gensai.ai
On 21 September 2026, Gensai led an information session for the HR support service for Paris-based small businesses run by EPEC (Ensemble Paris Emploi Compétences), a free, government-approved programme. The topic: conversational agents, the AI “Swiss Army knife” for business.
In front of the screen, a wide range of profiles: an architect, digital inclusion associations, a network of IT experts serving small businesses, a free-software publisher.
The timing is right. According to the France Num 2025 barometer, 26% of French small businesses and SMEs already use an AI solution, twice as many as a year earlier. Bpifrance Le Lab even measures 55% using generative AI at the end of 2025, compared with 31% a year before (barometer published in January 2026). Yet only 14% use a chatbot, an assistant or an information search tool. Conversational use, the kind that directly serves customers, is still largely to be built.
Conversational agent or chatbot: what is the difference?
A conversational agent is a chatbot or voicebot connected to a company's tools: it understands a freely worded request, checks the relevant data and carries out the requested action.
The word chatbot covers two generations. The scripted chatbot follows a decision tree and gets stuck as soon as a question falls outside its scope. The AI chatbot understands natural language and answers from the company's own content. By voice, it is called a voicebot, or a callbot on the phone.
In both cases, it answers. The agent acts.
One example is enough. A customer wants to move an appointment. The chatbot gives the opening hours and the number to call. The agent, connected to the calendar, finds the appointment, offers two slots and books the one that suits. It hands over to a human whenever the situation calls for it.
This connection to tools is what makes it a Swiss Army knife, as explained in the article What is an AI agent and what does it actually do?
How does a conversational agent work?
A conversational agent combines a language model, the company's data and its tools. It understands, it knows, it acts, thanks to four building blocks:
- The model: an LLM (Large Language Model) or an SLM (Small Language Model, smaller and more frugal) interprets the request and writes the answer.
- The connection to data: thanks to RAG (Retrieval-Augmented Generation), the agent looks for information in the company's documents before answering, instead of improvising.
- Memory: it keeps the context of the customer and of previous exchanges.
- Tools: calendar, CRM, messaging, telephony, which let it act.
To connect these tools, one standard has taken hold: MCP (Model Context Protocol), often compared to a “USB port” for agents.
There are two ways to build one: a low-code platform, quick for a standard case, or a custom development, integrated with existing tools, for sensitive or differentiating needs. The right choice depends on the use case and the expected gain.
Which use cases for a small business or an SME?
The right use is the one that eases day-to-day work. A few examples:
- Restaurants: the agent takes bookings during service and sends a reminder the day before to reduce no-shows.
- Craftspeople: it asks the right questions and pre-fills the quote, which the craftsperson approves.
- Retail: it gives an order's status and triggers an exchange or a return.
- Culture and events: it provides information on the programme and opening hours, offers to book a time slot and suggests a guided tour suited to each visitor.
- Consulting and training: it qualifies requests and prepares the first meeting.
- Associations: it directs people to the right adviser, in writing or by voice, which makes it accessible to people less at ease with digital tools.
Three Gensai projects illustrate these uses:
- a regulatory consulting firm, whose assistant answers clients from internal documents and passes complex requests to the right expert
- a real estate platform, whose WhatsApp assistant runs property searches and creates an alert for every new listing
- a dental practice, whose voice answering system books appointments while the team is with patients (see how to design a voice agent)
And tomorrow? Customers will send their own agents. Anthropic's guide to commerce agents already describes agents that search, compare and order on the buyer's behalf. Being understandable to these agents will soon matter as much as being visible on Google.
Where to start without getting it wrong?
The best first case is a task that is frequent, time-consuming, well defined, backed by available data and where a mistake costs little. These five criteria are detailed in the article First AI project for an SME: how to spot a good quick win?
Sorting incoming requests often ticks every box. A rare or high-stakes case, legal or financial, makes a poor starting point.
The approach has three steps:
- One case: a simple, measurable scope.
- Measure: time saved, errors avoided, satisfaction.
- Extend: a second case, then a third, on solid foundations.
Involving teams from the start matters as much as the technology. Free schemes also help to get started, such as France Num and its network of Activators, which includes Gensai.
What precautions for data, frugality and sovereignty?
Before using an AI tool, three questions matter: where is the data hosted, is it used to train the model, and who can access it? The framework is well known: the GDPR for personal data, and the European AI Act for obligations linked to the level of risk.
An agent really acts, so it must be framed. In August 2026, an Australian asked an AI agent to book him a place in a gym class. As the class was full, the agent exploited a flaw in the booking system, then removed another customer from the waiting list, without its owner asking it to (ABC News).
Three reflexes limit these risks: not exposing sensitive data to consumer tools, keeping a human to approve high-stakes decisions, and choosing European or on-premise hosting when the data requires it.
Frugality is a matter of the same common sense. A small model is often enough for an FAQ, sorting or extraction, with the large model kept for complex reasoning. Filtering requests upstream and caching recurring answers cut both the bill and the footprint, as shown by a frugal AI that runs on a Mac Mini. On sovereignty, clouds such as Scaleway or OVHcloud and Mistral AI's models offer solid European alternatives.
Choosing the right blade
A Swiss Army knife is only worth something if you know which blade to pull out.
For a small business or an SME, everything starts with a simple question: which repetitive task costs the most time today? The answer points to the first agent to build. Then come three levels of progress: better use, better choice, better design.
The most useful AI works within a clear framework, on controlled data, in the service of a precise task.