Agentic media buying explained
Agentic media buying means an AI agent carries out some or all of the steps of buying advertising. This guide describes how that works in the campaigns documented so far. Most of them are tests.
The basic flow
Digiday’s account of an agency test with PubMatic shows the steps:
- Agency staff wrote a client brief.
- The agency’s agent, built on Anthropic’s Claude model, sent the target ad types and audiences to PubMatic’s system.
- A PubMatic agent returned a set of curated inventory.
- Agency staff approved it.
- The agency’s agent confirmed, and PubMatic’s existing technology completed the purchase.
Magnite describes a similar pattern. Its seller agent matches a buyer agent’s intent to audiences and inventory, and the campaign is placed once both sides agree.
What an agent needs access to
- A brief and limits. PubMatic says advertisers set objectives, guardrails, brand-safety requirements and creative parameters through an LLM interface.
- A connection to the seller or buying platform. This is usually Model Context Protocol (MCP), or the Ad Context Protocol (AdCP) built on top of it.
- Inventory and audience data. Seller agents expose products, audiences and prices. AdCP defines tasks for requesting products and audience signals.
- An account that can spend. AdCP’s documentation says payment, legal agreements and settlement stay with the account or platform unless a contract covers them.
- A way to place the order. Magnite wrote in January 2026 that early proofs of concept “simply push campaigns into ad servers.”
Where human approval sits
In the documented cases, a person approves before money is committed.
- Yahoo states that the user provides “the final approval before an agent proceeds with any task.”
- Scope3’s Anne Coghlan told AdExchanger at AdCP’s launch that large campaigns require human approval and creative must be approved.
- In the Butler/Till test, staff approved the inventory before purchase. Digiday reported that staff kept control throughout, with checks at each stage.
- AdCP’s documentation describes a governance check that runs before execution. Its example sends a request above a $20,000 auto-approval threshold to a person.
Not every announcement describes this step. Magnite’s announcements of its Orchestration product do not mention human approval.
The documented test campaigns
Butler/Till and PubMatic, December 2025 to January 2026. A connected TV, online and mobile app campaign for Clubtails, a Geloso Beverage Group brand. Digiday reported an 82% reduction in supply chain costs, 40% more impressions, a 30% lower CPM and a 98% average video completion rate. The figures came from the agency and PubMatic. An audit by Jounce found a made-for-advertising rate below 1%. The budget was not disclosed.
Magnite and Scope3, December 2025. The first test of Magnite’s seller agent, with Scope3 as buyer agent. MiQ said it also set a buy live that month. No results were published.
Yahoo DSP, Newton and RPA. Yahoo says a trafficking agent built by the two companies has executed programmatic guaranteed buys. No results were published.
Broadsign and Draft Digital, May 2026. An out-of-home campaign over AdCP for a Dutch charity lottery on Global Netherlands inventory. The companies said agents worked with human oversight. No results were published.
Magnite and Amnet France, September 2026. A connected TV campaign for an unnamed automotive brand. Magnite’s buyer agent traded with Magnite’s seller agent. Magnite reported about 70% less setup time.
Where sources disagree
- Setup time. PubMatic’s launch release cited an 87% reduction in early tests. Digiday reported that PubMatic said 98% for one client. PPC Land reported 80 to 90% from PubMatic’s first-quarter 2026 results. The baselines are not published.
- Autonomy. PubMatic’s December 2025 announcement called the Butler/Till campaign “fully autonomous.” Digiday’s later account shows agency staff approving inventory.
- Cost. The Butler/Till test reported a 30% lower CPM. PPC Land reported that a June 2026 DataBeat study found conventional buyers had a 13.4% CPM advantage over AI agents. We did not review that study.
Limits and risks
- Scale. Magnite wrote that the industry will need systems that support hundreds of thousands of campaigns, and that it was “far too early to predict outcomes.”
- Self-reported results. Almost all figures come from the vendors involved.
- Fraud. PPC Land reported that researcher Augustine Fou doubts protocols that rely on honest declarations, citing past abuse of ads.txt.
- Manipulation. Aaron Andalman of Cognitiv wrote in AdExchanger that current agents lack the feedback loops to learn from outcomes, and cited an experiment in which an agent was talked into giving away stock.
- Reach. PPC Land reported that Google, Amazon, The Trade Desk and Microsoft were absent from AdCP’s founding roster. An agent using that protocol cannot assume it reaches their inventory.
Sources
- Butler/Till’s first agentic media buying tests cut media and supply chain costs, Digiday, March 17, 2026
- PubMatic, Butler/Till launch AdCP-enabled agentic AI campaign across CTV, Advanced Television, December 15, 2025
- PubMatic Launches AgenticOS, the Operating System for Agent-to-Agent Advertising, ExchangeWire PressBox, January 6, 2026
- PubMatic launches agentic advertising OS with live campaigns running, PPC Land, January 5, 2026
- PubMatic’s AI narrative can’t hide what the numbers actually show, PPC Land, May 14, 2026
- Why Magnite Built a Seller Agent — and What It Signals for AdCP, Magnite, January 6, 2026
- Magnite launches agent coordination layer for ad buying, B&T, June 12, 2026
- Magnite’s first EMEA agentic test cuts CTV setup time 70%, PPC Land, September 10, 2026
- AI buying agents gain Samsung, iHeartMedia and DIRECTV ad space via Magnite, PPC Land, September 26, 2026
- Yahoo DSP Advances Its Buying Platform with New Agentic AI Capabilities, Yahoo, January 6, 2026
- Welcome to the Agentic AI Era: How Yahoo DSP Is Redefining Programmatic, Yahoo, January 22, 2026
- The Ad Context Protocol Aims To Make Sense Of Agentic Ad Demand, AdExchanger, October 15, 2025
- AdCP documentation: introduction, AgenticAdvertising.org, accessed November 16, 2026
- Broadsign and Draft Digital’s AI agents run first fully agentic OOH campaign, PPC Land, May 31, 2026
- The AdCP Hype Problem: Why Standardized AI Workflows Don’t Equal Better Media Outcomes, AdExchanger, January 26, 2026