Agentic AI Could Lock Publishers Out Before Auction

By Elena Vasquez

Among video buyers spending $50 million or more annually, 63% are using or planning to use AI agents for inventory discovery and evaluation, according to IAB research. That makes selection a pre-auction problem, not merely a pricing question. In a Sell Sider column, Jonathon Shaevitz argues that identity-informed agents can decide inventory is not worth any bid before any auction begins.

What makes agentic AI different from earlier shifts?

Anthony Katsur, who runs IAB Tech Lab, expects “several false starts” and years of experimentation and alignment across platforms, agencies and publishers. Razorfish described the mood around agentic AI at CES in January as inevitable but incremental.

Digiday’s January research found that 54% of 142 brand and agency professionals said their companies did not use agentic AI in their workflows at all. Shaevitz uses the result to argue that the skeptical case has the stronger position.

Shaevitz has spent 20 years watching the industry declare a shift imminent about once a year. Mobile was called inevitable around 2007, yet it did not pass desktop in US digital ad spend until 2015. Cookie deprecation was announced in January 2020 and abandoned in July 2024.

He rejects claims that publishers are already behind as salesmanship. Earlier shifts required someone to build and finance new integrations, while paying parties disagreed about what to build. Agentic AI reverses that sequence because buyers already paid for programmatic infrastructure. Unlike earlier shifts, agentic AI can arrive through tools that buy-side teams already built, financed, integrated and operate, without waiting for a universal standard across platforms, agencies and publishers.

Sunk integration cost means spending already committed and unrecoverable. Shaevitz calls agentic AI the first shift in which buyers have already paid that bill. Fully autonomous media buying may remain years away. AI agents, software that performs buying tasks, can already evaluate inventory, inform bidding and optimize spend through existing programmatic infrastructure.

Where the money is already moving

According to Shaevitz, Microsoft ended a Dentsu relationship of more than a decade on April 8, 2026. The company moved global media planning and buying to Publicis in a mandate that trade press estimates at $700 million to $1.2 billion.

The appointment accompanied a broader announcement about embedding agentic AI across Microsoft’s marketing workflow, with agents on Microsoft Fabric informed by Epsilon deciding which audience segments mattered and deploying media against them.

Asked why Publicis kept winning without pitches, Arthur Sadoun said clients now understood “it’s about identity.” Publicis also agreed in May 2026 to acquire LiveRamp. The FTC cleared the transaction in October, subject to a five-year neutrality order.

Once it closes, one of the two largest buyers of publisher inventory will own more of the identity layer, the data used to identify audiences, that its own agents reason over. Shaevitz traces a direct market chain: agents buy media, media clears in auctions, and publishers receive payment in those auctions.

Why auction exclusion can happen first

Shaevitz uses bid shading to show how quickly new buying logic can reshape auctions. Bid shading, a strategy for avoiding an excessive price in first-price auctions, where the highest bid sets the price, spread across supply-side platforms, or SSPs, and then demand-side platforms, or DSPs, within months.

Exchanges began moving toward first-price auctions in 2017. Google announced Ad Manager’s move in March 2019 and completed it four months later. The strategy spread through existing integrations, without new infrastructure or industrywide standards. Publishers saw bid shading’s effect through clearing-price reports and adjusted their pricing floors.

Machine learning powered much of that behavior. The agentic difference lies in the decision: shading optimized the price of a bid already being made, while identity-informed agents decide whether inventory deserves a bid at all.

That distinction creates exposure before pricing. Publishers risk losing opportunities before their inventory reaches an auction. The issue is not whether agents eventually pay a premium. The issue is whether publisher inventory survives evaluation and enters the auction at all.

How autonomous are agents today?

IAB surveyed 360 digital video buyers in February and March. Ninety-six percent saw a role for agentic AI in programmatic advertising and real-time bidding (RTB), the auction-based buying of ad impressions in real time. Sixty-three percent expected that role to be an optimization layer, or software that improves bidding decisions.

The IAB research says adoption centers on decision support, or software guidance before a purchase decision, rather than the purchase itself. The rate for AI-based deal execution and negotiation in video campaigns, meaning agents handling those transactions, stands at 28%. Forty percent of buyers require human approval before an agent acts. Shaevitz describes a practical divide: agents operate freely where no money is committed and remain gated where spending follows.

Katsur’s false-start warning applies most directly to autonomous execution, meaning agents taking purchasing action without human approval. The decision-making layer is already gaining traction. Among video buyers spending $50 million or more a year, 63% are using or planning to use agents for inventory discovery and evaluation, compared with 30% of smaller buyers. Shaevitz says the largest budgets are already directing agents toward inventory worth considering.

What can publishers do before standards settle?

Shaevitz says publishers need not wait for agentic standards to mature or fully autonomous transactions to arrive. Much of the necessary preparation can improve performance in current auctions. The following three actions do not depend on which standard wins:

• Create stable identifiers across properties, meaning consistent labels for the same inventory assets.
• Structure first-party attributes, or publisher-collected data, so bid requests carry clear context.
• Authorize data access explicitly and put the terms in writing. Bid-request signals can remain visible to bidders that do not win, making authorization a contract question rather than a technical one.

According to Shaevitz, these actions can strengthen inventory evaluation and auction match rates today. The share of bid requests that produce a matched bid can improve regardless of how quickly autonomous buying develops.

Shaevitz draws a direct lesson from 2019. Publishers that understood their pricing floors before the auction change performed well. Publishers that first learned about the change through revenue reports did not. Agentic evaluation creates the same sequence. Publishers must make inventory understandable before agents decide it lacks value.

The near-term test is not whether an agent will pay a premium for publisher inventory. It is whether that agent considers the inventory worth evaluating before an auction begins.