In December 2025, PubMatic and Butler/Till launched what they called the first fully autonomous CTV campaign, with an AI agent handling negotiation, execution, and optimization without human intervention.
Weeks later, NBCUniversal and FreeWheel completed the first AI-agent-led programmatic guaranteed deal, running live during NFL playoff games. By January, Magnite and MiQ had executed one of the first Ad Context Protocol test buys. Amazon launched its own Ads Agent. Yahoo DSP rolled out agentic capabilities through a new Model Context Protocol framework. And at CES 2026, competing standards from the AdCP coalition and IAB Tech Lab’s Agentic RTB Framework made clear that the industry wasn’t debating whether AI agents would buy media, but how. By mid-2026 the momentum had only accelerated, even as a majority of advertisers flagged concerns about brand safety and loss of control as autonomous agents took the wheel.
The advertising industry has declared 2026 the year of the AI agent. But the core challenge these systems are trying to solve, teaching machines to make individual-level buying decisions in milliseconds, at scale, without human judgment, is not new. A small cohort of engineers and product leaders working in early programmatic advertising spent years building the algorithmic foundations that today’s AI agents depend on.
Harel Amir is one of them. Over the past fifteen years, Amir has built and scaled the kind of decision-making systems that the advertising industry is now racing to automate with artificial intelligence, first at Israeli retargeting pioneer myThings, then at content discovery platform Outbrain, and now at digital intelligence company Similarweb.
Teaching Machines to Evaluate Individual Users
In the late 2000s, most digital advertising still relied on audience segments, broad groupings of users sorted by demographics or browsing behavior. Advertisers would buy access to a segment and hope the averages worked in their favor. Efficient relative to what came before, but crude by today’s standards.
At myThings, Amir and his team took a different approach. Rather than evaluating segments, they developed a system that calculated conversion probabilities at the individual user level: a real-time prediction of how likely a specific person was to take a specific action. This wasn’t a refinement of existing targeting. It was a different model entirely.
That shift had real commercial consequences. By moving the unit of analysis from the segment to the individual, myThings was able to offer advertisers a cost-per-action model where they paid only for actual conversions. The risk sat with the platform, not the buyer, which required deep confidence in the underlying algorithms.
Major U.S. retailers adopted the technology, including Walmart, Best Buy, and Microsoft. Google’s DoubleClick division, not typically in the business of endorsing competitors, published a case study analyzing myThings’ approach, a rare validation from the industry’s dominant player. The company’s growth reflected the technology’s impact: second place in the Deloitte Technology Fast 500 for EMEA, first in Israel, and the 428th spot on the 2014 Inc. 5000 list.
The connection to agentic buying is hard to miss. Today’s AI agents are attempting to do at a broader scale what Amir’s systems did for display retargeting: evaluate signals in real time, make autonomous bidding decisions, and accept financial accountability for outcomes, all without a human in the loop.
“People talk about agentic AI as if it appeared out of nowhere. It didn’t," Amir said. “Every autonomous buying agent runs on the same core logic: predict an outcome for this specific user, calculate what that prediction is worth, and bid accordingly. We were doing that at myThings in 2009. What’s different now is that the agent doesn’t need a human to set the rules."
Bridging Two Advertising Economies
Amir’s next challenge was a different but related problem. At Outbrain, the leading content discovery platform, he was tasked with building the infrastructure that would connect the company’s native advertising ecosystem, which operated on a cost-per-click basis, with the broader programmatic marketplace, which transacted on cost-per-impression.
These weren’t just different pricing models. They reflected different ways of valuing attention. Translating between them required predictive algorithms that could estimate click-through probability in real time and convert that estimate into a viable impression-level bid. The margin for error was thin: overbid, and the platform loses money on every impression; underbid, and the inventory goes unsold.
Amir and his team engineered the prediction models and the bidding logic that made this translation work. The resulting infrastructure, known internally as the Outbrain Extended Network, or OEN, allowed thousands of independent publishers to access programmatic native advertising demand at scale for the first time. For many of these publishers, operating outside the reach of Google and Meta’s direct sales channels, the OEN opened up a meaningful new revenue stream.
The OEN became a central revenue engine for Outbrain, contributing a significant share of the company’s revenue through its growth years and supporting its initial public offering on the NASDAQ in July 2021.
What the industry is now calling AI-driven curation and supply path optimization is, in many ways, an attempt to replicate at scale what the OEN did for native advertising: intelligently match supply and demand across incompatible systems, using algorithmic decision-making to bridge the gap.
From Building the Black Box to Opening It
If the first two chapters of Amir’s career were about teaching machines to make autonomous decisions, the current chapter is about making those decisions legible. As VP, General Manager and Head of Product for Ad Intelligence at Similarweb (NYSE: SMWB), he now works on the measurement and market intelligence layer, the tools that allow advertisers, agencies, publishers, and ad tech companies to see how programmatic markets actually behave.
His trajectory mirrors the industry’s own unresolved tension. The same ecosystem that spent a decade building ever more sophisticated automated buying systems is now grappling with the fact that very few people can explain what those systems are actually doing.
“I spent the first part of my career building systems that made decisions faster than any human could follow," Amir said. “Now I build the tools that help the market not just understand what those systems did, but actually shift budgets and strategy based on what the data reveals. Cross-channel budget allocation, media mix monitoring, auditing what the various buying engines can’t see on their own. That’s the layer that’s still missing from the agentic conversation."
As the industry moves toward a future where AI agents negotiate, bid, and transact without human intervention, and where AI-powered chatbots are themselves emerging as new advertising surfaces, Amir occupies an unusual position: someone who built the first generation of autonomous buying systems and now builds the intelligence infrastructure to hold the next generation accountable. It is also where his current work is increasingly focused: on how advertising behaves, and how user intent forms, inside AI chat itself. The technology works. The question is whether transparency can keep pace.
About Harel Amir
Harel Amir is a product and business leader in digital advertising technology with over fifteen years of experience building programmatic infrastructure, predictive algorithms, and market intelligence platforms. He currently serves as VP, General Manager and Head of Product for Ad Intelligence at Similarweb (NYSE: SMWB) and as an independent consultant advising companies and agencies on digital advertising strategy. Previously, he held senior leadership roles at Outbrain (now Teads, NASDAQ: TEAD) and myThings, where he was responsible for building and scaling core advertising technology systems. His early career included service in the Israel Defense Forces’ Unit 8200, Israel’s premier signals intelligence unit.