Home / Use cases / Methods

Methods

An AI agent that answers from your content, without making things up

A useful agent does not replace your craft, it absorbs the question you get asked over and over. In our builds, the model only writes prose. A price, an availability or an address is computed by the server, then reviewed before it reaches the screen.

  • Closed corpus with cited source
  • Server side calculation
  • Handover to a human

Your AI diagnostic Book a meeting

Key challenges

A conversational agent prototype looks impressive in a demo. What decides whether it belongs in production is what happens the day it gets something wrong in front of a customer. Four concerns shape our method, from the first draft to deployment.

  • Made up answers

    A language model can produce a plausible and wrong answer with the same confidence as a correct one. The point is not to reduce that risk, it is to stop it from reaching the screen on anything that commits the company.

  • Data that commits the company

    A price, an availability, an address or a contractual condition should never come straight out of the model. It is computed server side from your real data, and the model only turns it into a sentence.

  • The safety of exchanges

    Conversations live outside the web root, with a short retention period and usage quotas. What is not collected cannot leak.

  • The cost of usage

    An agent running without limits consumes without anyone noticing. The cost per exchange gets measured from deployment, not when an unexpected bill shows up.

What we do

  • 1

    Framing the scope

    Writing down what the agent is allowed to answer and what it must never produce, before a single line of code. The scope shrinks first, and only extends afterwards.
  • 2

    Closing the corpus

    Connecting the agent to your real content, exhibitor list, catalogue, documentation or terms of sale, and preventing it from answering outside that material.
  • 3

    Taking the calculation out of the model

    Moving everything that commits the company, price, stock, schedule, to the server. The model writes the sentence, it never decides the figure.
  • 4

    Writing the handover point

    Defining precisely when the agent must pass the conversation to a human, and enforcing it through code rather than through an instruction to the model.

What stays in your hands

  • An agent in production

    A conversational assistant deployed on your website or channel, connected to your corpus and real data.

  • Server side guard rails

    A closed list of possible answers, engaging data computed outside the model, and a review filter before display.

  • Source citation

    Every answer points to where it comes from in your corpus, so your team can check it and correct it.

  • Usage monitoring

    A dashboard of real exchanges, to enrich the content over time and keep an eye on the cost.

Expected results

  • A traceable answer

    Every answer relies on an identifiable piece of your corpus, never on an improvisation from the model.

  • A controlled commitment

    What the company must honour, a price or an availability, no longer depends on a generated sentence but on a verifiable calculation.

  • A team relieved of repetitive requests

    Questions that keep coming back get an immediate answer, and your team can focus on the requests that truly need a human.

How we move forward

  1. 1

    The scope, written down

    We list with you the frequent questions, the answers the agent is allowed to give, and the topics where it must stay silent or hand over.
  2. 2

    The corpus, closed

    We prepare and organise the content that will serve as the source, with an import possible from your existing documents.
  3. 3

    The guard rails

    We code the checks that stop the model from producing a piece of data that commits the company, and the filter that reviews the answer before display.
  4. 4

    Deployment and measurement

    The agent goes live on a limited scope, gets tested on real questions, then is monitored for usage and cost.

Who it is for

This method suits organisations that keep receiving the same questions, at a trade show, on a catalogue, a service offer or a documentation set, and want to answer them without exposing a model that can invent a price or a promise. It fits a first agent as well as the redesign of an existing assistant that has become hard to monitor.

In the same family

Frequently asked questions

Can an AI agent get things wrongYes, a language model can produce a wrong answer with confidence. That is why everything that commits the company, price, availability, condition, is computed server side and checked before display, rather than freely written by the model.
Which language model do you useThe choice depends on the context, the volume of exchanges and the client's hosting constraints. We assess with you a large general purpose model, a lighter specialised one, or a combination of both depending on the task.
Can the data stay in FranceYes. For the uses that require it, the language model can be hosted in France, and conversations are isolated outside the web root with a short retention period.
Does the agent comply with the European AI ActTransparency is part of the framing from the start, telling the user they are talking to an agent, documenting the allowed scope, and keeping a record of what the system is allowed to do.
What does running an AI agent cost day to dayThe cost depends on the volume of exchanges and the model chosen, with no generic figure valid for everyone. We set usage quotas from deployment and track real consumption rather than promising a number upfront.

Let's talk about your context

We always start by understanding the real work. One conversation is enough to know whether we can help, and how.

Tell us what you need Book a meeting