Artificial Intelligence / Use Cases / Personal Assistant
A privately hosted AI assistant for your business
It only reads your documents, within the scope of each user's permissions. And the device can remain on your premises.Your employees are already asking questions of a public assistant, sometimes using company data. The privately hosted assistant provides better answers because it is familiar with your documents, and it does so securely, since it only reads what the user is authorized to view. It operates on Teams, cites its sources, and passes the question on when it doesn’t know the answer. Where it operates—at Microsoft, on your premises, or both—depends on the level of confidentiality of your documents.
Three situations we encounter in almost every company
None of these are technological problems. They are issues related to documents, rights, and habits —and that’s why a personal assistant bought from a catalog won’t solve them.
Your employees are already using a virtual assistant—and it's not yours
A personal account, a web browser, and the text of a contract pasted in to create a summary. No one is doing this with malicious intent: the tool is useful, and the company doesn’t offer one. The resulting output is neither tracked nor covered by your contracts.
The real competitor to a privately hosted assistant is the public assistant that everyone already has.
The assistant answers correctly, but misses the point
He found the procedure—the one from 2019, in a file that no one had put away. He responds with the confidence of a computer program. An assistant reviewing ten different versions of a document chooses one of the ten, with no way of knowing which one is the right one.
Organization is the goal; the assistant is the result.
No one can say what the assistant will be allowed to read
A system that has been open to the entire company for the past four years is now providing answers to the entire company. Pay records, acquisition files, annual performance reviews—no one could find them before, but now all you have to do is ask. The assistant doesn’t create leaks; it reveals the ones that already existed.
The rights check comes before the assistant.
What the assistant does, in everyday language
Six things—and the sixth is what sets him apart from a public defender: he knows when to stop.
The assistant provides answers based on your documents, citing the source
The procedure, the contract, the minutes, the product sheet: each answer cites the document and the passage from which it is taken. If you have any doubts, you can check with just one click.
The assistant sees only what the person sees
The permissions are those of your shares, managed by Entra ID. Two people who ask the same question will receive two different answers if their permissions differ—and that is by design.
The assistant prepares, summarizes, and compares
A draft response to a client, a summary of a 40-page file, the differences between two versions of a contract. The employee reviews them and makes a decision.
In Teams, not in a new tool
The question appears in the window your teams already have open. Adoption happens where people work, not in yet another app.
The assistant keeps track of it
Who asked what, which documents were used to respond, and which responses were flagged as incorrect. This is what makes it possible to correct the issue and respond to an audit.
The assistant steps in when there is no answer
No relevant documents, a question outside the scope, a request that involves the company: he says so, and he tells you who to contact. An assistant who makes things up is worse than an absent assistant.
A concrete, step-by-step example
A project manager asks on Teams : “What is the commissioning procedure for this client, and who needs to sign off on it?” Here’s how it works in the hybrid structure we most often implement.
What Remains Human
An assistant is a digital collaborator: a scope of responsibilities, certain permissions, a point at which they hand things over, and someone who supervises them. Here’s what they never decide.
Decisions That Cannot Be Delegated
They are recorded in the assistant’s file before the assistant begins work and reviewed during each quarterly evaluation. An assistant who does not have their own file is not ready, regardless of how well they perform during their demonstration.
- What he is authorized to read. The scope of the documents is determined by the data managers, not by the IT department, and never by the assistant.
- Verify the source before taking action. An answer is a lead backed by evidence, not a decision. The one who takes action reveals the source.
- What goes out. The assistant prepares the draft; a human reviews it and sends it. No response leaves the company without him.
- Correcting incorrect answers. A report is reviewed by a person, who corrects the document or the scope—not the assistant.
- Choosing and replacing the model. Tested based on your questions, compared, and decided upon. Every few months.
The data used, and what they require
An assistant is only as good as the documents he reads. Half the project happens right here, before the first question is even asked.
Sources
- SharePoint and OneDrive: procedures, templates, reports, documentation.
- Teams : channels and project files, when they are within the scope.
- Confidential materials: contracts, files, plans, results—kept on your premises when they need to stay there.
- Dynamics 365, theERP: the customer profile, the order, and the receipt, through their respective interfaces and based on the user's permissions.
- Entra ID : the groups and rights that determine everything else.
What they must have
- A single authoritative version, or a clear rule for determining which one is authoritative. Ten versions yield ten answers.
- Rights that tell the truth. A “for everyone” initiative launched four years ago is being revisited ahead of the opening.
- One owner per corpus, who responds when an answer is incorrect and decides what falls within the scope.
- Machine-readable documents: A contract that has been scanned at an angle cannot be read. It either needs to be scanned again, or it is excluded from the process.
- A scope that starts small. One department, one dataset, twenty people. The rest will follow after the first review.
Where the assistant works, and which one we would look at first
Three possible architectures, one of which is ruled out. The choice depends on how confidential your documents are, and it is made on a document-by-document basis, not for the entire set.
At Microsoft
Possible, often sufficientCopilot already respects your access rights: it only reads what the user is authorized to read in your tenant. When your documents can be read there, that’s the fastest response.
At your facility
PossibleEverything is in your datacenter, including the interface. When nothing should be exposed, including the orchestration. More complex to operate; the private RAG page explains this in detail.
Partly at Microsoft, partly at your office
RecommendedTeams and Entra ID at Microsoft; the confidential dataset, the search, and the model are hosted on your premises. Use of Teams ensures the confidentiality of your premises.
Next to the machine
Things to AvoidAn assistant working on your documents doesn't need milliseconds or a workshop. Unless it's on a site without a connection, in which case it's a special case.
Why Hybrids—and When We’ll Say Something Else
Three criteria are decisive for this use case. The other eight are less important here: the comparison tool breaks them down.
| Criterion | What tips the scales | Toward what |
|---|---|---|
| The Sensitivity of Documents | If the entire corpus can be read within your Microsoft tenant, Copilot is all you need. If part of it cannot—such as contracts, files, or trade secrets—that portion remains on your premises. | Copilot only, or hybrid |
| The Interface and Identity | Your teams use Teams and log in via Entra ID. Forcing them to use a second interface would result in losing half of their users. | Teams : Online or hybrid—rarely entirely in-person |
| The volume of requests | For a few hundred questions per day, the online service is the least expensive. Beyond that, the local model becomes the least expensive. The threshold is calculated over a three-year period. | Hybrid or local beyond the threshold |
The conclusion “Copilot in your licenses is enough; no infrastructure project needed” is included in our comparison tool, and we display it when it’s true. A privately hosted assistant isn’t a reason to buy a server.
What the assistant connects to
An assistant that only reads files is still just an enhanced search engine. Its value lies in its ability to access and trigger actions within your systems—always within the user’s permissions.
Infrastructure, When It Is Needed
It is only here that hardware comes into play, and only in one scenario: when volume, latency, or confidentiality make local execution preferable. In such cases, we also scale the necessary computing infrastructure.
Servers, Storage, Identity, Operations
A Dell PowerEdge server with one or two NVIDIA GPUs for inference, sized based on the number of concurrent users and the model size as measured in the prototype; storage for the corpus and its index; Azure Local as the foundation when Microsoft tools must remain in place; Foundry Local or an open-weight model. Prepared prior to delivery, connected to Entra ID, and maintained thereafter.
How We Carry Out the Project
There are five steps, and the first one is the one everyone wants to skip. It’s also the one that retains its value even if you decide to stop using the assistant later on.
The Framework: Rights, Body of Work, Scope
The access map, site by site. The initial data set and its owner. What can be viewed on the server and what must remain on your premises. Real questions that people are asking today, as documented.
Architecture
Copilot (standalone, hybrid, or fully local) is determined based on the data map and cost threshold. If hardware is required, its specifications are defined here, but it is ordered after the prototype is completed.
The prototype, as shown in your documents
A test group, about twenty users, three weeks, using existing equipment or borrowed equipment when space is an issue. We measure the accuracy of responses to the identified questions, latency, and how people use the system.
Validation and Industrialization
Hardware (if applicable), integration with your systems, logging, tested handoff to human agents, and role-based training. Rollout on a department-by-department basis.
Day-to-Day Operations
We review the flagged responses, verify the rights, test the template, and replace it when necessary, and monitor usage—either by us or by your team, whom we train.
Estimated timeline: nine to fourteen weeks from the first assessment to the launch of the first corpus for about 100 users, not including hardware delivery time, if any. This is a rough estimate, not a commitment: fine-tuning is part of the process. The first assistant isn’t profitable; the third one, built on the same foundation, is.
Scope and Limits
What an assistant doesn't do, and what we don't promise. Written here so it won't be discovered by the steering committee.
The assistant can make mistakes—with confidence
Sometimes a model generates a plausible but incorrect answer. The cited source is the solution: it allows you to verify the answer with a single click. Without a source, an answer is worthless, and the assistant is programmed to point that out rather than make something up. No setting can eliminate this risk entirely.
The quality of the documents determines the quality of the responses
Ten versions of a procedure yield ten answers. Organizing them, designating the authoritative version, and assigning one owner per corpus are all part of the project—often half of it. This work retains its value no matter what happens to the assistant.
Rights Are Being Put to the Test
What was left too wide open becomes clear in a single question. The rights statement precedes the opening; without it, we do not open. And it is redrafted with each review, because the shares shift.
The cost is incurred as the service is used, or paid for in the form of equipment
Online, each query costs something, and the bill comes in the third month. Offline, the equipment and its operation incur monthly costs regardless of usage. We calculate the costs for both options over three years, and someone reviews the bill afterward.
A local model is smaller than a public model
It often answers your questions better—because it reads your documents—but does a poorer job with everything else. It’s replaced every few months. We test it on your questions before choosing it—and before replacing it.
What We Won't Do
A demonstration as a selling point: the set of documents used there is specific to the demonstration, carefully selected, and up to date—and your environment doesn’t look like that. We evaluate your documents and address your questions—or we won’t make a decision.
Related cases and their corresponding pages
When another use case is the right one, and the technologies behind it.
Assistant for Company Documents
When the entire corpus is in SharePoint and can be read there: Copilot or an agent, after the rights have been cleared. The simplest case.
View the case A similar casePrivate RAG
When the online service is prohibited by a policy: everything must be done on-site, including the interface. This is the most demanding scenario.
View the case ArchitectureHybrid Artificial Intelligence
Where the border runs, what crosses it, and what it costs to maintain it.
View the page TechnologyMicrosoft Copilot
Roll it out smoothly: first, clarify the rights; next, determine actual usage; and finally, take action.
View the page TechnologyCopilot Studio
An employee's job description: what they know, what they are authorized to do, when they hand over responsibility, and who supervises them.
View the page ArticleManage the use of artificial intelligence that your employees have already adopted
Why Bans Don't Work, and Where to Start.
Read the articleLet's talk about this use case
Tell us what your teams are looking for, and where it's stored
The questions they ask a public assistant today, the documents that should be used to answer them, and those that must not be disclosed. We’ll return to the architecture we’re studying, the first dataset, and what a three-week prototype would allow us to measure.
What we offer is the opportunity to meet the engineers who will do the work. Assessment, prototyping, and industrialization are all part of the project. The initial discussion, however, costs nothing.
Renens, Sion, Châtel-Saint-Denis
Microsoft Solutions Partner and Dell Technologies Gold Partner. The support specialist, permissions, integration, and—when needed—the server, all handled by the same contact person.
Renens VD +41 21 806 37 15
Sion VS +41 27 552 00 22
Châtel-Saint-Denis FR +41 26 322 59 05