711 239 085 info@gilsys.com
es ca en
Abstract background

AI automation agency based in Barcelona (Osona)

Process automation with artificial intelligence

We automate processes in which someone on the team has to read, understand, or classify information before acting, combining AI with custom programming and n8n.

Intelligent automation

AI can read and understand what always used to require manual work

AI makes it possible to analyze information without predefined patterns, and we can program custom workflows with the conditions each process needs.

Automating a process has always worked well when it can be expressed as rules: "if the order status changes to shipped, send the email." This is what we have been doing for years with automations and APIs.

The limitation was in processes that start with information that has to be read and interpreted. That step had to be taken by a person, which is why the process could only be partly automated. For example, opening each delivery note that arrives by email, finding the order number, and entering the quantities in the ERP.

AI solves that step, and the rest of the workflow is still traditional automation: reliable, repeatable, and inexpensive.

Traditional or AI?

Traditional automation: same input, same output

Not every process needs AI: the key question is whether information has to be understood and interpreted.

For example, when a message arrives through the website's contact form, the workflow is always the same: it sends an alert to the mobile phone of the person in charge, schedules a reminder in the CRM to get in touch within 48 hours, and the system automatically replies to the sender with an email confirming receipt. If nobody has replied after two days, the system alerts the assigned person again.

There is no interpretation in this workflow: the form already indicates which field each piece of data belongs to, and fixed rules are simply applied. It is the basis of custom automation and API development.

Speed

No need to wait for a model to respond.

Reliability

The same data always gives the same result.

Security

The information is not sent to any AI service.

Control

Every step can be reviewed and audited.

AI automation: when interpretation is needed

Going back to the previous example, suppose all kinds of emails arrive in the inbox: enquiries, incidents, advertising and, among them, requests for quotes, each written in its own way. A rule that searches for the word "quote" would miss some and get others wrong.

With AI, each email is read, and the ones asking for a quote are detected and recorded in the CRM with the contact, the company, and the details of the request. From there, the workflow can continue with the same rules as the form: the alert, the 48-hour reminder, and the automatic reply.

Data protection

The model has to be chosen carefully, because not all models are suitable for every kind of information. When data cannot leave the company, the option is private AI.

Cost and time

Each query to the model has a cost and takes a few seconds. As a reference, in our own SEO tool, asking the AI to propose a text costs less than half a euro cent, and each decision by Jev, a model that only classifies, about two hundredths of a cent.

Measured results

It is tested with real emails before going into production.

Supervision

Critical steps, such as sending an invoice or making a charge, can wait for a person's approval before being executed (human-in-the-loop).

Some use cases and how to solve them with automation

Sometimes integrating AI is essential; in other cases, it is not recommended.

  • Generating the PDF technical data sheet of each product, with its layout

    How it is solved: Traditional programming

    Why: The layout follows exact rules, with every piece of data in its place on every sheet, and an image-generating AI cannot guarantee that.

  • Exporting ERP data to the gestoría (the accounting firm that handles taxes), the bank, or another platform

    How it is solved: Traditional programming

    Why: An API integration or a defined exchange format gives an exact result.

  • Checking that a Spanish tax ID (NIF), an IBAN, or a date is valid

    How it is solved: Traditional programming

    Why: There are validation algorithms: the letter of a NIF is the remainder of dividing the number by 23, and an IBAN is validated with its check digits (modulo 97). AI adds nothing and could make mistakes.

  • Sending an alert when a product's stock falls below the minimum

    How it is solved: Traditional programming

    Why: It is a fixed rule: if the stock is below the minimum, the alert is sent.

  • Reading expense receipts or supplier invoices, each in its own format

    How it is solved: Specialized service and traditional programming

    Why: The format changes, but there are services specialized in reading documents, such as Azure Document Intelligence or Google Document AI.

  • Knowing whether a customer writing to technical support is happy or angry

    How it is solved: AI

    Why: It requires understanding the tone of the message, not searching for words.

  • Detecting whether a message contains the company's financial information before it goes out

    How it is solved: AI

    Why: The same information can be expressed in many ways.

  • Finding a supplier's product in the company's own catalog even if it has a different name

    How it is solved: AI

    Why: It requires comparing by meaning, not by the exact text.

  • Detecting the pages of a website that drop in Google and explaining why

    How it is solved: Both

    Why: Measuring how much each page has changed is a fixed rule; explaining why and how to improve it requires AI. This is the example explained below.

TaskHow it is solvedWhy
Generating the PDF technical data sheet of each product, with its layoutTraditional programmingThe layout follows exact rules, with every piece of data in its place on every sheet, and an image-generating AI cannot guarantee that.
Exporting ERP data to the gestoría (the accounting firm that handles taxes), the bank, or another platformTraditional programmingAn API integration or a defined exchange format gives an exact result.
Checking that a Spanish tax ID (NIF), an IBAN, or a date is validTraditional programmingThere are validation algorithms: the letter of a NIF is the remainder of dividing the number by 23, and an IBAN is validated with its check digits (modulo 97). AI adds nothing and could make mistakes.
Sending an alert when a product's stock falls below the minimumTraditional programmingIt is a fixed rule: if the stock is below the minimum, the alert is sent.
Reading expense receipts or supplier invoices, each in its own formatSpecialized service and traditional programmingThe format changes, but there are services specialized in reading documents, such as Azure Document Intelligence or Google Document AI.
Knowing whether a customer writing to technical support is happy or angryAIIt requires understanding the tone of the message, not searching for words.
Detecting whether a message contains the company's financial information before it goes outAIThe same information can be expressed in many ways.
Finding a supplier's product in the company's own catalog even if it has a different nameAIIt requires comparing by meaning, not by the exact text.
Detecting the pages of a website that drop in Google and explaining whyBothMeasuring how much each page has changed is a fixed rule; explaining why and how to improve it requires AI. This is the example explained below.

Specialized services

Confidential document analysis with Azure Document Intelligence and Google Document AI

Between OCR detection and a general-purpose AI model there is another option for reading documents: AI services trained specifically for documents.

They are AI models trained to read documents, not to hold conversations. They recognize the text, like OCR, but also the structure of the page, such as tables and checkboxes, and the meaning of each piece of data: on an invoice, the supplier, the date, the line items, the tax base, the VAT, and the total. They return each piece of data with its position in the document and a confidence level, and they do not write anything: they only extract what is there.

They are integrated with traditional programming through their API, so the rest of the workflow does not necessarily need AI.

Prebuilt models or models trained on the company's documents

There are prebuilt models for invoices or expense receipts. If the company's documents always arrive in the same format, a custom model can be trained so that it knows in which area each piece of data is.

The confidence of each piece of data decides what a person reviews

The workflow could record the high-confidence data in the ERP and flag the doubtful data in a dashboard, with an alert, for a person to review.

Secure processing of financial data

Azure Document Intelligence and Google Document AI process documents in the region chosen, including Europe, and encrypt them in transit and during processing. Nor do they keep them: Azure deletes them after 24 hours, and Google processes them in memory or, when working in batches, keeps them for one day at most. Google also states that it does not use them to train its models. Before starting any project, we review the terms of the external services, as explained in AI agents; if the documents cannot leave the company, a model on private servers is considered.

Processes that can be automated with AI

These are some basic ideas; we can implement advanced custom automations as each project requires.

Incoming emails, messages, and files

An email, a Telegram message, or a file. The AI interprets what is being requested, extracts the relevant data, and creates a task in an ERP, for example in Odoo.

Forms and free-text messages

What a customer writes in their own words in a web form or a message becomes a record, with the information broken down into fields that can be integrated into a CRM.

Automatic summaries of emails, meeting minutes, and incidents

A long email thread, the minutes of a meeting, or the incidents opened during the last week. The AI extracts the information and sends notifications according to the defined workflow, for example an alert to the person in charge by email or Telegram.

Repetitive texts and translations

Draft replies, product descriptions, or translations of product sheets, which a person reviews and approves before they are sent or published.

Delivery notes, orders, and certificates

Each supplier sends them in its own format. A specialized service or an AI model extracts the fields, and the workflow checks them against ERP data to record them where they belong.

Spreadsheets in different formats

Excel sheets that each department or customer fills in with its own format, with columns that change name or order. The AI interprets them and a workflow loads the data into the right program or consolidates it into a shared database.

Not sure which process to automate first?

In an AI consulting engagement we analyze the company's processes and data and propose where to start, with an assessment and a roadmap.

See AI consulting

AI automation example: SEO and GEO monitoring of a website

Periodic collection of metrics

The workflow downloads data from Search Console, Google Analytics, PageSpeed Insights, Bing Webmaster Tools, and Cloudflare. From the information received, we know the state of the website, which search engine and AI crawlers read it, and the first pages of Google for certain keywords.

The effect of each change, calculated

For each page, it compares the impressions, clicks, and position before and after its last published change.

The AI interprets what has improved and what has dropped

On pages that drop by more than a set percentage, the AI reads the content block by block and analyzes the searches the page appears for and the pages ranking above it, to identify the differences with its own previous versions and with the competition.

Automatic report creation

If pages that can be improved are identified, the workflow prepares a report with the suggestions, shown in a dashboard where the marketing manager can analyze the proposals and apply them manually at their discretion.

This is a very simplified example that illustrates how our own SEO analysis and improvement tool works conceptually.

Scope of automations: channels, programs, sensors, and APIs

An automation can receive information and act on any system accessible through the company's network or the internet.

Telegram, WhatsApp, and SMS

A request that arrives via Telegram or WhatsApp can be turned into a task, and at the same time these communication channels allow automations to inform us of certain events.

Email and files

Incoming emails or files that arrive in a folder (hotfolder) can be classified and then turned into tasks, tickets, or whatever the project requires. In the other direction, the system can send emails or write files with periodic reports or reports triggered by specific events.

ERP, CRM, and internal programs

Automations can read from and write to the company's programs: create a task in an ERP such as Odoo, update a record in a CRM, or generate draft invoices at the end of the month.

Sensors and machines

Data from the machines in a plant can also trigger an automation: if one has been stopped for more than ten minutes, the maintenance manager receives an alert. It is the basis of AI for industry.

Webhooks from third-party tools

Many tools can notify other tools when certain events occur: a payment received, a form submitted, or a change in the code repository. The automation acts at that moment: for example, when the payment gateway confirms a charge, the order is marked as paid in the ERP.

Remote APIs

We can connect to remote APIs that act as external services. For example, we can automatically export to Google Sheets a list of all the tasks completed by each employee every day, broken down by project, hours, and client, so that the finance or human resources department has access to this cloud document.

At Gilsys we have also streamlined our internal management by integrating AI automations; we can recommend them from first-hand experience.

They are similar and easy to confuse

AI automation or an agent?

The practical difference is that the agent has autonomy (limited by the developer) to make decisions and carry them out.

AI automation

Workflows are decided in advance and are always based on the same conditions. The AI handles specific steps within them: reading, classifying, or summarizing. Cheaper, more predictable, and easier to check: if something fails, it can be identified. For example, each delivery note that arrives is read, checked against its order, and recorded in the ERP if the order exists, or a notification is sent if it does not match.

A delivery note arrives by email

The AI reads it

It extracts the order number and the quantities

It is checked against the order

Fixed rules

Match: it is recorded in the ERP
No match: a notification is sent

Always the same steps, in the same order

AI agents

The agent is free to decide what to do: it has access to tools and uses them according to its own judgment and the guidelines it has been given. Two similar cases can follow different steps. For example, if a customer asks the agent on Telegram about the status of an order, the agent can check the ERP, see that it is delayed, look up the carrier's tracking, and reply; when asked again, if the order has not been entered correctly, it may connect to the ERP, find the warehouse manager's contact details, and send them an email. It is important to define an agent's guidelines carefully and set its limits. At Gilsys we use it with our own Odoo: from Telegram we request information, create tasks, and receive alerts, and the agent works with the ERP through an MCP server.

A customer asks about their order

AI agent

Decides which tools to use and in what order

Tools it has access to

ERP
Carrier tracking
Contacts
Email

Within the guidelines and limits set for it

For defined workflows, we use AI automation. To give AI the freedom to solve tasks with the company's tools, an agent.

The comparison with a chatbot is in AI agents, and how an assistant is built into an application, in AI built into applications.

n8n and custom programming: how we build automations

n8n is an automation tool that can be installed on the company's own server. We use it in some of our automations and combine it with AI agents and even with custom-programmed modules. We have been programming since 2004: visual tools are useful to get started or to conceptualize, but we usually tackle it with code, as explained in why choose Gilsys.

n8n, on a company server

We install n8n on a company server, physical or in the cloud, inside a Docker container that also packages the scripts and custom code each workflow needs. Each workflow is shown as a diagram of steps, which can be more or less complex depending on each automation. Installed this way, the free edition of n8n is not paid per execution, unlike its cloud version.

Programming where n8n falls short

Complex workflows, with many conditions, are custom-programmed, without a graphical interface, to control every detail. And when a program has no connector or a step needs its own logic, we program it: a code node, a custom API, or an intermediate layer that connects to a legacy system, through system integration.

The AI model, depending on the data

We integrate commercial models such as ChatGPT, Gemini, or Claude when the information is not sensitive and, for highly sensitive data, we can integrate open-source models in a private AI setup within the company.

Debugging to find errors

Every run is logged: we can determine what came in, what the AI decided, and what was done. A visual tool makes it easier to get started, but when a workflow fails, someone has to know how to debug it: read the error, check each API's response, and fix the data or the source code. Workflows retry on their own when a service does not respond and send an alert if the error persists.

An n8n workflow that collects Telegram messages, extracts the text from PDFs, and replies with an AI agent
One of our internal workflows in n8n: it collects Telegram messages, extracts the text from incoming PDFs, and an AI agent, with an Anthropic model and tools connected via MCP, replies through the same channel.

How long does it take for automation to pay for itself?

Our calculator shows how much repetitive tasks cost today and how long it takes for a development project or an automation to pay for itself.

  • Monthly cost and 3-year projection
  • Hours freed up per year
  • Months to recover the investment
Calculate savings

Frequently asked questions about AI automation

How accurate is AI at reading documents?

It depends on the document and on the variety of formats that have to be processed. It is important to process real company documents and compare the result with what a person would produce. If it does not reach the agreed level, it is adjusted before going into production. At first, we always recommend that a person review the processing to validate its accuracy rate, and, depending on the type of automation, 100% of cases have to be validated, so that responsibility always lies with a person, not with an AI.

To integrate automations in a company, is it necessary to switch software?

Not always: it depends on the programs to be connected. If they have an API, an accessible database, or generate files, it is usually solved with custom-programmed connection modules, without modifying any existing application.

Which process should be automated first with AI?

A specific process that is repeated often and in which someone currently has to read or classify information manually, not the automation of the whole company. Once that process works and the savings have been measured, the next one follows. If it is not clear which one, AI consulting identifies it with an assessment by area. The signs that a process is worth automating, and when it does not pay off, are explained in process automation.

How long does it take to get a first automation?

A basic automation can be developed in a week, but conceptualizing it, defining the requirements, testing it, and putting it into production usually takes longer, as explained in how we work on an AI project. By basic we mean reading what comes in through a channel (an email, a Telegram message, a file), having the AI interpret it and decide, and having the workflow record it as a task in the ERP. Anything that chains several workflows together, such as deadline tracking or automatic reports, takes longer.

How much does it cost to automate a process with AI?

There are two costs: developing the workflow, which is quoted once, and model usage. As an order of magnitude, developing a basic automation costs €500 to €1,000, and an advanced one €1,000 to €5,000, depending on how many systems have to be connected and how many steps the AI handles. Those that chain several workflows or require a custom dashboard can exceed that figure. With a commercial model, usage is paid according to consumption, and with a model on private servers, it is the cost of the machine. These are approximate figures: each quote is based on an analysis of the process and is free and with no commitment.

Do you use Zapier, Make, or Power Automate?

We have used them and, for connecting two tools in a few steps, they are practical. In processes with many steps or exceptions they become cumbersome to maintain, and with n8n on the company's own server or with our own scripts there is more control and more independence from the platform, which is what we usually do.

Does the automation have to be rebuilt if a better AI model comes out?

No. The workflows we build make it easy to switch models, connecting, for example, to a model gateway such as LiteLLM instead of to a specific provider, so switching models is a configuration change, not a programming one.
Abstract background

Which manual task can be automated with AI?

With a real example, such as an email, a document, or a spreadsheet that is processed manually today, we can see which part AI can do.