A conversational agent is software that talks with your customers in writing, on your website, on WhatsApp or in your inbox. It uses artificial intelligence to understand the request, and a knowledge base to answer it. It differs from a voice agent, which handles phone calls, and from the old menu-driven chatbot, which only walked through a decision tree.
This is the most frequent inbound request at Tandem, and also the one that most often goes wrong. This article sets out the four decisions that determine the outcome, the figures that matter when making the call, and the line to hold with ticketing.
What exactly is a conversational agent?
A conversational agent understands a freely worded request, retrieves the information from your content, then writes an answer in natural language. Three building blocks make it up. A language model for interpretation and writing. A knowledge base for the facts. An orchestration layer that decides when to answer, when to call a tool and when to hand over.
The difference with a rule-based chatbot fits in one sentence. The rule-based chatbot follows a script, and it stalls as soon as the customer steps outside it. The conversational agent pursues a goal, keeps the context of the exchange and knows how to transfer the conversation. That transfer capability separates a useful tool from a frustrating answering machine.
I published this architecture in detail in my breakdown of the building blocks of an AI agent on LinkedIn. For a support-facing conversational agent, the two blocks that decide the outcome are the data and the orchestration. The choice of model comes a distant third.
This is also what Tandem sees on the AI agents that actually run in production. The projects that last are the ones whose data scope was clean before the first deployment.
Why so few customers use the company chatbot
The most useful figure on the subject comes from Gartner. In their most recent service interaction, only 7% of customers used the company's chatbot. Yet 49% said they would have done so had the company offered one. The gap traces back to the memory of bad experiences.
The same survey covered 3,566 customers, in February and March 2026. It shows that only 27% of customers would be willing to try a chatbot again after a negative experience. And customers turned out to be roughly three times more likely to use a consumer generative AI tool than the brand's chatbot to sort out their problem.
Gartner already noted in August 2024 that only 14% of customer service issues are fully resolved in self-service. In 45% of the failed cases, the customer felt the company had not understood what they were trying to do.
Which channel to choose: website, WhatsApp, Instagram or Messenger
The channel is chosen where your customers already write to you, and that decision drives every later one. An agent on the website serves discovery and pre-purchase, with an anonymous visitor and a short conversation. An agent on WhatsApp serves the ongoing relationship, with a known number, a history and customers who reply at any hour.
Tandem documented that second case in the field report of a WhatsApp agent at a small producer, where the agent answers in three languages outside opening hours. The channel there was set by the customers themselves, who were already writing on WhatsApp before an agent existed.
Instagram and Messenger follow a different logic from WhatsApp campaigns. The messages themselves are not billed by Meta, and the cost concentrates on the agent platform and the model calls. For a consumer brand that already receives its questions as direct messages, this is often the cheapest entry point.
What a conversational agent really costs
On WhatsApp, the billing mechanics changed substantially. Meta's developer documentation states that since July 2025, billing applies to the delivered message rather than to the conversation. Non-template messages sent during the 24-hour customer service window remain free, and that window reopens with each customer message.
That 24-hour window is the good news for customer support. An agent replying to customers who contacted it therefore works largely inside the free tier on Meta's side. What gets billed are the messages you initiate, and their rate depends on the category and the recipient's country.
| Line | What triggers the bill | Order of magnitude |
|---|---|---|
| Agent platform | Monthly subscription, often per conversation or per seat | Tens to a few hundred euros per month |
| Model calls | Every question asked, every search in the knowledge base | A few euro cents per conversation |
| Outbound messages | On WhatsApp only, outside the 24-hour window | Varies by category and country |
| Upkeep | Reviewing failures, updating the knowledge base | The most underestimated line |
The knowledge base decides answer quality
The knowledge base is the central element of a conversational agent. Your pages, your frequently asked questions, your product sheets and your internal documents are split into passages there, then indexed for search. When a customer asks a question, the agent retrieves the passages closest to the request and writes from them.
This mechanism has a name, and Tandem devoted a whole article to it: RAG, or how to give a business AI its context. Its practical consequence is direct. An agent wired to up-to-date documentation answers correctly, and the same agent wired to last year's export invents with confidence.
Two settings change everything in production. The first lets the agent say it does not know, rather than forcing it to produce an answer. The second locks the scope of the question in progress, so intent classification cannot push the customer into another journey halfway through their request. I hit that second fault on every tool I tested.
Where the conversational agent stops and ticketing begins
The line is drawn on the nature of the request. A conversational agent handles what settles inside the exchange itself, a question of availability, an opening time, an order status, a pointer to the right page. A ticketing tool takes over as soon as a request has to be tracked, assigned to someone and followed over time.
The two plug into each other, and Tandem compared the solutions for the latter in our comparison of AI ticketing tools. The common mistake is asking the conversational agent to play the ticketing role, which produces lost requests with no trace.
The handover to a human is designed before the first conversation. On the chatbot-building tools I tested, live transfer to a person is often missing by default. It has to be wired to a third-party messaging tool. Check this point before choosing the platform, because it governs the 50% of requests the agent will not resolve.
What French companies that already deployed are doing
Adoption tipped over in 2026. The France Num barometer surveyed more than 9,000 companies, through Crédoc. It establishes that four in ten French small and mid-sized businesses now use at least one AI solution, up 14 points in one year. Among the declared uses, chatbots and assistants come second at 24%, behind content generation. And only 19% pay for their solution.
Three named deployments give the measure of what is achievable. La Redoute put an agent in the messaging of its mobile app in autumn 2024. The case study published by Microsoft puts at 60% the answers handled there by the agent. Boulanger built its first agent in house in four months, led by its Chief Data and AI Officer Marie Cappelaere, according to the specialist retail press.
The third is public. In June 2026 the French State rolled out its interministerial assistant built with Mistral AI to all civil servants. The trial had covered 10,000 agents. the interministerial digital directorate measures a time saving of up to 16% on document synthesis tasks. That is an honest order of magnitude, far from promises of halving headcount.
Should you deploy a conversational agent in your company?
Yes, on two conditions, and they fit in few words. You already receive a volume of repetitive written requests, and you accept treating escalation to a human as a core feature rather than an admission of failure. On that basis, a conversational agent absorbs the simple requests and gives time back to the team.
Hold off if your requests are mostly complex, if your documentation is out of date, or if nobody will have time to review the failures. An agent wired to a wrong knowledge base degrades the customer relationship faster than it relieves it. Gartner's figures are a reminder that a bad first experience is very hard to win back.
The right starting point remains a narrow scope on the channel where your customers already write. Tandem frames this kind of project during an AI audit and then deploys it in autopilot. And if your requests arrive by phone rather than in writing, our voice agent guide is the one to read first.
