Between “AI agent” offerings and RPA software now presented as “AI-augmented,” SME managers struggle to know what to buy. According to Enguerrand Boulnois, president of Katalyx, an AI studio based in Aix-en-Provence, the choice depends less on the technology than on the nature of the task to be automated. He sums up his method in one question: does the task require understanding, or just executing?
RPA Executes a Rule
RPA, for robotic process automation, refers to those software robots that replicate clicks and data entries on a computer. A robot follows a fixed path: open a specific screen, copy a specific field, paste it elsewhere, click “validate.” It is fast, reliable, and inexpensive, on one condition: that the task is always the same and the data always arrives in the same format.
Recopying orders from a customer portal into an ERP, i.e., an enterprise management software, extracting the same indicators each month, or sending a reminder on a fixed date fall, according to Mr. Boulnois, under RPA or a simple integration between two software packages. Adding generative AI to this would cost more and introduce uncertainty where it is not needed. The limit is known: at the first change in screen or format, the robot breaks, and as soon as the task involves reading free-form text or making a small decision, it cannot do it.
The AI Agent Understands a Context
An AI agent starts from an objective, not a predefined path. It reads an email, a photo, a PDF, or a technical datasheet, grasps its meaning, chooses the useful action, and carries it out in the company’s tools. Enguerrand Boulnois cites several cases handled by Katalyx. For a social and solidarity economy company that resells second-hand toys, an application recognizes a toy from a simple photo, finds its brand and reference, then indicates where to store it, whereas a team used to enter each item manually.
For the support department of a management software company, the question was what proportion of customer requests, written in free form, could be categorized and prepared automatically. No fixed rule covers these situations. However, the agent has its limits: it can make mistakes, its cost depends on the volume processed, and its operation must be monitored. This is why, according to him, a serious agent keeps a human in the loop for any decision that commits the company: it proposes, a person validates.
Three Questions Before Buying
The consultant first suggests checking if the inputs are always in the same format. If yes, an integration or RPA is sufficient. If they arrive in the form of emails, documents, or photos, an AI agent becomes relevant. He then invites measuring the cost of an error: the more costly it is, the more human validation must be present. An agent that prepares and a human that validates often remain, in his view, more cost-effective than a robot that breaks without warning.
The last question concerns frequency. A task done ten times a month justifies neither a robot nor an agent, and it is better to train the team to use an AI assistant. A task repeated hundreds of times a week, however, deserves a dedicated tool. In most SMEs supported by Katalyx, the chosen solution is neither “all RPA” nor “all agent,” but a workflow where each tool does what it knows best: an integration to transport data, an AI agent to read and prepare, a human to decide. Before any investment, Enguerrand Boulnois recommends taking an inventory of repetitive tasks, with their volume and cost, because this list will indicate, in his view, where AI truly brings a gain and where a simpler tool is sufficient.
