This page is machine-translated from French. Read the French original.
Use Cases
The scenarios below describe typical implementations: what is installed, how it is organized, and what needs to be monitored. They can be combined.
Trust and accounting expertise
The problem. Thousands of documents per client — invoices, statements, contracts, correspondence — and recurring questions whose answers are buried in a file.
The setup. - A workspace per customer, This isolates each user's documents: a question asked in one space cannot access another user's documents. - Mode Request Enabled so that the assistant refuses to respond when the file does not contain the information, rather than providing unnecessary details. - Prompt system specifying the applicable reference framework and the expected level of detail. - Reference documents (internal chart of accounts, procedures) pinned in each space.
This is what it looks like. «"What charges were recharged to customer X in Q3?", with extracts of supporting documents — verifiable with one click.
Something to keep an eye on. The Query mode reduces inventions but does not eliminate them: the amounts must be cross-checked with the cited document.
Law firm and legal services
The problem. Searching a document collection (standard contracts, case law, previous briefs) costs billable time, and confidentiality excludes online AI services.
The setup. - Facility entirely local — Template, index, and documents on the firm's infrastructure. — Spaces organized by area of expertise (labor law, commercial contracts, litigation) or by case file, depending on the firm's size. Roles distinguishing between partners, associates, and administrative staff. Content control to block the entry of identifiers or data that have no place in a conversation.
This is what it looks like. Comparison of clauses between several contracts, preparation of summary notes, research in previous reports.
Something to keep an eye on. All production intended for a client goes through human review: it is a preparation tool, not a final draft.
industrial SME
The problem. The technical documentation — plans, instructions, quality procedures, intervention histories — is scattered and known to only a few people.
The setup. - A space technical documentation, a space quality, a space support.- Automated import from internal wikis and existing repositories. - Expertise SQL The agent is connected to the ERP system in read-only mode to cross-reference a document-related question with production data. Planned tasks producing a weekly summary of incidents.
This is what it looks like. A technician asks a question in natural language and obtains the exact procedure, with the document reference.
Something to keep an eye on. Obsolete documents perform just as well as others — document governance remains necessary.
Real estate agency
The problem. Mandates, leases, property condition reports and diagnostics are piling up, and prospecting is consuming a considerable amount of time.
The setup. - A space wallet for goods, a space legal for the templates of the deeds. Web widget on the public website, connected to a space containing only publishable information, with a limit on the number of conversations per day. - Agent of prospecting for the search for qualified contacts.
This is what it looks like. Answering common visitor questions without requiring staff involvement, and accelerating file preparation.
Something to keep an eye on. The widget is public: the space to which it is linked must not contain that disseminable information, and the list of authorized domains must be provided.
Public administration
The problem. Sovereignty and data protection requirements that exclude online services, and a significant regulatory volume.
The setup. - Instance on internal infrastructure or national datacenter, without network output. - Organizations To compartmentalize services, with single sign-on connected to the existing directory. Event log Retained for traceability. - Content control configured according to internal policy.
This is what it looks like. Research in the regulatory corpus, assistance in drafting responses, synthesis of files — without any data leaving the scope.
Something to keep an eye on. The event log records structuring actions; the entity's specific archiving obligations must be checked separately.
A cross-cutting scenario: internal support
Applicable to all the above structures.
- A space fueled by internal procedures, HR FAQs, and memos.
- THE web widget published on the intranet, or access guest for employees without an account.
- A planned task which identifies each week the questions that have not been satisfactorily answered, in order to identify the gaps in documentation.
Choose your starting configuration
| Priority | Recommended configuration |
|---|---|
| Absolute confidentiality | Local model + native embeddings + local vector database |
| Maximum quality of responses | External supplier, selected and documented |
| Balance | Local model for everyday use, router models towards a more powerful model for complex cases |
| Large volume of documentation | Enterprise vector database rather than local one |
| Many simultaneous users | Server deployment, appropriately sized hardware, or external provider |
