Most Portuguese companies talking about AI in 2026 have not yet changed anything in their operational routine. They have a presentation, they have a pilot, they have a board meeting where someone said "we have to do something with AI". What they do not have is a different process on Monday morning. This article is not a neutral survey of possibilities — it is an argument for a specific sequence, because the order of implementation determines whether AI produces results or merely produces reports about itself.
The problem is not the technology. It is the sequence.
In 2025, only 11.5% of Portuguese companies with ten or more employees used artificial intelligence technologies — up 2.9 percentage points on 2024, but still a clear minority. Among small companies (10 to 49 employees), that figure drops to 9.4% (INE, 2025). Most Portuguese industrial SMEs are therefore looking at AI from outside the window.
The figure nobody quotes next to this one: according to Eurostat, the main obstacle to AI adoption in EU companies is not cost — it is the lack of skills and knowledge, cited by 70.9% of companies. Cost does not even appear at the top of the list. This changes the diagnosis. The problem is not access to the technology. It is knowing what to do with it — and in what order.
AI does not fix a broken process. It accelerates it — with all its defects included.
Companies that start with the tool before they have clean data, documented processes and defined indicators spend budget automating chaos. The result is faster chaos. AI applied to operations only changes the routine when it enters a system that already works — and makes it faster, more accurate or more predictable. It does not replace the system. It amplifies it.
What you need before you start
Before evaluating any tool, audit the current state of the operation across these eight dimensions. If you fail on more than three, resolve those first — then move on to AI.
Clean historical data: a minimum of 12 months of transactional data with no significant gaps — sales, production, stock, returns. Not "more or less complete". Complete. A model trained on data missing three months of summer learns that summer does not exist.
Documented process: every flow you want to automate has to be written down — who does it, when, by what criterion, what happens if it fails. If the process exists only in the head of the planning manager, AI will not discover it by inference.
KPIs defined and measured today: if you do not know the current baseline value, you will not know whether AI has improved anything. Predictive analytics without a baseline is decoration.
Active ERP or system of record: AI needs a source of truth. Spreadsheets shared on OneDrive are not enough — and Excel files with formulas that only Pedro can interpret, even less so.
Named data owner: someone who validates the quality of the input data. It does not have to be a data scientist. It has to be someone who knows the business well enough to say "this number cannot be right".
Compliance with the AI Act: Regulation (EU) 2024/1689 has been in force since August 2024. The prohibitions have applied since February 2025. Identify whether the use cases you are considering are of limited or high risk — that defines transparency and documentation obligations that are not optional, regardless of the size of the company.
Budget for change, not just for software: include training, adaptation time and internal support. The software is consistently the cheapest part of any AI implementation. What costs is changing what people do with the results the system produces.
Top management approval with a defined success criterion: "let's try AI" is not approval. "We will reduce stock-outs by X% by the end of Q3" is. Without a success criterion defined from the outset, any result looks acceptable — and none is measurable.
Where AI changes the routine: by operational dimension
Production and the shop floor
In a garment factory with 120 employees in the Vale do Ave, production planning is done by the planning manager based on accumulated experience and an Excel sheet that no one else can read. When he is off, the factory loses two days. This is not an extreme case — it is the pattern we see repeated across dozens of companies. And this is exactly where AI has immediate impact, not because it replaces the manager, but because it externalises the knowledge that was locked inside him.
Predictive analytics applied to production does three concrete things: it anticipates breakdowns based on sensor patterns before the machine stops, suggests production-order sequences that minimise setup times between references, and flags OEE deviations before the shift ends — not in the following day's report, when there is nothing left to do. The difference between receiving the alert at 2 p.m. and receiving it at 5.30 p.m. is one production shift.
The unavoidable prerequisite is real-time data capture per production order and per operator. Tools such as KORA Productivity create that structured foundation on the shop floor. Without it, there is no production AI — there are estimates with a more sophisticated interface.
Stock and warehouse management
ABC analysis already exists in any decent ERP. AI goes a step further: it dynamically recalculates priorities based on real seasonality, returns history per reference, actual supplier lead times — not the contractual ones, which rarely match — and order patterns per customer. In a distribution warehouse along the Lousada-Paços de Ferreira corridor, this means picking is not the same in October and in March, and the system knows this without the warehouse supervisor having to configure it manually every time the season changes.
What changes in the routine: the purchasing manager stops calculating reorder points by hand. The system proposes, with legible justification. The buyer validates or rejects — and the system learns from the decision. After six months, the automatic suggestions have an acceptance rate that says everything about the quality of the input data: if it is below 60%, the problem is not the model, it is the outdated lead times in the ERP.
Confirm that the warehouse management system records movements with date, time and operator. Without that detail, the model cannot distinguish a stock-out from a discontinued reference.
Sales force and demand forecasting
Sales forecasting done by an experienced salesperson has a structural problem that no one likes to admit: it is good for the customers they know well and systematically pessimistic or optimistic for the rest, depending on the salesperson's temperament. AI levels this. It analyses purchasing patterns across the whole portfolio, identifies customers at risk of churn before they cancel — typically three to four months in advance, when the order pattern starts to change subtly — and suggests upsell actions based on the purchase history of customers with a similar profile.
In footwear — where an international buyer visits twice a year and the collection has between 800 and 1,200 SKUs with three axes of variation (colour, size, last) — AI applied to order forecasting per reference can significantly reduce production-planning error. It does not eliminate the salesperson's judgement. It gives them a second pair of eyes with perfect memory and without the bias of someone who has a quota to hit by the end of the month.
The prerequisite that most often fails: the sales history per customer and per reference has to be in the ERP, not in the salespeople's emails nor in a shared folder with proposal files. If the commercial information lives outside the system, the forecasting model will be only as good as the data someone remembered to export.
Human resources and absenteeism
Unpredictable absenteeism costs more than chronic absenteeism — because the chronic kind is already built into capacity planning. Predictive AI applied to HRIS identifies patterns before they become a problem: employees at high risk of leaving in the next 90 days, teams with consistently unbalanced load between shifts, times of year with a history of concentrated incidents. It is not surveillance — it is the difference between managing an absence on Friday at 7 a.m. and anticipating the need for reinforcement the week before.
Platforms such as pplPortal already incorporate predictive AI modules for turnover and absenteeism. What changes in the HR manager's routine: they stop reacting to absences and start managing labour availability in advance. In sectors with an ageing workforce — garment, footwear — that anticipation is worth more than any reporting dashboard.
The detail the manuals do not mention: predictive absenteeism models lose accuracy quickly if attendance data has gaps of more than two consecutive weeks. A poorly executed time-clock system migration can render months of history useless and force the model training to restart from scratch.
Finance and invoicing
Cognitive document capture — supplier invoices, delivery notes, credit notes — is one of the AI use cases with the fastest and easiest-to-measure ROI. A company processing 400 invoices per month manually has a cost per document that can be calculated precisely before any implementation. AI reads, classifies, validates against the order in the ERP and routes only the exceptions for approval. What changes in the accounting routine: manual posting disappears. Human validation concentrates on the cases the model cannot resolve with sufficient confidence — typically between 5% and 15% of the volume, depending on the quality and consistency of the documents received.
The mistake we see frequently: the company implements cognitive capture but does not define the exception approval workflow. The result is an inbox of "documents to review" that no one processes systematically, and the benefit of automation disappears into the human bottleneck that was left unresolved.
Prioritisation matrix: where to start
| Use case | Data complexity | Operational impact | Time to visible result | Critical prerequisite |
|---|---|---|---|---|
| Cognitive invoice capture | Low | Medium | 4–8 weeks | Defined approval workflow |
| Absenteeism forecasting | Medium | Medium-high | 3–6 months | 12 months of attendance history with no gaps |
| Automatic stock replenishment | Medium | High | 6–12 weeks | Up-to-date lead times in the ERP |
| Sales forecasting per reference | High | High | 3–5 months | 2+ years of history per SKU |
| Predictive maintenance | High | Very high | 6–12 months | Sensors installed + historical breakdown data |
| Financial anomaly detection | Medium | Medium | 4–10 weeks | Clean accounting data in the ERP |
Start with the case with the lowest data complexity and enough operational impact to justify the internal effort. Cognitive document capture and automatic stock replenishment are, in most Portuguese industrial companies, the entry points with the best effort-to-result ratio — and they produce a measurable result quickly enough to keep management support for the next project.
Three mistakes that cost more than the software
Mistake 1 — Starting with the model, not with the data. The team chooses an AI tool before auditing the quality of the input data. The model trains on inconsistent data and produces suggestions no one trusts. After three months, the system is installed and ignored. Devote the first four weeks exclusively to cleaning and validating the historical data. Only then evaluate tools — in that order, no exceptions.
Mistake 2 — Automating without defining the exception criterion. The system starts making decisions automatically, but no one has defined when a decision should be escalated to a human. When the AI gets it wrong — and it will — there is no recovery process, and the error propagates until someone notices. Before activating any automation, document: "if the model's confidence is below X%, the decision goes to person Y, who has Z hours to respond". Without this protocol, automation creates a new operational risk where before there was only inefficiency.
Mistake 3 — Ignoring the AI Act because it is "just an SME". Regulation (EU) 2024/1689 has no exemption by company size. If you use AI in HR decisions — assessment, recruitment, performance management — you are in high-risk scope, with transparency and documentation obligations that apply regardless of whether you have 50 or 5,000 employees. Classify each use case before implementing. The INFOS AI service includes this compliance assessment as part of the start-up process — because discovering the regulatory framework after the system is in production is significantly more expensive than doing it beforehand.
The right sequence is not glamorous. Clean data, documented processes, defined exception criteria, verified compliance — and only then the tool. The companies that skip these steps do not fail for lack of technology. They fail because the technology arrived before the system was ready to receive it.
Frequently asked questions
What is the main obstacle to AI adoption in Portuguese companies?
According to Eurostat, 70.9% of companies cite the lack of skills and knowledge as the main barrier — not cost. The problem is not accessing the technology, but knowing what to do with it and in what sequence to implement it to generate real operational results.
How many Portuguese companies use AI in operations in 2025?
Only 11.5% of companies with ten or more employees use AI technologies — 2.9 percentage points above 2024. In SMEs (10 to 49 employees), that figure drops to 9.4%. Most Portuguese companies are still outside this transformation.
What does "AI does not fix a broken process" mean?
It means that implementing AI in a disorganised process only accelerates the existing chaos. AI amplifies systems — it does not fix them. If the data is dirty, the processes are undocumented or the indicators do not exist, the tool merely automates inefficiency at greater speed.
What are the mandatory prerequisites before implementing AI?
Complete historical data (minimum 12 months), documented processes, defined and measured KPIs, an active ERP, a named data owner, compliance with the AI Act, a budget for change (not just software) and management approval with measurable success criteria.
How does AI change the routine in production?
Predictive analytics anticipates breakdowns through sensor patterns, suggests production sequences that minimise setup times and flags OEE deviations in real time — not in the following day's report. It externalises knowledge that was concentrated in one person, eliminating critical dependencies.
What is the role of the data owner in the implementation?
They do not need to be a data scientist. They should be someone who knows the business deeply and validates the quality of the input data, identifying numbers that make no operational sense. It is the guarantee that the model trains on credible information.
How much does it really cost to implement AI in a company?
The software is consistently the cheapest part. The real costs are in training, adaptation time, internal support and process change. Budgeting only for the tool is a mistake that leads to the failure of the implementation — the real investment is in organisational change.
Sources
- Instituto Nacional de Estatística (INE) — Statistics on the use of artificial intelligence technologies in Portuguese companies, 2025
- Eurostat — Labour Force Survey and data on obstacles to AI adoption in European Union companies
- Regulation (EU) 2024/1689 (AI Act) — Regulation on artificial intelligence, in force since August 2024
- Comissão Nacional de Proteção de Dados (CNPD) — Guidelines on the compliance of AI systems with Portuguese and European legislation