+39 348 9323674 +39 011 9222277 info@webcreation.it
Italiano Tell us about your projectGet a quote

Email·Orders·Invoices·ERP

AI agents for business
and process automation

Describe your process

Real case

We put our first AI agent to work in our own business.
It is the one answering you on this page.

On webcreation.it requests came in at all hours, and anyone writing in the evening or on the weekend waited until the next day. In the meantime many of them looked elsewhere.

We put to work an AI agent that knows our services and our price list, answers questions, understands what you need and collects the number of anyone who wants to be called back. Ten minutes after the last message the full conversation lands in our email, ready for the phone call.

  • Evenings and weekendsanswers outside office hours too
  • 10 minutesfrom the last message to the conversation in our email
  • Real pricesin the answers, taken from our price list
In-house project webcreation.it Try it now

AI agents, what they are
and what they do in a business.

AI agents are programs that use a language model to understand a task, decide the steps and use the company's tools, such as the ERP, email, databases and the online store, to get it done. When a case is doubtful they ask a person for confirmation. The difference from a chatbot comes down to this: the agent does not just answer, it does the work.

Almost every company has a process that absorbs hours every day without adding value: orders arriving as PDFs that have to be retyped into the ERP, emails to read and sort, documents to check, reports to compile by hand copying from three different programs.

An AI agent does this work in place of people. Unlike a normal rule-based automation, it understands texts written in different ways, documents with varying layouts and ambiguous requests; unlike a chatbot, it does not just answer but carries out actions in your systems: it reads, extracts, compares, fills in, enters, alerts.

We design it with a precise boundary: what it can read, what it can write and what it must submit to a person before acting. Automating a process starts at €8,000 and takes 1-2 months; for smaller automations, hour packages are enough. When you only need to answer customers, a business chatbot is enough; to connect Claude or ChatGPT to your data there is the MCP server. If the goal is to help colleagues find information, the way to go is a ChatGPT for business on your documents. All the services are on the AI for business page.

AI agent automating a business process

Processes you can automate

Examples of AI agents
for business process automation

These AI agent examples are the automations that bring the fastest return, because the processes are frequent, repetitive and today take up skilled people. For specific sectors there are pages on AI for accountants, AI for lawyers and AI for e-commerce.

Orders · Confirmations · Delivery notes

Customer orders into the ERP

The agent reads orders received by email or as PDFs, recognizes the customer and the items even when codes are written wrong, checks prices and availability and creates the order in the ERP. Doubtful lines get flagged instead of guessed.

Let's talk

Supplier invoices · Documents

Documents to record and check

It extracts data from non-electronic invoices, delivery slips, certificates and forms, even scanned or photographed, combining OCR and AI. It compares them with orders and price lists and prepares the entry. It reads Italian XML e-invoices directly, without OCR.

Let's talk

Mailbox · Requests · Tickets

Emails to sort and prepare

It classifies incoming emails by type and urgency, assigns them to the right person and prepares a draft reply with the customer and order data. Whoever answers reads it over and sends.

Let's talk

Reports · Analysis · Proposals

Recurring reports and analysis

It gathers data from the online store, the ERP and advertising platforms, compares it with the previous period and writes a report with the anomalies and the proposed actions, ready every Monday morning.

Let's talk

How an agent works

Clear permissions, tracked
actions, a person who approves.

An AI agent that writes into your systems has to be designed like a new hire: you give it access only to what it needs and check its work until you trust it.

Architecture of an AI agent with human approval

Minimum access to systems

Each agent has its own credentials and limited permissions: it reads only the folders and tables it needs and writes only where required. The AI provider keys stay on the server, never in the browser.

Approval for important actions

Entering an order, sending an email to a customer, changing a price: actions with an impact go through a human confirmation. When the agent proves reliable, it is given more autonomy.

A log of everything it does

Every step is logged: what it read, what it decided, what it wrote and with what outcome. If something goes wrong it can be reconstructed from the log, and errors become new rules.

A concrete example

Copilot, not autopilot.

For WooCommerce we are developing a plugin that applies this approach to online advertising.

Google Ads · Meta · WooCommerce

An agent that analyzes campaigns and proposes changes

It reads the data from the Google Ads and Meta campaigns connected to the online store, cross-checks it with sales and margins, spots waste and opportunities and proposes changes one by one. No change goes live without the approval of whoever manages the budget.

Let's talk →

How it works

Intelligent agents,
how an AI agent works

The concept of an intelligent agent comes from AI textbooks: a system that perceives its environment, decides and acts to reach a goal. The textbooks distinguish reflex agents, model-based agents, goal-based agents, utility-based agents and learning agents. Today's AI agents put a language model at the center and give it tools to act within company systems. These are the pieces we design every time.

  • Goal and instructions

    What it has to achieve, with which rules and in which cases it must stop. This is the part we write together with the people who do the work today, because it contains the checks nobody ever put in writing.

  • The language model

    It reads texts and documents, understands the request and decides the next step. For routing, small and inexpensive models are enough; for the hard cases we use more powerful models.

  • The tools

    The actions the agent can take: look up a customer, read an order, create a draft in the ERP, send an alert. We connect them with APIs, custom connectors or an MCP server.

  • Memory and context

    The data of the case being worked on, the customer's history, the reference documents. The agent sees only what the task needs, not the whole archive.

  • Control and logging

    Thresholds, human confirmations and a log of every step. It must be possible to check, correct and stop an agent in production at any moment.

  • Several agents together, only when needed

    In long processes one agent routes, another extracts the data, a third checks. We use them when a single agent is not enough, because every extra agent adds costs and points to monitor.

Building an AI agent

Build an AI agent yourself
or have it built?

With platforms such as n8n or Make, building AI agents for a simple task is within reach of anyone with a bit of practice. The difficulties come with real data, exceptions and security.

You can build it yourself if

  • The task concerns one person and their tools.
  • The data is limited and not confidential.
  • An error is visible right away and can be fixed by hand.
  • You have time to maintain it when the connected programs change.

It is better to have it built if

  • The agent writes into the ERP or the online store, or sends emails to customers.
  • It works on personal data, invoices or contracts.
  • It has to cope with high volumes and documents that differ widely.
  • You need an action log and someone who responds when something stops.

Often the best route is a mix: we build the delicate part and leave you a workflow you can change on your own.

How we build it

How we build
an AI agent for your business

The hard part of an agent is not the model but the edge cases of your process. That is why we always start from real documents and examples.

  • Mapping the real process

    We reconstruct the process with the people who do it every day: steps, exceptions, checks that today exist only in people's heads. The result is the specification the agent has to follow.

  • Testing on a sample of documents

    Before connecting any system we have the agent work on a sample of your real documents and measure how many cases it handles correctly and how many it flags.

  • Integration with your tools

    ERP, online store, mailbox, shared folders, spreadsheets. We use the available APIs, an MCP server or a custom-written connector when the system has no API.

  • The right orchestration tools

    For simple workflows we use automation platforms such as n8n or Make; for complex logic, sensitive data or high volumes we write the agent's code, with different models for different tasks.

  • Safeguards and edge cases

    Checks on the extracted data, thresholds beyond which a person is needed, defenses against instructions hidden in documents and emails. An agent has to stop when it is not sure.

  • Gradual launch and measurement

    It starts alongside your people, with the agent proposing and the person confirming. We measure errors and time saved and widen its autonomy only on the cases that prove reliable.

Before you automate

AI agent or rule-based
automation?

Not every process needs artificial intelligence. In process automation, if a fixed rule is enough, it costs less and makes fewer mistakes.

You need an AI agent if

  • Documents arrive in different formats: PDFs, emails, photos, tables.
  • Requests are written in people's own words and need interpreting.
  • You need to compare information across several systems before deciding.
  • The process has many exceptions that an experienced person handles today.

A rule-based automation is enough if

  • The data always arrives in the same structured format.
  • The decision can be expressed with a simple condition, if this then that.
  • The volume is low: a person takes less time than designing the automation would.
  • The error cannot be undone and nobody can check the result.

Often the best solution is a mix: fixed rules for the standard cases and the AI agent only where understanding is needed.

Pricing

How much does an AI agent
cost to automate a process

The price is fixed after analyzing the process and testing on the documents. The quote also includes an estimate of the monthly model cost on your volumes.

Small automations

from €350

Hour packages

5 hours €350 · 10 hours €600 · 20 hours €1,100

  • A simple workflow on n8n or Make
  • Notifications and routing
  • Connecting two tools
  • Tweaks to existing automations
  • Support on agents already running
  • Night work at a higher rate
Most requested

Virtual assistant

€3,000 to €7,000

ChatGPT for business · RAG

Timeline 3-6 weeks

  • Assistant for your employees
  • Search across company documents
  • Two integrations (CRM or ERP)
  • Dedicated web interface
  • User permissions and log
  • Training for the people who use it
Custom

AI programming

€5,000 to €15,000

API · MCP server · Plugin

Timeline 6-8 weeks

  • AI inside your software
  • Connectors and MCP servers
  • AI plugins for WooCommerce
  • Data pipelines
  • Multiple coordinated models
  • Your code, documented

Optimization and tasks

from€8,000

AI agents · Workflows

Timeline 1-2 months

  • Automation of a complete process
  • Reading emails, orders and invoices
  • Data extraction from PDFs
  • Request routing
  • Integration with your ERP
  • Human approval where needed

The monthly cost depends on volumes.

Every document read and every action decided uses requests to the model. We estimate them on the test sample, choose the cheapest model that does the job well and set spending limits with the provider.

Tell us about the project →

On request, payment can be spread over installments with no interest. 10% discount for payment in advance on amounts over €3,000.

Method

From the manual process
to the agent in production

Process analysis

One or two sessions with the people who do the work, collection of sample documents, volume estimate. A written proposal with scope, fixed price and monthly costs.

Testing on documents

The agent works on a real sample without touching your systems. We measure accuracy and flagged cases, and decide together whether to go ahead.

Integration

Connection to the ERP, email and online store with minimum permissions, approval rules, an action log and alerts in case of errors.

Side by side

In the first weeks the agent proposes and a person confirms. We correct, measure the time saved and widen its autonomy where it is safe.

Let's talk

Which process takes up
the most hours every week?

Tell us how it works today: where the data comes from, which programs it ends up in, who checks it. We will tell you whether an AI agent is worth it and how much it costs. If you would rather talk it over, leave your number with "I want to be called back".

AI

AI chat · AI agents

online · instant replies
Hi, I'm Webcreation's AI assistant. Describe the process you want to automate and I'll tell you how we would build it.

Frequently asked questions

Frequently asked questions
about AI agents

What are AI agents?

They are programs based on a language model that receive a goal, decide the steps and use tools such as the ERP, email or databases to complete it. An AI agent works within defined permissions and asks a person for confirmation on important actions.

What are some examples of AI agents in a business?

The most common ones read orders received by email or as PDFs and enter them in the ERP, extract data from invoices and documents, sort emails while preparing draft replies, and compile recurring reports from the online store, the ERP and ad campaigns. An example you can try right now is the agent on this page, which answers questions about our services and passes us anyone who wants to be called back.

What is the best AI agent?

There is no single best agent, because an agent is only as good as the process it works on. OpenAI, Anthropic, Google and Microsoft offer tools to build them and platforms such as n8n or Make let you assemble them; the choice depends on the systems to connect, data confidentiality and volumes. In every project we test the models on a sample of your documents and choose the cheapest one that does the job well.

How much does an AI agent cost?

Automating a complete process starts at €8,000, with a timeline of 1-2 months. More limited AI integrations range from €5,000 to €15,000, and small automations are done with hour packages from €350. Model usage is separate, and we estimate it on your volumes.

AI agent or chatbot, what is the difference?

A chatbot converses and answers. An agent carries out a task: it reads documents, makes decisions within defined rules and acts in your systems, for example by entering an order in the ERP. A project often uses both.

Can you build an AI agent yourself?

For simple tasks, yes, with platforms such as n8n or Make and a bit of practice. When the agent writes into company systems, handles personal data or has to cope with high volumes, it is worth having it designed, with permissions, logging and checks on edge cases.

Can it connect to our ERP?

Yes, if the ERP has APIs, database access or exports. When there is no ready-made interface we write a custom connector or an MCP server with access limited to the data required.

Which data from an invoice or an email reaches the model?

Only what the task needs. We can remove or mask personal data before sending, use providers with servers in Europe, or a model installed at your company when documents must not leave.