Guide
The world of AI Agents
Understand, compare, choose — the guide for business owners in Switzerland.
By Samuel Lavado · Published 3 September 2026
What is an AI Agent?
An AI Agent is an autonomous software system that carries out concrete tasks on behalf of a business owner: qualifying a new contact, replying to a client, sorting an email, blocking a slot in a calendar — by directly accessing the tools you use every day (email, calendar, client records) and chaining several steps together within the scope you set for it.
It is not a tool that waits for your instructions at every step. It is a digital colleague that reasons from your documents and your business rules, plans the actions required, carries them out, then reports back to you on the result.
Its autonomy is always limited: it acts within a precise scope that you control. When a situation falls outside that scope — an unusual request, an ambiguous case, an emotionally tense client — the AI Agent pauses and hands you the full context. The aim: to keep errors and frustration to a minimum.
If you run a law practice, an estate agency, a construction company or a service business in Switzerland, an AI Agent can take on a significant share of the administrative tasks that keep you from focusing on your real work.
Conventional chatbot, conversational assistant, AI Agent: what is the difference?
The term “chatbot” has become blurred: it can refer to a conventional chatbot (a program that follows a predefined scenario) or to a consumer conversational assistant. These are very different technologies. Here is a more useful distinction.
| Conventional chatbot | Conversational assistant | AI Agent | |
|---|---|---|---|
| How it works | Follows a predefined script | Generates natural-language answers | Reasons, plans and carries out actions |
| Understands context | No — recognises keywords | Yes, but may make up an answer if information is missing | Yes — understands your business context and checks the data you authorise before answering |
| Handles the unexpected | No — gets stuck outside the script | Partly — answers, but cannot act | Yes — finds a solution or hands over to you |
| Accesses your tools | No | Not by default | Yes, if you allow it — calendar, email, client records |
| Chains tasks together | No | No — one answer at a time | Yes — plans and carries out several steps |
| Relies on your data | No | No (except manual copy-paste) | Yes — consults the documents and history you authorise |
| Logs its actions | No | No | Yes — actions and decisions are traceable |
Conventional chatbot
How it works
Follows a predefined script
Understands context
No — recognises keywords
Handles the unexpected
No — gets stuck outside the script
Accesses your tools
No
Chains tasks together
No
Relies on your data
No
Logs its actions
No
Conversational assistant
How it works
Generates natural-language answers
Understands context
Yes, but may make up an answer if information is missing
Handles the unexpected
Partly — answers, but cannot act
Accesses your tools
Not by default
Chains tasks together
No — one answer at a time
Relies on your data
No (except manual copy-paste)
Logs its actions
No
AI Agent
How it works
Reasons, plans and carries out actions
Understands context
Yes — understands your business context and checks the data you authorise before answering
Handles the unexpected
Yes — finds a solution or hands over to you
Accesses your tools
Yes, if you allow it — calendar, email, client records
Chains tasks together
Yes — plans and carries out several steps
Relies on your data
Yes — consults the documents and history you authorise
Logs its actions
Yes — actions and decisions are traceable
A conventional chatbot answers questions that were planned in advance. “What are your opening hours?” → pre-programmed answer. If the question falls outside the script, it goes round in circles or points to a form. Useful for a basic FAQ, limited as soon as the conversation becomes real.
A conversational assistant is built on an LLM (Large Language Model — a language model trained to understand and generate text). It understands natural language and gives relevant answers, but it only answers: it does not check your calendar, send an email on your behalf, or qualify a contact according to your criteria. And when information is missing, it may produce a plausible but false answer — without warning you.
An AI Agent combines the same type of LLM with access to your tools and the ability to act, from the outset. A potential client writes to you at 10 pm with a detailed request? The AI Agent understands the request, checks your availability, qualifies the contact according to your criteria and offers them a slot, without disturbing you. If the request falls outside the intended scope, it notifies you instead of improvising. Before answering, it consults the document base you have approved for it using RAG (Retrieval-Augmented Generation — in plain terms: it looks up the information in your own files before forming its answer, rather than making it up).
How does an AI Agent work, in practice?
An example of a standard AI Agent. It follows a six-step cycle with every interaction:
MentilAI deployment strategy
Before deployment, MentilAI carries out an audit of your client communication processes. The AI Agent is built on the basis of this audit, validated together. Its scope of action is documented in a configuration specification: tasks covered, documents it may cite (chosen with you), default operating mode, actions classed as sensitive that require your approval.
Once validated, the AI Agent goes into production with two possible modes:
Co-pilot
The AI Agent analyses and proposes; you approve before anything is carried out. Default mode during the start-up phase, to build confidence gradually.
Autonomous
The AI Agent carries out validated tasks on its own within its scope. Actions classed as sensitive stay under your control.
You choose which mode applies to each type of task, and can adjust it at any time. A review is planned at the end of the start-up phase to assess the switch to autonomous mode for eligible tasks.
The precise commitments — timelines, scope of action, level of traceability — are set out in the configuration specification drawn up for each project.
In short
No technical skills required: just a little of your time at the start, to define the task. You receive an operational AI Agent, tested on your real cases. You move from “I approve” to “I delegate” — at your own pace, within a clear framework.
AI Agents configured by sector
Every sector has its own workflows, regulatory constraints and client expectations. A high-performing AI Agent is not a generic tool applied to your trade — it is a configuration built around you that reflects your rules, your vocabulary and your priorities.
MentilAI configures AI Agents for four business sectors in Switzerland:
Legal
Document follow-ups, sorting of enquiries: fewer back-and-forth exchanges, files that are more complete by the first appointment.
Real estate
Financial qualification, follow-up of viewings: serious enquiries are handled first, even outside office hours.
Construction
Technical pre-qualification, scheduling of visits: unrealistic enquiries are filtered out before the site visit.
Services
Bookings, flow management: no more lost bookings during busy periods.
These proposals give a direction. In the end, you define the role of your AI Agent — according to your priorities, your constraints and the reality of your day-to-day work.
Does one of these sectors sound like yours?
45 minutes to understand your needs and explore the possibilities.
What an AI Agent does not do
An AI Agent is not infallible. Hallucinations (answers that seem correct but are not) still occur. In a survey cited by Microsoft Research (2025, p. 18), 40% of employees surveyed said that, in the previous month, they had received AI-generated content that looked useful but turned out to be incomplete or inaccurate. That is why MentilAI AI Agents are designed with safeguards: human validation of critical actions, a defined scope of action, traceability of decisions.
An AI Agent does not replace your judgement. It does not make strategic decisions for you, and does not handle situations that call for empathy, creativity or a fine reading of human relationships. It does not negotiate the terms of a property instruction and does not argue cases in court.
The AI Agent is designed to handle routine requests reliably. For complex situations, it prepares the ground and hands over to you — your client notices no drop in the quality of the exchange.
The question is not “will AI replace my profession?” The useful question is: “Which repetitive tasks cost me time that I could spend on what really gives my expertise its value?” That is exactly the logic of delegating the invisible.
What the data says
+448%
increase in AI job postings in the United States since 2018, against −9% for non-AI IT roles
Job postings published in the United States, 2018 to 2025, UMD-LinkUp AIMaps data (University of Maryland), cited by BOND.
88%
of early adopters report a positive ROI (return on investment) on at least one use case
Survey of 3,466 business leaders worldwide, 2025, for Google Cloud.
62%
of organisations are already experimenting with AI Agents (23% are deploying them at scale, 39% are testing them)
McKinsey's annual global survey of executives and managers, November 2025 edition.
40–60 min
saved per day and per user of AI tools
Microsoft Research, 2025 — p. 17
Time reported by ChatGPT Enterprise users surveyed (Chatterji et al., 2025), cited by Microsoft Research.
These figures come from international studies on varied populations. They do not automatically apply to every Swiss business owner: results depend on the sector, the volume of enquiries, the quality of the data available and the rigour of the implementation. In practice, ROI appears when you target a specific flow — for example sorting incoming enquiries, qualifying contacts and booking appointments.
Data protection and Swiss law
The FADP (Swiss Federal Act on Data Protection — the Swiss law, based on the same principles as the European data protection regulation, the GDPR) applies directly to any processing of data by AI. The FDPIC (Federal Data Protection and Information Commissioner) makes this point, in substance, in its publication “AI and data protection” of 24 September 2025:
Your clients must know whether they are dealing with a machine — and to what extent their data is used.
For more information, see our MentilAI privacy policy.
Frequently asked questions
The words of AI, in plain English
The vocabulary of this guide, explained in one sentence and illustrated with an example from the everyday life of an SME.
The tools
- AI Agent — AI: Artificial Intelligence
In plain English: software that understands a request, consults your documents and acts in your tools, within a scope that you set.
Example: a request for a quote arrives at 10 pm at a construction company; the AI Agent asks the relevant questions and proposes a time for a site visit.
- Chatbot
In plain English: a conversation program that follows a scenario written in advance and recognises keywords. Outside the scenario, it gets stuck.
Example: on a shop's website, it answers “Open from 9 am to 6 pm” when asked about opening hours, but cannot handle a complaint.
- Conversational assistant
In plain English: a consumer conversation tool that understands your sentences and writes answers, but does not act in your tools.
Example: in a law firm, an employee has it draft a letter. It writes it well, but may cite an inaccurate legal provision without flagging it.
How they work
- Natural language
In plain English: everyday language, with its turns of phrase and its imprecision — as opposed to the fixed commands of a piece of software.
Example: a client of the estate agency writes “Would you have a 3.5-room flat not too far from the station?” — the AI Agent understands, with no form to fill in.
- LLM — Large Language Model
In plain English: the engine that understands and produces text. It has learned from vast volumes of text, not from your business.
Example: for a shop, the same LLM writes the text of the offer of the month or summarises a client's email; it handles language without knowing your stock.
At MentilAI: the models that process your data run in Switzerland, unless you expressly request otherwise.
- Knowledge base
In plain English: all the documents you entrust to the AI Agent — prices, procedures, terms. “No source, no answer”: it relies on them.
Example: a construction company puts its list of services, its usual lead times and its quotation terms in it; the AI Agent answers on the basis of these pages.
- RAG — Retrieval-Augmented Generation
In plain English: before answering, the AI Agent looks for the relevant passages in your documents, then writes its answer from them, instead of making things up.
Example: a client asks a law firm what a first consultation costs. The AI Agent finds the fee in the approved price list and quotes it, with its source.
- API — Application Programming Interface
In plain English: an access door provided by a piece of software, which lets another program read or write information, according to defined rights.
Example: through the API of an estate agency's calendar, the AI Agent sees the free slots and books a viewing — you grant only these two rights.
Note: the MCP. The MCP (Model Context Protocol) is a common rule, designed for AI. Without it, each piece of software has its own door, with its own rules, and a custom access has to be prepared for each one. With the MCP, every piece of software presents its door in the same way, like a universal travel adapter: a single cable fits every socket. The MCP also tells the AI Agent what it is allowed to do there, for example “read appointments” or “book a slot”. It relies on APIs; it does not replace them. Example: if an estate agency's calendar and its property management software both offer the MCP, the AI Agent connects to them in the same way, instead of going through two separately designed accesses.
Measuring
- Use case
In plain English: a specific task to which AI is applied, with a beginning, an end and a result that can be measured.
Example: in a construction company, “sorting the quote requests received by email” is a use case; “putting AI everywhere” is not.
- ROI — Return on Investment
In plain English: what an investment brings in, compared with what it costs — in money or in working hours saved.
Example: a shop compares the monthly cost of its AI Agent with the hours its team no longer spends confirming bookings.
The law
- FADP — Federal Act on Data Protection (Switzerland)
In plain English: the Swiss law that sets out how a business may collect, use and store data about people, and what those people can demand.
Example: an estate agency that receives a prospective tenant's application must inform them of the purpose for which their data is collected and of any recipients.
- GDPR — General Data Protection Regulation (European Union)
In plain English: the European counterpart of the FADP. It may concern a Swiss business that targets customers in the European Union.
Example: a Geneva shop that also delivers to France may have to take the GDPR into account as well as the FADP.
- FDPIC — Federal Data Protection and Information Commissioner (Switzerland)
In plain English: the federal authority that supervises the application of the FADP, gives advice and publishes recommendations, notably on AI.
Example: a law firm wonders whether its clients must know that they are writing to an AI Agent; the FDPIC answers this in its September 2025 publication.