AI agents in your business: what they do and how to have one built
An AI agent is software that uses a language model to carry out one well-defined task in your business software, such as reading a document, linking an email to the right file or filling in a work order from what someone dictates. Aziri builds custom AI agents for businesses, always checked by a person.

Some of our clients
What is an AI agent?
An AI agent is a program that uses a language model to complete a task on its own: it reads something that comes in, decides what needs to happen and fills in the result in your software. A chatbot answers a question; an agent does something with the answer.
This makes automation possible with data that is incomplete or unstructured, such as a customer email, a scanned weighbridge ticket or a dictated report. If a regular app, a form or an integration solves it, we do not build AI.
"Agentic AI" is the umbrella term for AI that takes steps itself instead of only answering. Aziri builds agents that each have one clear task, because a language model given three or four tasks at once often no longer completes them all.
- Something comes inAn email, a scanned ticket or a dictated report.
- The agent reads and decidesWhich customer, which file, which fields.
- A proposal in your softwareFilled in, with what it is unsure about.
- A person confirmsOnly then does it go to a customer or the books.
What does an AI agent consist of?
An AI agent consists of five parts. The language model is only one of them; the rest determines whether the agent is usable in your business.
| Part | What it does | Example |
|---|---|---|
| Language model | Reads and understands text, images or speech and proposes an answer or action. | Claude (Anthropic), OpenAI, Gemini (Google) or Mistral, through their business services. |
| Integrations | Give the agent access to your software, so it can look up and fill in data. | Looking up a customer and item in your ERP, creating a task in your application. |
| Instructions and limits | Set out what the agent may do, when it has to stop and what it never does. | Drafting a quote, yes; emailing a customer itself, never. |
| Context from your business | Your own items, customers, agreements and procedures, so the agent works the way you do. | The price list, the site details or the procedure for extra work. |
| Review screen | Shows the agent’s proposal, with what it is unsure about, so an employee confirms. | A red flag on a field with low certainty. |
What does an AI agent do in a company?
In a company, an AI agent mainly takes over reading and retyping. A short summary per use; the page behind it tells you more.
Dictating a work order or site visit
The technician or owner says what he did or saw, and the app puts it on the work order or in the file.
Reading documents
Floor plans, delivery notes, weighbridge tickets or order emails: the agent pre-fills the fields and shows how sure it is.
From request to draft quote
An agent turns a request into a draft quote with your own items, and translates it into French or Dutch.
An assistant that searches your own knowledge
Questions about customers, agreements or procedures answered from your own emails, reports and records, with the source included.
Handling email
The agent reads an incoming email, recognises customer, site or transport and prepares a proposal in your application. Sending to customers is left to people.
All uses by industry
Construction, installation, transport, wholesale, manufacturing, rental and office work: what AI can do in your business.
What is the difference between an AI agent, a chatbot and an n8n flow?
A chatbot talks, an automation flow follows fixed rules, and an AI agent reads what has no fixed form and then carries out a task.
| What it does | Works well with | Choose it when | |
|---|---|---|---|
| Chatbot | Answers questions in a conversation. | Frequently asked questions, explanations, texts. | People mainly want to know something. |
| Automation flow (n8n, Zapier, Make) | Moves data from one package to another, following fixed rules. | Data that always has the same form. | The rule "if X, then Y" always holds. |
| RPA (software robot) | Mimics clicking and typing in screens. | Old software without an integration. | No integration exists and the screen does not change. |
| AI agent | Reads an email, ticket or report, decides which fields or tasks are needed and fills them in. | Varying formats, free text, speech. | Someone now has to read or listen to do the work. |
In practice we combine them: a flow fetches the email, the agent reads it and your application shows the proposal.
A standalone agent or an agent in your own software?
A standalone agent
Enough for one task with little impact if it goes wrong now and then. Online you will find agents and automations you can have set up for a small fixed fee, and for such a task that may be enough.
An agent in your own software
- A screen to review proposals
- Rights per user
- An integration with your ERP or accounting, such as Exact Online
- Maintenance when a language model or package changes
The code and the data are yours. That is more work than a standalone flow, and so a different price.
With us, an agent proposes and a person confirms. It fills in fields or creates a task or draft, but sends nothing to customers itself and books nothing without approval. Only where it is very sure and a mistake has no consequences does the app accept its proposal straight away.
Which task do you choose for your first AI agent?
For your first AI agent, choose a task that comes back often, takes noticeable time now and that someone can check. A good first task meets these five conditions:
- It comes back often every week, such as order emails, work orders or delivery notes.
- Someone now has to read, listen or retype to complete it.
- The data is reachable: the mailbox, the folder of scans or the ERP allows an integration.
- You know what a good result is: which fields have to be filled in and what is correct.
- A mistake can be seen and corrected before it reaches a customer or the books.
Retyping from emails, tickets or reports usually meets all five. Measure in advance how long it takes now, so you know afterwards what the agent delivered.
How do you keep an AI agent under control?
An AI agent stays under control if it has one task, only makes proposals and a person decides where a mistake has consequences. At Aziri we build it like this:
- One task per agent, with instructions that say what it may not do.
- The user’s rights: the agent sees and changes no more than the employee it works for.
- Certainty visible: if the agent is in doubt, the app shows it, and it would rather leave a field empty than make something up.
- It does not overwrite what someone filled in themselves.
- Testing on real examples from your business before it goes live, and again when a language model changes.
If an agent works directly with customers, you make clear that AI is involved. For the uses on this page, the European AI Act asks little more than that; more in AI Act for businesses.
What does it cost to have an AI agent built?
Having an AI agent built by Aziri costs from €2,000. The price depends on what has to be built around the agent: the review screen, the integrations with your ERP or mailbox, and how many kinds of documents or emails it has to know.
On top of that comes language model usage, a third-party cost that is very small per document or email. We estimate it per month in advance. If you do not yet know where AI adds value for you, we look at that in a conversation or in a digital scan of your processes.
When is an AI agent not a good idea?
When the process itself is not yet settled
AI speeds up a process, but it does not resolve confusion. If nobody knows who should follow up a request, we sort that out first in a regular application.
When your documents always look the same
If an order form always comes from the same supplier in the same format, classic text recognition or an integration is enough. That is cheaper and more predictable than a language model.
When a mistake cannot be checked
An agent can get things wrong. Where a mistake goes straight to the customer or into the books, we always build in a check by a person, and we do not let an agent send emails itself.
When the volume is too small
Retyping ten invoices a month costs less than building an agent. It only pays off when someone spends hours on it every week.
How does Aziri build an AI agent?
Aziri builds an AI agent in four steps, as part of the custom software we make. A first working version is usually ready within 2 to 6 weeks.

Mapping the work
What comes in, who reads it, what is retyped and where. We measure how long it takes now.
Testing on real examples
We run the agent on emails, documents or recordings from your business, because only then do you see how often it is right.
Building it into your application
The agent goes into the software your people work in, with a review screen and integrations with your ERP or accounting.
Live and follow-up
An employee confirms the proposals. If a language model changes, we test again on the same examples.
Where does an AI agent fit in your software?
An AI agent fits in the application where the work already happens. An agent that helps fill in work orders belongs in a digital work order; an agent that reads transport documents belongs in transport planning software. How such a project runs is explained under custom software and our approach. All examples by industry are on AI in your business.
Frequently asked questions
Question not listed? Ask us directly.
What is an AI agent?
An AI agent is software that uses a language model to complete a well-defined task on its own: it reads an email or a scanned ticket, for example, decides what needs to happen and fills in the result in your software. At Aziri an agent always makes a proposal that a person confirms.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation. An AI agent carries out a task in your software, such as reading a document and filling in the fields, or linking an email to the right site. Often a user does not even notice there is a language model behind it.
What is agentic AI?
Agentic AI is the umbrella term for AI that takes steps itself instead of only answering: looking things up, deciding and filling something in in another system. An AI agent is one concrete application of it.
What does it cost to have an AI agent built?
From €2,000, depending on what has to be built around the agent: the screen to review proposals, the integrations and the number of kinds of documents or emails. The usage cost of AI services is small and we estimate it in advance, based on your volumes. Preferably you get one fixed price for the build.
Can we build AI agents ourselves?
For a simple task, yes, with tools such as n8n, Make or Microsoft Copilot Studio. If the agent has to work inside your own software, with rights per user, a review screen and integrations with your ERP, custom work is usually needed. Even then your team can build along while we review the structure and the code: build it yourself or together.
Why one task per agent?
A language model given three or four tasks at once often no longer completes them all. Small agents with one goal are more reliable, cheaper to run and easier to check.
How quickly is a first AI agent live?
A first working version is usually ready within 2 to 6 weeks, as part of the application. We first test the agent on real documents or emails from your business, because only then do you see how often it is right.
What if the language model changes or becomes more expensive?
We build so that we can switch models without rewriting your application. The model can be set per task. If a model changes, we test again on the same examples.
Can an AI agent work with my existing software?
Yes, if that software allows an integration. The agent reads emails or documents, for example, and puts the result in your ERP, your accounting or the application we build for you. We have connected to 70 systems so far, such as Exact Online, Odoo and SAP.
How reliable is an AI agent?
That depends on the task and the documents, so we test it on real examples from your business before the agent goes live. Where a mistake has consequences, an employee approves the result. When reading documents, we show how sure the agent is, so you know where to look.
What happens to our data?
Your data is stored in your own application. For processing, only what is needed goes to a model provider’s business service, which does not use it to train models. The code and the data are yours. Which language model we use and where it runs is discussed during the analysis, so you know where your data is processed. With the business services of the large model providers, your data is not used to train their models by default.
Discuss what an AI agent can do for you
Tell us where things get stuck. We will tell you honestly whether AI adds anything there, or whether a regular app or integration will do.
We reply quickly.
