
Business automation with AI agents
AI agents for business: development and integration
An agent looks up the information it needs, decides the next step, and carries out the action, with the tools and permissions assigned to it.
What is an AI agent?
An artificial intelligence (AI) agent is a program that works with a user account and a set of permissions, like any other member of the team: it checks the ERP or the CRM, decides what needs to be done, and does it. What sets it apart from a chatbot is the last part: it does not just answer, it acts. This is also known as agentic AI.
An example: a customer complaint comes in. The agent looks up the customer's orders, sees that one shipped late, opens a ticket in the customer service software, and prepares the reply email for someone to review and send.
Looks up
The data it needs, in the system where it is stored: a customer in the CRM, an order in the ERP, a price list, a document, or a previous ticket.
Decides
The next step, following the instructions it has been given: which department a request goes to, whether documentation is missing, or whether a person needs to see the case.
Executes
The operation, with the tools it is authorized to use: creating a ticket, updating a record, moving a file, generating a document, or preparing an email.
AI agent, automation, and chatbot: what each one does
All three tools are useful, but for different problems. A classic workflow (n8n, Make, or Zapier) handles "if A happens, do B" very well, when every possible case can be written as a condition. An agent is useful when the intermediate steps are not defined in advance or the information arrives unstructured and has to be interpreted before deciding. In many projects the most efficient approach is to combine the two: the AI interprets and decides, and automation reliably executes what is already fully defined.
Main goal
AI agent: Handle an entire process
Classic automation: Run a predefined workflow
Chatbot: Answer questions
How it decides
AI agent: Reasons based on the information in each case
Classic automation: Fixed rules: "if A happens, do B"
Chatbot: Generates an answer
Process steps
AI agent: Can vary with each run
Classic automation: Defined in advance, always the same
Chatbot: Does not execute steps
Unstructured information
AI agent: Interprets it and turns it into data
Classic automation: Needs data that is already structured
Chatbot: Looks it up
Takes actions
AI agent: Yes, with the authorized tools
Classic automation: Yes, the ones that have been programmed
Chatbot: Usually not
Exceptions
AI agent: Can handle them or pass them to a person
Classic automation: Each case has to be programmed
Chatbot: Not applicable
When to use it
AI agent: Processes that require interpreting and deciding
Classic automation: Stable, well-defined processes
Chatbot: Looking up documentation and information
| AI agent | Classic automation | Chatbot | |
|---|---|---|---|
| Main goal | Handle an entire process | Run a predefined workflow | Answer questions |
| How it decides | Reasons based on the information in each case | Fixed rules: "if A happens, do B" | Generates an answer |
| Process steps | Can vary with each run | Defined in advance, always the same | Does not execute steps |
| Unstructured information | Interprets it and turns it into data | Needs data that is already structured | Looks it up |
| Takes actions | Yes, with the authorized tools | Yes, the ones that have been programmed | Usually not |
| Exceptions | Can handle them or pass them to a person | Each case has to be programmed | Not applicable |
| When to use it | Processes that require interpreting and deciding | Stable, well-defined processes | Looking up documentation and information |
What tools an AI agent works with
Tools are what allow it to act
An agent can only do what it has tools for. They are provided through APIs, web services, databases, webhooks, message queues, or custom software: the same system integration work we do between business applications. When a program has no API, an intermediate layer can be developed to provide one.
An MCP server to use them from ChatGPT or Claude
MCP (Model Context Protocol) is the standard that AI agents and assistants use to connect to other programs. For companies that already have an API, or that commission one from us, we also offer an MCP server with a limited set of functions, so they can be used in natural language from their ChatGPT or Claude accounts, in the chat they already use every day. We explain it in how AI connects to the company's software.
A model capable of using tools
The agent needs an AI model that knows how to use tools. This capability is called "tool calling": instead of just writing an answer, the model asks for a specific function to be run, such as looking up an order in the ERP or creating a ticket, and receives the result to keep working.
Commercial models such as ChatGPT, Gemini, and Claude include it. There are also open-source models, which can be installed on private servers.
Agent teams for complex processes
When a process is complex, instead of one agent that does everything, we develop several that share the work and pass the result on to each other, with frameworks such as CrewAI or in-house development. Each agent has its own role and accesses only the systems it needs, so it can be reused in other processes: one for internal queries, another for writing, another for in-depth research, another for programming or designing mockups. With CrewAI, they can also cross-check the result with each other.
AI agents in practice
The agent understands the request or the data it receives and decides which operation to perform according to the permissions it has been assigned.
Our ERP, from the AI assistant
Our Odoo has an MCP server that makes it possible to work with it from AI: looking up data and performing operations in natural language, limited to the functions that were chosen to be made available.
Our ERP, with two-way communication through Telegram
A message to the agent can ask for information or create a task: the agent understands the message, decides which operation is needed, and performs it in Odoo. Alerts and notifications arrive through the same channel.
Public tender finder with Jev
An agent periodically reviews the public tenders that are published, and Jev decides whether they match the company's profile and to what percentage. For those that match, an automation downloads the tender documents, a language model summarizes them and estimates how difficult it would be to bid, and the person in charge of tenders decides whether it is worth it. See the case step by step.
Plant data in manufacturing
An agent queries machine, product, and monitoring data through MCP and, depending on the type of operation, carries out one task or another, such as choosing which alert or workflow to trigger. It has been developed and tested in our own environment, with test data, as we explain in AI in manufacturing.
A practical example with invoices and delivery notes
What an AI agent does when a case does not fit
In processes with financial or legal consequences, the system is prepared so that the AI does not have to improvise, and a person validates what matters.
Edge cases, planned in advance
Before going live, the real cases are gathered and what to do with each one is defined. If the invoices to be imported always arrive in the same format, a service such as Google Cloud's Document AI can be trained to know which area holds each piece of data, and the risk of error is low. If the documents vary a lot from one another, each case has to be studied.
What does not match is set aside, and the rest continues
If a line does not match when a delivery note or an invoice is imported, what is correct is processed and that line is left pending. An alert is sent through the agreed channel (email, SMS, or Telegram), and the document is flagged in a dashboard for a person to review. How each exception is handled is decided with the company in each project.
A final check with Jev
When the process ends, a decision AI such as Jev reviews the result against specific criteria and returns the probability of an error. If it exceeds the set threshold, the result goes to review.
Responsibility lies with a person
For actions with consequences, such as recording an invoice in the accounts or paying a supplier, a person always approves. Validating costs little, and even when AI is involved, responsibility for those actions must lie with a person or a department.
Permissions, security, and privacy of an AI agent
Least privilege
The more actions an agent can take, the more important it is to control its permissions. If it only needs to look up orders, it should not be able to delete them. We define which systems it uses, which operations it performs, under which identity, what requires confirmation, what limits it has, what gets logged, and what happens when there is an error. For example, in our SEO tool, the AI works with tiered permissions (read, propose, apply), and a person decides which proposals are applied.
What information the model receives
An agent should not receive all the available data when it only needs part of it to do the task. Before developing it, we analyze what information it will work with: customer data, emails, internal documents, financial information, contracts, or personal data.
Each provider's terms, reviewed in advance
Before recommending an AI service, we review its terms at that moment, even if we have worked with it before. We check three things: where the data is processed, whether it is used to train models, and how long it is kept.
One example is Azure Document Intelligence, which we recommend for reading invoices or payslips. It processes documents in the European region where the service is created, deletes them within 24 hours at most, does not use them for training, and runs within the company's own Azure subscription. Where documents are processed and how long they are kept is explained on Microsoft's page on data, privacy and security for Document Intelligence.
Another alternative would be Google Cloud's Document AI. We do not depend on any specific provider, so we can offer alternatives in each project. What matters is understanding how to connect the tool being built with the platform that processes the information.
When data cannot leave the company
In certain projects, a private AI infrastructure is advisable, with the models on a company server or in a private cloud with GPUs, or a hybrid architecture in which certain operations do not leave the company.
Input and output filters
Before a request reaches the main model, a fast model that only decides, such as Jev, checks whether it complies with the agreed usage rules and stops malicious instructions (prompt injection) or attempts to obtain data that is not allowed. The answer goes through the same check before it reaches the user.
Frequently asked questions about AI agents
Do agents work completely autonomously?
Is an AI agent the same as RPA?
How long does it take to build an AI agent?
How much does an AI agent cost?

Which manual process could an agent handle?
We analyze what steps the process follows today and with which programs, and point out which parts an agent can do and which ones a person should keep doing.


