
What is artificial intelligence and how does AI work?
AI only pays off when it improves a decision. Start there, not with the technology.
Artificial intelligence in short
- Artificial intelligence (AI) is technology that learns to discover patterns and relationships in data by itself, using algorithms and a lot of computing power. You do not program the rules; the system learns them from examples.
- There are three types of AI: narrow AI (one task), general AI (reasons like a human) and super AI (surpasses humans). Everything in use today is narrow AI.
- An AI model is developed in 10 steps, from defining the goal to continuously evaluating performance. With ready-made models you skip three of those steps, but the other seven remain.
- The EU AI Act entered into force on 1 August 2024 and requires human oversight of high-risk AI systems.
- AI creates value when it makes a decision faster, more accurate or more consistent. Diamond Decisions starts with that decision and works back to data and technology.
Artificial intelligence impresses, but there is nothing magical about it and it will not replace people any time soon. There are, however, concerns that AI could go off the rails or start imposing its will. That is one reason it has to remain controllable under the AI Act: the human as supervisor. In business and government, AI is now used on a large scale, with mixed success. This article covers the definition, how AI works, the applications, the benefits, the accidents and legislation, ethics and sustainability.
What is artificial intelligence?
Many definitions of artificial intelligence are in circulation. We use the most concise one. In short: we help the computer develop a capacity to learn by exposing it to large amounts of data, which it analyzes thoroughly with algorithms.
Artificial intelligence independently learns to discover patterns and relationships in data, using specific algorithms and usually unprecedented computing power.
So you cannot really program AI; it programs itself. That is the big difference with traditional software, where a programmer has to think of every possible situation in advance. This self-learning character is what defines AI. In essence it is nothing more than automating the analysis of large or complex amounts of data. AI adds most value in two areas:
- Thinking, learning and deciding: automatic classification, clustering, predicting events, planning, solving problems and, if desired, deciding independently.
- Acting, coordinating and making: robots, autonomous vehicles, automatically writing texts, creating images, 3D printing and translating.

AI can now do almost everything a human can, only often poorly. But in specific areas and for complicated problems it performs many times better and faster than people. AI algorithms see things that are invisible to the naked eye. That also explains the AI race: if your neighbors, competitors or other countries use AI successfully, they can outsmart you.
What AI means: deciding and acting independently
Artificial intelligence is an applied science in which computers take over certain human tasks, such as learning, reasoning, analyzing, solving problems and deciding. You speak of AI when a system can perform tasks independently in complex environments, without constant guidance from a user or trainer.
AI systems are also extremely good at differentiating and personalizing. Where a supply chain used to take a few hundred fairly general decisions once a week, algorithms now take a hundred thousand very specific micro-decisions every five minutes, for example whether the stock of an item needs replenishing. Scalable, fine-grained, and it removes arbitrariness from the process. That is why we see AI first and foremost as a way to improve decisions.
The types of AI: narrow, general and super AI
The literature broadly agrees that there are three types of AI: Artificial Narrow Intelligence (ANI), Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI).
- The narrow variant can perform only one task properly. Think of a game of chess or Go, or predicting the weather or the harvest.
- The general variant, AGI, can also reason abstractly, just like a human. That allows it to respond to unexpected situations: it can improvise.
- The super variant, ASI, surpasses human intelligence in every area, even that of the smartest people on earth. Nobody knows yet what it would be capable of.

Kaplan and Haenlein add a second classification of AI systems: the analytical variant has cognitive abilities with which it can learn; the human-inspired variant also has emotional intelligence; and the humanized variant is socially intelligent as well, so it can communicate and interact with others.
Whatever happens, people must control artificial intelligence and not the other way around. In our view, every AI system that functions as a system technology should have a red button or pause button, one that society can press and not only the company that built the system.
The economic significance of artificial intelligence
The global market for AI is huge. Good software developers, machine learning engineers and data scientists are scarce and therefore expensive. Venture capitalists, banks and private investors are pumping billions into start-ups that develop new products and services based on AI applications. These find their way into virtually every industry. The overview in figure 3 is certainly not complete, and the number of applications keeps growing.

Governments also contribute financially, because they see that AI can help with public safety, better healthcare, streamlined infrastructure, energy savings and better education. The smart city and a more sustainable society will hardly get off the ground without AI and big data.
How does AI work?
The technology behind AI consists of computer algorithms with which you train AI models. Many of those algorithms have existed for more than 40 years. What is new is the explosive growth of data and computing power in recent years. That is what finally gave AI wings.
Algorithm versus AI model
You can choose from a manageable number of algorithms: neural network, genetic algorithm, linear or logistic regression, decision tree, random forest, DBSCAN, K-means clustering, hierarchical clustering, nearest neighbor, gradient descent and the Bayesian network. Each has its own field of application. Two terms are often confused:
- The algorithm: the mathematical method that determines how the model learns from the data. Think of it as the recipe for learning.
- The AI model: the result of training with the algorithm, also called a machine learning model. It contains the specific parameters, such as weights or coefficients, that were learned from the training data.
You train an AI model with an algorithm and a dataset until the results are good enough for the purpose you have in mind. The result will never be perfect, because AI models work with probabilities and statistics. Just like a human, a model draws its own conclusions and therefore makes mistakes.
How do you develop an AI model in 10 steps?
- Define the goal: describe which problem you want to solve and which decision or analysis you want to automate: the use case. Is the problem suitable for AI? Is it legally allowed? Do you have good data? What is the business case?
- Collect as much relevant data as possible: look for all data that may be relevant to the use case, including big data and open data.
- Clean the data: for AI, data quality matters perhaps twice as much as for dashboards and reports. Merge duplicate records, complete the data and correct it. Anonymize or pseudonymize personal data in this phase.
- Model the data: pour the data into a form that is suitable for training, usually one large flat file with extra calculated columns, such as the number of days between two orders. This is the final dataset.
- Analyze the data thoroughly: inspect the data to the bone. This gives you insight into biases that may be hidden in the data and prevents you from walking into well-known pitfalls.
- Choose the type of algorithm: for classification you often use decision trees, neural networks or logistic regression; for clustering, DBSCAN or K-means; for optimization, such as routes, genetic algorithms or gradient descent.
- Train the AI model: train the model with the chosen algorithm and, on average, 60 to 80% of the final dataset. With some algorithms you give the model instructions and feedback (supervised learning).
- Test and validate the model: use the remaining data to check whether the model does what you trained it for. If it is wrong too often, the cause may be data quality, bias in the dataset, the choice of algorithm or the training itself.
- Take the model into production: this is perhaps the most critical phase. Will users trust the model? How do they learn to interpret the results? The change-management and business side of AI is usually neglected, yet it is decisive for success.
- Evaluate performance continuously: monitor the model once it is in production. Human oversight is essential and, for high-risk systems, mandatory under the AI Act. Regularly compare a sample of the outcomes with the judgment of an expert.
It is no longer always necessary to develop and train AI yourself. Many AI models are available off the shelf, for example for classifying objects in a photo or estimating the sentiment of an e-mail. With such a ready-made model you can skip steps 6 to 8. But do not count yourself rich: the other seven steps remain necessary, and that is where the biggest risks of accidents lie. The data you send to the model must have the same structure as the data it was trained on, and that takes work. A data warehouse helps.
How Diamond Decisions applies these steps
Most AI projects that disappoint started with the technology or the data. We turn it around and start with step 1: the decision that has to improve. In Diamond Inspirations, management agrees on which decisions matter most. In business modeling we map those decisions, the processes and the KPIs. Only then do we build the data platform and analytics, including AI where it adds value, and we train the decision makers who have to work with the results.
Traditional software versus AI
The table below shows the main differences between traditional software and artificial intelligence.
Traditional software
Artificial intelligence
How it is made
Rules written by a programmer
Rules learned from data
What it needs
A complete specification
Many good examples
Outcome
Always the same answer
A probability, sometimes wrong
New situations
Fails if not foreseen
Can generalize from examples
Explainability
Rules can be read
Often hard to explain
Maintenance
Adjust the rules
Monitor and retrain
Examples of artificial intelligence
AI often happens behind closed doors, but through research and our work for clients we have come across many interesting examples and use cases:
- With robotic process automation (RPA) and AI you can automate administrative processes in finance, logistics and other sectors ever better. This is called hyperautomation.
- AI is a powerful weapon against fraud and organized crime, from credit card fraud to money laundering. Banks, insurers and benefits agencies use algorithms to nip potential fraud in the bud.
- Data-driven information security: detecting malware, phishing and fake news, and predicting and automatically resolving IT failures with self-healing systems.
- Call centers and customer service departments use AI assistants and chatbots that take over a large part of the workload.
- HR departments use AI to screen and assess candidates, for example with automatic CV screening.
- In healthcare, AI supports diagnoses, helps radiologists analyze X-rays and builds prediction models for diseases. Robots are increasingly used in elderly care to combat loneliness.
- In retail, machine learning provides reliable demand forecasts, minimal stock and dynamic pricing. Supermarkets use it in the fight against food waste.
- In education, adaptive learning tools make personalized learning possible.
- In logistics, the applications range from order-picking robots and smart route planning to Internet of Things sensors on containers and models that predict demand and delivery times.
- Drones detect gas leaks, inspect dikes and crops, and assess fire, storm and roof damage for insurers.
- Regional grid operators save costs by using AI for smart maintenance and for quickly locating the cause of a power failure.
- In the port of Rotterdam, an AI-driven risk analysis determines which containers are opened and checked in the fight against drug smuggling.
The 5 major AI applications: from NLP to computer vision
AI applications can roughly be divided into five main categories:
1. Natural language processing (NLP)
NLP is about computers interpreting, understanding and producing human language. Speech recognition is a specific application that converts spoken words into text. NLP is mainly used by virtual assistants and chatbots in call centers and customer service departments.
2. Generative AI
Generative AI builds on NLP but goes further. Based on a prompt, the system writes a complete article on any subject. ChatGPT, Microsoft Copilot and Google Gemini are well-known examples. Generative AI can also create images, sound, music and video from instructions.
3. Traditional data mining
Data mining focuses on finding relationships and patterns in structured data with machine learning, statistics and database techniques. The aim is to gain new insights that are hidden in the data.
4. Text mining and analytics
Text mining finds relationships, patterns and correlations in unstructured data such as text. Here too, the aim is to gain new insights and knowledge.
5. Image and facial recognition
With computer vision you imitate and surpass human eyesight: computers identify objects and people in images and video, or detect damage. Facial recognition by intelligent cameras is increasingly used in public spaces, at airports and in buildings for surveillance and investigation.
The 20 benefits of artificial intelligence in one diagram
At first glance, the benefits of AI seem to lie mainly in saving labor: robotic process automation and physical robots in industry. But if you see AI purely as a way to cut labor costs, you sell the concept short. Applied properly, you can list the benefits for your organization from four angles: people, organization, society and nature.

With AI, organizations can forecast demand much better, optimize their sales channels, improve preventive maintenance, increase production and quality, personalize services and improve customer experiences. In general, algorithms can speed up, improve and partly take over the entire decision-making process.
Accidents with artificial intelligence
Do not only count the blessings of this technology. Some accidents with AI can be seen as teething problems, but they are instructive. Seven well-known examples:
- In 2016 Microsoft launched a chatbot, Tay, that turned into a true racist within 24 hours. The cause was an army of trolls on Twitter that fed the self-learning chatbot a flood of racist and sexist messages.
- Several government agencies use AI to estimate the risk of fraud with benefits and taxes. In some cases this led to evident discrimination and privacy violations.
- In 2018 the first fatal accident with a self-driving Uber test car took place in the United States. In the dark, the system failed to recognize a pedestrian crossing the road in time.
- In 2021 a fire broke out in a British distribution center of an online supermarket after three order-picking robots collided.
- In 2018 the face of a well-known Chinese businesswoman appeared in an advertisement on the side of a city bus. Facial recognition software wrongly identified her as a jaywalker and publicly shamed her.
- In 2017 Facebook stopped an experiment after two chatbots developed a language of their own that only they understood.
- In algorithm-driven stock trading, known as flash trading, an algorithm sometimes runs wild and leaves everyone at their wits’ end.
The moral of the story: depending too much on artificial intelligence makes you vulnerable.
Legislation, ethics and sustainability
AI can put fundamental rights under pressure. Surveillance by governments can violate privacy rights. Far-reaching robotization at home, in healthcare or at work can affect personal autonomy and human dignity. The use of AI models in job applications can conflict with the principle of equality. And algorithmic censorship on social media, such as a shadow ban, can violate freedom of expression.
Algorithms can restrict free speech.
Algorithms in the courtroom, also called robot judges, can limit access to an independent and impartial judge and undermine the right to a fair trial. Bias in the data behind government decisions can lead to discrimination. And citizens keep their right to a transparent and well-founded explanation, also when a decision is driven by an algorithm. The EU AI Act, which entered into force on 1 August 2024, sets rules for this, with the strictest requirements for high-risk systems.
AI can have positive effects on sustainability, such as sharply reducing food waste. At the same time it is in the news because of the large amounts of drinking water and electricity that data centers use. Here too you have to proceed carefully.
As management, put a magnifying glass on the development and use of AI in your organization. That takes time and good advice, and in practice both are often skipped.
Talk to the author
Daan van BeekDiamond Decisions
Discuss your question with Daan
Tell us about the choices you face. In a no-obligation conversation we discover together which decisions make the difference in your organization.
Schedule an introduction →Step 1 of our approach
Diamond Inspirations
The spark for decisions with impact: in an inspiration session you discover which decisions make the difference in your organization.
Explore Inspirations →Frequently asked questions
01
What is the difference between AI and machine learning?
Artificial intelligence is the umbrella term for systems that take over human tasks such as learning, reasoning and deciding. Machine learning is the most widely used technique within AI: algorithms that learn patterns from data. Almost all current AI applications are based on machine learning.
02
What is embedded artificial intelligence?
Embedded AI means that AI and machine learning algorithms are built into devices, robots or components. Think of cameras, mobile phones, thermostats, self-driving cars, refrigerators and spam filters.
03
What does an AI-first strategy mean?
AI-first means that in everything you do, you check whether algorithms can make decisions more accurate, faster or more efficient. When there is a problem, you do not start with a report or dashboard, but first look at AI as a solution.
04
How do you apply AI responsibly?
Check five basic principles: there is no bias, AI promotes the well-being of people and the planet, AI does not harm citizens, AI serves people and not the other way around, and the algorithms are explainable (explainable AI).
05
What is a bias?
A bias is a prejudice. The results of an AI system can be influenced by training data that contain existing prejudices and reinforce them further. People themselves are not free of bias either.
06
What is a training dataset?
A training dataset is a large collection of digital information with which you train an AI system. Usually 60 to 80% of the available data is used for training and the rest for testing and validating the model.
07
How does a small or mid-sized company start with AI?
Start with one decision that is taken often and has a clear effect on margin or customers, such as purchasing, pricing or planning. Check whether you have reliable data for that decision and test a ready-made model on a small scale first. You do not need your own data science team to get started.
Further reading
- Machine learning: how computers learn from data.
- Generative AI: what it is and how to use it.
- Data science and predictive analytics: predicting with data.
- Data quality and the data warehouse: the foundation under every AI application.
- Data ethics: 5 principles of responsible AI.
Sources
- European Union: The EU Artificial Intelligence Act (in force since 1 August 2024).
- Kaplan, A. and Haenlein, M. (2019): Siri, Siri, in my hand: who’s the fairest in the land? On the interpretations, illustrations, and implications of artificial intelligence. Business Horizons, 62(1).
About the author
About Diamond Decisions
Diamond Decisions helps organizations become better by enabling better decisions. We don’t start with data, dashboards or AI, but with the decision that needs to improve, and work back from there to KPIs, data and technology. We do that in four steps: inspiration and alignment, business modeling, analytics and data platform and data-driven decision makers. That way BI and AI don’t remain an IT project, but deliver value every day.

