Faster Buying Won’t Fix Programmatic Waste

In mid-August, Georgia-Pacific made clear why it was not ready to hand programmatic media buying to AI agents. A day later, Butler/Till disclosed a test suggesting exactly why other advertisers may want to do so.

The Butler/Till experiment, conducted with iHeartMedia, involved a four-week streaming-audio and podcast campaign for an unidentified agricultural company. The media spend was modest, just under $10,000, but the results were difficult to dismiss. Butler/Till said the agentic campaign reduced CPMs by 42% compared with the advertiser's direct-buy benchmark. On the podcast side, 48% of impressions were delivered in premium, non-skippable mid-roll placements, compared with 33% allocated to mid-roll under the conventional plan.

Georgia-Pacific, meanwhile, is looking at much of the same technology and reaching a different decision. Its concern is not that buying agents are incapable of operating programmatic campaigns. It is that an agent introduced into an inefficient media market will inherit the inefficiencies.

Taken together, the two positions make agentic media buying considerably more interesting than another story about AI reducing manual work.

They expose the question underneath the technology: what exactly are we accelerating?

Two Decisions, One Market

The Butler/Till test matters partly because of what it did not attempt to automate.

A human planner still established the campaign brief. The agency's buying agent took that brief to an iHeartMedia selling agent, with the two systems configured to interact through an MCP server. The agents handled the transactional exchange within parameters established by people. Strategy did not disappear, nor did the advertiser surrender an unrestricted budget to an autonomous system and wait to see what happened.

That is much closer to the likely near-term form of agentic media than the more theatrical descriptions of autonomous advertising suggest. Agents are beginning to take over parts of execution while humans define objectives, permissions, and financial boundaries.

Butler/Till has been working toward this for months. An earlier test with PubMatic used an agency agent built with Anthropic's Claude to communicate with PubMatic's systems. That campaign reportedly reduced supply-chain costs substantially and lowered CPMs.

The iHeartMedia test goes further because the transaction occurred directly between an agency and a publisher rather than passing through the conventional sequence of intermediaries.

There is a genuine economic proposition here. If a buyer agent can find inventory, evaluate an offer and transact directly with a publisher's agent, some of the infrastructure created to help humans navigate fragmented advertising markets becomes less necessary.

That possibility is already attracting the large agency groups. WPP Media is testing a buyer agent for video. Omnicom told investors in April that it had conducted real media buys for several clients through agent-to-agent transactions. Its chief technology officer, Paolo Yuvienco, described shortening the media supply chain as part of the objective.

None of this proves that agentic buying is already better at scale.

The Butler/Till audio campaign was a small pilot involving one client, one publisher relationship, and less than $10,000. Its 42% CPM improvement is a company-reported comparison with an existing client benchmark, not the result of a large randomized study. The 48% mid-roll result also compares delivered inventory with the allocation in a traditional plan, rather than a simultaneous control campaign operating under identical conditions.

I would not read those limitations as a reason to disregard the test. They tell us what the test is: an encouraging demonstration that a more direct transaction architecture can work and can produce attractive economics under controlled conditions.

Scaling it is another problem.

Georgia-Pacific Wants the Waste Gone First

Georgia-Pacific is approaching the issue from the opposite end of the stack.

The company brought much of its programmatic buying in-house in 2019 and spent years working on the demand side. More recently, Paras Shah and his team turned toward the supply feeding those systems. Georgia-Pacific had been working with more than 30 supply-side platforms, many offering access to substantially overlapping inventory.

Its response was not to add another automation layer. It began reducing the supply footprint and using technology from SWYM to evaluate inventory before bidding rather than relying primarily on post-campaign optimization.

The distinction is practical. A conventional programmatic system may discover after an impression has been purchased that the inventory was poorly viewable, fraudulent or otherwise low value. Georgia-Pacific wanted more of those judgments made before its money entered the auction.

According to results reported by Digiday, CPMs declined by between 17% and 44% depending on the brand and implementation period. Viewability improved, as did video completion rates. The reporting put the reduction in Georgia-Pacific's SSP footprint at roughly 80%.

Those numbers, like the Butler/Till results, originate from the participants rather than an independent audit. What makes Georgia-Pacific's position more consequential is that broader industry data points in the same direction.

The Association of National Advertisers' Q1 2026 Programmatic Transparency Benchmark found a remarkable spread between stronger and weaker advertisers. The higher-performing group converted 54% of programmatic spending into qualified impressions. The lower-performing group managed only 32.1%.

Fees were not the major reason for the difference. Transaction costs varied by only 2.4 percentage points between the two groups, while media productivity losses differed by 19.4 points. The better performers also operated with more concentrated supply footprints.

They were not paying more for the privilege. Their average CPM was lower.

That changes how the agentic media conversation should be framed. A sophisticated buying agent placed on top of an undisciplined supply chain does not begin with a neutral environment. It begins with whatever inventory rules, commercial arrangements, measurement weaknesses, and optimization objectives the advertiser has already accepted.

The agent can become extremely good at navigating those conditions without ever asking whether they should exist.

Automation Inherits the Objective

Programmatic advertising has spent years demonstrating that optimization is inseparable from the target being optimized.

A system instructed to minimize CPM can become exceptionally effective at finding cheap impressions. Cheap impressions and valuable impressions are not synonymous. A system rewarded for delivery can optimize delivery. A supply platform rewarded for volume has little reason to solve the advertiser's duplication problem unless the commercial model pushes it in that direction.

Adding an LLM or an agentic layer does not remove those incentive structures.

In fact, greater autonomy may make the objective more consequential. A human trader working through a questionable process can make questionable decisions at human speed. An agent can make them continuously, across more inventory and with less friction.

This is where the Georgia-Pacific hesitation becomes more than conservatism. Shah is effectively treating the programmatic environment as an input to the agent rather than assuming the agent will repair the environment.

There is already a governance parallel emerging among the agencies building these systems. WPP Media's video buyer agent, for example, currently keeps financial commitments and campaign activation behind explicit human approval. The company is also working with publishers and standards organizations on the rules governing how buyer and seller agents communicate.

That work can sound secondary beside a 42% CPM improvement. It is probably closer to the center of the problem.

Once two software agents can propose, negotiate and eventually execute a media transaction, governance has to reach beyond whether the model produced an acceptable answer.

The organization needs to know which inventory the agent may consider, whose data it can use, what economic objective it is pursuing, and where authority returns to a person. Those choices determine the system long before the transaction occurs.

What Changes for Agencies

Agentic buying could still alter the competitive structure of media agencies, although perhaps not in the way the largest groups would prefer.

Scale has historically helped large media agencies absorb the complexity of programmatic advertising.

They have the personnel to operate intricate buying systems, negotiate broad publisher relationships, and maintain proprietary technology that smaller competitors struggle to reproduce. Agents can lower some of that operating burden.

Butler/Till is a useful example precisely because it is an independent agency, not a global holding company. If a relatively small team can use agents to negotiate directly with publishers and automate procedural buying work, part of the operational advantage associated with agency size becomes cheaper to replicate.

That does not eliminate scale. It changes where scale pays.

A global agency with better data, stronger publisher access and a more mature control environment may be able to give its agents better information and a broader set of useful choices. WPP's emphasis on proprietary intelligence reflects that logic. Omnicom's effort to shorten the supply chain does too.

The competitive question therefore moves upstream from who can perform the transaction toward who can design the environment in which the transaction occurs.

That favors organizations that know which supply paths they want, which inventory they distrust, and how they define a successful impression before an agent starts optimizing. It also increases the importance of knowing whose interests are embedded in the systems being used. Media buying has never been free of conflicts between advertisers, agencies, technology platforms, and publishers. Software autonomy does not dissolve those relationships. It can make them harder to see.

Faster Buying Raises the Cost of Bad Decisions

The Butler/Till and Georgia-Pacific stories appear to describe two positions on agentic media buying. In practice, they are much closer than that.

Butler/Till created a relatively narrow transaction between a buyer agent and a publisher agent, with human strategy around it, and found evidence that the direct path could improve economics. Georgia-Pacific has spent years narrowing and improving the system it would eventually ask an agent to operate.

One is testing execution after defining the lane. The other is still rebuilding the lane.

For marketers, that is a more useful way to think about agentic media than asking whether AI should replace a media buyer. The immediate issue is whether the advertising operation is ready to delegate decisions at much greater speed.

The ANA data suggests many are not. A market in which strong advertisers convert 54% of programmatic spending into qualified impressions while weaker ones convert only 32.1% already contains a large execution gap. Giving both groups faster automation is unlikely to close it by itself.

The first generation of programmatic advertising promised efficiency through automation and then created an industry devoted to recovering transparency from the resulting complexity. Agentic media buying offers a chance to avoid repeating that sequence, but only if advertisers resist treating autonomy as the starting point.

Before the budgets become large, marketers need to know what their agents will be optimizing and what kind of market they are being allowed to enter. Otherwise, the most impressive achievement may be a system that reaches the wrong answer with unprecedented efficiency.