Model family · in regular use

Generative AI agency in Nice

The models that write, summarise, classify and extract — the building block of most of our automations.

What we do with it

Generative AI at our clients, concretely

We connect them to your existing tools rather than adding a platform: triage and assisted replies to incoming requests, extraction from received documents, meeting notes. Every use case is costed in hours saved before it starts.

The limits

When we recommend something else

If the task is neither frequent, repetitive nor measurable, we say so and the scoping stops there.

Our work

Projects we delivered with Generative AI

  • Paradigm

    Platform for a Nice business centre: registered address, digitised mail, messaging and assistant.

    Generative AI: Extracting the useful information from each letter, and the assistant's answers.

    Read the Paradigm case study
  • LLM Monitor

    SaaS that measures how visible brands are in AI answers.

    Generative AI: Product core: brand normalisation and classification of cited sources.

    Read the LLM Monitor case study
  • Ordelya

    Restaurant management platform: back office, tablets, kiosks, screens and website.

    Generative AI: Menu translation with a language model.

Method

How we bring Generative AI into a project

We start from one specific flow your staff currently handle by hand, such as messages landing in a shared inbox or attachments whose content has to be retyped. Working with the people who do that job, we gather real examples, anonymised where needed, and record the expected result for each: the right category, the right field, an acceptable wording. That collection becomes the reference for the whole project.

The choice of model follows from the nature of the text and how sensitive it is. We compare several candidates on those examples, note where each one goes wrong and keep the simplest that reaches the required standard. Prompts, output formats and automated checks are written and version-controlled like the rest of the code, in a repository in your name.

In production, every run is logged: what was sent to the model, what it proposed, what a member of staff corrected. Those corrections are added to the example set and show where the prompts need to change. You receive a monthly breakdown of usage costs per use case, along with documentation that lets your own team or another provider take over the whole system.

For your business

What it changes in practice

The first effect shows up in reading and retyping work: sorting a shared inbox, copying details from a purchase order, drafting a first reply. These tasks do not vanish, but they shift from producing to reviewing. The time freed up goes to conversations that call for judgement, and handling becomes more consistent, since every request follows the same path whoever happens to be in the office that day.

The other change concerns control over your information. Rather than leaving each person to paste internal documents into whichever online assistant they prefer, the business sets a defined way of working, with a known place where processing runs and a record of what is handled. Because the model sits behind its own interface, you are not tied to one vendor: if its terms change, another can step in. Staff usually take to it without much resistance, because nothing new appears on their screens.

Our reading of it

What we take away from Generative AI

Generative AI is useful where a task is repetitive, textual and tolerant of human review: sorting incoming requests, summarising minutes, drafting a reply, pulling fields out of a badly structured document. It is the wrong tool as soon as the answer has to be exact and verifiable without anyone reading it over.

The model is chosen after the need, never before: we connect the OpenAI API or the Claude models depending on the task, the language and the cost per call, keeping the option to switch. And for any question that concerns your own documents, a model alone is not the answer: document search retrieves the relevant passages first and only then lets the model write, which makes every statement traceable.

Decision guide

Should you choose Generative AI?

Generative AI or conventional rule-based automation for handling incoming requests?

If your requests arrive in a fixed format with predictable fields, written rules are enough: they cost nothing per use, they are exact and they are easy to audit. Generative AI earns its place when the content is free text, phrased differently by every customer, and its meaning must be understood to triage, summarise or draft a reply. We often combine both: rules for the straightforward cases, and the model only for what rules cannot read.

A model hosted by a major provider, or one running in Europe or on our own servers?

It depends on what the text you send actually contains. For sales emails or documents without sensitive personal data, a hosted model gives the best writing quality with no infrastructure to run. As soon as health data, legal files or information under professional secrecy are involved, we favour an open model hosted in the European Union or on your own servers. Output quality can be slightly lower, but you keep full control of the data.

An off-the-shelf general assistant, or integration into the tools we already use?

A subscription to a chat assistant suits occasional, individual tasks: rewording an email or sketching an outline. It does not connect to your mailbox, your business software or your shared folders, and each person uses it differently. A bespoke integration handles one specific flow, the same way every time, with an output you can check. That is why we connect the model to your existing tools rather than adding yet another platform.

Client questions

What clients ask us about Generative AI

Can the AI get things wrong in what it writes or extracts?

Yes. A generative model can produce information that sounds plausible but is false, or misread a field in a document. That is why we only let it act automatically where a mistake has no consequence, and keep a human check for everything else: the reply is drafted, and a member of staff reviews it before it goes out. Extracted data is also tested with simple checks, such as a date format or a total that must add up.

How do we know whether a use case is worth pursuing?

We first estimate how much time the task takes you today: how often it happens, the volume and the handling time. This estimate in hours saved is done before any development starts. If the task is rare, barely repetitive or impossible to measure, we tell you and stop there. Walking away from an AI project is better than paying for a tool nobody will be using six months later.

Will our staff have to learn new software?

Usually not. We build the model into what your team already uses: their mailbox, their business software, their shared folders. A request arrives already sorted, an incoming document is already summarised, meeting notes appear once the meeting ends. The main change is a new habit: reviewing and approving what the tool proposes rather than producing everything from scratch.

What should we do about the AI tools our staff already use on their own?

Banning them with no alternative rarely works: the use carries on, you just stop seeing it. We start by listing what is being used and for which tasks, then separate what can stay as it is, what should move to a controlled tool and what must no longer leave the business. A short usage policy, written for your line of work, goes with that review.

Does a generative AI project create obligations under the GDPR?

Yes, as soon as personal data passes through the model. The processing must appear in your record of processing activities, the people concerned must be informed, and a data protection impact assessment may be required depending on the data. We identify these points during scoping and hand you the processing documentation: purpose, data sent, where it runs and how long logs are kept.

Going further

Generative AI sits within our Artificial intelligence

A technology does not make a project. The method around it decides the outcome: scoping, prototypes, short cycles, and a clean way out.

Neighbouring technologies

A Generative AI project in mind?

The first conversation is free. We will tell you plainly whether this technology is the right one, and if not, which is.

Updated on