French independent agencies push agentic AI into campaign setup (2 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.

After a first installment on the holdcos' strategies, we move to the independent media agencies, with an update on Values.media, Biggie, Haiku, Jellyfish and CoSpirit.

Values.media first equipped itself with a secured internal agent through French vendor Cominty, used as a closed environment to run different models without exposing sensitive campaign data.

Rolled out to part of the consulting and trading teams, the assistant runs on preconfigured prompts covering the most frequent use cases, with gradual, voluntary adoption.

That group of users meets regularly for training and to share best practice. "There is no instruction manual for our jobs. It is up to each person to volunteer, experiment, test and learn, and up to the agency to build the framework to share use cases and create momentum around these new practices. We think that is the fastest way to bring everyone along," said Emmanuel Crego, managing director of the group.

The agency quickly went further, developing with the same partner agents able to connect directly to buying platforms. First tests ran on the DSP Xandr, with full campaign creation from prompts (budgets, targeting, line items and optimization options) plus access to reporting.

"On Xandr, the tests showed an ability to reproduce prompted instructions faithfully, with no unpredictable behavior. When the agent lacks information, it asks the user rather than making arbitrary decisions," said Olivier Lavecot, innovation director at the agency.

A Meta connection is also available, though stability work remains before large-scale rollout. The goal now is to extend those connectors to other social business managers and DSPs, particularly as part of a migration to The Trade Desk.

"Operational gains are hard to quantify at this stage. Teams have not precisely measured the time saved, because the test phases involve more human checks," Lavecot said.

Campaigns have genuinely been set up by an agent, but the agency keeps strict human supervision, with systematic checks in the platform interfaces before anything goes live.

The logic stays pragmatic: simplify existing workflows, particularly on complex campaigns where several hours of configuration can drop to a few dozen minutes, without trying to reinvent the job or build new AI products.

The agency also holds an independent stance toward third-party tools, to avoid stacking up channel-specific products that would recreate silos. That is why SSP-deployed agents do not, as things stand, interest Values.media much.

The main issue today is the economics and adoption across the ecosystem. Until ad sales arms and platforms standardize their APIs or their agentic logic, the gains will stay partial. Values.media, the first French media agency to buy a TV campaign through an agentic system, hopes French broadcasters will move quickly to build their own.

Crego does not expect headcount cuts in the short term, but rather a shift in the trader's role toward supervision, negotiation and defining the rules agents will have to apply.

Biggie's approach stacks specialized agents inside a proprietary interface, Wallace, used daily by the large majority of its teams.

Agentic AI comes in at several levels: brief enrichment and strategic analysis, display campaign setup still being structured, creative analysis tools, and correlation reporting to steer optimization.

Biggie was also the first media agency to test Supply Finder, an agentic copilot tried on about 10 test campaigns. For a travel advertiser, the agent generated up to 90 variations per destination, which would have been impossible to do manually on a roughly €10,000 campaign.

"The result is significant performance gains, around 40% more qualified visits," said Baptiste Pechery, head of media trading at Gamned. Pechery said he is also considering testing a similar product from C Wire.

Like WPP Media, Biggie is moving forward with Concord, comparing the performance of a campaign set up manually inside the DSP with one run from Concord in natural language. "We are starting with DV360 on the open web and we cannot wait to see the results," Pechery said.

He sees agentic systems as a promising route for media plans that are as heavy as they are granular, "with hundreds of line items per campaign that take forever to configure."

He also sees a good answer to the hyper-fragmentation of some media environments, such as CTV, where buyers sometimes have to go through nearly as many buying interfaces as there are ad sales arms.

Beyond digital execution, Biggie is considering broader uses, including offline environments such as TV and out-of-home, with tools able to spot buying opportunities based on inventory availability and data granularity. "But we are not there yet," said Erwan Lohezic, managing director of Biggie.

Biggie is clearly in advanced rollout, with a stated intention to industrialize agentic AI and turn it into a commercial product. Wallace is already offered to some clients as a license, in a model close to SaaS, combining human time and machine time in recurring revenue.

The strategy comes with a clear-eyed view of the social impact: probable cuts in support headcount, gradual disappearance of some junior roles, and a rise in highly technical and analytical profiles, a shift the agency considers inevitable over time.

Jellyfish recently validated a first deployment built on Amazon Ads' MCP protocol, with the goal of moving away from interfaces traders operate by hand toward control through natural-language requests executed by agents.

At the center sits Pencil, Jellyfish's proprietary platform, developed inside the Brandtech group. Originally designed as an ad creation tool, Pencil has gradually become a control tower for the media workflow.

"The platform now covers the whole chain: media planning generated in minutes from a brief, real-time insight extraction and, soon, automated campaign setup on Meta and Google," said Rafael Lasnier, head of programmatic France at Jellyfish.

Jellyfish's strategy rests on a hybrid architecture. The business intelligence is concentrated in Pencil, while platform connections rely on the market's emerging standards. The agency already uses MCP to talk directly to Amazon Ads' API and plans to test AdCP through Scope3.

"The idea is to code our own intelligence layer on top of those standards. We use Pencil to drive these protocols in a differentiated way, which gives us an agility the platforms' standard interfaces do not offer," Lasnier said.

The first test, on a Monoprix activation, shows the operational potential. Combining Pencil agents and MCP, Jellyfish completed a full campaign setup in barely an hour:

"Up to 65% of the operational setup load was absorbed by AI," Lasnier said, adding that the time saved on mechanical tasks is reinvested in refining strategy, quality control and performance analysis.

Next comes a test with Scope3 on direct deals through AdCP. The end goal is clear: "Pencil has to orchestrate every channel, from social to search to video."

Agent autonomy is deliberately fenced in. At Jellyfish, AI can execute certain technical actions, but strategic arbitration, budget management and brand suitability stay strictly in human hands.

Lasnier insisted on clear governance. Business rules (brand safety, exclusions, advertiser constraints) are written directly into what the teams call the agent's "constitution." Auditability of decisions is also a prerequisite for trusting these systems.

On economics, the first ROI is mostly operational efficiency: a setup that used to take half a day now runs in about an hour. "In the short term, though, part of that gain is offset by the investment needed to train the agents, test the protocols and train the teams," Lasnier said.

The shift is already changing skills. The trader's job is moving toward profiles able to steer and supervise agent systems rather than configure platforms by hand. Jellyfish has introduced two hours of training a week for all its teams to speed that up.

Longer term, the agency also expects business models to change, with hybrid approaches combining SaaS licenses, token-based usage billing and possibly performance-based pay. In that scenario, access to agents will probably become a market standard. Differentiation will then hinge on how well agencies can inject their media expertise into these architectures and orchestrate increasingly open ad ecosystems.

At Haiku, agentic AI is advancing step by step, with one clear priority: structure the data before automating media execution. The independent agency sits between exploration and first proofs of concept, aiming for partial rollout in 2026.

"With no dedicated R&D department, maturity has come in stages, driven by a cross-functional AI committee and reinforced recently by a strategic audit run by Amadia," said Elie Boussira, one of the agency's co-founders.

That qualitative phase identified the main pain points and ranked use cases, while setting a more structured view of AI inside the organization.

The priority for the coming months is structuring data flows around Alfred, the agency's proprietary tool, which won gold in the media agency innovation category at the Open Garden Awards.

"The challenge is harmonizing heterogeneous metrics, TV GRPs, digital CPMs, data from social or the big platforms, to build a reliable foundation for future agents," Boussira said.

That step also involves connecting office tools and media platforms through APIs, plus internal work on translating KPIs across the agency, seen as essential to keep automated analysis consistent.

For Haiku, in other words, the initial investment sits less in agentic AI itself than in normalizing the data.

Operationally, the agency is building several productivity bricks in-house: an e-assistant to aggregate data, automated dashboard generation, reporting summaries and optimization recommendations.

In parallel, Haiku is testing Concord with a view to quick, concrete effects. Originally built for programmatic, the tool is being explored here on social and search environments, to assess how relevant it is beyond its historical scope.

"This test phase is as much about understanding what the agent can really do as about getting teams used to new ways of working," Boussira said.

Governance stays tight. Agent autonomy is deliberately limited: budget commitments and setup changes remain under human control.

That caution answers identified risks, particularly errors tied to misuse of AI or badly calibrated configurations. In that light, the agent is seen first as a decision-support and operational-safety tool, able to generate alerts and cut human error.

On staffing, Haiku does not expect cuts. The goal is productivity, freeing traders for higher-value work while platforms such as Google and Meta already automate a growing share of trading.

Over the medium term, Boussira expects agentic AI to become a market standard: "differentiation will not come only from the technologies used, but from how they are integrated into a broader approach mixing data, advice and creative."

For an independent, the shift is also a reassurance play with advertisers, showing that technological innovation and premium service can coexist outside the big networks.

Before building agents, CoSpirit's teams first mapped possible AI use cases across their jobs to prioritize the ones that could generate operational value quickly. That phase identified several concrete applications, particularly around data use and performance analysis.

The first agents are now in production. They plug directly into Cockpit, the reporting and media analysis environment the agency built. Fed by the internal data warehouse, Cockpit already gives teams and advertisers a set of dashboards to track campaign performance. The agentic layer now takes it past those static dashboards.

"In practice, users can query the data in natural language to get specific analysis a standard dashboard cannot produce on its own: unusual variable crossings, highly contextual questions or one-off analysis," said Eric Boyer, managing director of CoSpirit.

The agent interprets the request, generates the matching SQL query, hits the data warehouse and returns a structured analysis in seconds. The point is not to replace dashboards but to add conversational exploration of the data.

Technically, the product was built entirely in-house. CoSpirit uses cloud and open-source components, but its teams designed the architecture, the business logic and the agent orchestration. "That lets the agency keep control of its technology stack and its data security," Boyer said.

The first use cases fall into three areas. Performance analysis and media planning comes first: agents let teams explore historical data quickly to inform investment decisions and operational trade-offs.

The second is business intelligence, making data easier to reach for non-technical profiles who can query campaign performance directly. Third, some operational or administrative tasks can be prepared by agents, always under human supervision.

For now the agent works mainly on the data warehouse exploitation layer. Longer term, CoSpirit plans to bring more measurement into Cockpit, including post-tests, attribution models and MMM, to enrich the analysis and feed strategic recommendations and budget decisions more directly.

The agency is also watching the emergence of interoperability standards such as AdCP, which could open the way to more automated interactions with platforms and ad sales arms.

Autonomy stays deliberately limited. The agent works today as an analytical assistant able to understand a request, query the data and produce a structured answer. Committing decisions, such as launching campaigns, budget trade-offs or targeting choices, stay strictly under human control.

To frame the work, CoSpirit set up an internal AI committee bringing together technical and business experts. The committee oversees the projects, prioritizes use cases and keeps deployments aligned with operational needs.

Technical guardrails are in place too: human validation for sensitive actions, traceability of processing, test environments before production and strict control of data access.

The first benefits are mostly speed of access to information. "An ad hoc analysis that used to take several hours of extraction and formatting can now be generated in seconds," Boyer said.

He also sees better data democratization: more profiles can use the available information directly without systematically going through technical experts.

CoSpirit does not expect headcount cuts, but rather a gradual transformation of the jobs. The most repetitive tasks should be increasingly automated, letting teams focus on strategy, data interpretation and client relationships.

Within two to three years, the agency expects a majority of operational digital tasks to be agent-assisted. Media strategy, the trade-offs between branding and performance, and understanding advertisers' business issues will stay largely human.

Like most of the players interviewed, CoSpirit expects agentic AI to become a market standard. Differentiation will hinge less on having agents than on data quality, the robustness of the software stack and how well agencies can inject their media expertise into these new systems.