AI agents poised to replace DSPs for indie media buying

AI Agents and the DSP Seat Fee

The pitch is simple: an independent agency stops paying DSP seat fees, routes spend through Claude or ChatGPT, and lets an AI agent plan and buy across Meta, Google, and Amazon. A Digiday podcast floated this as the coming model. Butler/Till is already calling it "agentifying workflows."

The question is who owns the layer between an advertiser's budget and inventory. Demand-side platforms (software agencies use to buy ads across many sites and apps) have owned that layer for a decade. The question is whether an AI agent takes it, and if so, who the agent answers to.

What is actually plausible now

The Meta-Claude integration exists, but it is a wrapper on one walled garden. A true cross-platform agent needs to hold state across campaigns, pace spend in real time, enforce brand safety, stitch attribution, and carry contractual indemnification. None of that is in the current demo. Nobody has shown a flight running thirty days unattended with no human override.

What does happen first: the account manager breaks before the DSP contract does. Ops teams pushing Meta spend through Claude hit rate limits, attribution holes, and brand-safety gaps that a human was quietly papering over. The DSP software survives. The managed-service relationship around it does not. Once those campaign-ops bodies are cut, there is nobody left to defend the seat fee in the budget review.

Who wins and who loses

The walled gardens, meaning Meta, Alphabet, and Amazon, win. Their first-party interfaces become the execution layer the agents plug into. Every dollar an agent routes through Meta's own API skips The Trade Desk (the largest independent ad-buying platform) entirely. That is not disintermediation of the platforms; it is the three platforms you already could not audit becoming the execution layer too.

The Trade Desk is the exposed name. Its stock carries a multiple built on dominant independent DSP status. Fee pressure and agent hype do not need working autonomous agents to bite. They need CFOs at independent agencies to question platform minimums in 2027 budget reviews, which is exactly when that conversation happens. Independent agencies are already saying on the record that they resent the seat fee and see Claude as the exit.

There is a real CFO objection here. A DSP fee is a known percentage, and because a DSP does not own the inventory it buys, its measurement is not conflicted. An agent that allocates freely across Meta, Google, and Amazon grades its own homework inside each walled garden. The payback only works if the agent is cheaper after you price in the third-party verification you now have to bolt back on.

Our call

The Trade Desk's full-year 2026 platform spend growth, reported on its February 2027 earnings call, comes in below its full-year 2025 growth rate. Medium confidence. TTD's open-web scale cushions a near-term hit, but the griping is on the record and the walled gardens are actively courting agent integrations to capture that redirected spend.

For independent DSPs, the defensible ground is the one thing the walled gardens cannot replicate: you do not own the inventory, so your measurement is not conflicted. That is the argument to make. The seat fee was never the moat.

The pitch on a Digiday podcast is clean enough to put on a slide: an independent agency drops its stack of buying platforms, routes spend through Claude or ChatGPT, and lets an AI agent plan, negotiate, and buy across Meta, Google, and Amazon. No more DSP seat fees. No more Frankenstein stack. Just deposit the dollars and let the machine allocate.

Here's what's actually being decided, and it isn't "do DSPs die." It's who owns the layer that sits between an advertiser's budget and the inventory. A DSP (demand-side platform, the software agencies use to buy ads across many sites and apps) has owned that layer for a decade. The question is whether an AI agent eats it, and if so, who the agent answers to.

How hard is this to undo? For an agency, easy. Routing Meta spend through Claude today and back through a DSP next quarter costs almost nothing. For the public DSPs whose revenue is the seat fee, the repricing is hard to undo once it starts, because agency budget cycles don't swing back.

What sets the deadline? Nothing hard. But the 2027 agency budget reviews are where the "why are we paying platform minimums" conversation actually gets had.

---

The Skeptic. For this to work, the agent needs to hold state across campaigns, pace spend in real time, enforce brand safety, stitch attribution, and carry contractual indemnification. Meta's Claude integration does none of that. It's an AI wrapper on one walled garden. "Agentifying workflows" at Butler/Till is automation with a fresh coat of paint. The reason the stack is a Frankenstein is that the ad world is genuinely fragmented at the data and contract level, and no prompt resolves a publisher's floor price or a supply-path audit. Show me the flight that ran thirty days unattended with no human override. To a non-specialist: the demo buys ads in one store and stops there.

The Operator. The account manager breaks before the DSP contract does. In the first half of 2027, ops teams at shops like Butler/Till will try to push Meta spend through Claude and immediately hit rate limits, attribution holes, and brand-safety checks their current verification vendors were quietly papering over. The DSP software survives. The managed-service relationship around it does not. Then the second-order effect: trafficking and campaign-ops headcount gets cut, and once those bodies are gone, the seat fee has nobody left to defend it in the budget review. For a non-specialist: the robot doesn't replace the tool, it replaces the person who used to run the tool.

The Market Analyst. The Trade Desk is the exposed name and the market knows it. Its stock carries a premium built on being the dominant independent DSP. If agent buying even partially commoditizes the seat, the multiple comes down. Anthropic and OpenAI don't make money on media buying, so the equity winners are Meta, Alphabet, and Amazon, whose first-party interfaces become the actual execution layer the agents plug into. Consensus treats this as a five-year story. The agency budget cycle means it hits profit-and-loss statements inside eighteen months. For a non-specialist: the money moves back to the big three platforms, and the independent middleman loses pricing power.

The CFO. The seat fee is the easy target, but it's not the real cost. A DSP fee is a known percentage. An agent that allocates "wherever it wants" across Meta, Google, and Amazon is a budget with no referee, grading its own homework inside each walled garden. Who audits whether the agent overspent on the platform that trained it? The independent DSP's one honest feature is that it doesn't own the inventory it buys, so its measurement isn't conflicted. Hand execution to an OpenAI or a Meta interface and you've traded a transparent fee for an opaque one. The payback only works if the agent is cheaper after you price in the verification you now have to bolt back on.

---

Where the council splits. Two real disagreements.

First, the Skeptic versus the Market Analyst on timing. The Skeptic says this is a panel topic, in production nowhere, and the gap between the Meta demo and an autonomous cross-platform buyer is years of unglamorous plumbing. The Market Analyst says it doesn't matter whether the agent is fully autonomous, because the repricing of the DSP seat starts the moment CFOs believe it's coming, and belief moves on the budget cycle. The engineering cycle is irrelevant to the budget conversation.

Second, everyone versus the clean story. "DSPs are dead" obscures the messy part: the agent has to run on somebody's rails, and today those rails are Meta's, Google's, and Amazon's APIs. If agents win, the open web buyer fragments while the walled gardens consolidate more volume. That's not disintermediation. That's the three platforms you already couldn't audit becoming the execution layer too.

What it hinges on. One belief: can an AI agent deliver cross-platform buying with transparent measurement and brand safety, or only single-platform buying that routes money deeper inside walled gardens? If the former, independent DSPs reprice hard. If the latter, the walled gardens win and the "agent revolution" is just a nicer front door to Meta and Amazon. Nothing in the Meta-Claude integration demonstrated the former.

The council leans one way. The seat fee is under real pressure and the buy side is openly griping about it. But the beneficiary is not the agent and not the agency. It's the platform the agent is pointed at.

---

Prediction: The Trade Desk's full-year 2026 platform spend growth rate, reported on its February 2027 earnings call, will come in slower than its full-year 2025 growth rate.

Confidence: Medium. Fee pressure and agent hype are real, but TTD's open-web scale cushions any near-term hit.

Why: Independent agencies are now saying out loud, on the record, that they resent DSP seat fees and see routing Meta spend through Claude or ChatGPT as a way around them, and Meta has already shipped that integration. That griping doesn't need working autonomous agents to bite; it needs CFOs at independents to question platform minimums in 2027 budget reviews, which is exactly where "why are we paying this seat fee" gets asked. The walled gardens are courting agent integrations precisely because every dollar an agent routes through Meta or Amazon's own interface is a dollar that skips the independent DSP. The opposite outcome, growth re-accelerating, would require TTD to prove the agent threat is empty at the same moment its largest customers are publicly shopping for the exit, which is the harder story to tell.

Revisit by 2027-02-28: We're right if The Trade Desk's 2026 full-year revenue growth rate is below its 2025 full-year growth rate. We're wrong if 2026 growth matches or exceeds 2025.

One more thing worth saying plainly. If you run an independent DSP, the defensible ground is the one feature the walled gardens can't copy: you don't own the inventory, so your measurement isn't conflicted. Lean into that. The seat fee was never the moat.