+33 (0)1 87 66 00 65 · Monday to Friday, 9am–6pm Free audit (15 min)
Under the bonnet

How a Blue Lemon Agent agent works

The market sells "an LLM". An agent that actually works inside a company has four layers, and the model is the one that decides least. Here is what each does, and which one you will have to build yourself if you go without us.

A language model watches nothing, triggers nothing and enforces no ceiling. It reads, it classifies, it writes. Everything that commits your company — what goes out on its own, what waits for approval, what alerts whom — is decided by the three layers around it. That is where our work is, and that is what you are buying.

The four layers

In the reverse of the market's order: starting with what everyone sells, ending with what makes the difference.

  1. 1

    The language model

    It understands a request written in plain language, classifies it, pulls what is needed from a document and drafts a reply. That is all, and it is already a great deal. Mistral, Lucie, EuroLLM, Teuken: at this level of task the sovereign models do the job, and a model running on your own machine is enough.

    Who provides it An open model, run locally or on your isolated resource. Interchangeable.

  2. 2

    The rules and the guardrails

    What goes out on its own and what waits for a signature. Ceilings, pre-approved categories, the cases that stay human, the log of who decided what. A ceiling enforced by a model is not a ceiling: it would hold ninety-five times out of a hundred and give way on a money decision, without being reproducible. This layer is deterministic code, not an instruction written to the model.

    Who provides it Us. Configured with you, revocable by you, logged.

  3. 3

    The connectors and the scheduler

    What gives the agent something to look at, and a reason to look now. Email, ERP, CRM, ticketing, document base, staff directory — plus the clock that fires the morning report and the memory of what has already been seen. Without this layer there is no monitoring: there is a conversation.

    Who provides it Us. This is where most of the deployment effort goes, and what varies from client to client.

  4. 4

    The supervision

    What makes the whole thing checkable: the trace of every action with its source, the screen showing what the agent did, the button that stops it. It is the layer you notice on the day something went wrong — and on that day it decides whether you understand in ten minutes or in three days.

    Who provides it Us. Included in the subscription, not an option.

So what does the model actually do?

The four behaviours our demonstrators show, and what really produces them.

How the work divides between the language model and the layers around it
The behaviour What does it The model's part
Continuous monitoring, daily report Scheduler, connectors, memory of what has been seen Draft the summary
Activatable rules, ceilings Rules engine, audit log, configuration screen Classify the request into a category
Routing, reminders, summary up the line Staff directory, timers, approval workflow Draft the summary
Spotting a pattern (bank details, domain, due date) A query, a string comparison, a subtraction of dates Explain the result in plain language

None of these four lines is an emergent capability to be hoped for from the right model. It is ordinary, well-understood software, built during deployment — and that is why we can commit to it.

What this architecture changes for you

The choice of model does not lock you inThese behaviours depend on no particular model. Changing model does not break them and forces no rebuild.
A small sovereign model is enoughThe only task given to the model is to understand and classify. A model running on your premises does that — so sovereignty costs nothing in capability.
A ceiling is a real ceilingAmounts, thresholds and always-human cases are enforced by code, not suggested to a model. They do not give way one time in twenty.
The effort is visible in the quoteConnectors and rules are what vary from client to client, and what the audit prices. Nothing hides behind the word "AI".

Your questions, our answers

Do your agents use ChatGPT?
It depends on the package. The Essential pack relays to a market model, and the pricing page says so explicitly. The sovereign packages run an open model — Mistral, Lucie, EuroLLM, Teuken — locally or on an isolated resource hosted in France. In both cases the rules, the connectors and the supervision are ours.
What happens if the model gets it wrong?
It gets wrong what it does: understanding, classifying, drafting. A misclassification sends a request to the wrong queue, and that shows in the log. It cannot grant a discount above a ceiling or send something that required approval, because those decisions are not entrusted to it.
Can we take our rules with us if we change provider?
Yes. The rules, thresholds and approval workflows are your settings, exportable like the rest of your data. Clause 21 of our terms provides for their return within 30 days, in a usable format.
How long does all this take to put in place?
Several weeks to several months, depending on how many systems have to be connected. The model installs in a day; it is the connectors and the rules that take the time, and that is exactly what the free audit lets us price before anyone commits.

Want to know which layers you already have, and which are left to build? Book the free 15-minute audit