Gauthier Huguenin on Comptoir IA: putting AI agents to work

From sales to deploying AI agents

Before Kirako, I spent around ten years in commercial roles. I started in B2B telecoms, then joined a healthcare startup. At France's leading telephone AI company, I later worked as a Customer Success Manager deploying voice agents.

That role placed me between the customer's teams, the developers and the product. A business wanted to answer calls, handle a request or connect its software. My job involved turning that expectation into a precise process, then checking what happened when the agent was taking calls.

In the podcast, I emphasise project scoping. An overly broad request, unwritten rules or a poorly defined promise can complicate a project before anyone chooses a model. That experience explains how I approach work at Kirako: establish who does what, which information they use and what should happen when a case falls outside the agreed scope.

The first requests I received went beyond voice. There were incoming emails to process, guest messages to handle and actions to coordinate across tools. I was already experimenting on my own servers. Kirako grew out of those experiments and requests. My background on the About page gives more context on that transition.

What I mean by an AI agent in this interview

I compare setting up an agent to onboarding a colleague. They need an assignment, procedures and access. You also need to explain which decisions they can make and which must go back to someone else.

In the deployments we discuss, an agent is a program that uses an AI model and tools to carry out a task. It can run on a server independently of my computer. Instructions describe the expected work; connectors let it read or modify authorised systems.

Email makes the distinction tangible. Drafting a reply and sending it are separate actions. For certain recipients or situations, I want to review the message. That rule needs to be part of the agent's operation from the start.

The colleague analogy helps identify these requirements. It does not give software a person's judgement. Within the Kirako team, agents have named roles. They remain systems I configure and supervise.

Prospecting: prepare the commercial work and follow replies

Nicolas asks which agents clients request most often. At the time of the interview, prospecting accounts for a substantial part of those requests: finding relevant companies, enriching available information, preparing an opening message and keeping the CRM updated.

My sales background helps here. A contact list is of limited use if nobody knows why those companies were selected, who has already replied or which conversation needs attention. The agent has to fit a sales process the team understands.

In the episode, I explain that reporting and follow-up depend on the customer's process. Some want a daily report. Others mainly need the information in their CRM. We agree with them on where the work should appear and how often.

I describe one architecture in the article on B2B prospecting with Lemlist and Sales Navigator. It also sets out the division of responsibilities: the agent prepares and tracks the work, while commercial approval and LinkedIn actions remain human tasks in that setup. The podcast gives the context; the article explains the operating boundaries.

Airbnb: an early arrival requires information from several sources

The short-term rental case I discuss with Nicolas concerns an owner managing several properties. Guest messages, cleaning providers and repairs required daily attention. Information was spread across different channels.

One of the agent's tasks was to compare arrival requests with the cleaning schedule. A small adjustment could be accepted within the owner's agreed allowance. A much earlier arrival required another check. If schedules conflicted, the decision had to go back to the owner.

This shows where integration work sits. A polite reply is insufficient when you do not know whether the property will be ready. When damage or a dispute arises, the agent needs to assemble the relevant information and leave the decision to the owner.

I am describing a deployment experience in this interview, rather than a comparative study. A result at one property does not establish how independently an agent will operate elsewhere. Arrival rules, providers and software differ. This coordination work is part of our service for rental managers and short-term rental businesses.

Building Kirako in public

Early in the conversation, we discuss X. I followed technology discussions there before sharing my own experiments. When I started my business in spring 2026, I began posting about tests, discoveries and work in progress.

Those posts led to professional conversations and enquiries. They also helped Nicolas discover my work. The podcast gives those conversations more room: enough time to explain a case, revisit a limitation and answer an objection.

I understand why Kirako's organisation might make someone want to automate everything. In the interview, however, I describe advising prospects against copying my entire setup in one go. An established team has its own habits, customers and responsibilities. It needs a manageable first project. The starting question can be simple: which repetitive task takes time today, and who will check that it is being handled correctly?

There is still work after launch

Towards the end of our technical discussion, we talk about maintenance. A server can run into problems, a model provider can become unavailable and an agent can produce an incorrect response. Its operation needs to be observable and recoverable.

I also discuss documentation and the customer's involvement during deployment. Having someone available to clarify a rule or approve access helps avoid building on assumptions. The business should understand what was delivered and how to take over its management.

That is the recurring concern throughout this episode: start with a real situation, define the agent's assignment and follow what it does. Our AI agents and automation service follows that approach. Choosing a tool such as Hermes comes afterwards, according to the intended operation.

Watch the episode

The full interview is available in Nicolas Guyon's Comptoir IA video. The chapters listed in its description cover the definition of an agent at 9:56, customer cases at 30:22 and costs at 1:02:30.

Thank you to Nicolas for the invitation and his questions. They bring the discussion back to what a business owner needs to understand: what the agent does, which information it requires and when a person should take over.