Is Agentic AI simply an extension of the generative AI projects undertaken until now? The answer is no. The focus is no longer on local optimization through a single use case, but on the complete reinvention of core business processes to extract tangible and measurable value.
An agent is not a rebranded assistant. An assistant answers or generates content. An agent, however, aims for a specific business objective, makes decisions, interacts with systems, and executes actions. It maintains a state, manages exceptions, and escalates when it reaches its limits. This difference is not cosmetic: an agent is not designed to “answer,” but to act within a process.
From GenAI to Agentic AI: A Paradigm Shift
The approach of Agentic AI marks a clear break with the previous GenAI wave, which was often tackled in a fragmented manner, on a use case-by-use case basis.
The debate is moving from “conceptual spheres to the operational reality of large corporations,” as recently evidenced by companies such as Orange, SNCF Voyageurs, and the Beaumanoir group. Today, the subject is handled “from the top,” directly by Executive Committees (Comex), sponsored by the general management or the Chief Transformation Officer. The debate is moving from “conceptual spheres to the operational reality of large corporations,” as recently evidenced by companies such as Orange, SNCF Voyageurs, and the Beaumanoir group. Today, the subject is handled “from the top,” directly by Executive Committees (Comex), sponsored by the general management or the Chief Transformation Officer.
This change in scale is significant. It represents a true paradigm shift, according to Hanan Ouazan, Partner & Generative AI Lead at Artefact. The objective is no longer to incrementally improve an algorithm, but to rethink and automate processes end-to-end to generate a direct, measurable return on investment (ROI).
Hanan Ouazan also points to the limits of the previous approach: “Even if you have strongly optimized your algorithm locally, it remains local. You are dependent on many other aspects of the process that are also very important to achieve the final ROI”. By focusing on the entire process, Agentic AI promises to overcome the “disappointing effects” of isolated optimizations.
Agentic AI also breaks with classical automation. Automation executes a defined sequence of tasks. Agentization introduces arbitration, controls, the capacity to interpret a context, and adapt the path to achieve a goal. We are no longer talking about a fixed workflow, but a system that navigates a business process.
Setting the Course: The “Why, Where, How” Method for Enterprise Alignment
To manage such an ambition, a clear strategic framework is essential. To structure this transformation, the expert proposes a simple yet powerful methodological framework: the “Why, Where, How” triptych. The approach involves sequentially answering three fundamental questions to align the entire organization.
- The “Why”: Defining the High-Level Strategic Objective. This step establishes the scale of the ambition and justifies the necessary investments. This “Why” can take various forms, such as SNCF Voyageurs’ client-centric “compass” or Beaumanoir’s prioritization of immediate “professional efficiency” for its teams.
- The “Where”: Defining the Playing Field. Once the objective is set, the company selects the priority functions and processes where value can be captured most effectively.
- The “How”: Defining the Concrete Means. This involves establishing the technological platform, governance, and organizational structure that will support the transformation. This governance is critical: Agentic AI requires clarifying who defines the business rules, who supervises the agents, and who authorizes sensitive actions.
These organizational decisions, more so than the technology, enable scaling. The framework ensures that the initiative is firmly anchored in the company’s strategy.
Choosing Battles: Identifying Prerequisites and Eligible Processes
The success of an Agentic AI project depends on a rigorous assessment of three essential prerequisites:
- Process Knowledge: It is imperative to have a clear mapping of current operations, as there is often a significant gap “between the process as written on paper and the process as it is operated in the company”. Functions like finance or legal, where “sovereign” stakes necessitate strict documentation, are often more favorable starting points.
- Data and Information System Maturity: Automation often runs into legacy systems that were not designed to be interoperable. An agent must not only be able to read information but also write it to execute actions, which demands open systems. This capability requires a precise architecture: an orchestrator, context memory, connectors to business systems, and a supervision layer. The goal is to add a transversal “nervous system” that links processes and Information Systems (IS), rather than reinventing what already exists.
- Organizational Maturity: A structure capable of overseeing new processes, integrating human intervention (“human in the loop“), and managing change is indispensable.
Considering these three dimensions is decisive for a smooth deployment phase.
A Pragmatic Deployment Roadmap
Once eligible processes are identified, the deployment methodology should be resolutely pragmatic.
The first step is to prioritize processes by trading off potential value (expected gain) against transformation complexity (technical, human, organizational). A crucial step is then to rewrite the target process “in an ideal world,” without constraints, and subsequently measure the gap with reality.
However, Hanan Ouazan warns against a major pitfall: “trying to immediately recreate the process of tomorrow”. This approach is doomed to fail. The best practice is to build a roadmap in “chunks,” starting with the most accessible pockets of value.
Oney provides a concrete example with its Agentic Proof of Concept: the BPCE subsidiary focused on the specific objective of “streamlining the personal loan subscription journey”. Instead of reinventing everything, they started by automating the analysis of the 42 required documents. This single step saves three days on the overall process and generates immediate gain.
This incremental approach secures the project but does not eliminate all obstacles. Anti-patterns observed in the field include: building an agent before clarifying the process, giving too much latitude to a critical agent without safeguards, or assuming an agent will “replace” a role without re-organization. Conversely, imagining an agent is a finished product when it requires continuous supervision and adjustment is another pitfall.
Challenges Beyond Technology: Governance, Legacy, and Tool Maturity
Agentic AI projects face challenges that extend far beyond the purely technological framework.
One is the extremely rapid evolution of tools, which Hanan Ouazan likens to “trying to take a picture of a TGV passing at 300 km/h”. Betting on technologies still under construction is risky, hence the importance of “choosing your battles” wisely.
A second major impediment is organizational. Process transformation is inherently transversal and clashes head-on with siloed organizations where “everyone steps on each other’s toes”. As illustrated by SNCF Voyageurs’ challenge of “reaching the entire organization,” these initiatives are doomed without a clear mandate carried at the highest level.
Finally, the maturity of current tools is a concern. The Artefact expert identifies a “huge gap” between solutions like enterprise chatbots and complex development platforms intended for expert profiles. Hanan Ouazan believes there is a need for “low-code” tools to enable business users to create agents, as well as model-agnostic, “end-to-end” supervision platforms.
Standards for Agent Interoperability and Supervision
Regarding protocols like Agent to Agent (A2A) or MCP, they are still in their early stages, and the market clearly lacks maturity. The expert notes that a CAC40 player using various models (Mistral, OpenAI, Google, Azure) cannot currently supervise all their systems. According to him, vendors are highly anticipated to resolve this problem of interoperability and large-scale management. They must be able to offer solutions that “scale the construction of agents in a reliable and robust manner” and allow for the “reliable governance” of these agents.
In the absence of unified protocols and supervision platforms, companies are currently forced either to make a “big bet” (betting everything on a single model) or to manage “interfaces,” which severely limits their ability to scale. The main technological challenge lies in the lack of standards and mature management solutions for the interoperability and supervision of agents from multiple vendors.
The landscape is set to evolve rapidly in the next 12 to 24 months: the emergence of low-code platforms dedicated to business users, the standardization of transactional connectors, the reinforcement of observability capabilities, and the appearance of specialized agents integrated directly into daily tools.
The question will no longer be “is it feasible?” but “what sequence of processes should be transformed and how can we supervise it effectively?”.
Watch the replay of the presentation:
