To create AI agents today you no longer need a team of developers. An AI agent is a program built on a language model that understands a request, decides what to do, looks up the information it needs and completes the task: it answers a customer, opens a ticket, books an appointment, updates the CRM. With a no-code platform, the people who know the process, in customer care, sales or HR, can build one themselves and have it running in a few days.
Companies are moving fast. According to the Google Cloud AI Agent Trends report published in April 2026, 52% of organisations already use AI agents and 39% have more than ten in operation. The gap is with small companies: in Italy, the Politecnico di Milano AI Observatory counts at least one AI project in 71% of large enterprises and in only 8% of SMEs.
This guide explains what an AI agent is, which technologies it uses, how to create one in seven steps, with a concrete example, the use cases by department, the most common mistakes and the rules to respect.
Table of contents
- What is an AI agent?
- Why companies are creating AI agents now
- The technologies behind an AI agent
- How to create an AI agent in 7 steps
- A concrete example: an order status agent
- Use cases by department
- The most common mistakes
- Governance and the AI Act
- Where to start
- FAQ on how to create AI agents
What is an AI agent?
A traditional chatbot follows a script: if the user writes X, it answers Y. An AI agent has a goal instead of a script. It reads the request, works out what is needed, chooses which tools to use and in which order, and checks the result before replying. If a customer writes "I ordered a sofa two weeks ago and I still have no news", a chatbot shows the FAQ on delivery times. An agent asks for the order number, looks it up in the management system, sees that the shipment is stuck at the warehouse, tells the customer the new delivery date and, if the customer asks, opens a complaint.
Three things make the difference. The agent understands free-form language, because it runs on a large language model. It knows your company, because it reads your documents through a knowledge base. And it can act, because it is connected to other systems through APIs. Not every agent needs all three: an agent that only answers questions from the manuals is already useful, and you can add actions later.
If you want to go deeper into the difference, read our article on chatbots and AI agents. For a broader overview of how they are used in companies, see the page on enterprise AI agents.
Why companies are creating AI agents now
The Google Cloud AI Agent Trends 2026 study, based on 3,466 senior executives in 24 countries, gives a clear picture. 52% of organisations use AI agents, 39% have deployed more than ten, and 74% of executives report a return on investment within the first year. The return comes from simple things: fewer repetitive requests reaching operators, answers at night and at weekends, leads followed up in minutes instead of days.
In Italy the market is growing but unevenly. The Politecnico di Milano AI Observatory valued the Italian AI market at €1.8 billion, up 50% in a year, but projects are concentrated in large companies. For an SME the obstacle is rarely the technology. It is the cost and time of a development project. No-code platforms remove that obstacle: the first agent costs a few days of work by someone who knows the process, not six months of development.
The technologies behind an AI agent
You do not need to master these technologies to create an agent, but knowing what they do helps you choose a platform and understand why an agent answers well or badly.
The language model (LLM), such as GPT, Claude, Gemini or Mistral, is the engine that understands the request and writes the answer. A good platform lets you choose the model and change it later without rebuilding the agent.
The knowledge base with RAG (Retrieval-Augmented Generation) is what makes the agent know your company. You upload manuals, product sheets, policies and FAQ; when a question arrives, the system finds the relevant passages and the model answers from them instead of from its general knowledge. Our complete guide to RAG explains how it works.
Tools and integrations let the agent act. Webhooks and APIs connect it to the CRM, the ticketing system or the e-commerce back end. The Model Context Protocol (MCP), an open standard, is becoming the common way to expose these tools to an agent: you describe what the tool does once, and any compatible agent can use it.
Conversation flows handle the steps that must always happen in the same way, such as collecting data for a booking or a legal disclaimer. The agent improvises where it helps and follows a fixed path where it must.
Finally, channels and handover: the agent has to reach customers where they already are (website, WhatsApp, Messenger, phone) and pass the conversation to a human operator when needed.
How to create an AI agent in 7 steps
These are the steps we see in successful projects. The order matters more than the speed.
- Define one goal and one metric. "Answer questions about order status and reduce tickets on this topic by half" is a goal. "Improve the customer experience" is not. Pick a frequent, repetitive request with clear rules, and decide how you will measure success before you start.
- Choose the type of agent. An agent that only answers questions from documents is quick to build and low-risk. An agent that also acts, changing a booking or issuing a refund, needs integrations and stricter rules. Start with the first type even if you plan the second.
- Prepare the knowledge. Collect the documents the agent must know and clean them: remove old price lists, outdated procedures and duplicates. The quality of the answers depends on these documents more than on the model. Upload them to the knowledge base and decide who will keep them up to date.
- Write the behaviour. In the system prompt describe who the agent is, its tone, what it must answer, what it must refuse and when it must ask for help. For questions that need a fixed answer (prices, legal terms) use exact answers instead of generated ones.
- Connect tools and systems. Add only the integrations the goal requires: the order management system for order status, the calendar for bookings, the CRM for leads. Each integration is an action the agent can take, so give it the minimum permissions it needs.
- Plan the handover to people. Decide when the conversation moves to an operator: on explicit request, after two failed answers, or for sensitive topics such as complaints. Operators must see the whole conversation, so the customer does not have to repeat everything.
- Test, publish, monitor. Test with real questions taken from past tickets, including the badly written and ambiguous ones. After launch, read the transcripts every week for the first month, add missing content and correct wrong answers. An agent improves with use only if someone looks after it.
With a no-code platform such as Crafter.ai, steps 3 to 6 happen in a visual interface: you upload documents, write the prompt, draw flows and connect channels without writing code. A developer is needed only for integrations with systems that have no API or ready-made connector.
A concrete example: an order status agent
Take an online furniture shop that receives about a hundred messages a day, half of them asking "where is my order?". Each one takes an operator three or four minutes: find the order, check the courier, reply.
The team decides to create an agent for this request only. In the knowledge base it uploads the shipping policy, the delivery times by region and the returns procedure. It connects the agent to the order management system with a read-only integration, so the agent can look up an order but not change it. In the prompt it writes that the agent must always ask for the order number and the email used for the purchase, must never promise a delivery date that the system does not show, and must pass the conversation to an operator if the customer mentions damage or wants to cancel.
The conversation becomes: the customer writes, the agent asks for the order number and email, checks the status, replies with the courier, tracking link and expected date. If the order is late, it apologises and explains why. If the customer writes "the sofa arrived scratched", it passes the conversation to an operator with a summary.
After a month the team reads the transcripts and finds two gaps: questions about assembly service and about delivery to upper floors. It adds two documents. That is the whole project: one goal, one integration, clear rules and someone who reads what happens.
Use cases by department
Customer care is where most companies start. The agent answers frequent questions, checks orders and bookings, opens tickets and hands over complex cases. We discuss costs and results in the article on AI agents for customer care.
In sales and marketing, the agent qualifies website and WhatsApp contacts with a few questions, recommends the right product and books a call only for qualified leads. It works especially well on WhatsApp, where customers reply more than to email: here is our guide to the WhatsApp chatbot for business.
In HR, it pre-screens candidates, schedules interviews and answers employees on leave, expense reports and benefits, taking the answers from the company handbook.
In e-commerce and retail, it helps customers choose, tracks orders and starts returns. In IT and operations, it answers internal help desk questions, guides users through standard procedures and opens tickets with the details already filled in. In healthcare, insurance and finance, it books appointments, gives information on policies and products and handles a first triage, with a fast handover to people for anything sensitive.
If you want to see where automation has the greatest effect across the company, our guide on how to improve business processes with AI goes through it process by process.
The most common mistakes
Most failed agent projects fail for the same few reasons, and none of them is the model.
The first is too many goals at once. An agent that has to handle orders, returns, complaints, product advice and job applications from day one answers all of them badly. Start with one request and add the others when the first works.
The second is outdated documents. If the knowledge base contains last year's price list, the agent will quote last year's prices with great confidence. Someone must own the content and update it.
The third is no way to reach a person. Customers accept an agent that answers quickly. They do not accept being stuck in a loop with no way out. A visible handover to an operator is part of the design, not an extra.
The fourth is no measurement. Without a metric decided at the start, after three months nobody knows whether the agent works. Count requests handled without an operator, handovers, and the topics the agent could not answer.
Governance and the AI Act
When anyone in the company can create an agent, someone has to decide which agents go online, which data they can read and who checks their answers. Without these rules a company ends up with several agents giving different answers to the same question.
There is also the law. Under the EU AI Act, the transparency obligations apply from August 2026: users must know they are talking to an AI system. Personal data processed by the agent are subject to GDPR. Choose a platform that hosts data in Europe, lets you show a disclosure message and keeps a record of conversations and access.
Where to start
Look at your support inbox or your tickets from last month and find the request that comes back most often. That is your first agent. Build it for that request only, measure for a month how many conversations it closes without an operator, then extend it.
If you have never built an assistant before, our guide to no-code conversational AI explains the basics. Crafter.ai is a no-code platform for creating AI agents with LLMs, RAG on company documents, visual flows, MCP integrations and handover to operators, with data hosted in Europe. You can compare the plans on the pricing page.
FAQ on how to create AI agents
How much does it cost to create an AI agent?
It depends on volume, channels and integrations. On a no-code platform the main costs are the subscription or pay-per-use fee, the time to prepare the documents and the monitoring in the first months. Agents that need custom integrations with old systems cost more because of the development involved. Our article on how much a chatbot costs breaks down every item.
How long does it take to create an AI agent?
A first agent that answers from documents takes a few days on a no-code platform. An agent connected to several company systems takes weeks, mostly because of the integrations and the testing.
Do you need technical skills to create an AI agent?
Not for most projects. Someone who knows the process can build and update the agent on a no-code platform. You need a developer for integrations with systems that have no API or ready-made connector.
What is the difference between a chatbot and an AI agent?
A chatbot follows a script and answers. An AI agent has a goal, understands free-form requests, looks up information in your documents and can act on other systems, for example opening a ticket or updating the CRM.
What is MCP and why does it matter for AI agents?
The Model Context Protocol is an open standard for connecting AI agents to external tools and data. You describe a tool once (look up an order, create a ticket) and any compatible agent can use it, which makes integrations faster to build and easier to reuse.
Related links:
