Copilot Studio
An agent is a digital employee: a scope of responsibility, permissions, and a supervisor.You don’t just “assign customer service” to an agent. You assign them a routine, tedious, and verifiable task —and you spell out in black and white what they know, what they’re allowed to do, when they hand off the task, and who reviews it. The rest isn’t about technology; it’s about organization.
Three tasks that no one should have had to do by hand
They have one thing in common: they come up often, the answer already exists somewhere, and they take up the time of people who are paid to do other things.
The same request, three hundred times a year
A request for access, a certificate, a duplicate invoice. The form exists, the rule is known, and yet someone is processing them one by one.
Three hundred times fifteen minutes: ten days of work per year.
One in four customer questions has the same answer
Where is my order? What are your delivery times? How do I cancel? The answer is in writing, it's accurate, and it's updated every day.
Meanwhile, requests that require human attention are waiting.
Human Resources answers phone calls based on what's on the intranet
The vacation policy, the procedure for obtaining a certificate, the reimbursement form. Everything is posted online. No one can find it.
A document that cannot be found does not exist, regardless of its quality.
What an agent does—and what he doesn't do
Six points, stated without fanfare. The word “intelligence” does not appear even once: what matters is the scope.
On your documents—and only yours
Catalog, procedures, knowledge base, internal regulations. It cites its sources, which makes it possible to verify—and correct—the document rather than the agent.
Do, not just talk
Open a ticket, check a delivery time onERP, book a time slot, or update an address. That’s what sets a helpful agent apart from a chatty search engine.
Knowing When to Stop
Two unsatisfactory responses, an angry remark, a complaint, a request not listed in the catalog: he forwards the entire conversation, not just a case number.
In " Teams," on the website, in the contact center
The same agent, the same scope, the same rights—regardless of where the agent is queried. It is created only once.
A weekly chart, not a printout
Questions received, answers provided, human review rate, topics on which it fails. It is this last figure that helps improve your documents.
He doesn't make up his own rights
He sees only what the person speaking to him sees—and nothing more. An employee who does not have access to a file will not be granted access simply by asking the agent nicely.
A staff member's job description
This isn't a technical diagram, and that's by design. An agent is managed just like any other employee: we document what it knows, what it can do, when it stops, and who reviews it. Here is the profile of an agent we often use.
What He Knows
The product catalog, the 180 published and validated articles in the knowledge base, the terms and conditions, and the delivery times provided by the carrier.
- And the order status, displayed in real time in theERP
- Nothing else: neither shared leadership nor internal communication
What he is allowed to do
Three actions, listed here—and not one more.
- Provide the order status to the person who placed it
- Open a claim file, with the conversation attached
- Send a duplicate invoice to the registered address
When he steps down
Immediately, and without trying again.
- Two responses in a row deemed unsatisfactory
- An angry remark, a threat, a mention of a lawyer
- Any questions regarding pricing, discounts, or cancellations
- Anything that's not in the catalog
Who supervises him?
A specific person, not “the IT department.” That person receives a spreadsheet every Monday and decides on the corrections.
- Questions received, answers provided, reviewed by a human
- The ten topics he failed on: that means ten articles to write
A sheet like this fits on one page and can be reviewed in three minutes by a department manager. An agent who does not have their own copy is not ready to be put on duty, regardless of how well they perform the demonstration.
The first three agents who are making progress
All three share the same characteristics: the task is routine, the solution already exists, and the results are verifiable. That’s what makes them reliable—and profitable as early as the first quarter.
The agent who answers pre-sales questions
Lead times, compatibility, delivery terms, availability. He consults the catalog and knowledge base, and creates a request as soon as a genuine need arises.
It attracts viewers at 10 p.m., who never call back the next day.
The Administrative Affairs Officer
Vacation, expense reports, equipment, access. It responds to your requests, and it forwards the request to the appropriate department when the response isn't sufficient.
It frees up the time the HR department spends repeating what's already written.
The agent who lays the groundwork
He identifies the caller, understands the reason for the call, resolves simple issues, and forwards the others— using what he has already learned—so the customer doesn't have to repeat themselves.
It reduces the number of calls without compromising perceived quality.
What we do, and in what order
Building the agent is the quick part. The second step is the one that takes time, and it’s the one that determines whether the agent’s answer is right or wrong.
Choosing the Right First Task
Three criteria, and all three must be met: the task is frequent, it is tedious, and its result is verifiable. If any one of the three is missing, we choose something else.
We start with your actual volumes, not a list of ideas.
The Source of Truth, and How It Is Organized
That’s the real issue. An agent only knows what he’s given: if your procedures exist in three versions across two different shared folders, he’ll give the wrong answer with the same confidence as if it were the right one.
We identify the authoritative documents, set aside the others, and establish a process to keep them up to date.
Shares and Rights
What the agent is authorized to trigger, and on what data. The agent sees what the person they are speaking to sees—and nothing more—and every action is logged.
The Shift to Human-Centered Approaches, and Safeguards
Written before the system goes live, not after the first incident: what triggers the transfer, to whom, and how. And what the agent will never say, no matter what the question.
Monitoring — An Agent Reviews Their Work
Every week: what it received, how it responded, and what was handled by a human. The tasks it fails at are the list of documents to be written.
Without this feedback loop, an employee gradually loses effectiveness as the company changes.
How long, honestly?
For the first agent. The second one costs much less, because the organizing has already been done.
The scenario: a single agent, one task, two sources of documentation, two or three actions to trigger, and a business manager available half a day per week. Well-organized documentation cuts the second phase in half; scattered documentation doubles it.
Selecting a Task
Volumes, risk, expected return. The job description is provided here.
The Source of Truth
Organizing documents. By far, the task that takes the most time.
Construction and Testing
Actions, rights, safeguards, and misuse tests.
Supervised Opening
First internally, then to a segment of the public, and finally to everyone.
Seven to eleven weeks for the first agent to go live. A second agent based on the same documentation: three to four weeks. That’s why we never judge profitability based on the first one.
An agent connected to disorganized documents responds "false" with the confidence of a computer program.
Organization is the goal. The agent is the result.
This is the opposite of what a demonstration promises, where an agent can be set up in ten minutes to work with a selected set of documents. At your company, that set of documents doesn't exist yet.
A wrong answer is more costly than no answer at all
A customer who gets no response will call back. A customer who receives the wrong answer will act on that answer—and you’re the one who’ll bear the consequences.
The benefit remains, even if the agent is abandoned
The cleanup effort benefits research, customer service, the onboarding of new hires, and Copilot. That’s what makes this project low-risk: the time-consuming part is useful anyway.
What You Need to Know Before Signing
Five points that determine both the budget and the schedule.
You pay as you go, and that changes the design
An agent uses credits with each exchange. It is not a subscription per user: it is a counter, and it increases as the agent becomes more successful.
The practical result is this: we don’t deploy the agent everywhere as a matter of principle. We deploy it where the task occurs frequently and where the benefit is measurable—and we monitor usage just as we monitor a phone bill.
What Is the Responsibility of an Agent, and What Is the Responsibility of a Development
Responding to documents and triggering three simple actions: that's an agent. Orchestrating a seven-step process with business calculation rules and hierarchical validations: that's an application.
We'll be honest with you, even when the answer is "this isn't the agent you need."
Your access rights will be put to the test
An agent doesn't create leaks—it reveals the ones that already existed. Sharing information openly across the entire company becomes a response for the entire company, and this usually becomes apparent on the very first day.
That is why the state of rights is part of the second phase, before any opening.
You remain responsible for the response
What an agent says is binding on the company, just as if an employee had said it. Hence the written safeguards, the record of every exchange, and the rule that prohibits discussing prices, discounts, and cancellations from the outset.
The first agent isn't profitable, but the third one is
The first one pays for organizing the documents and setting up the monitoring system. The subsequent ones reuse all of that and are built in three to four weeks.
To judge the process based solely on the first agent is like judging a workshop by its first piece.
What plugs into it
An isolated agent hits a plateau. Here's what takes their abilities to the next level.
Microsoft Copilot
Copilot Studio builds agents for a task. Copilot assists your teams in their day-to-day work. Both are based on the same prerequisite: organized documents and appropriate permissions.
The Copilot Page The SourceCustomer Service
The knowledge base that powers the agent, and the file it opens when it hands things over.
The Customer Service Page When it's not an agentPower Apps
For multi-step processes, hierarchical validations, and business calculations: that's where you need an application, not a conversation.
The page Power AppsCopilot Studio is a use case first and a tool second.The document assistant,the customer service agent, and shared inbox processing are all written as projects, incorporating the human elements and the architecture that goes with them.
Tell us which task tires you out
The one that keeps coming up, for which the answer already exists, and which takes up the time of someone who is paid to do something else. We’re back with a scope, a workload, a timeline, and a job description for a staff member.
Preparing a proposal is free. Design studies and audits are included in the scope of work. Three agencies: Renens, Sion , and Châtel-Saint-Denis ; a Microsoft partner for sixteen consecutive years.
Search for one of your procedures
Pick an internal policy that everyone is familiar with—vacation time, expense reports— and look it up using your own tools, just as a new employee would.
If you find two different versions—or none at all—you've just figured out, in five minutes, what your first agent would have said.