Agentic media buying has left the display-and-CTV sandbox. Independent agency Butler/Till and iHeartMedia completed what they describe as the industry’s first agentic streaming-audio campaign, a four-week pilot that ran across streaming audio and podcast inventory in July and August, Digiday reported on 20 August. The four-week pilot ran on a budget of just under $10,000 for an agricultural-sector client.
The architecture matters more than the budget. The pilot paired two AI agents — one built by the agency, one by the publisher — communicating over an MCP server. A human planner handed Butler/Till’s agent the campaign brief; that agent then negotiated delivery with iHeartMedia’s agent. What distinguishes this from the agency’s earlier agentic pilots is the path: this was a direct agency-to-publisher transaction, where previous tests ran through intermediaries such as the SSP PubMatic. Butler/Till has now run agentic pilots with a dozen clients across CTV and digital display, and it is not alone — WPP Media and Omnicom are testing comparable tooling.
The reported results favoured the machines. Streaming audio CPMs came in 42% below the client’s traditional direct-buy benchmark. Podcast delivery skewed premium: 48% of podcast impressions landed in non-skippable mid-roll placements, against 33% in the traditional plan. Kristie Murphy, Butler/Till’s associate director of programmatic, framed the milestone as channel expansion — “we’re moving away from the one-off test phase,” with audio showing agentic buying “scaling beyond one channel.” Chief strategy officer Scott Ensign stressed the governance point: AI executed the buying “within human-defined strategy and governance.” For iHeart, chief business officer Lisa Coffey stated the seller’s motive plainly: the company wants advertisers to activate audio, podcast and broadcast “as easily as large digital platforms.” iHeart plans to open broadcast radio to agentic transactions later this year, and its Audiograph data solution launches through Amazon’s DSP in mid-September.
The numbers in this piece
01The channel is the news — and so is the missing middleman
Two things changed at once here, and each deserves separate attention.
First, the channel. Audio has always been a relationship-sold, insertion-order medium at heart — streaming brought programmatic pipes, but podcasts and broadcast remained heavy with manual process. That friction is precisely what has kept audio undersized relative to its share of consumer time, a gap audio sellers have complained about for a decade. If agents can compile, negotiate and optimise an audio plan from a brief, the friction argument for underinvesting in the channel weakens. Coffey’s framing is the tell: the publisher’s goal is for audio to be as easy to activate as the walled-garden platforms. Agentic buying, on this evidence, is a demand-unlocking play for hard-to-buy channels — which is the opposite of a threat for the sellers of those channels, provided they are equipped to transact.
Second, the path. The agent-to-agent transaction was direct: agency system to publisher system, no exchange in the middle. It is one pilot, and SSPs continue to provide aggregation, decisioning and clearing functions a point-to-point integration does not replicate at scale. But the demonstration stands: when both sides field capable agents speaking a common protocol, the transaction can complete without an intermediary. The pilot’s economics will invite hard questions about where intermediary margin fits in such a path. For publishers large enough to field their own agent infrastructure, disintermediation becomes a genuine strategic option; for everyone else, the question becomes which partners will provide that layer, and at what take rate.
The mid-roll number carries its own lesson for sellers. The agent did not just buy cheaper — it allocated nearly half the podcast delivery into the most premium placement class, against a third under traditional planning. Machine allocation followed value density rather than habit. Sellers whose premium inventory is well-described in machine-readable terms should expect agents to find it; sellers whose best placements live in a rate card PDF and a seller’s institutional memory should expect agents to miss it.
02Why this matters for publishers
| Agentic buying is compiling into a cross-channel norm, not a display experiment | Display, CTV and now audio in a dozen-client programme at one agency — with WPP Media and Omnicom running parallel tests — is the pattern of an emerging default, and every publisher's channel is on the roadmap. |
|---|---|
| The direct agent-to-agent path is proven at small scale | A publisher with its own buying-facing agent transacted without an SSP in the chain. That reshapes the make-or-buy question for publisher ad stacks well beyond audio. |
| Agents buy what they can read | The premium mid-roll skew shows machine buyers concentrating spend where inventory value is legible to them. Inventory described only in decks and rate cards is invisible at the moment of allocation. |
| The efficiency headline will travel faster than its context | A 42% CPM improvement from one small pilot will be quoted at sellers in negotiations regardless of its sample size. Sellers need their own account of what agentic efficiency means for their pricing before buyers supply one. |
03What publishers should do
04The bottom line
One four-week, five-figure pilot proves nothing about the market by itself — but the direction across pilots is now unmistakable: agentic buying is moving channel by channel toward everything, and it rewards sellers whose inventory, deals and value story are legible to machines. Audio’s lesson generalises. The publisher that co-built the pipes got premium utilisation up and a first-mover position in a channel buyers have historically found hard to buy; the intermediary was absent from the transaction entirely. Publishers should take the pilot as a costed preview of their own channel’s future and spend the interval usefully — structuring feeds, opening deal APIs and choosing who represents them to the agents — because when the brief arrives by machine, the inventory that answers is the inventory that gets bought.