AI applied to industrial management: from data to automated decision-making

In 2025, only 9.4% of small Portuguese companies with 10 to 49 workers were using artificial intelligence — against 49.1% of large ones (INE, 2025). The difference is not one of ambition or budget. It is one of data architecture. Most Portuguese industrial SMEs have all the data: production orders, cycle times, returns, absenteeism, raw-material consumption. What they lack is the layer that turns that noise into actionable signal. And here is the thesis this article defends: the main obstacle to AI in Portuguese industry is not technological — it is the illusion that the data is already ready. It is not. And starting with the algorithm before resolving that illusion is the most expensive and most repeated mistake we see on the ground.

The problem is not a lack of data — it is a lack of clean data

Anyone who has already walked into a distribution warehouse along the Lousada-Paços de Ferreira corridor knows: the data exists in three different places and none agrees with the other two. The ERP says one thing, the warehouse manager's Excel sheet says another, and the picking system has a third number. Applying AI to this reality without resolving the data layer is like installing surveillance cameras pointed at the floor.

AI applied to industrial management does not start with a machine learning model. It starts with the decision of which ERP table is the source of truth for stock.

Before evaluating any platform, audit four dimensions — and be honest about the results, because this is where most projects fail before they even begin.

Completeness. What percentage of production records have all fields filled in? In textile factories in the Vale do Ave, it is common to find production orders without a real completion time: the operator closed it in the following shift, the time was recorded wrong, and the history becomes unrecoverable. A blank field is not a missing data point — it is a poisoned one, because the model will impute a value and no one will know.

Consistency. Is the article code in the ERP the same as in the quality system and in the dispatch file? In footwear from Felgueiras, the same SKU frequently has three different nomenclatures depending on the system — one inherited from the customer, one internal, one from the component supplier. Nomenclature discrepancies kill any predictive model before it runs for the first time.

Granularity. Do you have data at the individual operation level or only at the order level? To predict breakdowns, you need sensor readings — not daily totals. A daily total of energy consumption says nothing about the degradation pattern of an injection press. It only says that the factory consumed X kWh.

Latency. Does the data arrive in real time, in a daily batch, or in a weekly file? A production planning model fed by yesterday's data has limited value when the customer calls to request an earlier delivery. Latency is not a technical detail — it is the difference between a decision and a historical reconstruction.

Only after this inventory does the conversation about algorithms make sense. The article AI applied INFOS: turning data into automated decisions explores in depth the data architecture required for this step.

What has changed to make this relevant now

Three things converged between 2022 and 2025 that had not previously existed simultaneously. Cloud computing costs fell to the point where a demand forecasting model became accessible to an SME with 40 employees — what was an €80,000 project in on-premises infrastructure now fits into a monthly subscription. Vertical ERPs began to expose native REST APIs that allow analytical layers to be integrated without six-month integration projects. And the proliferation of IIoT lowered the cost of instrumenting old machines to values that fit within a normal maintenance budget — a vibration and temperature sensor kit for a 20-year-old press now costs the equivalent of two days of unplanned downtime.

What has not changed: the quality of historical data. And that is why technological convergence does not automatically translate into adoption. Eurostat (2025) points to the lack of skills and knowledge as the main obstacle to AI adoption in 70.9% of EU companies — but on the Portuguese ground, what we see is that the skill is lacking precisely in data management, not in understanding the algorithms.

The role of the AI Act in the architecture decision

Regulation (EU) 2024/1689 — the AI Act — has been in force since August 2024. The prohibitions on high-risk systems have applied since February 2025. This has practical consequences that rarely appear in vendors' proposals.

AI systems that make decisions about workers — shifts, performance assessment, selection — are classified as high risk and require technical documentation, registration in the EU repository and mandatory human oversight. Demand forecasting or predictive maintenance models applied to machines, without decisions about people, typically fall into low or minimal risk. The distinction seems clear on paper; in practice, a shift optimisation system that uses individual productivity data sits in a grey area that your legal team needs to classify before any RFP.

The choice between an internally developed model and a solution from a certified vendor also has compliance implications: who is responsible for the AI system before the regulator? The SaaS vendor that sells the model, or the company that integrates it into its decision processes? The answer affects the contract, not just the architecture. The GDPR (and Law 58/2019, which enacts it in Portugal) adds an additional layer when the data processed includes worker information — which is almost always the case in production systems.

Real technical options: the honest map

There are four distinct approaches to applying AI to industrial management. They are not maturity phases — they are options with different trade-offs. Choosing the wrong one for the right size is the most expensive mistake we see repeated.

Why most SMEs start in the wrong place

The pattern we see repeated: the company buys an ML platform because the competitor "already has AI", without having resolved the integration between the ERP and the quality system. Six months later, the project is stalled, waiting for data that never arrives clean. The right investment for a factory with 60 employees and two years of history in the ERP is almost always the second row of the table — predictive models in the cloud — fed by a well-configured Qlik Sense as a visualisation and validation layer.

There is an even subtler mistake that rarely appears in retrospectives: buying the right approach, but for the wrong problem. An SME in clothing in the North that implements demand forecasting based on its own sales history, without integrating the end customer's sell-through data, is forecasting its own production based on past orders — not on real demand. The model will be accurate and useless at the same time.

Trade-offs by company size

Size matters — but not in the way you might think. The problem for small companies is not the cost of the models. It is the maintenance cost: who retrains the model when the product line changes? Who validates the results when the model starts to drift? In an SME without dedicated IT, the honest answer is "no one" — and an unsupervised model is worse than a well-maintained Excel sheet.

The factor no one puts in the proposal: the cost of retraining

A demand forecasting model trained with 2022–2024 data in a clothing factory in the North will start to err when the parent brand changes the collection or when a new fabric reference is introduced. Retraining is not automatic in most entry-level SaaS implementations. Ask the vendor: who retrains, how often, and at what cost? If the answer is vague, the total cost of ownership — the real ROI — is underestimated in the proposal.

There is an operational detail that implementation manuals rarely mention: the model degrades faster in companies with double seasonality. A footwear factory that serves men's buyers in August and women's buyers in February has two collection cycles with distinct demand patterns. A model trained on a full 12-month cycle may have good average annual performance and systematic errors at the peaks — which are precisely the moments when planning matters most.

Use cases with the greatest return in Portuguese industry

Demand forecasting and production planning

In footwear in Felgueiras, planning a collection involves 800 to 1,200 SKUs across three axes — colour, size, last. A forecasting model that integrates sales history, customer sell-through data and seasonality reduces the forecast error and frees up capital tied in stock. The prerequisite is having batch traceability implemented in the ERP — without it, there is no reliable consumption history by reference, and the model will forecast based on aggregates that hide the real variability by product axis.

Predictive maintenance via IIoT

In a plastic injection factory in the Aveiro-Marinha Grande corridor, the presses have vibration and temperature sensors installed. Integrating that data via OPC-UA into an analytics platform allows degradation patterns to be detected before the breakdown. The value is not in the algorithm — it is in reducing unplanned stoppages, which in batch production have a direct impact on OEE. KORA Productivity captures this data in real time and feeds the operational efficiency dashboards. What is rarely said: predictive maintenance only works if corrective maintenance is well recorded. If maintenance work orders do not have a documented root cause, the model has no signal — it has noise with dates.

Absenteeism and shift management with predictive AI

Absenteeism in the textile and clothing industry in the North is a structural problem — an ageing workforce, shift work, seasonality. Predictive absenteeism models, when trained with 24 or more months of history, can anticipate peaks two to three weeks in advance, allowing rosters to be adjusted before the problem arises. pplPortal includes predictive AI for turnover and absenteeism — but it only generates value if the attendance records are digitised and consistent.

The most sophisticated predictive absenteeism model on the market fails if clocking-in is still done on paper and entered into the system on Friday afternoons.

What works in practice: three sector patterns

Pattern 1 — BI first, AI later

The most consistent pattern in the successful implementations we follow: the company begins by consolidating the data into a well-configured industrial BI, defines the KPIs it wants to monitor, and only then — with 12 to 18 months of clean data — moves on to predictive models. The temptation to skip this phase is strong when the AI vendor promises results in eight weeks. In practice, the eight weeks are for setup; reliable results arrive when the data arrives. And the BI phase is not a bureaucratic prerequisite — it is the moment when the company learns which data it actually uses to decide, which is different from the data it collects out of habit.

Pattern 2 — A single, well-defined use case

A medium-sized textile factory in the Vale do Ave that wants to apply AI should start with a single use case: forecasting yarn consumption by knit reference, integrated with the ERP's purchasing planning. Not a "digital transformation with AI" project — a specific problem, with a clear success metric (reducing raw-material shortages), and an internal owner who is accountable for the results. This focus is what separates the projects that reach production from those that die in the pilot. The internal owner does not need to know machine learning. They need to know what a shortage is, when it happens, and how much it costs — and to have the authority to change the process when the model suggests an early purchase.

Pattern 3 — Structured human oversight

The projects that fail silently have a common trait: the model starts producing suggestions, no one validates them systematically, and after three months no one knows whether the model is still working well. Set up a monthly review process — the digital gemba: the production manager compares the model's suggestions with what actually happened, identifies deviations, and escalates for retraining when necessary. Without this cycle, AI degrades silently. And when someone finally notices, the damage has already been done in purchasing, stock or planning decisions that no one questioned because "the system said so".

How to measure success post-implementation

Define the metrics before the start — not afterwards. The four most relevant for AI applied to industrial management are as follows.

The forecast error (MAPE) measures the accuracy of demand and planning models. A MAPE below 15% in stable categories is a good benchmark for most Portuguese industrial sectors — but in footwear with 800 SKUs and double seasonality, demand the MAPE by product cluster, not the aggregate, which can hide systematic errors in the higher-volume references.

The suggestion adoption rate tells you what MAPE does not: whether operators trust the model. Below 40% acceptance, the model is not reliable — either the interface is poor, or the operators know something the model does not. Both hypotheses warrant investigation before retraining.

The time between anomaly and detection is the central indicator for predictive maintenance: how many days in advance the system detects the degradation pattern before the actual stoppage. This number should be compared with the average repair time — if the system detects 6 hours in advance and the repair takes 8 hours, the value is limited.

The variation in OEE is the aggregate indicator that captures the combined effect of availability, performance and quality. Track it in Qlik Sense with the right industrial KPIs — and always compare with the equivalent period, not with the previous month, to isolate the effect of seasonality.

The most common measurement error

Measuring the success of the AI project by the number of models in production. That is not it. It is by the impact on the operational metrics that existed before the project. If the OEE has not improved, if the raw-material stock has not been reduced, if absenteeism has not been anticipated more effectively — the project was not successful, regardless of how many dashboards are running.

Compliance and cybersecurity: what you cannot ignore

The NIS2 Directive (Directive (EU) 2022/2555, transposed into Portugal by DL 65/2025) classifies operators of critical industrial infrastructure as entities subject to reinforced cybersecurity requirements. AI systems connected to production networks expand the attack surface — a predictive maintenance model that communicates with PLCs via the industrial network is a potential entry vector for ransomware. Check the network segmentation with the cybersecurity officer before connecting any AI system to the OT network.

There is a point that rarely appears in risk assessments: AI models trained with production data are, themselves, sensitive information assets. A demand forecasting model trained with three years of order history from an international customer contains strategic information about the commercial relationship. The retention and access policy for these models must be covered by the information security policy — not just access to the raw data.

The security of an industrial AI system is not assessed by the model — it is assessed by the network it runs on and the access it has to the shop floor.

Next steps: what to do this week

1. Audit the data quality in the ERP: select an entity — production orders or stock movements — and check the completeness rate of the critical fields over the last 24 months. If it is below 85%, the AI project starts here, not with the algorithms.
2. Identify a use case with an owner: choose an operational problem with a clear metric and a person accountable for the result. Without an owner, there is no project — there is an eternal pilot.
3. Classify the AI Act risk: determine whether the system you want to implement involves decisions about workers. If so, involve legal before proceeding to RFP.
4. Check the ERP-MES integration: confirm that the ERP exposes the necessary data via API or structured file. A well-configured ERP MULTI or a QAD Adaptive ERP with the integrations active is the technical prerequisite for any AI layer.
5. Define the oversight process: before putting any model into production, document who validates the suggestions, how often, and what happens when the model errs. This document is more important than the choice of platform.

For the broader context of how AI fits into the industrial digital strategy, read the article AI applied to business management: a guide to getting started with real use cases. For the operational perspective on process automation in the Portuguese context, the operational guide to process automation in Portuguese industry complements what was left undeveloped here. And for the KPIs the operations director should monitor daily, the respective article provides the measurement framework against which AI models should be evaluated.

The difference between a company that "has AI" and one that uses AI to decide better is, almost always, the quality of the work done before the first model runs. And that work is not glamorous: it is cleaning tables, aligning nomenclatures, convincing the warehouse manager to close the orders in the right shift. Those who do that work first reach the model faster — and with results that last.

Sources

• INE — Statistics Portugal. Survey on the Use of Information and Communication Technologies in Enterprises 2025. Lisbon: INE, 2025.
• Eurostat. ICT usage in enterprises — Artificial intelligence. Luxembourg: European Commission, 2025.
• European Commission. Regulation (EU) 2024/1689 of the European Parliament and of the Council — AI Act. Official Journal of the European Union, 12 July 2024.
• European Parliament and Council of the EU. Directive (EU) 2022/2555 — NIS2. Official Journal of the European Union, 27 December 2022.
• McKinsey & Company. The State of AI: How organizations are rewiring to capture value. McKinsey Global Survey, 2025.

Frequently asked questions

Why do most Portuguese industrial SMEs fail to implement AI successfully?

The main obstacle is not technological, but data quality. Most SMEs have data scattered across multiple systems (ERP, Excel, picking systems) that do not match one another. Without resolving this data layer — completeness, consistency, granularity and latency — any AI model fails before it starts.

What does "poisoned data" mean in the context of industrial AI?

These are blank or incorrectly filled fields in production records. A worker who closes a production order in the following shift with the wrong time creates a poisoned data point. The AI model will impute an arbitrary value, compromising the entire forecast. An empty field is not just a gap — it is an active error.

How do you audit data quality before implementing AI?

Assess four dimensions: Completeness (what percentage of records have all fields?), Consistency (is the article code the same across all systems?), Granularity (do you have data at the operation level or only daily totals?) and Latency (does the data arrive in real time or in a weekly batch?). This audit determines whether the project is viable or doomed.

What is the difference between a demand forecasting model and a performance assessment system under the AI Act?

A demand forecasting model is classified as low or minimal risk. A system that makes decisions about workers — shifts, performance assessment — is high risk and requires technical documentation, registration in the EU repository and mandatory human oversight. The distinction affects legal compliance and contracting.

Why is data latency critical for production decisions?

A model fed by yesterday's data cannot respond to requests for earlier delivery made today. Latency is not a technical detail — it is the difference between an actionable decision and a historical reconstruction. For dynamic planning, the data needs to be available in real time or in a daily batch at most.

What changed between 2022 and 2025 that makes AI viable for industrial SMEs?

Three convergences: cloud computing costs fell drastically (forecasting models now fit into monthly subscriptions); vertical ERPs expose native REST APIs (without six-month integration projects); and IIoT sensors became affordable (instrumenting old machines fits into the maintenance budget). But the quality of historical data did not improve.

Who is responsible for the compliance of an AI system under the AI Act — the vendor or the company?

It depends on the contract and the architecture. If it is a SaaS solution from a certified vendor, there are shared responsibilities. If you developed it internally, the company is responsible. This distinction affects not only legal compliance, but also the contract, the insurance and the audit. It should be clarified before any RFP.