Five years ago we said AI would arrive "soon" in Portuguese factories. We were wrong — not because AI did not arrive, but because the pace was slower and the impact more limited than we predicted. Today, in 2025, only 11.5% of companies in Portugal (10+ employees) use AI, and among small companies that figure drops to 9.4%. Source: INE, 2025. The gap between the conversation at conferences and the reality on the shop floor remains huge. But that is no reason to postpone the decision — on the contrary, it is a reason to take it now, while there is still time to learn without pressure. The difference between you and the 11.5% who have already moved ahead is not the technology: it is the courage to start with a real problem, not with an innovation budget.
What has changed since 2020 is not the technology. It is the cost of inaction. NIS2 requires auditing of IT suppliers. The European AI Act classifies your processes into risk categories. Your competitors — even the small ones — are starting to use AI to forecast demand, optimise stock, and measure OEE in real time. If you wait for a "best practice" standard before acting, you will find that the standard forms without you. So the question is not "when should I invest in AI". It is: "what decisions do I need to take by the end of 2025 so that in 2027 my AI strategy is not a backlog to catch up on".
AI is an architecture decision, not a software one
Many industrial companies see AI as a module you buy and install, like a chatbot or a dashboard. That view costs dearly later on. The AI that matters — the kind that changes operations — lives in the data you collect every day. If your data is scattered (ERP here, MES there, spreadsheets somewhere else, history on paper), no AI model will work. Full stop.
What you need to decide now is: what is your data architecture? Do you have an ETL that centralises production data? Do you have an API strategy that lets systems talk to each other? Do you have a data lake or, at least, a single repository of truth? If you answered "no" to two of these questions, AI is premature. What is not premature is designing that foundation now, while migration costs are still manageable.
We see this repeatedly in implementations: a textile factory in the Vale do Ave with 80 employees buys a demand forecasting model because "AI is the future". Three months later, it discovers that the historical order data sits in three different systems and no one knows which is the right one. The model fails. AI gets the blame. The blame lay with the architecture. The software investment was lost. The team's time, spent. Internal trust in AI, shaken for five years.
There is no AI without integrated data. There is no integrated data without an architecture decision. That decision has to be taken in 2025 so that AI works in 2027.
Before contacting an AI supplier, ask your CTO or IT Director three questions: (1) Do we have a single repository of operations data? (2) Can our core systems (ERP, MES, mobility) share data via API or ETL? (3) Who is responsible for the quality of that data? If the answers are vague, that is the 2025 project — not AI itself, but the foundation that supports it.
Generative AI is a shortcut that costs more than it saves
There is a legitimate fascination with corporate chatbots and generative-AI-based assistants. A production manager dreams of a system that reads procedures, regulations and failure histories, and suggests corrective actions. It is appealing. It works in demos. Then, in reality, a corporate chatbot handles 60-70% of routine questions well. For the rest, it either errs with confidence (hallucination — the model invents a plausible answer), or asks for human confirmation (and then the time saved evaporates). If your document management system is chaotic (loose files, inconsistent names, old versions in circulation), the chatbot inherits that chaos and amplifies it.
A garment maker in Famalicão implemented a chatbot to answer workers' questions about quality procedures. Six months later, it discovered that the system had automatically "corrected" references to ISO standards — because the model thought they were typos. The result was compromised compliance. The chatbot was switched off. No one wanted to use AI again for two years.
Generative AI is powerful for well-structured documentation and mature processes. If you are still struggling with disorganised files and procedures that no one follows, start by organising. Use a Document Management system with cognitive capture and workflow, not a chatbot. Then, once the foundation is solid, generative AI multiplies the value. The sequence matters more than the technology.
The metric that matters is the one you do not measure
Industrial companies in Portugal measure with rigour: output per shift, rejects, cycle times. But there is one metric that rarely appears on dashboards — the time people spend on tasks a machine should do. A warehouse supervisor spends two hours a day checking picking and resolving discrepancies. A planner spends three hours adjusting orders because the forecast failed. A controller spends four hours reconciling data between the ERP and spreadsheets. None of those hours is logged as "inefficiency". They show up as "normal operation".
That invisible time is where industrial AI (not generative — data AI) creates real value. A model that forecasts demand with 15% less error reduces manual planning adjustments. A route optimisation system reduces picking time. An automatic ETL eliminates manual reconciliation. The saving is not in software — it is in person-hours freed for work that really makes a difference.
Calculate how much that invisible inefficiency costs. Start with a week: ask your warehouse supervisor, your planner, your controller to record how much time they spend on repetitive or manual validation tasks. Multiply by 52. Then multiply by the average cost per hour (salary + overheads). Compare with the AI investment. If the payback is under 18 months, it is a decision. If it is longer, perhaps AI is not yet the priority — perhaps it is to optimise the manual process first, or redesign the workflow to eliminate the need.
This is uncomfortable because it forces you to admit that the current operation is inefficient. But it is honest. We see many companies that buy AI to "be modern" without measuring what they change. Then they end up disappointed. Change because you have a measurable problem, not because you have a budget.
Regulation and risk: the decision you cannot postpone
The European AI Regulation (AI Act, Regulation (EU) 2024/1689) came into force in August 2024. The prohibitions have applied since February 2025 and the rules for general-purpose models since August 2025. This is not a recommendation — it is law. If you use or plan to use AI to make decisions about people (recruitment, performance appraisal, sales route management), you are subject to transparency, documentation and audit obligations. Source: European Commission.
At the same time, NIS2 (Directive (EU) 2022/2555, transposed in Portugal via Decree-Law 65/2025) requires auditing of IT suppliers that have access to critical systems. If your AI supplier cannot demonstrate compliance with NIS2, the risk stays with you. It is not a contractual detail — it is a responsibility of your board of directors. An external audit that finds an undocumented AI system can result in significant fines.
Map where you use or plan to use AI in your operation. Classify each case according to the AI Act — prohibited, high risk, limited risk. Require suppliers to provide NIS2 and AI Act compliance documentation. Appoint an internal lead for AI governance (it can be your IT Director or a board member, but there has to be someone). This seems like bureaucracy. It is. But it is bureaucracy that the law imposes. Ignoring it is a risk exposure that no insurance covers.
Skills: the resource no one has
In Portugal there is a shortage of industrial AI specialists. It is not that data scientists are lacking — it is people who understand both AI and textile production, or food distribution, or plastic injection moulding. That person is rare. And expensive. According to Eurostat, 70.9% of EU companies identify the lack of skills as the main obstacle to adopting AI. Source: Eurostat, 2025. In Portugal, that figure is probably higher — because the talent market is smaller and competition from larger technology hubs is intense.
The solution is not to hire a PhD in machine learning and expect them to learn your industry in three months. It is to train someone from your current team — someone who already knows the processes, the operational constraints, the customers — in applied AI. Or it is to bring in an external partner (consultancy, software supplier) who has already worked in similar contexts. A footwear factory in Felgueiras that trained its planning engineer in demand forecasting models managed, within 18 months, to reduce forecast error from 22% to 11%. The engineer was not an AI specialist — they were someone who understood the seasonality of women's footwear and had technical curiosity.
Industrial AI skills cannot be bought off the shelf. They are built over 18-24 months with partners who have already walked the path.
The decision you need to take now is: who will manage AI internally? If you do not have that person, start looking (or training) now. Because in 2027, when AI is critical, you do not want to be dependent on an external consultant who only works 3 days a month.
Start small. But start now.
The trap is thinking that the AI strategy has to be ambitious from the outset. It does not. Start with a well-bounded use case: optimise demand forecasting for one product family. Or reduce picking time in a specific warehouse. Or automate data reconciliation between two systems. A project with a clear scope, with a success metric defined up front, with a 3-6 month timeframe. If it works, scale up. If it fails, learn and adjust. That is the only way to build internal trust in AI without burning budget.
Do not start with an 18-month, €500,000 "digital transformation via AI programme". Start with "let's reduce demand forecast error by 20% over the next 6 months" — something your planner understands, something your CFO can measure, something your team can support without being pulled away from critical work.
2025 is the year this decision becomes urgent. Not because AI is magic — it is not. But because the regulation is here, the competitors are moving, and the cost of falling behind is starting to become visible. What you need to do by the end of 2025 is simple: choose a real problem, define the metric, identify the partner, and start. The rest builds from there.
Frequently asked questions
Why do only 11.5% of companies in Portugal use AI if the technology has been around for years?
The article explains that the pace of adoption was slower than predicted and the impact more limited. The gap between the conversation at conferences and the reality on the shop floor remains huge. Even so, that does not justify postponing the decision — on the contrary, it is the ideal moment to start with a real problem, while there is still time to learn without pressure.
What has changed since 2020 that makes a decision about AI urgent now?
The cost of inaction has changed. NIS2 requires auditing of IT suppliers, the European AI Act classifies processes into risk categories, and competitors are starting to use AI to forecast demand, optimise stock and measure OEE in real time. If you wait for a "best practice", the standard forms without you.
Can I buy AI software and install it like any other module?
No. The article warns that this view costs dearly later on. The AI that changes operations lives in the data collected daily. If the data is scattered (ERP, MES, spreadsheets, paper), no AI model works. AI is a data architecture decision, not just a software one.
What is the first step I should take before contacting an AI supplier?
Ask your CTO or IT Director three questions: do you have a single repository of operations data? Can the core systems share data via API or ETL? Who is responsible for the quality of that data? If the answers are vague, the 2025 project is to build that foundation, not to implement AI.
Is generative AI (chatbots) a good solution for our operations?
Generative AI is a shortcut that costs more than it saves. A corporate chatbot handles 60-70% of routine questions well, but errs with confidence on the rest (hallucination). If your document system is chaotic, the chatbot inherits and amplifies that chaos. Use it only when the documentation is well structured and the processes are mature.
How should I measure the real value that AI can bring to my factory?
Measure the invisible time people spend on repetitive tasks: checking, manual adjustments, reconciliations. That time does not show up as inefficiency, but as "normal operation". Industrial AI creates value by freeing those person-hours. Start by recording a week of repetitive tasks and multiply the annual cost.
If I do not have integrated data, should I wait before investing in AI?
You should not wait for AI, but you should act now on data integration. If you answered "no" to two of the three questions about data architecture, AI is premature. However, designing that foundation now, while migration costs are manageable, is the right decision for 2025.