IAB Europe's Dimitris Beis breaks down agentic advertising infrastructure in a March 12 video, outlining three scaling levels, LLM risks, and what an ecosystem of autonomous ad agents would actually require.
IAB Europe on March 12, 2026 published a short explainer video in which Dimitris Beis, the organization's Data and Innovation Strategist, walked through what scaling agentic advertising actually involves - from individual agents driven by large language models, through multi-agent coordination, and up to what a full agentic ecosystem for digital media buying might one day look like. The video, running under ten minutes, arrives at a moment when the advertising industry has been debating agentic infrastructure intensely for months, with protocols, registries, and deployment frameworks multiplying across the ecosystem.
The explainer is structured around three levels of scale. Each presents distinct technical requirements, and Beis is clear that the challenges compound as the scope expands.
According to Beis, when the industry talks about agents, it is referring to "partially or fully autonomous systems driven by large language models." These systems act using tools, and one increasingly common mechanism for giving agents access to tools is through Model Context Protocol servers - the open standard originally developed by Anthropic that has now become a central building block for advertising technology integrations across platforms including Amazon, Yahoo DSP, and Google.
At the single-agent level, the core challenges stem from the LLMs themselves. Beis identified four specific problem categories: context limitations, the risk of hallucination, a lack of observability at the inference layer, and the risk of prompt injection - which he described as "their brains being hacked with the right sequence of input tokens." These are not abstract concerns. Security researchers documented prompt injection vulnerabilities in agentic browser environments in October 2025, and Google Cloud's 54-page agentic AI framework, released in November 2025, devoted substantial space to addressing them.
IAB Europe is itself beta-testing a single agent internally. The use case is specific: the organization maintains a large volume of documents and Transparency and Consent Framework data that updates frequently, and members submit questions about these materials regularly. According to Beis, a prompt-driven interface - combining a basic set of functions with conversation context - can do "pretty useful things." His example: asking the agent what the main retail media challenges in Europe are, at which point a librarian agent fetches the relevant report and compiles an answer.
This in-house deployment illustrates a recurring theme in the video. Agents can be made competent, Beis argued, but only by "providing the LLM with the right kind of scaffolding" - and the conditions for success are "pretty specific to the use case." There is no universal calculus. One open question that follows: if an agent replaces a human, who is now accountable for what it outputs?
The second level is multi-agent systems. This design pattern, Beis said, has become common precisely because of LLM limitations. As task complexity increases, the practical response is to distribute the work across multiple agents rather than demand more from one. What emerges is a need for agents to communicate and coordinate - which requires what Beis called a "delegation framework" or "delegation blueprint."
Such a framework defines how tasks are broken down and assigned, how agents communicate with each other, whether they can query one another directly, whether they are aware of other agents, when to escalate to a human, and how guardrails are applied. IAB Europe's own internal delegation framework, built around its document assistant, is deliberately simple. The organization controls both agents and trusts them fully - they are not third-party systems. For that kind of narrow, controlled application, constructing a functioning delegation framework is tractable.
Beis went a step further, demonstrating an experiment in which delegation instructions are injected directly through the prompt. In his test, each sub-agent is told to assess its own capability on a component of the overall task, then bid on whether it can execute that component. The test used three sub-agents. Each returned a bid of 10 for the subtask where it had a competitive advantage, without being aware of what the other agents could do.
The experiment also illustrated fragility. When the underlying model was changed, the bids shifted - and the agents became more confident about subtasks where they lacked competitive advantage. The point was not that the approach failed, but that systems built on prompt injection are brittle by nature. Switching the model changes behavior in ways that are not immediately predictable.
The third level - an agentic ecosystem, meaning a network of agents with different owners - is where the real complexity begins. According to Beis, orchestrating activation in digital advertising is already a complex task involving multiple stakeholders. Adding autonomous agents with different owners to that picture introduces a new layer of requirements.
Among those requirements: discovering what agents exist and what they can do. IAB Tech Lab has been working on this through what it initially called an agentic registry, which reached ten registered entries this week - adding Amazon, Burt Intelligence, Optable, Dstillery, and HyperMindZ.ai to an expanding list. Trust and security, accountability, policy enforcement - these are all, in Beis's description, items on "a very long list of questions" for which the industry does not yet have full answers.
He acknowledged this directly: "These are non-trivial questions of course when you look at how fast agentic advertising has evolved and how early we are still when we look at the overall impact agentic advertising could have on digital advertising."
The Deals API, which reached version one recently and will in future versions support revisioning, is one example of existing standards being extended to support orchestration. IAB Tech Lab formally named its umbrella agentic initiative AAMP - Agentic Advertising Management Protocols - on February 26, 2026, clarifying three pillars: execution, protocols, and the Agent Registry. The IAB Tech Lab's Agentic Real-Time Framework, which cuts latency by 90% according to the organization's own documentation and uses gRPC with protobuf serialization, is one component of the execution layer.
The core question the video poses is whether agentic advertising can actually scale. Beis argued against dismissing it based on current challenges, pointing out that the ecosystem is not starting from zero. "The ecosystem has had agents for a long time," he said. "They just haven't been driven by LLMs." Trust and verification - topics that appear new in the agentic framing - have been explored extensively in different forms. The question of whether an adtech platform with both buy-side and sell-side clients can truly act neutrally is not new; it predates LLMs entirely.
What is new, according to Beis, is two things: the need for a defined protocol for orchestration, and a potential new degree of autonomy across orchestration and activation that LLMs make possible.
The answer to whether agentic scales may also depend on whether it develops as something closer to a private marketplace or an open market. For threats like Sybil attacks or prompt injection, short-term solutions might include pre-verifying all participating parties and agents, adding monitoring and reporting mechanisms, or delegating more actively back to humans. Longer-term responses could include private or public reputation systems, potentially run by trusted third parties. These are not hypothetical futures - the Ad Context Protocol, launched October 15, 2025, already requires publishers to register sales agents publicly, and mandates that buyers verify publisher domains to confirm what agents are authorized to sell.
Beis's framing for what success would look like is measured. He did not argue that all digital advertising spend should eventually flow through agents. His position was more restrained: "I think it's perfectly fine if some substantial amount of spend in some markets ends up going through agents and it's just another option for buying and selling digital media." That framing is significant in the context of an industry that has seen the agentic conversation sometimes drift toward maximalist claims.
The IAB Europe explainer arrives against a backdrop of rapid practical deployment. PubMatic launched AgenticOS on January 5, 2026, with live campaigns already running on connected television inventory and early testing showing campaign setup time reductions of 87% and issue resolution improvements of 70%. Magnite embedded a seller agent into SpringServe in December 2025 and announced the integration formally on January 6, 2026. Yahoo DSP launched three agentic capabilities on January 6, including a campaign activation agent operating through Model Context Protocol and an always-on troubleshooting agent for pacing issues. Amazon Ads moved its MCP Server from closed to open beta on February 2, 2026, at the IAB Annual Leadership Meeting.
Most recently, Optable's Audience Agent was integrated into PubMatic AgenticOS and announced on March 12, 2026 - the same day as the IAB Europe video - positioning it as one of the first live demonstrations of AdCP operating across real programmatic infrastructure. In the Optable-PubMatic model, publisher first-party data stays with the publisher while still powering buyer decisions, a design choice Beis's video implicitly supports when he emphasizes scaffolding over raw autonomy.
The scale of investment behind this buildout is substantial. McKinsey identified $1.1 billion in equity investment flowing into agentic AI during 2024, as noted in prior PPC Land coverage. Job postings related to agentic AI increased 985% from 2023 to 2024. Gartner, however, predicted in June 2025 that over 40% of agentic AI projects would be canceled by end of 2027 due to escalating costs, unclear business value, and inadequate risk controls.
The IAB Europe video does not resolve these tensions. It was not designed to. What it does is lay out the structural logic of why agentic advertising is genuinely difficult to scale - and why that difficulty does not make scaling impossible. Whether the ecosystem treats agentic infrastructure as a catalyst for addressing foundational questions about trust, transparency, and accountability, or as a distraction from them, is a question Beis left deliberately open.
Who: IAB Europe, represented by Dimitris Beis, the organization's Data and Innovation Strategist. IAB Europe is the European trade association for the digital advertising industry, responsible among other things for managing the Transparency and Consent Framework used across European programmatic advertising.
What: A ten-minute explainer video published on YouTube that breaks down the structural components of agentic advertising - individual LLM-driven agents, multi-agent systems with delegation frameworks, and full agentic ecosystems - and examines what would be required for each level to scale. The video covers technical challenges including hallucination, context limitations, observability gaps, and prompt injection, and discusses possible approaches to trust, security, and accountability in a future agentic ecosystem.
When: The video was published on March 12, 2026, at a point when multiple agentic advertising platforms, protocols, and standards bodies are already in active deployment rather than theoretical planning stages.
Where: Published on IAB Europe's YouTube channel. The subject matter concerns the global digital advertising industry, with particular relevance to European market participants navigating both agentic AI and regulatory environments including the GDPR and Digital Services Act.
Why: The digital advertising industry is building agentic infrastructure faster than common frameworks for evaluating it have emerged. The video matters because it offers a structured, technically grounded view from a European standards body at a moment when the questions of trust, accountability, and ecosystem governance are more urgent than the question of whether the technology itself functions.