IAB Europe on April 14, 2026 hosted a two-hour virtual event titled "Agents! Innovation, Live From Europe," bringing together three companies to demonstrate how agentic AI is being put into production inside the European digital advertising ecosystem. The event, moderated by Dimitris Beis, Data and Innovation Strategist at IAB Europe, replaced the organization's regular monthly AI working group meeting for the month. Three keynote presentations were followed by a panel discussion and Q&A. The recording is available on demand at the IAB Europe website.

The timing is notable. The digital advertising industry has spent the better part of 2025 and early 2026 debating protocols, registries, and frameworks for agentic advertising - with much of the visible experimentation concentrated in the United States. This event was a deliberate attempt to give European practitioners direct contact with working implementations, not roadmaps.

PubMatic: MCP servers, EMEA campaigns, and the cost of building agentic infrastructure

The first keynote came from Mohsin Pervez, Senior Director, Platform Solutions and Operations, EMEA at PubMatic. He traced the company's journey through what he described as several sequential milestones.

According to Pervez, PubMatic partnered with Nvidia towards Q4 of 2024 to install accelerated computing hardware across its data centers, laying the performance foundation for what followed. The second step, he said, was standards alignment - pointing to IAB Tech Lab's work on agentic specifications as providing confidence to publishers, given that the same body created the OpenRTB standard in 2008 or 2009.

The third milestone was a live agentic campaign in the United States, executed with independent agency Butler/Till in December 2025. That test used a Claude-based agent communicating with PubMatic's AgenticOS infrastructure. According to Pervez, the campaign produced 40% more impressions than comparable non-agentic activity, a cost efficiency of 5.5x for Butler/Till, and an average video completion rate of 98%. The company then built on these results by launching its AgenticOS on January 5, 2026 - described as an operational layer that brings agentic capabilities together rather than a physical operating system.

What is new in 2026, Pervez explained, is the expansion of that infrastructure into an AI Acceleration Program run in collaboration with IAB Europe - the webinar itself is part of that program. A further milestone he highlighted: the first agentic campaign delivered in EMEA, confirming that agentic execution is no longer limited to the US market.

On the technical side, PubMatic demonstrated its MCP server integration through a live Claude demo. The MCP server sits behind authentication, meaning access requires an active advertiser, publisher, or DSP relationship with PubMatic. Once connected, users can create media buys, discover audiences, build CTV campaigns, and troubleshoot deals entirely through natural language - without logging into PubMatic's interface directly. Pervez demonstrated creating a CTV campaign targeted to the UK with a budget of $5,000 and flight dates from May 20 to June 20, watching the campaign appear live inside PubMatic's system within seconds.

The audience and media buying capabilities have been extended in 2026. According to Pervez, inventory discovery alone was available at the start of the year; audience discovery has since been added, allowing buyers to find and attach audiences to existing campaigns or new auction packages. A reporting suite is also being made available through MCP, meaning deal-level analysis, audience performance, and vendor data can be queried without accessing PubMatic's UI.

For publishers, PubMatic has built seven to eight agents covering deal management, deal troubleshooting, AI benchmarking insights, DSP-level demand insights, an email assistant, and yield optimization. The deal troubleshooting function demonstrated during the webinar identified a low match rate between bids sent and bids winning for a test publisher - analyzing 1.2k incoming bids, 2k PubMatic winning bids, and explaining a 1% DSP bid adherence rate as the root cause of poor delivery, all within seconds.

The cost question came up during the panel. Pervez acknowledged that PubMatic's AI costs are lower than general-purpose LLMs because its models are trained specifically on ad tech terminology and PubMatic operations - they cannot draft emails or perform unrelated tasks. He said the investment is materializing commercially, with spend increasing as partners run agentic tests, but that breakeven is likely four to eight months away at current rates.

PubMatic's agentic positioning has attracted scrutiny alongside recognition, given that the company's Q4 2025 reported revenue was $80 million, down 6% year-over-year, while its closest competitor Magnite reported $205.4 million, up 6%. The strategic bet on agentic infrastructure is substantial, and the EMEA expansion confirmed at this event extends the scope of that bet to European markets.

Human Security: 7,000% agent traffic growth and the new audience mix

Moomal Shaikh, Director of Product Management at Human Security, opened her presentation by positioning the company as a cybersecurity firm first - one that measures and disrupts digital fraud across the full customer journey, from ad impression through account creation and transaction.

The numbers Human Security cited for its infrastructure are significant. According to Shaikh, the company processes approximately 20 trillion weekly transactions across 3 billion devices, running over 2,500 signals and more than 400 real-time algorithms simultaneously. The detection methodology combines technical indicators, behavioral signals, reputation models, and statistical and machine learning frameworks.

The core argument Shaikh made is that the traditional distinction between human traffic and IVT (invalid traffic) is no longer sufficient. A third category now exists: AI agents. The challenge of distinguishing human from automated activity has grown substantially more complex as agentic systems have proliferated, and Human Security's data gives some indication of the pace of change.

According to Shaikh, Human Security observed a 7,000% increase in agent use between August and December 2025. That figure covers all agent activity the company could measure across its integrations - not just advertising-related traffic. It reflects a genuine shift in how consumers are beginning to interact with the internet, with agents performing research, discovery, and increasingly purchase-oriented actions on behalf of users.

The implications for advertising are structural. If an e-commerce site blocks all bots to prevent fraud, it may also be blocking an AI agent attempting to make a purchase on a user's behalf. The binary distinction between human and bot that has governed fraud prevention for over a decade does not map cleanly onto agentic traffic. Some of that traffic has high commercial intent. Some does not. Distinguishing between them requires data at a level of granularity that most existing measurement systems do not provide.

Human Security has responded with two product areas. The first is what the company calls agentic visibility - a JavaScript tag deployed in the page header that classifies traffic across humans, general IVT, sophisticated IVT, LLM crawlers, and AI agents. The company notes an important limitation: if an agent does not execute JavaScript, the tag will not capture it. The company argues its server-side integrations compensate for this through behavioral signal analysis and its broader measurement depth. In a sample shown during the webinar, over 700,000 page requests were analyzed over seven days, with 98.6% of those attributed to the top five AI agents - a concentration Shaikh described as consistent with real usage patterns she has observed.

The second product is called Agentic Trust, a server-side integration that allows publishers and site owners to set granular permissions for which agents can perform which actions. The system can distinguish between OpenAI's GPT user (a crawling agent) and OpenAI's ChatGPT agent (a task-executing agent), and can assign different permissions to each. A publisher could allow Claude to perform transactions while restricting a different system to scraping only. The policy configuration mirrors subscription access models - what a paywalled site would allow a paying subscriber versus a casual visitor.

Human Security also mentioned a partnership with Tollbit, a company that places a monetization layer between agents and publishers, allowing publishers to charge for agent access to their content. Tollbit's bot-human-agent classification relies on Human Security data, illustrating how the company is positioning its data as infrastructure for others building in the agentic space.

A question from Beis during the panel asked whether the agentic browser activity shown in the sample data matched broader patterns in Human Security's real data. Shaikh confirmed it did, adding that she expected the figures to continue increasing.

Plan.Net Studios: a blockchain-based European agentic economy, open source

The third presentation came from Yves Bollinger, General Manager at Plan.Net Studios, the production unit within the Serviceplan Group. Bollinger described a project begun approximately two and a half years ago - well before agentic AI became a mainstream topic - based on a vision that agents would eventually trade, collaborate, and transact across the internet.

The observation that prompted the project, according to Bollinger, was that existing agentic infrastructure was dominated by US and Chinese platforms. What was missing was an open, European-built system. The result is what Plan.Net Studios calls the Zumi Suite - an end-to-end open-source stack for agentic commerce, hosted in Germany and designed to be EU-compliant.

The suite has three components. Mazumi is an agent-to-agent protocol built on the Cardano blockchain. It handles agent identity (each agent receives a verified unique ID), agent discovery (how agents find other agents to delegate tasks to), compliance with EU regulations, and payments via smart contracts. According to Bollinger, over 27,000 transactions have been executed on the Mazumi protocol, and over 101 agents are registered on it. Payment through the protocol is automatic once a task is confirmed complete; refund mechanisms for failed tasks are also encoded on the blockchain. The integration process to add the Mazumi node to an existing agent takes approximately 15 minutes.

Kodumi is a runtime environment for deploying agents. It runs in Germany, is GDPR and EU AI Act compliant, and supports every major agentic framework - including CrewAI, OpenAI Agents, and Python network frameworks. It is designed to scale agent capacity without requiring additional server deployments.

Soukumi is the marketplace where agents are listed, hired, and managed. In phase one, humans hired agents through the marketplace. Phase two, launched a couple of months before the webinar, allows agents to hire other agents. According to Bollinger, approximately 35 agents are currently listed on the marketplace. Notable participants include TWWI, Statista, and DPA - the German national news agency - which runs agents on the platform.

The most recent development, launched five weeks before the April 14 webinar, is a trio of specialized agentic co-workers built on top of the Zumi Suite. Elena is a strategy and project management agent. Hana is a research specialist. Alex handles coding and data visualization. The three agents share a single memory pool, meaning context built with one is available to the others, and that accumulated knowledge improves over time.

During the live demo, Bollinger sent a single email brief to Elena asking for a competitor teardown of European direct-to-consumer skincare brands across Germany, Austria, Switzerland, and the Nordics. Elena structured the task, delegated research to Hana, and Hana returned a 35-page PDF report alongside competitive analysis, social media intelligence, and an Excel sheet. Alex then produced an interactive dashboard and an executive PowerPoint deck. The total cost was approximately 50 euros in token and credit charges. The process took roughly 25 minutes. Three agents worked asynchronously without any additional human prompts after the initial email.

Bollinger described two working modes. Standard mode uses Claude Sonnet 4.6 and is faster; deep work mode uses Claude Opus 4.6, draws on more data sources, and takes longer. The agents can be reached via email, WhatsApp voice message, or Microsoft Teams - requiring no specialized interface or dashboard knowledge.

The agents can access premium data on Soukumi, including GWI data for target group analysis and DPA data for news analysis. According to Bollinger, every significant claim in an output is cross-checked across independent sources before delivery, and every output distinguishes between evidence-based conclusions and inferences, marking claims as high, medium, or low confidence.

The system went fully productive two weeks before the webinar. During the panel, Bollinger was asked about the decision-making loop. He said triage (checking what the agents already know) must complete in under 30 seconds. Execution runs asynchronously. A human can intervene at any point through the Kanban-style board inside Soukumi. Structured reflection after each run stores observations in a searchable semantic memory, compounding the agents' knowledge over time.

IAB Europe's taxonomy and the current state of adoption

Beis closed the presentations with a short analysis of how to classify agentic offerings - a framework IAB Europe is developing to help members understand what they are adopting.

The framework identifies three types of integration. Platform-native agents operate within a single platform on first-party data and functionality. MCP-based integrations port existing tools and APIs to the agentic layer. Agent-to-agent (A2A) systems coordinate between autonomous agents across platforms. Of these, MCP currently dominates - all 10 entries in IAB Tech Lab's Agent Registry as of March 11, 2026 operate under MCP. No A2A entries appear in that registry yet.

IAB Tech Lab formally named its umbrella agentic initiative AAMP on February 26, 2026, providing the standards scaffolding that PubMatic and others are building against. Beis used IAB Europe's own internal multi-agent system, called Ibot, to pull registry data during the presentation. Ibot has nine specialized agents, including an orchestrator, a librarian, a statistician, a decoder, and an explorer. The explorer can plug into open agentic registries.

According to Beis, the IAB Tech Lab Agent Registry currently shows 15 agents marked as generally available, with one still in beta. The breakdown by category shows ad servers with the largest presence, followed by supply-side platforms, then analytics, DMP, and data providers, and one order management solution. The registry also lists itself - a detail Beis acknowledged with some amusement.

Beis noted that the preference for MCP over A2A reflects practical considerations: MCP is easier and faster to ship than proper A2A, and A2A raises unresolved trust questions. But he said live experimentation with more advanced orchestration is already happening - the webinar itself offered three examples.

IAB Europe's March 2026 explainer on agentic advertising scale set out three levels - single LLM agent, multi-agent coordination, and full agentic ecosystem - providing the structural backdrop for the classifications Beis presented. The April 14 event extended that work from theory into demonstration.

The session closed with a reminder that IAB Europe's Impact of AI and Digital Advertising survey is live, with a deadline of Friday, May 15, 2026. A report based on the survey is expected in June.

Why this matters for the marketing community

Several aspects of the April 14 event carry practical weight beyond their immediate demo value.

PubMatic's confirmation of the first EMEA agentic campaign is significant because European markets have different regulatory constraints - GDPR, the EU AI Act, and TCF compliance - that US-first implementations may not address cleanly. The fact that PubMatic's MCP server carries a GPP and TCF ID (number 76) in the IAB Tech Lab Agent Registryaddresses this directly, connecting agentic infrastructure to existing European privacy compliance mechanisms.

Human Security's 7,000% agent traffic growth figure between August and December 2025 is a measurement, not a projection. For any marketer or publisher evaluating their traffic, the absence of agentic visibility tools means a growing proportion of site activity is currently unclassified. Whether that activity represents commercial intent or simple scraping has direct implications for conversion attribution, audience building, and publisher monetization. HUMAN Security's fundraising and integration activity reflects the scale of investment going into solving exactly this problem.

Plan.Net Studios' demonstration carries a different kind of significance. A competitive teardown that previously required a strategist a week of work was delivered in 25 minutes at a cost of 50 euros. That is not a marginal efficiency gain. For agencies handling new business pitches, client briefing cycles, or ongoing competitive monitoring, the economics change substantially. Bollinger's point that the bottleneck is accountability rather than capability - who is responsible for agent outputs, where do they come from, how are costs tracked - identifies the real outstanding challenge precisely.

IAB Europe's 2025 AI whitepaper documented that 91% of digital advertising professionals have experimented with generative AI. The question now is how many have moved from experimentation to production with agentic systems specifically. The April 14 event presented three companies at different stages of that journey - one SSP with live EMEA campaigns, one cybersecurity firm with commercial agentic classification products, and one European agency group with a fully productive agentic workforce.

Timeline

Summary

Who: IAB Europe hosted the event, moderated by Dimitris Beis. Speakers were Mohsin Pervez (PubMatic, Senior Director, Platform Solutions and Operations, EMEA), Moomal Shaikh (Human Security, Director of Product Management), and Yves Bollinger (Plan.Net Studios, General Manager).

What: A two-hour virtual showcase titled "Agents! Innovation, Live From Europe," presenting live demonstrations of agentic advertising infrastructure from three companies, followed by a panel discussion on autonomy, cost, infrastructure, and commercial opportunity.

When: April 14, 2026, at 15:00 CET.

Where: The event was held virtually and is available on demand via the IAB Europe website. The technologies demonstrated span US and EMEA markets, with Plan.Net Studios' infrastructure hosted in Germany.

Why: IAB Europe organized the event to give European advertising professionals direct access to working agentic implementations, replacing the monthly AI working group meeting for April 2026 and providing a contact surface between the European ecosystem and innovation in agentic advertising that has until recently been led primarily by US-based deployments.

Share this article
The link has been copied!