PubMatic and Amnet Deploy Claude LLM for France's First Agentic Ad Campaign

PubMatic and Amnet launched France's first agentic advertising campaign on 31 March 2026, using Anthropic's Claude LLM to automate bid optimization and creative selection without human intervention between impressions. The deployment marks the first known instance of a large language model making autonomous spend decisions in a European programmatic environment.

The campaign operates as a closed loop: Claude ingests real-time performance data from PubMatic's sell-side platform, adjusts bid parameters across inventory sources, and selects creative variants based on contextual signals—device type, time of day, publisher environment. Amnet, Dentsu's programmatic arm, provided the client brief and success metrics. PubMatic provided the infrastructure. The LLM writes its own optimization instructions every 90 seconds, a cycle time approximately 40 percent faster than the previous generation of rule-based systems.

The significance is operational, not conceptual. Agentic systems—software that sets its own goals within defined parameters—have existed in algorithmic trading for two decades. What changed is cost and accessibility. Running Claude at campaign scale became economically viable in late 2025 when Anthropic reduced inference pricing by 60 percent and introduced batch processing for ad tech workloads. PubMatic's infrastructure now supports sub-second API calls at a cost structure that fits inside typical programmatic margins. This is the first French deployment, but five other European markets have pilots scheduled for Q2 2026.

The privacy architecture matters for replication. Claude operates on anonymized performance vectors—CTR, conversion rate, cost per acquisition—without accessing user-level data. The system never sees names, device IDs, or browsing history. It sees only aggregated signals that comply with GDPR's "legitimate interest" standard for optimization. This design choice limits the model's contextual depth but makes regulatory approval straightforward. Amnet confirmed the campaign cleared France's CNIL without a formal review, a procedural detail that shortens deployment timelines for other buyers.

Two constraints define near-term adoption. First, agentic systems require high-volume campaigns to generate statistically valid signals for the LLM. Amnet has not disclosed the daily impression count, but industry participants estimate the minimum threshold at 5 million impressions per day. Smaller campaigns cannot feed the model quickly enough. Second, the system lacks explanatory depth. When Claude adjusts a bid, it can summarize the decision in natural language, but it cannot produce a causal chain that satisfies a CMO's quarterly review. The model optimizes for outcomes, not narratives.

Operators should watch for three developments. First, whether PubMatic publishes performance benchmarks comparing agentic campaigns to rule-based controls. The company has committed to releasing aggregated lift data by June 2026. Second, whether other SSPs announce Claude integrations. PubMatic holds no exclusivity, and The Trade Desk has publicly tested agentic bidding with a different LLM. Third, whether luxury and travel verticals adopt the technology. Both sectors prize brand safety and contextual precision, which current agentic systems handle poorly. A hospitality or fashion brand running an agentic campaign in France would signal that the model's contextual limitations have been solved.

The campaign is live. Amnet has not disclosed the advertiser, the budget, or the KPI thresholds that would trigger a manual override. Those details will matter when the first agentic campaign fails in a way that a human buyer would have prevented.