The Model Is No Longer Just a Cog in the System
Differences among large language models remain real: token cost, context window, sector-specific specialization. But a brilliant model, poorly connected to the information system, rarely yields more value than a modest model properly linked to the business tools. This shift explains the proliferation of marketplaces for agents specialized by function or sector. On its Agent Store, the SiAgents platform offers over 400 industry AI agents covering finance, energy, the public sector, or compliance, designed to fit into an existing workflow rather than serve as an isolated conversational interface.
From Chatbot to Execution of a Business Process
An assistant answers a question. A copilot helps a colleague move faster on a task. An agent, however, executes multiple steps in a single process : it searches for information, analyzes a document, queries one or more internal tools, produces a recommendation, requests a validation, and then triggers an action. This complete chain, not merely the single answer generated, is the value-building block that companies are now aiming to industrialize.
Measuring ROI at the Process Level, Not at the Model
Tracking the number of requests or tokens consumed does not reveal the value created for a business function. The relevant indicators relate to the transformed process:
- turnaround time for a case or a request;
- error rate before and after partial automation;
- number of steps actually handled by the agent;
- volume of files processed per employee.
A study by Anthropic conducted with consulting firm Material among more than 500 technical leaders, published in September 2025, indicates that 57 % des organisations utilisent déjà des agents pour des flux de travail en plusieurs étapes, including 16 % at cross-functional scale. They would be 81 % looking to tackle more complex use cases by 2026, with 39 % explicitly citing multi-step processes.
Proprietary Company Data and Processes, the Only Real Differentiator
The same models remain accessible to direct competitors. Differentiation lies elsewhere: proprietary data, internal rules, history of handled cases, and the sequencing of tasks from one service to another. A well-designed agent inherits this accumulated capital; a generic agent, not. This is the business layer, hard to copy, which makes an agentic system truly tailored to an organization.
Keeping Human Arbitration at Critical Stages
Fully automating a process remains rare, and rarely desirable. In finance, compliance, energy, or the public sector, certain decisions must remain the responsibility of a human collaborator: the validation of a sensitive file, interpretation of a borderline case, regulatory arbitration. The late-2025 partnership between OpenAI and four major consulting firms, BCG, McKinsey, Accenture and Capgemini, to accelerate the deployment of agents at their clients, illustrates this logic: orchestrating several specialized agents around the same case, with a human validation point before the final action.
The choice of model continues to matter, but it weighs less and less in the equation. The value is now built in the architecture surrounding the agent (data, tools, rules, human oversight) and is measured at the level of the transformed process, rather than at the level of the model that powers it.
