Knowledge

AI agents for Swiss SMEs

An AI agent is a program that takes a goal, decides itself which tools to call and works in a loop until the goal is reached or a human steps in. Vectoryon rebuilds the processes at the heart of your company with you and AI. The AI agent is the tool for this. It works in your mailbox, file storage, accounting and own templates. In Vectoryon, companies create AI agents without programming, as a block in an automation. The workspace agent checks mail and calendar 1, 2 or 4 times a day and suggests new tasks.

Dominic Kunz, CEO8. Oktober 202611 min read

Vectoryon is a Swiss AI platform for SMEs by k-innovations GmbH, based in Bern. An AI agent works directly in Outlook, calendar, CRM and ERP. It reads requests from the inbox and assigns them. It prepares quotes from templates and checks receipts. Nothing goes out without approval by a person. The application, files and the AI models run in Swiss data centres. Swiss data protection law applies. Vectoryon costs CHF 49 per person per month.

A chatbot answers once. An AI agent chains steps such as reading an email, opening a file and writing a draft on its own. The term is often sold bigger than it is. In June 2025, Gartner expected over 40 percent of agent projects to be cancelled by the end of 2027, mostly due to unclear value and cost. This text explains what an AI agent really does, how it differs from an AI assistant and where a human decides. The definition follows the distinction Anthropic described in December 2024 in "Building effective agents": a workflow follows a predefined path, an agent chooses its own path. If you only want a list of tools, see AI tools at a glance.

What an AI agent actually does

An AI agent gets a task in plain language, for example: "Prepare the reply to the request from Meier Ltd and put the draft in the mailbox." It breaks the task into steps. It finds the email, reads the attachment, fetches the price list from file storage, writes the draft and files it. Each of these steps is a tool call: read mailbox, open file, create document.

The difference from a language model in a chat lies in this loop. A model answers a question once. An agent checks after each step whether the goal is reached and otherwise calls the next tool. If information is missing, it looks for it. If a call fails, it tries another way or reports the error.

Three building blocks make up an agent. First, the language model, which plans and writes. Second, the tools, each with a clear description of what it does and what it needs. Third, the context: the documents, the mailbox and the templates of the company. Without tools, it is a chatbot. Without context, it answers in general terms instead of about your company. How an agent reads large amounts of data without overflowing its context window is explained on the page Context Engineering.

What an agent is not: an employee who carries responsibility. It prepares. Whether the reply to Meier Ltd goes out is decided by the person who reads the draft.

AI agents in Swiss SMEs: five everyday workflows

AI agents pay off in an SME where work comes up often, the documents are already in the company and a person decides at the end. The Vectoryon AI agent can take over the following five workflows today. Each one says what the agent does and who approves.

First, the customer request. A request arrives by email in Outlook. The agent reads it with its attachment, creates the contact in the CRM, checks in the ERP whether material and capacity are sufficient and prepares the quote. Then it stops. The responsible person checks the draft and approves it. Only then does the quote go to the customer.

Second, preparing a meeting. Asked to «Prepare my 10 o’clock meeting», the agent fetches the appointment from the calendar, the latest emails from the customer and the open tasks. It writes a one-page summary. It changes nothing. The person reads it and decides what to use in the meeting.

Third, the task list. The ready-made automation «Project and task updates» reads email, calendar and file storage. It suggests which tasks are new, due or done, each with source and reason. A suggestion is only applied once the person confirms it.

Fourth, time tracking. Asked to «Log today’s hours on the project», the agent books the hours, shows the entry and asks if something is unclear. Deleting only works when the person clicks the confirmation.

Fifth, the work report in a trade business. A photo of the signed time-and-material report is uploaded. The agent reads date, names, hours and material and enters them in the project’s report table. From this it drafts the time-and-material invoice. Invoicing and sending stay with the company.

All five workflows have one thing in common: the AI agent prepares, a human approves. Every step is logged with time, system and approval. Where a decision needs expertise, such as a price outside the norm, it stays with the person in the SME.

AI agent, AI assistant, chatbot, workflow: the difference

A chatbot answers questions from a fixed knowledge base. It calls no tools and does not know the company’s documents. Typical example: the question and answer box on a website.

An AI assistant works on the company’s data and prepares a result on request: an answer, a draft, an analysis. It calls tools, but one step per request, and the person leads. This is the most common use in SMEs today. What an assistant does exactly is on the page AI assistant.

An AI agent takes a goal instead of a step. It chains several tool calls itself and works until the goal is reached. It can run on a schedule, for example reviewing the mailbox every morning and suggesting tasks.

A workflow, for example in automation software, follows a fixed process someone drew beforehand: if an email has an attachment, file it, then notify. Reliable, but rigid. An agent decides the process itself, which makes it less predictable. Anthropic therefore recommends starting with the simplest setup and using an agent only where the path is not known in advance.

In practice, the terms blur. Vectoryon calls its product an assistant because the person leads in the chat. Technically, an agent with a tool loop works behind it. The automations are the scheduled part of it.

In the video: how an AI agent works in four steps and how it differs from a chatbot (7 minutes, in Bernese Swiss German).

How an agent works at Vectoryon: tools, sandbox, approval

Each tool has a description and a schema that says which inputs it needs. The agent reads the description and decides whether the tool fits the task. Examples: search the mailbox, read the calendar, fetch a file from SharePoint, create a document from a template, change a row in a table.

The tool calls run in a sandbox, a sealed area that only knows the approved tools. The agent cannot run code that acts outside this area. Every call is logged: which tool, with which inputs and which result. This trail is the basis for tracing a run later.

Anything that leaves the company or changes data goes through an approval. An email to a customer, a change to an appointment, a document sent to an authority: the agent prepares, a person approves. This is not a safety net for emergencies. It is the normal process.

The data stays in Switzerland. Processing follows the revised Swiss data protection act. Access to Microsoft 365 runs through the Graph interface with OAuth and can be revoked at any time.

Creating AI agents without programming: how it works in Vectoryon

In Vectoryon, an AI agent is a block in an automation. The automation is built on the canvas, a surface where blocks are connected with lines. No programming is needed. Today, the company admin creates an automation, or Vectoryon builds it together with the company.

The first step is the trigger: a schedule such as every morning at seven, a webhook, a website form, a change in a table or a start from another workflow. The second step is the «AI agent» block. It gets an instruction in plain language, for example: «Read the new requests and draft a reply to each one.»

The third step is choosing the tools. On the block, you tick which tools the agent may use, such as search the mailbox or fetch a file from SharePoint. The output is free text or fixed fields that the next block uses.

The fourth step is approval. Before every step with an external effect, such as an email to customers, comes the «Approval» block. The automation then runs in test mode first. Only when the results are right is it switched live.

AI agents for SMEs: three examples

AI agents for SMEs pay off where work comes up several times a week, the documents are already in the company and a person approves at the end. Three examples from Swiss industries show what the Vectoryon AI agent does and where the specialist decides.

First example, the fiduciary office: the AI agent reads incoming receipts, assigns them to the right client and suggests booking entries. The account assignment stays with the fiduciary. The agent does not answer a valuation question in the annual accounts. That remains the specialist’s job.

Second example, the trade business: from a photo of a signed time-and-material report, the AI agent drafts the invoice. For quotes without a reply for five days, it puts a follow-up email draft in the drafts folder. The owner reads, adjusts and sends.

Third example, the scheduled mailbox review: every morning at seven, the AI agent reads the new emails, flags those with a deadline and files a list of open items. The time is fixed, the path through the mailbox is not. That sets an AI agent apart from a rigid workflow.

In a hotel, the same agent prepares replies to guest requests in several languages. What use under professional secrecy looks like is shown on the pages AI for doctors and AI for lawyers.

The three examples share three conditions: the work repeats, the documents are in the company and a person approves. If one of these is missing, an AI agent is of little use in an SME. A price negotiation or a building law decision stays with the specialist.

In the video: which office work goes to an AI agent first, three cases on a real screen and three settings for a safe start (4 minutes, in Bernese Swiss German).

Chatbot, assistant and agent compared

FeatureChatbotAI assistantAI agent
Inputquestionrequest for one stepgoal over several steps
Toolsnoneyes, one call per requestyes, several calls in a loop
Knows company datanoyesyes
Runs without a requestnonoyes, can run on a schedule
Who decidesnobody, only an answerthe personthe person, via approval
Predictabilityhighhighmedium, hence log and approval

The columns describe the typical case. Products mix the forms. Vectoryon is an assistant in the chat and an agent in the automations.

Where the limits are

  • An AI agent does not decide. It prepares and submits for approval. Responsibility for an email, a booking or a contract stays with the person.
  • It does not replace specialist software. No general ledger, no ERP, no CAD, no hotel system. It works alongside them and hands over by export.
  • It is only as good as the documents. If the price list is missing from file storage, it does not find it and says so. It does not guess.
  • An agent is less predictable than a fixed workflow. Where the process is known and never changes, a workflow is the better choice.
  • It does not act outside the approved tools. What is not connected as a tool, it cannot do, even if it can describe it.

Details

How do you limit what an AI agent may do?

An AI agent in Vectoryon can only call the tools selected on its block. An agent without the «send email» tool cannot send an email, even if the instruction asks for it.

Every «AI agent» block has limits for the number of rounds, the cost and the time per run. When a limit is reached, the run stops instead of continuing.

The admin approves in the company settings which mailboxes an automation may read. An agent in the chat works with the permissions of the signed-in person and sees nothing that this person may not open.

Every tool call is logged with input and result. After each run, you can check what the agent did and adjust an instruction or tool selection precisely.

What an AI agent costs

The Vectoryon workspace costs CHF 49 per person per month. This includes the agent, the workspace and the Microsoft 365 connection. The page Pricing calculates the cost for the whole team.

To set it up, you create templates, connect file storage and approve tools. How long this takes depends on the state of your documents. A company with well-kept file storage is ready in days. One with files across five drives needs longer.

Running costs arise per conversation with the language model. They are included in the price, up to a quota per person. If you send the agent through a thousand emails every morning, you need more. Usage is visible in the workspace.

Sources for this text

Anthropic, "Building effective agents", December 2024: anthropic.com/research/building-effective-agents. The distinction between workflow and agent and the advice to start with the simplest setup.

Gartner, press release of 25 June 2025: "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027". gartner.com/en/newsroom.

Search volume for "ki agent" and "was ist ein ki agent" in Switzerland: own measurement with DataForSEO on 27 August 2026, 260 and 180 searches per month.

Frequently asked questions

Is an AI agent the same as ChatGPT?

No. ChatGPT is a language model in a chat. An agent uses a language model but also calls tools and works on a goal in a loop. The tool comparison is on the page AI tools.

Can the agent send emails on its own?

It writes the draft and submits it for approval. Sending goes through a person. This applies to everything that leaves the company.

What happens if the agent makes a mistake?

Every step is logged. You see which tool was called with which input and what came back. The error becomes visible before anything goes out, because the approval comes first.

Does an SME need an agent, or is an assistant enough?

For most tasks, the assistant is enough: one person, one request, one result. An agent pays off for recurring workflows without a request, such as the daily mailbox review. Both are included in the Vectoryon workspace.

Where does the data go when the agent works?

It stays in Switzerland. Processing follows the revised Swiss data protection act. The service providers involved are listed in the privacy policy.

Are AI agents worth it for a small SME?

Yes, if some work comes up several times a week, such as answering requests, preparing quotes or logging hours. Vectoryon costs CHF 49 per person per month, details on the page Pricing. If a task comes up only twice a year or the documents are scattered, the agent is hardly worth it. It also replaces neither specialist software nor a specialist.