The majority of "AI in management" projects we see fail before they even start. Not for lack of an algorithm — for lack of clean data and a concrete decision that the model should automate. This guide gives you a five-step method to turn data you already have into automated decisions, with a simple criterion for knowing whether a use case is worth the investment or just innovation theatre.
The thesis, to save time for those who have already evaluated software: you don't need generative AI to get started. You need a repetitive decision, reliable historical data and someone to take responsibility when the model gets it wrong. In Portugal, only 11.5% of companies with 10 or more employees used AI technologies in 2025, up 2.9 points on the previous year (INE, 2025). But the number that matters is another one: among small companies (10-49 employees) adoption drops to 9.4% (INE, 2025). And most of those that tried failed at the first obstacle, which is not technical. It's about data.
What you need before you start
Before writing a line of code or requesting a demo, gather this. If you fall short on three of these points, the project will overrun.
- A specific, repetitive decision — "predict raw material shortages", not "optimise the company".
- Historical data covering at least 18 to 24 months, with daily or batch-level granularity.
- A business owner who signs off on the automated decision — not the IT department.
- A measurable success metric defined before you start: forecast error, time saved, cost avoided.
- A system that already captures the data in a structured way — ERP, WMS or shop-floor terminal.
- Clarity on which AI Act risk class your use case falls into.
- A data protection officer involved from day zero, not at the end.
Step 1 — Choose the decision, not the technology
The most common mistake we see: the company buys an "AI platform" and then looks for a problem to hand it. Reverse the order. List the decisions your team makes repeatedly, based on data, and where a mistake costs money.
In a garment factory near Famalicão, the most obvious candidate decision is not the most complex one. It's this: how many machines to allocate to each production order in the following week, given the efficiency history per operator and garment type. It repeats every Monday. A mistake delays the parent company — Inditex or Tom Tailor do not forgive a two-week delay on a collection. You have the data. It's automatable. And nobody would call it "artificial intelligence" over a coffee, which is precisely the point.
If you can't describe the decision in a sentence that fits into a text message, you don't yet have a use case — you have an ambition.
Step 2 — Audit the data before dreaming about the model
This is where most projects die. The lack of skills and knowledge is the main obstacle to AI adoption in EU companies (70.9%), ahead of legal uncertainty at 52.5% (Eurostat, 2025). But beneath that lack of skills there almost always hides a more mundane problem: the data is dirty, scattered or does not exist in the necessary detail.
Take an honest inventory:
- Where the data lives — in a single ERP, in spreadsheets, in the warehouse manager's head?
- What the rate of empty or inconsistent fields is over the last 12 months.
- Is there a single key linking production, stock and sales? In footwear, the SKU with three axes colour-size-fitting in a collection of 800 to 1,200 references is rarely consistent between modules.
- Do the records have reliable date and time, or do they depend on someone jotting it down at the end of the shift?
There's one detail the manuals don't mention that sabotages half of all audits: in many operations, the data exists but is wrong out of convenience. The operator records the closing time of an order as 5 p.m. because that's when they clock off, not when the garment was finished. The warehouse manager records the receipt of a batch of knitwear the next day because on Friday they were closing the month. This data looks clean in a spreadsheet — it has a date, it has a time, it has a quantity. It just doesn't describe reality. A model trained on it learns the lie and returns it with a scientific air.
If the shop-floor data comes from sheets filled in by hand at the end of the shift, fix that first. A real-time capture solution such as KORA Productivity generates the data at the moment the garment leaves the machine, not in the version reconstructed from memory two hours later. Without that foundation, the model learns from fiction.
Step 3 — Start with the most boring use case
Forget dazzling computer vision. The first real gains are in the tedious, repetitive processes. The four fronts where INFOS applied AI delivers concrete results are deliberately unglamorous:
| Front | Automated decision | Data required |
|---|---|---|
| Predictive HR (via pplPortal) | Flag turnover and absenteeism risk | Attendance history, rosters, appraisals |
| Cognitive document classification | Route and file invoices and documents without intervention | History of scanned and classified documents |
| Computer vision in quality | Detect defects on the line | Labelled images of conforming and non-conforming parts |
| Governance and explainability | Ensure every decision is auditable | Log of the model's inputs, outputs and criteria |
Note the right-hand column. None of these fronts work without structured historical data. That's why document classification is usually the first use case to produce results — the documents already exist, they are already scanned and the decision (which folder it goes to, who approves it) is binary and verifiable. Document Management turns that flow into a process that learns from every human correction: when finance reroutes a wrongly classified invoice, the model logs the correction and gets it wrong less often next time.
Step 4 — Keep the human in the decision (and not out of courtesy)
Automating the decision does not mean removing the person. It means changing their work: from deciding everything to validating exceptions and correcting the model. Explainable models, with privacy by design, exist for exactly this — they allow the manager to ask "why" and receive an answer that makes operational sense, not a number that comes out of a black box.
This point is not philosophy. It's compliance. The AI Act (Regulation (EU) 2024/1689) has been in force since August 2024, with the prohibitions applicable since February 2025 and the rules for general-purpose models since August 2025 (European Commission). Systems that affect workers — such as turnover prediction — fall into risk categories that require documented human oversight. Automating the decision not to renew a contract based on a turnover score is not aggressive. It's illegal. And the GDPR, with the national implementation of Law 58/2019, already required explaining automated decisions affecting people long before the AI Act existed.
Step 5 — Measure the before and after, or measure nothing
Define the baseline before you switch on the model. If you don't know what the demand forecast error was in January, you won't be able to prove that the model improved anything at all by July. The most elegant model in the world loses the next budget renewal if nobody can show the number that changed.
To consolidate forecasts, actual values and deviations into a single dashboard, a BI tool makes the result visible to the CFO in the month-end meeting — which is where AI projects live or die. It's not in the server room that the decision to continue is made, it's when the CFO sees that the forecast error fell from 22% to 12% and that this freed up capital tied up in stock. See how this plays out in another context in turning data into profit with Qlik Sense.
Common mistakes and how to avoid them
Five patterns recur in almost every project we've inherited half-built. Starting with the flashiest technology is the first: computer vision requires thousands of images labelled part by part, work that nobody budgets for. Start with document classification, where the data already exists. The second is ignoring data quality — before any model, carry out the Step 2 audit, because a model trained on dirty data automates its errors at scale and with a veneer of credibility that makes them more dangerous.
The third mistake is not defining the owner of the decision. If the model gets it wrong and nobody is responsible, the project dies at the first failure — appoint the business owner, not IT. Note that in many Portuguese family businesses, IT is a self-taught hero with 15 years of knowledge of the business but no time to take operational responsibility for every forecast; pushing the decision up onto them is a guarantee that it has no real owner. The fourth is forgetting the AI Act: classify the use case's risk before you start, because cases that affect people require documented human oversight and explainability. And the fifth — promising full autonomy right at the start. Begin with the model suggesting and the human validating. Only after months of measured accuracy should you reduce oversight, and only in low-risk exceptions.
The next step
Choose a repetitive decision, check whether you have 18 months of clean data to support it, and ignore everything else for now. Useful AI in Portuguese management does not start with a model — it starts with an ERP that already captures the right data and a person willing to sign off on the decision. The rest is engineering. If only half of your historical data survives the Step 2 test, the honest answer is not to buy AI — it's to sort out the capture first. To go deeper into concrete use cases, see the guide to AI applied to business management and the one on process automation in Portuguese industry.
Sources
- INE — Survey on the Use of Information and Communication Technologies in Enterprises, 2025.
- Eurostat — Use of artificial intelligence in enterprises, 2025.
- European Commission — Regulation (EU) 2024/1689 (AI Act), in force since August 2024.
Frequently asked questions
What is the main reason AI projects in management fail in Portugal?
Most fail for lack of clean, structured data, not for lack of algorithms. According to the article, the main obstacle is the quality of historical data and the absence of a concrete, repetitive decision to automate. Only 11.5% of Portuguese companies with 10+ employees use AI, and among small companies it drops to 9.4%.
How many months of historical data do I need to train an AI model?
A minimum of 18 to 24 months of historical data with daily or batch-level granularity is recommended. This period allows the model to capture seasonal variations and real patterns. Data with less coverage produces models that are unreliable for critical decisions in production.
Do I need generative AI to get started with decision automation?
No. The article is clear: you don't need generative AI to get started. You need a repetitive decision, reliable historical data and someone responsible when the model gets it wrong. Most of the initial gains come from simple use cases such as document classification or turnover prediction.
How do I know whether an AI use case is really worth the investment?
Apply the simple criterion: can you describe the decision in a sentence that fits into a text message? If you can't, you still have an ambition, not a use case. The decision must be specific, repetitive, data-based and with a measurable impact on cost or time.
What is the first step to implementing AI in automated decisions?
Choose the decision before the technology. List the decisions your team makes repeatedly based on data, where a mistake costs money. Reverse the common approach: don't buy an AI platform and look for a problem. Start with the concrete problem.
What does "dirty data" mean and how does it affect an AI model?
Dirty data is records that are incomplete, inconsistent or wrong out of operational convenience. For example, an operator records the closing time as 5 p.m. because that's when they clock off, not when they actually finished. Models trained on false data learn the lie and return it as a scientific forecast, sabotaging the project.
Who should be responsible for the automated decision in a company?
A specific business owner should sign off on the automated decision, not the IT department. This responsibility ensures alignment with real objectives, accountability when the model gets it wrong, and decisions based on business impact, not on available technical capacity.