On April 21, 2026, Microsoft Advertising published a blog post titled "Win across all three eras of the web." The post was written for advertisers. Its function was not to announce a product. It was to reframe the basic question of what advertising is for.
The post identified three simultaneous eras of the web: "Help me find it" — the human web of search and browsing. "Help me choose" — the LLM web where AI synthesizes but humans still act. And "Do it for me" — the agentic web where AI agents evaluate and transact on users' behalf.
Then it told advertisers what to do about the third era.
Tim Frank, Corporate Vice President of Microsoft AI Monetization, wrote that advertisers should stop optimizing for clicks and start optimizing for "selection." Not human selection — selection by AI systems. Being chosen by Copilot. Being selected by the automated buyer on the other side of the auction.
Microsoft is not alone in this. The Trade Desk launched Koa Agents in April 2026, with Stagwell as the first launch partner. PubMatic had already launched AgenticOS in January 2026. Infillion launched what it called the first "agent-native" advertising platform in December 2025, operating through the Model Context Protocol. IAB Tech Lab released an Agentic RTB Framework for public comment in November 2025, designed to enable sub-millisecond AI-driven bidding via containerized agents inside shared data center environments.
The infrastructure for an advertising market with no humans in the transaction is being assembled. The question is what happens when it runs.
Modern programmatic advertising already involves multiple automated systems negotiating in milliseconds. A brand's demand-side platform bids on impressions. A supply-side platform or exchange responds. An auction resolves. What Microsoft is describing is a configuration where the bidder on the brand side is itself an AI agent — one that has been given an objective function, a budget, and a set of constraints, and is now making bidding decisions without human review of each individual impression.
This is not hypothetical. It is documented in academic literature.
Researchers at Alibaba published a framework called MAAB (Multi-Agent Auction Framework for Auto-bidding) at WSDM 2022. The paper studied the competition and cooperation dynamics between auto-bidding agents — specifically, what happens when multiple advertisers all delegate their bidding to autonomous systems that compete in the same auctions. The finding was that these agents can exhibit complex strategic behaviors: they learn to cooperate in some configurations and compete aggressively in others, often in ways their human overseers did not anticipate or design.
The autobidding market exhibits dynamics that formal analysis has shown can be non-convergent. A 2024 paper, "Complex Dynamics in Autobidding Systems" by Leme, Piliouras, Schneider, Spendlove, and Zuo, presented at ACM EC 2024, proved that autobidding markets can exhibit bistability, periodic orbits, and quasi-periodicity — not because of market manipulation but because the game-theoretic structure of competing auto-bidding agents with budget constraints does not guarantee convergence to equilibrium. The paper established that the two-bidder case converges to equilibrium, but a three-bidder system can produce oscillations converging to a limit cycle. With more bidders — the actual market condition — the dynamics can simulate arbitrary boolean circuits and are formally unpredictable.
This is the three-layer problem: humans delegate objectives to autobidders, autobidders compete in auctions against each other, and the auction outcome determines prices and allocations that feed back into the autobidders' learning systems. No human reviews the impression. No human reviews the bid. No human reviews the outcome.
Microsoft's announcement described automated traffic growing eight times faster than human traffic. Agentic browser traffic, by Microsoft's own figures, was up approximately eight thousand percent year over year.
What this means in practice: the demand side of the advertising market is already populated substantially by automated systems making purchase decisions. On the supply side, Microsoft's Copilot is the interface through which many of those automated systems discover and evaluate options.
When Microsoft's AI selects which ads to surface and a brand's AI bids on which impressions to buy, the actual decision-maker is neither human. The two systems negotiate on behalf of humans who are not in the room.
This creates a specific structural vulnerability: arbitrage opportunities that multiply with each layer of automation.
In search advertising, this is not new. Search arbitrage — buying traffic from social or native placements and redirecting it to search feeds — has existed for years. Arbitrageurs use AI-optimized campaign management across multiple networks, feeding aggregated query data back into keyword optimization systems. The arbitrage exists because different platforms price the same or equivalent inventory differently, and automated systems are faster at exploiting the gap than human analysts.
The agentic layer adds a new dimension. When both sides of the transaction are AI systems with different objective functions, different training data, and different constraint sets, the surface for misalignment grows. A brand's autobidder might optimize for cost-per-acquisition in a way that a publisher's inventory management system interprets as a signal about content quality. A contextually relevant placement for a human reader might look like low-value inventory to a system optimizing for a different metric. The two systems converge on a price that neither would have arrived at with human oversight — and the delta between that price and the correct price is the arbitrage.
Research from the University of California Davis and the University of Iowa, published as arXiv:2210.06654v2 in May 2023, documented a related phenomenon in programmatic inventory: "dark pooling." Misinformation publishers shared publisher IDs with high-reputation sites, making their inventory indistinguishable to programmatic buyers. The mechanism was that reputable brands' advertising appeared in the same pools as QAnon-adjacent sites — not because brands chose to advertise there, but because the automated systems could not distinguish the inventory at the level of the auction. Eleven percent of all inventory pooling instances involved misinformation sites, with Forbes and GoDaddy appearing in the same pools.
The brand's programmatic budget filtered through a washing machine the brand never authorized, emerging clean of brand safety concerns while funding the opposite. The buyers and sellers never met. The humans never saw each other's inventory.
At Digiday's media buying summit in early 2026, agency strategists described a specific discomfort. They were broadly comfortable with AI systems handling insights, ideation, and summarization. They were resistant to AI agents operating at the actual transaction point — the moment when money moves and impressions are purchased.
The reason, as one executive put it: "The algorithm did it" is not an acceptable explanation to clients. The accountability gap in agentic advertising is not primarily a technical problem. It is a governance problem. When an autobidding agent operating at millisecond speed on a constrained budget makes a decision that produces a bad outcome — whether that is a brand safety failure, a dramatically overpriced campaign, or a legal violation in a regulated category — the chain of causation runs through objective functions, training data, platform auction mechanics, and system interactions that no human designed and no human can fully reconstruct.
According to executives speaking under Chatham House rules at Digiday's 2026 media buying summit, seventy-three percent of agencies report that their clients do not understand what agentic AI is. Forty-seven percent of those same clients are considering using it within the next twelve months. The gap between adoption and comprehension is not a marketing problem. It is an accountability problem. The people authorizing the systems are not the people who understand them. The people who understand them are not the people who bear legal or reputational responsibility for outcomes.
The IAB Tech Lab's Agentic RTB Framework was built by infrastructure suppliers — companies whose primary optimization target is fill rate and throughput. ExchangeWire noted in March 2026 that holding companies and agencies, the primary buyers of advertising, were not co-authors of the agentic standards now being presented to them. They are being handed protocols built to solve the infrastructure supplier's problem, not the brand advertiser's problem.
Gartner and WARC have projected that one hundred billion dollars in programmatic spend will flow through AI agents by 2028. The efficiency gains are documented. The revenue return on investment is not. Attribution complexity increases significantly when autonomous agents are making media decisions — because the causal chain between an ad exposure and a business outcome now runs through at least two autonomous systems with different objective functions before it runs through a human consumer.
The first wave of agentic advertising is being gamed by the same dynamics that polluted search: arbitrage, misaligned incentives, and the exploitation of gaps between what automated systems can observe and what they can correctly interpret. The difference is that the velocity is higher, the number of systems in the chain is greater, and the humans who could intervene are further removed from the transaction.
Microsoft's Tim Frank told advertisers to optimize for selection. The advertisers who are actually doing this — the ones whose autobidders are negotiating directly with Copilot's selection systems — are discovering that "being selected by AI" is a different problem from "being seen by humans." It requires understanding how machine selection systems are trained, what signals they weight, and what objective functions they are optimizing for. None of which is disclosed. All of which can be gamed.
The room where the selection happens is not empty because no one showed up. It is empty because the humans decided the transaction was fast enough and cheap enough to run without them.
They were right about the first part.
Sources