French media holdcos put agentic AI to work, mostly upstream of buying (1 of 2)

Where do the big French media agencies actually stand on agentic projects? Who works with whom? Who is testing what, and who is ruling what out?

Open Garden set out to answer those questions with a factual snapshot, built on interviews with holding companies and independent agencies, focused on real usage, the lines where humans stay in control, and the first measurable effects.

We start with the holdcos. Publicis Media declined our request, calling the subject "too confidential."

Dentsu stands out for how actively it works on emerging agentic protocols, starting with tests launched in September 2025 on Scope3's Brand Agent.

The first use case briefs the agent in natural language on a brand's requirements and objectives, then has it find affinity-based, brand-safe contexts that match each advertiser's criteria.

The method delivers results comparable to traditional ad verification tools while unlocking more reach and volume, according to the agency's early findings.

"The first results from this brand agent are enough to push Dentsu to keep going on agentic initiatives, particularly through the AdCP protocol," said Julien Domingues, head of trading at Dentsu France.

The agency also hopes to test the Buyer Agent that Scope3 launched in March. "On that front we stay dependent on the ad sales arms' roadmap. They still have to deploy their seller agents," Domingues said.

TF1 Pub, Prisma Media and CMI Media, among the most advanced publishers on these topics, should be among the first partners on campaigns run agent to agent. Dentsu has also started tests with Supply Finder.

Dentsu is working on MCP architectures too, including Amazon Ads', and on AdCP, so its teams can run campaigns from a single LLM connected to partner DSPs and platforms.

Tests are under way with Amazon Ads, Criteo and PubMatic. "For Criteo, the user authenticates with the Criteo platform, connects through an MCP protocol, then configures the campaign in natural language (brief, creatives, dates, key parameters) from our dentsu.Connect platform," Domingues said. Everything is then injected and executed automatically in the DSPs.

The main opportunity for agencies is operational efficiency. The digital chain is extremely fragmented once you add up every task an activation requires.

"AI can cut that complexity and save time on operations with low perceived value. But that efficiency only makes sense if it comes with stronger strategic counsel," Domingues said. The point is to free up time to make teams more senior, inform strategic choices and reposition the agency as a guide in a very dense environment.

Over the medium term, Domingues expects profiles to adapt, new skills to come in, expertise to hybridize across programmatic, social and search, and roles able to supervise and challenge agents to grow in importance.

The stakes for Dentsu are twofold: "gain operational efficiency on a complex digital chain, while keeping a differentiating advisory capability and deploying responsible, secure AI aligned with the most demanding standards," Domingues said.

At WPP Media, agentic work is industrialized and centralized. The launch of Agent Hub (formerly OMA, Open Media Assistant), a cross-functional agentic layer built into WPP's Open OS, marks the structuring step.

"Every locally developed agent has been rationalized, categorized and consolidated into a global architecture, with about 10 super-agents refined by the group's teams," said Adrien Lepage, managing director of WPP Media DTA (Data Technology Analytics).

The first use cases sit deliberately upstream of execution: brand DNA analysis, translating marketing targets into media audiences, and creative ideation from briefs. Adoption is there. "90% of staff in France use the platform," Lepage said.

On media execution the rule is blunt: "human in the loop, non-negotiable," Lepage said. Budget trade-offs, inclusion lists and investment safety stay behind human checkpoints. Automation targets operational efficiency and continuity of reasoning between planning and activation, not blind delegation.

WPP already reports 30% time savings and 10% to 15% performance gains in some mature markets, though France is still in the adoption phase.

For now, agentic AI is a decision-support tool more than an activation tool. But tests are planned in France with Concord, the agentic platform that automates programmatic and social campaign execution.

The goal is to compare the platform's performance with a trader's on KPIs such as time to launch, management cost and campaign ROI, then see whether the few percentage points of media fee Concord takes are offset by the time and performance gains it generates.

The group also expects its business model to shift, "with advanced thinking on usage-based models, potentially tokenized, at the border between agency service and SaaS," Lepage said. At this stage, agentic AI is treated as an internal standard rather than an external competitive edge.

Havas Media is taking a gradual, pragmatic route. Agentic AI sits first in a workflow optimization logic, through the continuation of the Arche transformation program, built to automate the main steps of the media chain from brief intake to invoicing, said Nathalie Bajeux, chief digital officer at Havas Media Network.

"Arche is due for industrial rollout across the group in 2026," Bajeux said. The first use cases focus on augmented media planning, drawing on campaign history, briefs and internal benchmarks.

Alongside the launch of Ava, a secure portal giving access to the main LLMs on the market, announced at CES by Yannick Bolloré, about 15 proofs of concept have run on programmatic setup through internal agents, including inside Converged with Match AI, with human validation kept at every step.

The aim is to become more platform-agnostic, reduce dependence on individual traders' preferences and, ultimately, move their role toward higher-value work. "The tests show reasonable fidelity to human intent, but also limits in how finely briefs are translated," Bajeux said. Like many others, she sees "real opportunities to optimize direct deals, which are as time-consuming as they are artisanal."

Havas is clearly positioned as a lab: no imposed quantitative targets, no fixed calendar, but a desire to learn before industrializing.

That is why the group has not committed to tests with third-party tools yet. "When we are able to pinpoint our needs and the direction we want to take, we will leave ourselves the option of working with outside partners," Bajeux said. Scope3's name came up during the interview, but the door appears open to anyone.

On economics, Bajeux flagged a key point: "the productivity gains agentic AI promises are largely absorbed by the technology, human and organizational investments it requires." Agentic AI does not mechanically trigger a shift in the model, she said, but fits into a performance logic already written into client contracts.

At Omnicom Media Group, agentic AI is not designed to replace talent but to amplify it, said Jérémy Fritsch, co-head of digital at OMG France. Provided three things are under control: "the quality of the data corpus, the way you talk to the machine, and critical thinking about what comes back."

That is how the agency intends to avoid the homogenization risk that comes with overusing agentic tools. "The real differentiator is still the human. The human formalizes the need and knows the right questions to ask, which brings you back to major issues of training, data culture and business understanding," Fritsch said.

AI suggests, the human decides, a pattern found at most media agencies. The human role stays essential to interpret results, challenge them and spot inconsistencies. "Critical thinking remains the real key to turning technological power into strategic advantage," Fritsch said.

The Omni platform holds most of the development. It aggregates first-party data, client econometric models and business tools, with modules for audience exploration and media planning.

As at every holdco interviewed, automated activation stays limited: setup and execution are still largely manual. Partly because programmatic guaranteed, less time-consuming to configure than finely tuned open auction campaigns, is a structural share of the group's spend.

"Advertisers' big upper and mid funnel priorities are handled today through PG and programmatic direct, with volume guarantees over a set period and negotiated, dedicated prices, so the impact of automation is very limited," Fritsch said.

No immediate impact on headcount is expected. "Low-value tasks exist, but they are a fraction of what our traders actually do, which is largely negotiation and advice," Fritsch said.

OMG has deliberately chosen not to depend on third-party agentic tools, to avoid losing control of the underlying logic. "We do not plan to jump on every initiative that emerges. Everything will be audited and assessed for relevance and ability to scale," Fritsch said.

Interest in agent-to-agent exchanges with ad sales arms exists, but runs into uneven platform maturity and the absence of widely adopted standards. AdCP is barely deployed so far, though Prebid's recent announcement could change that.

In the short term, agentic AI changes neither the business model nor the structures, but it prepares a deeper hybridization of skills and raises the bar for people able to understand and challenge the machine. "Agentic AI does not replace Omnicom's talent, it amplifies it," Fritsch said.