Marketing agents must prove value before gaining autonomy

Agentic marketing readiness means earning autonomy use case by use case

A marketing team can automate more work and still become less prepared to run it. If an agent depends on a model service the company cannot reliably access, acts beyond its permissions or costs more than the work it replaces, greater autonomy creates an operational problem. Readiness means being able to justify and supervise each use case before expanding it.

An agent is a software system assigned tasks within a workflow. At Ibexa’s June 17 Orchestration Day, participants in the roundtable “Prepare your business for the agentic era” described a future in which these systems help marketers manage “thousands of personalized campaigns.” In their proposed model, agents handle more execution so marketers can focus on strategic coordination. That shift depends on security, human oversight and data sovereignty: control over data under the relevant legal requirements.

Ibexa has a commercial interest in this shift because it is launching an orchestration platform to support it. The company used the event, previously its long-running Activation Day, to discuss that platform and its direction in agentic marketing. Participants included Laurent de Vivie, Marketing Director, BNP Paribas Wealth Management, and Benoit Delisse, VP Actito Product Unit, Ibexa, alongside the speakers discussed below. The event’s new name expresses Ibexa’s preferred operating model; each team still needs to assess whether that model works for its tasks.

Start with one repetitive task, then establish its value

A useful first task is work the team already understands well enough to evaluate. Marketing teams repeatedly produce campaigns under time pressure, often following familiar sequences. Grégory Bécue, CPO, Ibexa, recommends beginning with that everyday friction: “I‘m convinced that you have to solve everyday problems first,” he said. His recommendation supports adoption of the agent-based workflows Ibexa’s platform is designed to serve.

Reusing the previous year’s campaign offers a concrete starting point. A team could assign an agent to clone the campaign, then update its visuals and targeting through a sequence of tasks. Copying the campaign, changing its creative material and adjusting its audience each have a clear purpose and different ways to succeed or fail. Defining those steps gives the team specific outputs to inspect and keeps the assignment manageable for organizations unable or unwilling to undertake an organization-wide overhaul.

A manageable assignment also lets employees build confidence through experience. “Taking small steps, starting with a simple example and moving to increasingly complex use cases is also how you build trust and adoption within your company,” Bécue said. In his proposed approach, teams increase complexity as they gain experience with simpler work. Bécue presents less rushed execution, better productivity, higher-quality work and improved working conditions as the intended benefits; headcount reduction falls outside that objective.

To test those benefits, the team needs to prioritize a specific use case, identify the data it requires and choose a suitable model. The task determines what information the agent needs, while the expected benefit gives the team a basis for comparing execution costs. A narrow assignment makes that comparison easier and offers an early chance to reject weak business cases. Yet even a small assignment can create a substantial dependency on the model service supporting it.

Model sovereignty is an operational-continuity decision

That dependency matters because the workflow needs continued access to the model service. An agent may use a large language model, or LLM, to interpret instructions and produce responses; the campaign process then depends on the agent completing its work. The roundtable identified reliance on US-based AI models as a particular concern for European-based and multinational organizations. Regulatory or geopolitical decisions affecting access can therefore reach directly into marketing operations.

Jean-Claude Pitcho, Head of Partnerships France, Ibexa, described the potential disruption: “When you have agents who rely on some of these LLMs, your operations and your security suddenly come to a complete halt,” he said. Pitcho speaks for a company selling orchestration technology, so his warning also supports the case for managing these dependencies. For teams adopting agents, the implication is clear: model access belongs in planning as soon as a workflow depends on continued execution.

To address that exposure, the roundtable described many organizations moving toward sovereign or open-source LLMs hosted in private clouds. A sovereign arrangement aims to keep model use and data under the jurisdiction and control the organization requires, while an open-source model changes the basis for access and deployment. A private cloud provides a computing environment reserved for an organization. These choices serve different purposes, so teams need to assess how they work together against legal, security and continuity requirements.

BNP Paribas’ partnership with Mistral was the named example in this discussion of sovereign and open-source alternatives. The partnership illustrates an organizational response, while BNP Paribas Wealth Management’s participation in the roundtable provides a separate connection to the discussion. Other European-based and multinational businesses need to assess their own constraints, including who can decide whether they retain access to a model. Private hosting and model labels are useful only if the resulting arrangement provides the required control.

Once an arrangement meets those requirements, the model still needs to do the assigned work effectively. Guillaume Moigneux, Field CTO, Upson, argued that model size alone does not determine usefulness: “We‘re finding that a smaller model, with the right context, good data, and fine-tuned performance, is actually just as effective and delivers results at almost the same cost,” he said. Upson has a commercial stake in how organizations adopt and deploy this technology, and Moigneux is describing its experience. His comparison makes context, data quality and fine-tuning, additional training to adapt a model, conditions of the reported effectiveness.

Those conditions connect model selection back to the use case. A team needs to establish what information the model requires and how it will assess performance on the task. Performance checks need to use the context and data the model will receive in production. Continuity, task performance and expense therefore remain separate questions, each requiring its own assessment.

Keep workflows deterministic and expand agent permissions gradually

Reliable model access addresses an external dependency, but the organization must also control the agent’s actions within its workflow. A deterministic process follows specified steps and conditions, while a non-deterministic agent can produce different outputs or decisions across executions. The roundtable’s proposed approach keeps the process structured and gives agents bounded roles within it. This preserves flexibility within defined limits.

Those limits begin with documentation because the team needs a shared account of the work it intends to delegate. The workflow should identify tasks, responsible roles and the points where outputs need review. In the campaign-reuse example, the documented cloning, visual-update and targeting steps provide those assignments. Reviewers can then judge whether the agent has completed each one.

Documented roles also need rules for acceptable behavior. Semantic constraints are rules about the meaning of permitted actions and outputs, and validation checks whether a result meets those rules before the workflow continues. Permissions determine which resources and actions the agent can access. Together, these controls connect the intended assignment to what the agent can actually do.

Moigneux recommends putting those controls into practice by explaining the agent’s duties, setting objectives and monitoring execution as work begins. He also advises checking its access and actions before granting more autonomy. Early supervision therefore needs to examine individual stages, since a satisfactory final output can leave important questions about how the agent reached it. A team can then reduce intervention where performance is acceptable while retaining checks elsewhere.

Increasing permissions requires someone to decide whether observed execution warrants greater freedom. Moigneux’s guidance addresses that management decision, while Bécue’s description of Ibexa’s monitoring capabilities concerns the information available to support it. These are different parts of supervision: records let people inspect activity, and managers decide what that activity justifies. Even controlled execution, however, must still justify its cost.

Compare execution costs with realized value

An agent can perform an acceptable task at an unacceptable price. Moigneux illustrated this with a hypothetical campaign-cloning comparison: “If today my campaign that I‘m cloning requires the work of someone who costs me €25 to run, and my tokens are at €75, it’s not worth it,” he said. The figures describe a hypothetical rejection test. They make execution cost an explicit condition of the decision to automate.

Managing those costs is sometimes called tokenomics: assessing the economics of model usage, including charges for tokens, the units of text a model processes. Token charges and infrastructure costs need to be compared with the human labor cost of the same work. Moigneux also highlighted the complexity and potential expense of training agents. The assessment therefore needs to cover preparation and operating costs, with continued monitoring to prevent bills from escalating.

Cost monitoring is more useful when the team can connect spending to the work performed. Observability means being able to inspect a system’s activity while it runs, allowing reviewers to examine execution alongside cost and quality. Bécue described Ibexa’s commercial offering in those terms: “The observability suite we integrated into the product is precisely to verify what the agents are doing,” he said. His description explains how that capability is intended to support supervision.

Alongside execution records, Ibexa intends to display a time-saving value assigned during agent configuration. Bécue said: “Each time we define an agent, the person defining the agent will also be able to define the time saved by that agent’s execution. And so, this will be visible on a dashboard.” The person creating the agent supplies that value, so the dashboard shows an assumption about expected benefits. Actual time saved still needs to be measured through execution.

This separates three measurements that serve different decisions. An execution record shows what the agent did, a configured time-saving value expresses its creator’s expectation, and realized net value compares actual benefits with actual costs. A return-on-investment assessment requires the last of these. Keeping them distinct lets reviewers use the dashboard without treating an expected saving as an observed result.

The benefits side of that assessment also depends on the quality of the completed work. LLM judges use a language model to compare inputs and outputs for response quality and consistency, providing signals that can support review. Their assessments are quality signals, so correctness still requires appropriate checks. Even a favorable assessment leaves the team to determine whether the work delivered enough value to justify its expense.

Key takeaways for decision-makers

• Start with a familiar task: Marketing teams can test agents on work such as cloning a campaign, where outputs are easy to inspect. Define the required data, review criteria and expected benefit before expanding the assignment.
• Treat model access as a business dependency: Technology owners need to assess whether model access, hosting and jurisdiction meet continuity and legal requirements. Evaluate sovereign or open-source options against those requirements, then test performance using the task’s actual data and context.
• Expand permissions through observed performance: Workflow owners need documented tasks, clear access limits and validation checks at each stage. Grant agents greater autonomy as execution records demonstrate acceptable performance.
• Measure realized net value: Marketing and technology owners need to compare actual time savings and output quality with model, infrastructure and preparation costs. Keep configured savings estimates separate from measured results when assessing returns.