On this page
- 01An AI agent in plain language
- 02Chatbot, workflow or agent: the difference
- 03Three agents a small business can have built
- 04Worked example: what an email agent costs per month
- 05Where agents still stumble
- 06How an agent talks to your tools
- 07What the EU AI Act asks of an agent
- 08How to start with a first agent
An AI agent is software that gets a goal and picks its own steps to reach it: it plans, uses tools such as your mailbox or CRM, looks at what comes back and decides the next step. A fixed workflow always runs the same steps in the same order; a chatbot answers a question and stops. An agent keeps working until the task is done and adapts to what it meets along the way. For a small business an agent works best when it prepares the work and lets a person decide where it matters. The model usually costs cents per task; the build and the agreements on permissions and checks are the real work.
How we build them: AI automation and AI for business.
An AI agent in plain language
Think of the difference between a recipe and a cook. A workflow is a recipe: step one, step two, step three, and if an ingredient is missing, it goes wrong. An agent is a cook who knows what has to be on the table and checks the cupboard himself.
In practice an agent runs in a loop:
- A goal. "Answer the questions that came into info@ today."
- A plan. Which emails are there, what does each sender ask, what information do I need?
- Use a tool. Read the mailbox, look up an order in your system, open your price list.
- Check the result. Do I have the answer, or do I need to look something else up?
- The next step. Prepare a reply, or go back round the loop.
And where it matters, there is a person. A reply to a customer, a quote, a payment: the agent prepares it, you approve it. That is not a brake on the technology; it is how you put an agent safely into a business.
AI agent
A loop around a goal. It picks the next step itself.
The goal
- Plan
- Use a tool
- Check the result
- Next step
Fixed workflow
A straight line. The same steps every time.
- Form comes in
- Contact in your CRM
- Confirmation email
- Alert to the team
- Done
If the request differs, it stops or does the wrong thing.
Chatbot, workflow or agent: the difference
| Chatbot or assistant | Fixed workflow | AI agent | |
|---|---|---|---|
| What it does | Answers a question and stops | Always follows the same steps | Picks its own steps towards a goal |
| Example | ChatGPT, a chat on your site | Form, CRM, confirmation email | Read emails, look up, prepare a reply |
| Works with your tools | Only what you give or connect | Yes, set up in advance | Yes, and chooses which tool when |
| Cost per run | Per conversation | Low and predictable | Higher and variable, per reasoning step |
| Biggest risk | A wrong answer, given confidently | Stops silently at anything unexpected | Does something nobody asked for |
| Built with | A subscription or a custom chatbot | Zapier, Make, n8n, Power Automate | A language model with tools, permissions and a log |
The lines blur. A chatbot on your site that books an appointment in your calendar is already a small agent. A workflow with one AI step that classifies an email is not yet an agent: the steps are fixed, only one decision is made by a model. What a chatbot costs: how much does a chatbot cost; which workflow tool fits: Make, Zapier or n8n.
AI agent or workflow automation
Two ways to hand off work. They do not exclude each other: many agents use a workflow as one of their tools.
- How it works
- Fixed workflowAlways the same steps in the same order.
- AI agentPicks its own steps to reach a goal.
- Strong at
- Fixed workflowPredictable work: a form, an email, an alert.
- AI agentWork that differs each time: reading emails, finding information, weighing up.
- When something unexpected happens
- Fixed workflowStops, or does the wrong thing without noticing.
- AI agentAdjusts its plan, or asks a person.
- Cost per run
- Fixed workflowLow and predictable.
- AI agentHigher and variable: every model step costs tokens.
- Control
- Fixed workflowYou know exactly what it does.
- AI agentYou decide what it may do, and a person approves where it matters.
Choose a workflow where the work is predictable. Choose an agent where every case differs, and keep a person at the end.
Three agents a small business can have built
These are examples of what you can have built today, not projects we have already delivered for clients.
Reading emails and preparing a reply
The agent reads what comes into your shared mailbox and recognises what it is: a quote request, a delivery question, a complaint. For a delivery question it looks up the order and writes a reply with the right date. It sends nothing: the reply is ready and someone on your team approves it with a click, or edits it. A complaint goes straight to a person.
Fetching the numbers every Monday and emailing a report
The agent pulls revenue from your accounting, new leads from your CRM and your site's visitors, compares with last week and adds a short explanation: what rose, what fell, what stands out. At eight o'clock it is in your inbox. Honestly, this is more a workflow with an AI step than a full agent, and that is fine: the numbers come from your systems, the model only writes the explanation.
Looking up a lead from your form and adding it to your CRM
Someone fills in your contact form with a company name. The agent looks the company up, summarises what it does, puts the contact into your CRM with those details and suggests a first follow-up. Your salesperson decides whether and how to call.
Worked example: what an email agent costs per month
A small business gets 40 emails per working day in info@, or 880 a month. Per email the agent takes five reasoning steps averaging 4,000 tokens in and 400 out: read, classify, look up, write, check. That is 20,000 tokens in and 2,000 out per email. Model prices per million tokens in dollars, converted at 1.12 dollars to the euro.
| Per month, 880 emails | Claude Haiku 5.5 | Claude Sonnet 5.5 | GPT-6 Astra |
|---|---|---|---|
| Price per million tokens (in / out) | 0.10 / 0.50 dollars | 2 / 10 dollars | 10 / 50 dollars |
| Tokens in (17.6 million) | 1.76 dollars | 35.20 dollars | 176 dollars |
| Tokens out (1.76 million) | 0.88 dollars | 17.60 dollars | 88 dollars |
| Model cost per month | 2.64 dollars, about 2.40 euros | 52.80 dollars, about 47 euros | 264 dollars, about 236 euros |
| Time saved (3 minutes per email, 40 euros an hour) | 44 hours, 1,760 euros | 44 hours, 1,760 euros | 44 hours, 1,760 euros |
So the model is rarely the big cost. A smaller model for classifying and a larger one for writing keeps the bill low. What does count is the build: connecting your mailbox and systems, permissions, a log and a screen where your team approves. Say that is 60 hours of work at 85 euros an hour, inside the Belgian market range; then the build costs 5,100 euros and in this example pays for itself in about three months. Put in your own wage cost and build rate: the sum works the same way in any market. Work out your own volume with the AI cost calculator. The time saved is an example; measure your own before you decide.
Where agents still stumble
Agents make mistakes, like people, but differently. They can misread a question, get stuck in a loop or do something wrong with great confidence. Every reasoning step costs money too: an agent that looks things up five times costs more than a workflow that calculates once.
- Too many permissionsAn agent allowed to send emails can also send the wrong one.
- No logWithout a trail, nobody knows why it did what it did.
- Invented factsA model fills gaps with something that sounds like a fact. Give it your real prices, terms and procedures.
- Instructions hidden in the inputAn email can contain text that makes the agent do something other than you intended. That is why a person approves everything that goes out.
- Costs that run awayAn agent stuck in a loop keeps using tokens. Cap the number of steps per task.
So you start small. One agent with one task, only the permissions it needs, a person who approves before anything goes out, and a log in which you see what it did. If the work is the same every time, you do not need an agent: a workflow is cheaper and more predictable.
How an agent talks to your tools
An agent works only with the tools you give it: your mailbox, calendar, CRM, accounting. Those connections run through each system's API, or through MCP, an open standard that lets a language model use tools. More and more software offers such an MCP connection, automation platforms such as Make included. For you that means: the more of your systems have a connection, the less custom work an agent needs. If your work runs in an old package without one, or in Excel, it often starts with Excel to web app.
What the EU AI Act asks of an agent
An agent that talks to customers in the EU must, since 2 August 2026, make clear that they are dealing with AI, unless that is already obvious. And since 2 February 2025 you take measures to support the AI literacy of the people who work with AI. An agent that helps your team answer emails is usually not high-risk; an agent that helps decide on staff or credit can be. Read more in the EU AI Act for business and our news item on EU AI Act transparency obligations.
How to start with a first agent
- Pick a task that comes back every day and is slightly different each time. Answering emails, sorting requests, checking documents.
- Write down what a good result is. Ten real examples of how your team does it now.
- Give the agent read-only access and let it prepare proposals, never send them itself.
- Measure for six weeks: how many proposals does your team approve unchanged, how much time does it save?
- Only then extend: more tasks, or more autonomy where the risk is low.
How to set up such a pilot: AI implementation in your business. Which kind of AI fits which task: AI tools for business.
An agent is as good as its limits
Give an agent access only to what it needs. Let it send, pay or delete nothing without approval, and keep a log of what it did. If it works with customers, tell them clearly they are dealing with AI.
Questions about AI agents
What does AI agent mean?
An AI agent is software with a language model that gets a goal and picks its own steps to reach it, with access to tools such as your email, calendar or CRM. Agent here means: someone who acts on your behalf.
What is the difference between an AI agent and a chatbot?
A chatbot answers what you ask; an agent carries out a task. A chatbot can be part of an agent, for example when the agent prepares a reply a customer reads in a chat.
What is the difference between an AI agent and workflow automation?
A workflow always follows the same fixed steps; an agent picks its own. Workflows are cheaper and more predictable; agents cope better with work that differs each time.
How much does an AI agent cost?
The build is the biggest part and depends on the connections. The language model usually costs cents per task: in our example of 880 emails a month, between about 2 and 236 euros a month depending on the model. An agent that looks things up a lot costs more per task than a workflow.
Is an AI agent safe for my business?
Yes, if you give it only the permissions it needs, have a person approve before anything goes out and keep track of what it does. Without those limits every mistake becomes a mistake at scale.
Can I build an AI agent myself?
For simple tasks, yes, with the agent features in Make, n8n or Zapier, or the assistants in ChatGPT and Copilot. If the agent has to work with your own systems and sensitive data, with permissions per person and a log, it is better to have it built.