Automation & AI — Private copilot & RAG

Private AI copilot and RAG on your company data

A conversational assistant that answers from your documents and your processes, not from generic knowledge. With RAG (Retrieval-Augmented Generation), answers draw on identified passages from your document base and cite them, while respecting the access rights you have defined. We scope, build and operate your copilot for companies in Paris and Île-de-France, with governance and hosting matched to the sensitivity of your data.

Scope my AI copilot project01 83 81 17 48
What holds you back

The patterns we keep seeing.

01

Scattered knowledge, answers nobody can find

Procedures, contracts, proposals, project histories: the information exists, but it is scattered across SharePoint, mailboxes and file servers. Teams keep asking the same experts the same questions, and every search interrupts someone. A copilot connected to those sources returns the information the moment it is asked for, with the original document to back it up.

02

Consumer AI tools used without a framework

With no internal alternative, staff paste contract extracts or client data into public chatbots, with no access control and no traceability. A private copilot provides a framework: a defined document perimeter, authentication, logging of exchanges and a usage charter.

03

AI answers that cannot be verified

A generic assistant answers with confidence, including when it makes things up. Without a cited source, there is no way to tell a reliable answer from an approximation — unacceptable for a quote, a contract clause or a safety instruction. RAG makes it possible to cite the passages used: the answer can be checked before it is acted on.

What we cover

The scope of the service.

01 / ASSISTANT & RAG

A copilot connected to your sources

The design and build of the assistant: connection to your document sources, indexing, retrieval of the relevant passages and generation of cited answers.

  • Connection to your sources: SharePoint, OneDrive, Teams, file servers, business applications
  • Document indexing with scheduled updates
  • Answers written with citations of the source passages
  • Access from the browser or embedded in Teams
  • Actions triggered in your tools through n8n workflows
02 / ACCESS & PERIMETER

Access control and data perimeter

The copilot is designed to answer within each user's rights: authentication, filtering of the sources consulted, and sensitive corpora isolated or excluded.

  • User authentication through Entra ID
  • Filtering of the documents consulted according to the user's rights
  • An explicit document perimeter: corpora included, excluded or sensitive
  • Logging of questions and answers
  • Periodic review of rights and perimeter
03 / HOSTING & GOVERNANCE

Suitable hosting and governance

Where the language model and the index run is chosen according to the sensitivity of your data, and the flows are documented: who sends what, where, and why.

  • Hosting according to sensitivity: cloud or dedicated infrastructure (Hyper-V, VMware, Proxmox)
  • Mapping of data flows to the language model
  • AI usage charter and user awareness
  • A system designed to fit within the applicable obligations — legal qualification to be validated with your counsel
04 / EVALUATION & IMPROVEMENT

Evaluation and continuous improvement

Answer quality is measured: a reference question set serves as the evaluation baseline before go-live, then whenever the system evolves.

  • Reference question set built with your teams
  • Answer quality evaluated before opening to users
  • Feedback collection and iterative corrections
  • Usage tracking and gradual extension of use cases
How we proceed

From scoping to follow-up.

The exact scope, deliverables and timelines are formalised in the proposal, before any commitment.

1
Use-case scoping

Workshops with your teams: the questions actually asked, the corpora involved, data sensitivity and success criteria. The pilot's perimeter follows from this.

2
Architecture and hosting

Choice of language model, hosting mode and connectors according to data sensitivity and budget. Data flows are documented before anything is built.

3
Pilot on a limited perimeter

A limited corpus, a test group of users and an evaluation against the reference question set. The pilot validates answer quality before widening the rollout.

4
Controlled deployment

Gradual opening to your teams: access rights verified, usage charter circulated, and users supported as they get started.

5
Operation and evolution

Monitoring of the system, index updates, tracking of answer quality and new use cases added at the pace of your priorities.

Frequently asked questions — automation & ai

The answers describe how the service works. Quantified commitments are formalised in your contract.

It is a conversational assistant connected to your own documents and processes, deployed within a framework you control. RAG (Retrieval-Augmented Generation) first retrieves the relevant passages from your document base, then generates the answer from those passages, cited in support. Unlike a consumer chatbot, it answers within a defined document perimeter, with your access rules and logging of exchanges.
In the architecture we propose, your documents are used to retrieve passages at the time a question is asked, not to train a model. The exact conditions depend on the components and hosting selected: they are examined during scoping, documented, and formalised in your service contract.
No system can guarantee absolute confidentiality, and we do not promise it. We reduce the risk through concrete mechanisms: authentication through Entra ID, filtering based on existing rights, sensitive corpora isolated or excluded, logging of exchanges and hosting chosen according to data sensitivity. Residual risks are documented during scoping, so you can decide with full knowledge of them.
A language model can be wrong, even with RAG: that is why answers cite their sources and are checked before use. Quality is measured against a reference question set before go-live, and reported errors feed the corrections. For high-stakes uses — legal, financial, safety — human validation remains built into the process.
No. Scoping identifies the corpora that are already usable; a well-kept subset is often enough for the pilot. Poorly structured documentation degrades answer quality: where that is the case, a clean-up workstream is proposed and prioritised, in parallel with starting on the rest.
The number of use cases, the volume and diversity of the sources to connect, hosting requirements, the level of access control and the integrations with your tools. Add to that ongoing operation: index updates, monitoring and quality tracking. Pricing is established after scoping and formalised in the specific terms of your contract.
That depends on the number of sources to connect, the state of your documentation and the hosting requirements. The pilot approach puts a first perimeter in the hands of a test group before any wider deployment. The timeline and milestones are defined during scoping and formalised in your service contract.
We design the system to fit within the applicable obligations: informing the individuals concerned, retention periods, documentation of processing activities and data flows. The analysis of obligations under the European AI regulation is carried out during scoping, based on your use cases. Legal qualification remains to be validated with your counsel; we provide the technical documentation needed for that analysis.
Your documents remain yours and are not locked into the system: the index can be rebuilt from your sources. The configuration, documentation and logs are handed back to you under the reversibility terms set out in your service contract.
Related services

Frequently engaged together.

Business process automation with n8n

Automate repetitive tasks and connect Microsoft 365, your CRM, 3CX phone system, support and business tools with documented, maintainable n8n workflows.

Secure custom intranet and extranet development

Centralise documents, workflows, access rights and client or partner exchanges in a secure intranet or extranet, designed, built and maintained by Dalena.

Microsoft 365 and cloud architecture for companies

Migration, hardening and administration of Microsoft 365, Entra ID, Intune and cloud or hybrid architectures for companies in the Paris region.

An audit, then a clear plan.

Describe your need in a minute. We come back with an assessment and prioritised next steps.

Request a free audit01 83 81 17 48
Appeler Dalena au 01 83 81 17 48Demander un audit
Private AI copilot & RAG on your company data | Dalena