.svg){kind=small}
How to build AI agents: a 7-step guide with a real example
Want to know how to build AI agents that do real work in your company? Start with a clear distinction. A chatbot answers questions. An automation follows fixed if-then rules, always in the same order. An AI agent gets a goal, uses a language model to plan the next steps and works with approved tools and data to get there, such as a CRM query or generating a document.
This guide walks you through building your own AI agent in seven steps, shows which approach fits your situation and explains how our own quoting agent works day to day.
Do you really need an agent?
First check which approach fits your task:
- Fixed sequence, clear rules: AI automation with fixed workflows is usually the simpler and more stable choice.
- Questions about internal documents: An AI chatbot with RAG and source references is often enough.
- Several steps, varying inputs, access to systems: This is where an agent shines, because it decides for itself which tool it needs next.
How to build AI agents in 7 steps
1. Pick one task with a clear result
Start with exactly one recurring task whose result can be checked. "Support the sales team" is too vague. "Turn an inquiry into a draft quote" is a good brief: there is an input, a result and a person who can judge it. Write down when the task counts as done and who takes over when exceptions come up.
2. Define tools and permissions
List every action the agent should take and separate reading from writing. Give it only the permissions it needs for this one task, ideally through a dedicated account. Every tool needs a clear description: what it does, which inputs it expects and what it returns. The Model Context Protocol standardizes exactly this kind of connection, as our hands-on report on the Model Context Protocol (MCP) shows.
3. Sort out data access (RAG)
An agent is only as good as the knowledge it can access. Decide which documents, past cases or price lists it needs. For a small, stable knowledge base, a well-structured long prompt is often enough. For larger or frequently changing content, you need RAG: a search step passes only the relevant excerpts to the model for each request. Our post From long prompt to RAG explains when each approach pays off.
4. Choose how to build it
You can assemble an agent in a visual workflow tool, develop it with your own code and MCP, or configure it on a ready-made platform. What matters is what fits your data, your team and how you plan to run it. The comparison below helps you decide.
5. Add guardrails and human approval
Guardrails are rules that limit permitted actions and check inputs and outputs. Define when the agent stops or hands over to a person, for example when information is missing. A person approves actions with real consequences, such as sending a quote or changing customer data. Log every tool call so results stay traceable.
6. Test with real cases
Don't test with made-up examples. Use completed cases from your day-to-day work, including incomplete and unusual inputs. Compare the agent's results with what your team actually did, and decide beforehand which deviations are acceptable. Repeat the tests after every change to the prompt, model or tools.
7. Set up operation and monitoring
After launch, monitor errors, response times, costs and how much your team has to correct. Decide who fixes incidents and who approves updates. Version your tools so you can roll back to a working state if something breaks. Measure time saved net, meaning after subtracting the time spent on reviews and corrections.
n8n / Make vs. code with MCP vs. ready-made agent platforms
n8n / Make
- Fits when: the process is largely fixed and AI handles individual steps
- Effort: low to medium: visual editor, little code
- Control: medium: existing building blocks, custom code steps partly possible
- Data protection & hosting: n8n as cloud or self-hosted, Make as a managed service
- Typical pitfalls: workflows grow messy, error handling is missing
Code with MCP
- Fits when: the agent works across several systems, chooses tools flexibly and runs in production long term
- Effort: higher: development, testing and operation need a dev team
- Control: high: logic, tools, model choice and logging are in your hands
- Data protection & hosting: your choice: EU cloud, on-premise or your own hardware
- Typical pitfalls: starting too big, monitoring and permissions come too late
Ready-made agent platforms
- Fits when: you need an assistant quickly inside your existing software environment
- Effort: low to start: configuration instead of development
- Control: low to medium: the vendor decides on features and models
- Data protection & hosting: with the vendor: check region, data processing agreement and data flows
- Typical pitfalls: vendor lock-in, connecting your own systems can be limited
No approach is better across the board, and combining them often makes sense: a workflow tool triggers the process, an agent handles the steps that need flexible decisions. Our own agents, such as the quoting agent, access systems via MCP because they need controlled access to several tools.
Real example: how our quoting agent works
At Ahoi Kapptn!, a quoting agent has been running in production since 2025. Here is how it maps to the seven steps.
The task: Turn a project inquiry into a complete quote that can be reviewed. The result is clearly defined: a quote PDF in the CRM.
The steps:
- The agent creates a project specification.
- It compares it with past quotes.
- It estimates the effort on that basis.
- It generates the complete quote PDF and saves it in the CRM.
The tools for these steps are connected via MCP, so the agent only works with functions we have explicitly approved.
The approval: A person in sales reviews the content, pricing and commitments and approves the quote. Only then is it sent. The agent delivers the draft; the decision stays with the team.
The result: The quote PDF is in the CRM in under 60 seconds. Together with other admin workflows, such as invoice preparation, this saves around 100 hours of manual work per month. That is a figure from our own operations, not a forecast for your company.
The value doesn't come from the agent deciding everything on its own. It comes from the agent doing the groundwork and presenting the result for review at a fixed point. Toni, our AI phone assistant, follows the same principle: he answers calls, responds, routes or documents them, and uses an MCP layer for controlled access to systems. You can test Toni live at +43 677 61279177.
Common mistakes and how to avoid them
- Starting too big: An agent "for the whole sales team" can't be tested or measured. Start with one task.
- Too many permissions: Write access to every system makes every error expensive. Expand permissions only once the agent has proven itself.
- Messy knowledge: Outdated or contradictory documents lead to convincing but wrong results. Decide who keeps the sources up to date.
- No defined stop: Without a clear handover, an agent guesses when information is missing. It should ask or hand over to a person.
- Demo instead of reality: Prepared examples prove little. Real, messy cases reveal the weak spots.
- No plan for operation: Models, interfaces and data change. Without monitoring, you only notice a decline when someone complains.
Build it yourself or get help?
Building it yourself makes sense when the task stays internal, the data isn't sensitive and your team has time to learn. The AI Masterclass, taught by our CEO & Co-Founder Philipp Baldauf, covers the fundamentals. If your whole team should work confidently with AI, consider AI training for companies.
Getting help makes sense when the agent works across several systems, processes customer or employee data, needs approvals or has to run long term. In that case we build it as part of our AI agent development service and support it in operation. If you don't know yet which use case pays off, start with AI consulting.
How it works: after a free, non-binding intro call, a compact fixed-price discovery clarifies scope, timeline and budget. Development runs as a monthly team retainer with costs known in advance.
Related services
FAQ: how to build AI agents
For simple agents, yes. With workflow tools like n8n or Make, or with ready-made platforms, you can build an AI agent without writing code yourself. As soon as the agent works across several systems, processes sensitive data or has to run in production long term, you need development know-how for permissions, testing and monitoring.
In n8n, you connect a trigger, a language model and the tools the agent is allowed to use in a visual workflow. The steps in this guide apply in the same way: one clear task, minimal permissions, approvals and tests with real cases. You can run n8n in the cloud or self-host it.
It depends on the task: on interfaces, data quality, approvals and operating requirements. On top of that come ongoing costs for models, hosting and maintenance. A reliable figure is only possible once the use case is clear. In AI consulting, we work out value and effort before you invest in development.
Yes, if data flows, access rights, logging and hosting are planned from the start. EU hosting is one option. If your data has to stay inside your company network, the agent can run on local AI on your own hardware without external API calls. To see what the EU AI Act means for your use case, use our AI Act checklist. The hosting model alone does not replace a review of the specific use case.
That depends on how clearly the task is defined and which data and interfaces are available. A prototype often comes together quickly; testing, approvals and monitoring take longer. A narrowly scoped first use case keeps the path short.
As much as errors can be tolerated. An agent can usually research, cross-check and draft on its own. Sending, changing data or making payments should be approved by a person. With our quoting agent, the agent creates the PDF and sales approves it.
Start with a task whose result you can check
Bring a specific workflow to the intro call. Together, we'll check whether an agent, a workflow or a chatbot is the right fit.