, they employ highly controlled, deterministic models for specific tasks rather than general LLM agents, and these systems are distinct from the conversational AI typically associated with "agents."

Beyond the bid stream, AI's application in programmatic advertising is more nuanced. Creative work has seen the most disruption, with AI significantly speeding up iteration and reducing resource needs. However, even here, real-time adjustments within the bid stream are considered too risky for brands due to the potential for error. Instead, AI is proving valuable in the planning and reporting stages, assisting with audience ideation, generating campaign reports, and automating repetitive tasks. These applications leverage AI as a tool to enhance human capabilities, providing insights and efficiency gains without ceding critical decision-making authority. The development of new protocols like Agentic RTB aims to standardize communication and reduce latency by containerizing logic, but these are intended for structured, deterministic models, not the current generation of LLMs. The adoption of such protocols is further hampered by the lack of support from major platforms like Google, Meta, and The Trade Desk, who benefit from the current fragmented ecosystem, limiting the widespread adoption of universal standards that could empower smaller players.

Ultimately, the case against AI agents in programmatic buying is a call for measured adoption, focusing on areas where AI can augment human expertise rather than replace it entirely. While AI has the potential to revolutionize aspects of advertising, its current limitations in speed, control, and deterministic accuracy mean that human oversight remains indispensable for critical functions like real-time bidding and brand reputation management. The future of AI in this space likely involves a gradual integration into planning, analysis, and specific, well-defined optimization tasks, with a continued reliance on human expertise for strategic decision-making and risk mitigation.

Action Items

  • Audit AI agent use: Define 3-5 critical programmatic functions where AI agents are currently deployed and assess their performance against defined KPIs.
  • Develop AI agent governance framework: Establish 5-7 guidelines for AI agent deployment in programmatic, focusing on control, auditability, and risk mitigation.
  • Create AI training modules: Design 3-4 training sessions for media planners on responsible AI agent utilization, emphasizing critical evaluation and human oversight.
  • Evaluate AI-driven creative impact: Track performance of AI-generated creative assets across 10-15 campaigns to measure effectiveness and identify potential brand risks.
  • Implement AI for audience ideation: Utilize AI tools to generate 5-10 audience segment hypotheses per campaign, followed by human validation and refinement.

Related Episodes

Key Quotes

"I think there was like multiple agency execs who kind of made the point of like these ai agents right now are good for stuff that you would otherwise hand off to an intern."

This quote highlights a common sentiment among agency executives regarding the current capabilities of AI agents in programmatic ad buying. The speaker, Seb Joseph, suggests that these agents are best suited for tasks that are typically assigned to junior staff or interns, implying a limitation in their sophistication and reliability for more critical functions.

"The case against ai agents in programmatic advertising is no one can agree on what an ai agent is but like the creative side of things so we have ai generated creative now which seems like the next leap from dynamic creative optimization of you have different assets whether it be fonts actual text colors actual like graphic assets and dynamic creative optimization is we're just going to these are used up in different ways based on what the audience is or whatever other like instructions are given but now we have ai generated creative what's the parallel there that you would apply to programmatic buying."

Christopher Francia points out a fundamental challenge in discussing AI agents in programmatic advertising: a lack of consensus on their definition. He contrasts this ambiguity with the clearer evolution of AI in creative generation, moving from dynamic creative optimization to AI-generated creative, and questions how this progression might parallel programmatic buying.

"The bid stream is a very small window of time to communicate with it I think a dsp has maybe up to 100 milliseconds so your fastest llm can't even get close to that I don't know as gemini this morning gemini was just like a really optimized one can be 10 milliseconds now this maybe hallucination it's a little bit of hallucination because you have to think of it's not just speed but volume so you have to have it take that but also doing 100 000 a million requests per second."

Francia explains a key technical limitation for AI agents in real-time programmatic bidding. He details the extremely tight time constraints of the bid stream, noting that even advanced Large Language Models (LLMs) struggle to process the necessary volume and speed of requests within the available milliseconds, making them currently unsuitable for direct bid stream interaction.

"The agentic rtb framework is really there to standardize plugins doing whatever they need to do and recommending they co load them let them install them as containerized apps on the servers that are running the bidder or the multiple servers running the bidder so you reduce latency so if you do that then you're dealing with only internal latencies you're maybe gaining 80 efficiency from just making an outside call so that's really what that's for and that's a huge value I think that's actually a really good protocol for that but those containers are not going to be an llm that's at the current state of the industry it's just not fast enough and you couldn't containerize the size of a container for an llm would be insane."

Francia discusses the Agentic RTB framework as a solution for reducing latency in programmatic advertising. He explains that this framework aims to standardize and co-locate plugins as containerized applications on bidder servers, which significantly speeds up communication. However, he reiterates that LLMs, due to their size and processing requirements, are not yet suitable for this type of containerization within the current industry state.

"The issue is they're all trying to solve a problem that doesn't yet exist right we don't have agents demanding to talk to other computers right now I mean we can ask everyone in the room I don't think the majority here are using agentic ai daily to buy media all over the place it is is it I'm curious to show of hands is anyone using agentic ai for your ad buys not a single hand."

Francia argues that many proposed protocols for agentic AI in advertising are attempting to solve a problem that does not yet exist. He supports this by noting the lack of current adoption, as evidenced by a show of hands, indicating that media buyers are not yet widely using agentic AI for their daily ad buys, suggesting these protocols may be premature.

"AI is good for insights and summarization and helping you it's a tool helping you ideate and things like that so we found a lot of success with ai when it was like hey you know take a look at this you know information here what are some potential audiences that might be relevant to this so like audience search and things they might not get it all right but it helps us ideate we're so it can help us generate ideas other ways we utilize it is if you need to you know menial tasks like look i need to copy this campaign template 10 times and i need to make these three changes in it each time can you just do that for me and they can do it it might be three might be 300."

Francia clarifies the effective applications of AI in advertising, distinguishing them from direct programmatic buying. He states that AI excels at tasks like generating insights, summarizing information, and aiding in ideation, such as identifying potential audiences or automating repetitive campaign setup tasks, positioning it as a supportive tool rather than an autonomous agent.

Resources

External Resources

Articles & Papers

  • "The case against AI agents for programmatic ad buying" (The Digiday Podcast) - Featured segment discussing why programmatic ad buying should not be outsourced to AI agents.

People

  • Seb Joseph - Executive editor of news at Digiday, co-host of the podcast.
  • Christopher Francia - Director of product development and client performance at Attention Arc, makes the case against AI agents in programmatic ad buying.
  • Tim Peterson - Executive editor of video and audio at Digiday Media, host of the podcast.
  • Kimiko McCoy - Senior marketing reporter at Digiday, co-host of the podcast.
  • David Ellison - Son of Larry Ellison, owner of Paramount.
  • Larry Ellison - Founder of Oracle.
  • James Gunn - Overseeing DC Studios.
  • Greta Gerwig - Directing "The Chronicles of Narnia" movie for Netflix.
  • Christopher Nolan - Director who left Warner and moved to Universal.
  • Bob Iger - CEO of Disney.
  • Sam Altman - CEO of OpenAI.
  • Julia Bazer - COO of Bloomberg, hired by Microsoft to oversee its AI news product.
  • Jane Fonda - Voiced opposition to the Netflix-Warner Bros. Discovery deal.

Organizations & Institutions

  • Attention Arc - Company where Christopher Francia is Director of Product Development and Client Performance.
  • Netflix - Discussed in relation to its potential acquisition of Warner Bros. Discovery studio and streaming business.
  • Warner Bros. Discovery - Subject of acquisition discussions, potentially by Netflix or Paramount.
  • Meta - Mentioned for its Advantage Plus product and foray into signing content licensing deals with publishers for its AI chatbot.
  • Google - Mentioned for its Performance Max product, Gemini 3 model, and AI mode search engine.
  • OpenAI - Discussed in relation to its AI products, including ChatGPT, and its response to Google's Gemini 3.
  • Microsoft - Mentioned for hiring Julia Bazer to oversee its AI news product and signing content licensing deals with publishers.
  • New York Times - Mentioned for suing Perplexity for copyright violation.
  • Chicago Tribune - Mentioned for filing a lawsuit against Perplexity.
  • Conde Nast - Publisher that has signed a deal with Microsoft.
  • Omnicom - Mentioned in parallel to the Netflix-Warner Bros. Discovery deal regarding its acquisition of Interpublic Group.
  • Interpublic Group (IPG) - Mentioned in relation to Omnicom's acquisition.
  • Paramount - Mentioned as a bidder for Warner Bros. Discovery and for its acquisition of SkyDance.
  • SkyDance - Acquired Paramount.
  • Apple - Mentioned in relation to its app deployment process.
  • Amazon - Mentioned in relation to its RTB Fabric and potential bidding for NFL rights.
  • NBC Universal - Mentioned for its co-marketing deals, such as with "Wicked."
  • Disney - Mentioned in relation to Bob Iger's return as CEO and potential acquisition targets.
  • AT&T - Mentioned in relation to its acquisition of Time Warner.
  • Time Warner - Acquired by AT&T, became Warner Media, then Warner Bros. Discovery.
  • CNN - Part of the cable TV network business of Warner Bros. Discovery, not included in the Netflix acquisition.
  • Versant - Potential acquirer of the cable TV business of Warner Bros. Discovery.
  • The Writers Guild of America - Voiced opposition to the Netflix-Warner Bros. Discovery deal.
  • Electronic Arts - Mentioned as a party Netflix had looked at acquiring.
  • The Information - Reported on an internal memo from OpenAI CEO Sam Altman.
  • Anthropic - Discussed as going after enterprise and businesses with its products.
  • The Trade Desk - Mentioned in relation to its internal systems and RTB Fabric.
  • DV360 - Google's demand-side platform.
  • CM360 - Google's ad serving platform.
  • Google Ads - Google's advertising platform.

Tools & Software

  • ChatGPT - OpenAI's large language model, discussed in relation to its capabilities and potential advertising product.
  • Gemini 3 - Google's large language model, discussed as a competitor to ChatGPT.
  • Advantage Plus - Meta's product, discussed as a potential AI agent.
  • Performance Max - Google's product, discussed as a potential AI agent.
  • RTB Fabric - Amazon's technology for co-locating ad tech logic.
  • Agentic RTB Framework - Protocol designed to standardize and speed up programmatic processes.
  • Vector Database - Used for audience discovery platforms, enabling natural language lookups.
  • Elasticsearch - Technology used for tokenized vector database lookups.

Websites & Online Resources

  • Digiday.com/newsletters - Website to sign up for Digiday's daily newsletter.

Other Resources

  • AI Agents - Discussed in the context of programmatic ad buying, with arguments for and against their use.
  • Programmatic Ad Buying - The process of buying digital advertising space, a central topic of discussion regarding AI agents.
  • Intellectual Property (IP) - A key motivation for Netflix's potential acquisition of Warner Bros. Discovery.
  • User Context Protocol - A protocol aiming to create universal standards for communication between systems.
  • OpenRTB - A protocol developed to standardize programmatic advertising transactions.
  • JSON - Mentioned in relation to "demand JSON," a protocol that did not gain adoption.
  • AI Generated Creative - Creative content produced by AI, discussed as a parallel to programmatic buying.
  • Dynamic Creative Optimization (DCO) - A method of using different creative assets based on audience or instructions.
  • Embeddings - Used in User Context Protocol to translate audience descriptions into a format computers can understand.
  • Hallucination - A phenomenon where AI generates incorrect or fabricated information, a concern in AI agent use.
  • Monthly Active Viewer Stat - A metric mentioned as being created by Netflix.

---
Handpicked links, AI-assisted summaries. Human judgment, machine efficiency.
This content is a personally curated review and synopsis derived from the original podcast episode.