Butler/Till runs agentic media buying tests across multiple channels

Full analysis

An independent agency, Butler/Till, has spent six-plus months running AI agents that make programmatic media buys, starting in connected TV and spreading to online video, display, and streaming audio. After all that testing, the share of actual dollars flowing through those agents is still low single digits. Chief Strategy Officer Scott Ensign expects double digits eventually, once the guardrails mature.

Reversibility: Type 2 for the agency (you can turn the agents off Tuesday and go back to human traders). Closer to Type 1 for the industry, because once clients accept autonomous spend as normal, the trader headcount doesn't come back.

What's actually being decided: Not "does agentic buying work." The real question is who eats the cost of configuring, supervising, and taking the blame for an autonomous buy: the agency, the DSP, or the client.

Forcing function: None hard. This is a slow-burn adoption curve, not a deadline. That matters for the prediction.

The Market Analyst. In plain terms: the agents need a clean, fast machine to plug into, and only a couple of buying platforms have built one. The winners here are the DSPs with mature APIs and real event streams. Trade Desk with Kokai and Google's DV360 are the surfaces an agent can actually drive. SSPs that never built low-latency programmatic-guaranteed endpoints are the ones exposed, because agents route around slow, human-shaped inventory. Measurement is where it gets interesting: if an agent optimizes in real time, the flights and line items that attribution models were built around start to dissolve. That opens a fresh RFP cycle for real-time measurement. Low single digits is a real number, though. The market this creates today is a rounding error.

The Skeptic. Six experiments and low-single-digit spend after half a year is a pilot that hasn't graduated, not a proof of concept. Look at the channels Butler/Till picked: CTV, online video, display, streaming audio. Those are exactly where a human already mostly presses "go" on a pre-negotiated deal. Calling that agentic is generous. The honest question for Scott Ensign is what the agent did that a well-configured private marketplace and a pacing alert wouldn't. For informed outsiders: this is software making buys inside guardrails so tight it can barely move. The double-digit forecast rests on every client's legal and procurement team agreeing on who's liable when an agent spends money badly. That agreement doesn't exist anywhere at scale.

The Operator. The friction isn't the model. It's the reconciliation problem. When an agent makes a buy at 2am that a human trader would have flagged, someone has to explain that buy to the client the next morning. Right now nobody's job description says "own the agent's decisions." Trafficking, analytics, and account management get pulled into SLA conversations they were never hired for. Ensign's own line gives it away: these environments "have to be controlled on a client by client basis." That means no reusable playbook. Every client is a bespoke guardrail build. That's services work, and services work doesn't scale the way software margins do.

The Customer / End User. The client here is a brand marketer, and most of them didn't ask for this. They asked for lower cost and better outcomes, and "autonomous AI spending my budget overnight" reads as new risk, not new value. That's why the number is stuck in low single digits: the brands are voting with their approval workflows. The bespoke, client-by-client configuration Ensign describes isn't a product feature. It's the sound of every client demanding a custom leash before they'll let the thing off the porch. Adoption moves when one big brand greenlights it publicly and gives the others cover. Until then, everyone waits.

Where the council splits.

The Market Analyst sees a structural shift worth positioning for. The Skeptic sees a pilot dressed up for a conference dinner. Both are looking at the same low-single-digit number and reading it in opposite directions.

The deeper tension is the Strategist's moat question against the Operator's cost reality. If client-by-client guardrail configuration is a durable switching cost, the agency has a business. If it's just a feature toggle the DSPs will embed and sell direct, the agency spent six months building a case study that Trade Desk and DV360 will cash. The bespoke-configuration work is either the moat or the trap, and nothing in the announcement tells you which.

What it hinges on. Two beliefs. First, whether autonomous spend liability gets solved at the contract level. Until legal and procurement sign off on who owns a bad 2am buy, this stays a pilot. Second, whether the guardrail-configuration work is defensible or commoditized. Both point the same direction near-term: this stays small.

Before anyone builds a roadmap around agentic buying, verify the boring thing. Ask what an agent did in these tests that a good PMP and a pacing rule couldn't. If the answer is "moved faster inside a box we drew," that's automation with a better logo, not a new category.

The council leans Skeptic in the near term, Market Analyst over three years. The bridge between those two views is one public brand greenlight, and it hasn't happened yet.

Prediction: No independent or holdco agency will publicly report double-digit spend share through autonomous buying agents before the 2027 upfront/NewFronts cycle (spring 2027); the industry stays in single digits through then.

Confidence: Medium. Liability sign-off is unsolved and moves at legal-department speed.

Why: Butler/Till ran six-plus experiments across four channels for over six months and still moved only low single digits of spend, which tells you the bottleneck is client approval, not agent capability. That bottleneck is contract language on who owns an autonomous buy gone wrong, and legal and procurement teams don't resolve that on a demo timeline. For the number to hit double digits, a large brand has to publicly accept that liability and give peers cover, and nothing in the current tests suggests that's imminent. The opposite outcome, a fast jump to double digits, would require the liability problem to solve itself quietly across many clients at once, which is the less likely path.

Revisit by 2027-05-15: We're right if no agency (independent or holdco) has publicly disclosed double-digit spend share running through autonomous buying agents by the spring 2027 NewFronts. We're wrong if any agency reports crossing 10% autonomous spend share before then.