A textile factory in Famalicão with 120 employees spends 14 hours a week on processes that an algorithm could execute in 14 minutes. Nobody knows this because nobody measured it. AI will not cost less; it will cost differently.
Five years ago, when the topic was "AI in industry", the conversation revolved around robots and computer vision — the distant and very expensive future. Today, the reality is more modest and, therefore, more dangerous for those who delay: intelligent automation already pays off in administrative tasks, production planning and exception management. It is not glamorous. It does not appear on conference slides. But it pays.
What changed was not the technology. It was access. A trained language model is no longer the property of AI laboratories and has moved to accessible APIs. A webhook connecting the ERP to an automatic decision agent now costs in the hundreds, not tens of thousands. And, most importantly: the learning curve has flattened. An IT director who knows their ERP can instruct an agent to do what the accounting team does manually.
What stopped working: the "tomorrow" argument
For years, the conversation about AI in industrial SMEs was postponed with the argument that "the time was not yet right". The technology was expensive, immature, and required rare specialists. That argument has disappeared — not because the technology became perfect, but because it became good enough. Good enough means: more reliable than a manual process subject to human error, less expensive than keeping a person doing the same thing eight hours a day, and implementable in weeks, not years.
The uncomfortable truth is that most Portuguese companies that "have not yet implemented AI" are not waiting for better technology. They are waiting for internal permission to change mindset. A CEO who sees the shop floor as a place where people do things has difficulty conceiving that a machine decides whether an order is a priority or whether a timesheet has an irregularity. This resistance is legitimate — but it has an invisible cost.
The numbers reveal the reality: according to INE in 2025, only 18.2% of medium-sized companies (50-249 employees) in Portugal use artificial intelligence technologies, while in small companies (10-49) the percentage drops to 9.4%. Eurostat documents that the main obstacle to adoption is not the cost of the technology — it is the lack of internal skills (70.9% of EU companies cite this factor) and legal uncertainty (52.5%). Both are problems that can be solved. The real problem is psychological: the fear of not knowing what you are doing.
AI will not cost less. It will cost differently — in machines, not in people. The uncomfortable part is that it is easier to dismiss a machine than a person.
Where AI already pays: the production planning that gives no warning
Take a garment factory. It has 15 orders in progress, each with 3-5 variants (size, colour, fit). The planner — often a person with 20 years of experience and no formal training — looks at the backlog, considers the state of each machine, and decides the sequence. If a machine goes down, they replan by hand. If an urgent order from a large customer comes up, they interrupt and resequence.
An algorithm trained on 18 months of that factory's history does the same. But it does so without fatigue, without favouritism, without forgetting any constraint. And, most importantly: it does so in 90 seconds, not in 40 minutes. This does not mean the planner leaves. It means they stop spending 3 hours a day on manual planning and start spending 30 minutes on validation and exceptions. What do they do with the other 2.5 hours? That is a question most CEOs do not want to ask. Because the answer is: "I don't know". And that is frightening.
But there are companies that have already asked the question. A construction materials distributor in Porto, with 45 employees, implemented an automatic stock reordering agent in 8 weeks. The machine monitors consumption, supplier lead times, and triggers orders when the reorder point is reached. Previously, an employee did this manually, with errors of 12-15% (stock-outs or over-stock). Now, the error rate has dropped to 2%. The time the employee gained? They started planning more efficient delivery routes. The cost? Around €8 thousand in development and API. The ROI? Recovered in 4 months.
What AI does not do: it does not replace decision-making, it amplifies it
There is a very common mistake in the conversation about automation: thinking that AI takes decisions away. It does not. It takes away the repetitive work that preceded the decision. Take the approval of overtime in a factory with 180 employees. The HR manager receives 40-60 overtime records per week. Each one requires validation: did this person really work those hours? Is that level of overtime authorised? Is there anything abnormal? If 90% of cases are routine (yes, yes, no), an AI agent can approve those 90% automatically and leave the 10% exceptional ones for human decision. Result: what was 3 hours of manual work becomes 15 minutes of reviewing exceptions.
But — and this is critical — the agent needs rules. Who writes the rules? Who defines what is "abnormal"? Who has the authority to approve? These questions force the company to document what until now was tacit. Often, when we get here, we discover that the process was more chaotic than thought. That is uncomfortable. But it is also the starting point for improvement.
Five years ago, we said that AI would "solve everything". We were wrong. AI solves what is already well documented and what is repetitive. If your overtime approval process is a black box where "the boss knows when to approve", AI does not help. If your production planning is based on "what the planner thinks", AI does not help. AI only works when there are data and rules. Hence, true intelligent automation always starts with an audit of the process — and that audit is where 60% of companies fail.
The real cost: it is not the technology, it is the learning time
A company decides to automate the credit note approval cycle. There are 200 documents per month, each passing through 3 validations (accounting, sales, management). Today, it takes 8 days on average. With an AI agent that knows the business rules, it can be 2 days. How much does it cost? The technology, perhaps €3 thousand. The time of an IT director or consultant documenting the rules, training the agent and validating the results? Perhaps 120 hours. If the hourly rate is €80, that is €9.6 thousand. The real cost is not the machine. It is the work that precedes the machine.
Here is the point nobody says: most Portuguese medium-sized companies have no problem spending €3 thousand. They have a problem freeing up 120 hours of an IT director who is already at 110% with maintenance and support. That is the real bottleneck. It is not technological. It is organisational. That is why intelligent automation only pays off when there is someone to govern it. Not an AI director (it is still early for that in SMEs). But someone who, every week, looks at the agent, sees whether it is working as expected, and calibrates the rules. If that someone does not exist, AI ends up in a folder, and the company goes back to doing everything manually because "it didn't work".
AI does not fail because it is weak. It fails because nobody governs it after it is switched on.
When it starts to pay: the calculation that is missing
Ask your CFO: how much, in person-hours, does your expense approval process cost? How much does manual production planning cost? How much does managing exceptions for orders that arrive late cost? Most CFOs do not know the number. They know the cost of the payroll, but they do not know how much of that cost is work that a machine could do. When they finally calculate it, the answer is shocking. A factory with 100 employees may be spending €180 thousand a year on administrative work that is 80% repetitive. If AI can eliminate 60% of that work, that is €108 thousand. The investment in intelligent automation, done well, costs €25-40 thousand. It pays off in 3-4 months.
But — and here is the trap — those €108 thousand are not "savings". They are hours the person gains to do something else. If the company has nothing else for that person to do, then it has a problem: either it reduces headcount (which has a political and social cost), or it redeploys (which has a training cost). Both options are uncomfortable. That is why many companies prefer to keep the status quo: "AI doesn't pay off because we have nowhere to put the people". That is a legitimate decision. But it is a decision that must be conscious and documented. Because, if they do not make it, a competitor who does will have 15% more operational efficiency. And that, in terms of margin, is significant.
What changed in 2025: governance is no longer optional
Until 2024, implementing AI in an SME was almost a hobby — "let's try it, we'll see if it works". Today, with the MULTI ERP and integration platforms such as MULTI Connect, intelligent automation has become an investment that requires formal governance. The EU AI Regulation (AI Act, Regulation (EU) 2024/1689) has been in force since August 2024; the prohibitions have applied since February 2025 and the general-purpose model rules since August 2025.
What does this mean in practice? If your AI makes decisions about people (approval of hours, selection of candidates, performance evaluation), or if it processes personal data, the company has obligations of transparency, audit and documentation. It is not unnecessary bureaucracy — it is protection. Protection against biased decisions, against discrimination, against errors that nobody can explain.
A company that implements an AI agent without documenting the rules, without testing for bias, without keeping a record of decisions, is running a regulatory risk. If a person says "I was discriminated against by the machine", the company has to be able to explain why. If it cannot, the cost is not the CNPD fine — it is the lost trust.
Therefore, in 2025, the conversation about AI in SMEs cannot be "will it pay off?" but "how do we implement this in a safe and auditable way?". That change in mindset is the real turning point. Because when the company starts thinking about security and compliance, it also starts thinking about processes. And when it thinks about processes, it discovers where AI really pays off.
The next step: start small, measure everything
Do not start with "let's automate production planning". Start with something smaller, measurable, and without high regulatory risk. An agent that validates duplicate invoices. An agent that classifies complaint emails by priority. An agent that triggers alerts when the stock of a critical material falls below the limit. Something that, if it fails, causes inconvenience but not catastrophe.
Implement it. Let it run for 4-6 weeks. Measure: how many hours did it save? How many errors did it make? How much did it cost to maintain? Then, and only then, scale up to more complex processes. That discipline — of starting small, measuring, and scaling — is what separates the implementations that work from those that turn into a forgotten folder.
And, above all, appoint someone responsible. They do not need to be an AI specialist. They need to be someone who knows the business, who is willing to learn, and who has the authority to say "this isn't right, let's fix it". Because AI is not a project that ends when it is "switched on". It is a continuous process of calibration, adjustment, and improvement. Anyone not willing to invest in that continuous process should not invest in AI.
Frequently asked questions
How long does it take to implement AI automation in an industrial SME?
According to the article, implementation is possible in weeks, not years. The example cited of a construction materials distributor in Porto implemented an automatic reordering agent in 8 weeks. The technology became good enough to be implemented quickly, unlike what happened five years ago.
What is the typical cost of implementing an AI agent for automation?
The article states that a webhook connecting the ERP to an automatic decision agent costs in the hundreds of euros, not tens of thousands. In the specific case of the Porto distributor, the development and API cost around €8 thousand, with a return on investment in 4 months.
Will AI eliminate jobs in factories?
The article does not advocate elimination, but a redistribution of tasks. A planner who spends 3 hours on manual planning starts spending 30 minutes on validation, freeing up 2.5 hours for other functions. The question is what to do with that gained time, not the loss of employment.
What is the main obstacle to AI adoption in Portuguese companies?
According to the article, the main obstacle is not the cost of the technology, but rather the lack of internal skills and legal uncertainty. However, the real problem is psychological: the fear of not knowing what you are doing and internal resistance to changing mindset.
Can AI make decisions autonomously?
No. The article clarifies that AI does not take decisions away; it takes away the repetitive work that preceded the decision. In an example of overtime approval, AI approves 90% of routine cases and leaves the 10% exceptional ones for human decision.
What is needed for AI to work in a company?
AI only works when there are well-documented data and rules. If the process is a "black box" based on intuition, AI does not help. The article states that 60% of companies fail in the initial audit of the process, which is the real starting point.
What is the difference between the cost of AI five years ago and now?
Five years ago, language models were the property of AI laboratories. Today, they are accessible via APIs. The learning curve has flattened, allowing an IT director who knows the ERP to instruct an agent without the need for rare specialists.