LLM integration: a model in your tools, your data in France

Integrating an LLM means connecting a language model, whether ChatGPT, Claude, Mistral or an open model, to a tool your teams already use, for one specific task: reading, extracting, classifying or writing a first draft. At Bonne Idée, the model works on your data, with a measured cost and human review wherever a mistake is costly.

Your teams are probably already saving time with an assistant open in a tab. That gain is lost to copy and paste, it exposes your data, and a model that knows neither your business nor your customers gets things wrong too often to be relied on.

We start from a task: answering incoming requests from the CRM, reading invoices, purchase orders or contracts and filing the information in your tools, sorting and summarizing emails or tickets. We test several models on your real examples, keep the one that gives the best result at an acceptable cost, then connect it where the work gets done.

You get an AI feature inside the tool your teams already open, a running cost estimated on your actual volume, and an indicator that tells you whether the answers are good. We compare the models for you, and you know whether the use case holds up before you have to hire a specialist.

This is the first question in scoping, and it is settled document by document. A model hosted by a provider receives the text you send it, so everything depends on what that text contains.

Three categories, three ways of handling them

• Content with nothing sensitive in it: standard letters, product sheets, internal procedures. A hosted model, selected for its contract terms, is suitable.
• Content with personal or confidential data: customer names, amounts, contracts. We mask the identifying details before sending, or we use hosting in France.
• Content that must never leave: information under professional secrecy, health data, contract clauses that forbid it. The model is then installed on your own infrastructure, which costs more to run and is decided before building.

The result fits on one page: for each type of document, where it goes, to whom, and under what terms. That page is what lets your teams answer the customer who asks.

A hosted model is billed by the volume of text processed, both input and output. The monthly cost therefore depends on three variables, which we measure on your real examples rather than on an assumption.

1. Volume: how many documents, messages or tickets per month, and their average length.
2. The processing: what the model reads and what it writes each time. A short summary costs less than a fully written reply, an extraction less than an analysis.
3. The model selected: price gaps between models are wide, and the most expensive is rarely the only good choice for your task.

During the benchmark, we run the fifty or so real examples you give us through several models, measure what each one consumes, then extrapolate that to your monthly volume. You know the running cost before go-live, and a budget alert is set so that a rise in volume never comes as a surprise.

Here is how a typical integration unfolds, in a situation common in SMEs: customer requests arrive by email, and someone reads them, sorts them and re-keys them into the CRM.

• Scoping defines the task: read each request, extract the customer, the subject and the urgency, create the record in the CRM and suggest a first draft of the reply. The success indicator is the share of records that are correct without any correction.
• The benchmark compares several models on around fifty real requests, anonymized where necessary.
• Integration connects the selected model between the inbox and the CRM. The reply never goes out on its own: a member of your team reviews it and sends it.
• Measurement tracks the share of correct records, processing time and actual cost every month.

The same pattern applies to reading invoices, purchase orders or requests for quotation, as we describe for industrial SMEs and accounting firms.

Without measurement, there is no way to know whether an integration works. During scoping we set a simple indicator, the share of outputs usable without correction, and we measure it on a set of examples whose correct answer is known. The same set is used to retest each new version of the model before it goes live, and to decide whether to widen the scope or fix things first.

These indicators are presented to you every month, and your teams can still read them if you take the tool over without us. When the need goes beyond a single task and becomes an assistant you can query on your documents, that is an AI agent, built on the same principle as CoBIos, our agent for accounting and law firms.

The first question in any integration, what can be sent to a model, is settled document by document: we cover it above. The rest is written into the contract: a data processing agreement (DPA), hosting in France by default (or on your own infrastructure when the sensitivity of the data requires it), no model training on your data, and a record of processing activities.

GDPR and EU AI Act obligations, and what to do when the model gets it wrong, are set out on our AI guarantees and data sovereignty page.