Daily Pulse · March 17, 2026 · 3 signals
Who Needs the Middle?
The intermediary layer is under coordinated pressure.
Meta Built an AI Agent That Runs Its Own Ads Ranking R&D. It Works.
Change
Meta automated its ads ranking R&D at 5x productivity
→ Meta's modeling advantage now compounds on a timeline competitors can't match
Why it mattersOptimization gap widens faster than hiring can close
5×
Engineering productivity gain
2×
Model accuracy improvement
8
Ads ranking models managed
Meta's Ranking Engineer Agent (REA) is an autonomous AI system that manages the full machine learning lifecycle for Meta's ads ranking models: generating hypotheses, executing training jobs, debugging failures, analyzing results, and allocating GPU resources. It operates via a "hibernate-and-wake" mechanism – suspending during multi-day training runs and resuming automatically. In its first production deployment, three engineers used REA to improve eight ads ranking models. Previously, that work required two engineers per model.
The 2x accuracy improvement and 5x productivity gain are significant on their own. The structural implication is larger: the bottleneck in competitive advertising AI has always been engineering iteration speed. Training a new ranking model takes days. Analyzing results takes days. The cycle from hypothesis to deployed improvement was measured in weeks. REA compresses that cycle to automated loops that run continuously while engineers sleep.
Competitors running ads ranking R&D with human teams cannot close this gap by adding headcount. The iteration advantage is architectural. Meta publishes this as a technical blog post – the kind of announcement that's easy to process as "cool internal tool" rather than what it actually is: a permanent structural change to how fast Meta improves its ad auction.
▲ Wins
Meta – permanent widening of the ads ranking quality gap vs. every competitor still running manual R&D cycles. Advertisers whose ad performance improves as Meta's models get more accurate, faster.
▼ Loses
Competing ad platforms whose modeling teams rely on human iteration cycles. Any platform whose competitive positioning assumes it can out-engineer Meta over a 2-3 year horizon – that horizon just compressed to a continuous automated loop.
◆ Pulse Take
REA doesn't make Meta's ads team larger – it makes them faster than any competitor can replicate. Every week that passes, Meta's modeling advantage compounds in a way that headcount can't match.
Publicis Failed a Trade Desk Audit. Now It Won't Recommend Them.
Change
Publicis withdrew TTD recommendation after Kokai consent audit
→ Every DSP must now decide: is its AI layer a service or a risk?
Why it mattersAI automation becomes a compliance liability for DSPs
3
MSA violations found
$2.9B
TTD 2025 revenue
47%
TTD profit margins
An independent audit commissioned by Publicis and conducted by FirmDecisions found three specific MSA violations: The Trade Desk improperly applied its DSP fee on top of other fees; clients were automatically opted into additional tools without consent; and TTD refused to provide data validating that media and data costs were invoiced at cost without mark-up. TTD denied the findings, citing client confidentiality.
The third violation is the most consequential: Kokai – TTD's AI platform – was found to be automatically changing campaign settings without advertiser input. This is the same feature TTD sells as a competitive advantage. An audit finding reframes it as a compliance liability. Publicis has formally halted client recommendations.
The revenue and margin data shows this isn't happening to a weak business: $2.9B in 2025 revenue, 47% profit margins, $1.3B cash on hand. The challenge is structural. The largest holding company just institutionalized a buying alternative – and any agency that hasn't audited its own TTD MSA now has a reason to.
▲ Wins
Amazon DSP, direct publisher deals, and competing DSPs not named in the audit. Agencies that can credibly route client spend to alternatives while the audit narrative runs its cycle.
▼ Loses
The Trade Desk – a formal recommendation withdrawal by the world's largest holding company is the opening move of a structured revenue erosion campaign. Kokai's AI automation layer, previously a competitive feature, is now framed as a compliance risk in every agency pitch deck.
◆ Pulse Take
The audit didn't find TTD is expensive – it found Kokai was making campaign changes without consent. Every DSP now needs to decide: is its AI automation layer a disclosed service or a liability?
An AI Agent Replaced a DSP on a Live Campaign. It Outperformed.
Change
AI agent bypassed DSP layer and outperformed on CTV
→ The DSP's 'we make buying efficient' pitch just met a counter-example
Why it mattersDSP intermediary layer loses its economic justification
82%
Supply chain cost reduction
40%
Impressions lift, same budget
30%
CPM reduction
Butler/Till deployed PubMatic's AgenticOS – powered by Anthropic's Claude – to run a CTV campaign for Geloso Beverage Group in December 2025 through January 2026. The agent retrieved curated inventory directly from publishers (Samsung, Paramount, Vizio, Tubi), sent selections to agency staff for approval, and executed the buy directly with PubMatic. No DSP was in the stack.
Results, verified by Jounce and DoubleVerify: 82% reduction in supply chain and DSP tech costs. 40% more impressions from the same budget. 30% lower CPMs. 98% video completion rate. Less than 1% MFA rate. 80% of inventory above DoubleVerify benchmark.
The case study was published March 17. The campaign ran premium CTV inventory at better economics than the DSP stack it replaced. The 98% campaign setup time reduction – reported by MediaPost – means the efficiency case isn't just about fees. It's about the entire operational overhead the DSP layer was supposed to justify.
▲ Wins
Independent agencies with technical capability to deploy agentic buying. Publishers with clean, directly-accessible inventory. PubMatic's AgenticOS – earns a marquee case study as DSP-bypass infrastructure.
▼ Loses
Independent DSPs whose core value proposition is efficiency and data access. The campaign ran on Samsung, Paramount, Vizio, and Tubi – premium CTV inventory – without a DSP layer. The case study proves the layer was optional.
◆ Pulse Take
This isn't a proof of concept – it's a published case study with live client data proving 82% cost reduction and 40% impression lift. The independent DSP's pitch of 'we make buying more efficient' just met a peer-reviewed counter-example.
3 signals · March 17, 2026