How to Create an AI Agent: A Step-by-Step Guide for Companies

Creating an AI agent no longer requires programming skills. You need a clearly defined task, a suitable language model, good instructions, the right knowledge and the necessary tools. In this guide, we go through these building blocks step by step, in general and using Evoya Agents as an example. At the end, there is a sample agent you can rebuild right away.

Key points at a glance

• Start small: Your first agent should take over a single, recurring task with a clear input and a clear result.
• Six steps: Define the task, choose the model, write the instructions, connect knowledge, enable tools, test and share.
• Think about data protection from the start: For sensitive data, choose a model that runs in Switzerland. At Evoya, your data is never used for AI training.

What is an AI agent?

An AI agent is a program that uses a language model to complete a task on its own. Unlike a simple chat, it doesn't just answer questions: it works with your knowledge, uses tools such as email or calendars and carries out several steps one after another. People decide what the agent is allowed to do and check the results where it matters.

Before you start: choosing the right task

Most agents don't fail because of the technology, but because the task is too big or too vague. Good first agents have three properties:

• Narrowly defined: «Answer customer questions about delivery times» rather than «take over customer service».
• Recurring: The task comes up daily or weekly, so the setup pays off quickly.
• Clear input, clear result: An email, a PDF or a question goes in; an answer, a summary or a completed form comes out.

Typical candidates are answers to frequently asked questions, extracting orders and invoices, pre-sorting applications or emails, and recurring research.

Step 1: Define the task and the goal

Write down in two or three sentences what the agent should do, for whom, and how you will know it is doing a good job. Also define what it must not do, such as make binding commitments or quote prices that are not in the documents.

This short brief later forms the basis for the instructions and for your tests. In Evoya Agents, you can start with a preconfigured agent and adapt it to your task, or build your own agent from scratch.

Step 2: Choose the right model

The language model is the agent's brain. Not every model suits every task: for texts and summaries, fast and inexpensive models are often enough; for demanding reasoning, larger models are worth it. Just as important is where the model runs.

In Evoya Agents, you can choose from more than 30 models, from Swiss open-source models to ChatGPT, Claude and Gemini. Our rule of thumb: for personal data and confidential documents, a model that runs in Switzerland. For public research, ideas or drafts without sensitive data, international models too.

Step 3: Write the instructions

The instructions (also called the prompt) give the agent its role, its approach and its tone. Good instructions are specific and include examples. A proven structure:

Role: You are the information assistant of the municipality of Sampletown.
Task: You answer residents' questions about regulations and fees.
Approach: Answer only on the basis of the stored documents and name
the document the answer comes from.
Limits: If you can't find an answer, say so openly and refer to the
municipal office. Do not give legal advice.
Tone: Friendly, short and polite.

In Evoya Agents, you can add skills to the instructions where needed, giving the agent extra abilities for specific tasks.

Step 4: Connect knowledge

An agent is only as good as the knowledge it works with. Upload the documents it needs for its task, or connect existing storage so it always works with the latest version. Less is often more: ten current, well-structured documents beat a thousand outdated ones.

Technically, this usually works through RAG (retrieval-augmented generation): for each question, the agent searches for the relevant passages and builds its answer from them. Learn more in the glossary.

Step 5: Enable tools and integrations

Tools are what turn an assistant into an agent that acts: sending emails, looking up appointments, entering data into a CRM. Only enable the tools the agent really needs for its task.

In Evoya Agents, the agent accesses external systems through connectors, APIs and the Model Context Protocol (MCP), for example Microsoft 365, databases or file storage. Our guide Teaching AI agents to send emails shows what this looks like for email. For larger tasks, an agent can hand off subtasks to specialised subagents.

Step 6: Test, share, improve

Test the agent with ten to twenty real cases from your everyday work, including difficult ones and ones it should decline. Check: Is the answer correct? Does it name the source? Does it stay within its limits? Then adjust the instructions or the knowledge.

Once the agent works, share it with your team. In Evoya Agents, everyone then uses the same agent with the same knowledge and the same tools. An agent you have built can also be used in several channels: in the chat, on your website, in Microsoft Teams or Slack and in the email inbox. Plan a regular review, for example monthly, to spot new questions and outdated documents.

Example to rebuild: the quote assistant

An agent that prepares a draft quote from a customer enquiry, which a person then only needs to check and send:

The same structure works for a regulations information service in a municipality, an agent for internal IT questions or a daily research assistant.

Five AI agents Swiss organisations use today

• Insurance advice at insurando.ch: An agent in the website chat answers questions about basic and supplementary insurance, clarifies needs and generates leads, even outside business hours. View the project
• Order capture at ASTORtec: Orders from emails and PDFs are captured, checked and transferred to the ERP automatically, taking only 1–2 minutes per order instead of around 10. View the project
• More than 30 agents at BWS Uster: Tutors for teaching and administrative agents for the back office, for more than 200 users. View the project
• Website chat of the municipality of Küsnacht: Answers frequently asked questions around the clock, by text or voice, and forwards when needed. View the project
• Research assistant: Researches automatically every day and stores the knowledge so the team can search it. Read the guide

AI agents and data protection

As soon as an agent works with personal data, the Swiss Federal Act on Data Protection (FADP) applies. The Federal Data Protection and Information Commissioner (FDPIC) states that the FADP applies directly to AI. Clarify three points before you start:

• Where is the data processed? For sensitive data, choose a model that runs in Switzerland.
• What is the agent allowed to do? Only enable the tools and storage it needs.
• Where does a person check? For decisions with consequences for customers, employees or citizens, the agent makes a proposal and a person makes the decision.

To learn what the federal government is planning, read our post on the AI law in Switzerland.

No-code platform, n8n or custom development: which fits?

Our assessment from practice. The approaches can be combined: many organisations start on a platform and have individual integrations developed.

Frequently asked questions about creating AI agents

Do I need programming skills to create an AI agent?
No. On an AI platform such as Evoya Agents, you put together the model, instructions, knowledge and tools without code. Programming is only needed for custom integrations into special systems.

How long does it take to create an AI agent?
A preconfigured agent is ready in minutes. Your own agent with your knowledge and your own tools needs a bit more time for setup and testing, depending on the task.

What is the difference between a chatbot and an AI agent?
A chatbot answers questions. An AI agent completes tasks: it works with your knowledge, uses tools and carries out several steps on its own.

Which language model is suitable for an AI agent?
That depends on the task and the data. For sensitive data, we recommend a model that runs in Switzerland; for demanding tasks without sensitive data, large international models too. Test two or three models with the same cases.

Is my data used to train the AI?
Not at Evoya: your data is processed in Switzerland or the EU and never used for AI training. With other providers, it pays to check the contract terms carefully.

Can several AI agents work together?
Yes. A main agent can hand off subtasks to specialised subagents, for example one researches, a second writes, a third checks.

Which agent pays off first for you? Book a free introductory call → or try Evoya free for 14 days

Related reading: Evoya Agents – AI agents for your company, AI agent in the glossary, AI process automation – when we build the agent for you, and our AI projects – from idea to productive solution.