How does AI work, and what are the branches of AI?

Ask ten people what AI is, and you will hear ten answers: a robot, a chatbot, a recommendation on a streaming site, a self-driving car. All of them are partly right, which is exactly why planning for AI in an organisation can feel slippery. Before deciding where AI could help you, it is worth knowing two things: how it works at a basic level, and which branch of AI fits which kind of problem.

This article gives you both, without equations. It follows the glossary in the previous article, so the terms should already feel familiar.

The core idea: learning from examples

Traditional software follows rules that a programmer writes by hand. If the customer is over 60, apply the discount. If the file is larger than 10 megabytes, reject it. This works well when you can describe the rules, and fails when you cannot. Nobody can write down, rule by rule, how to recognise a cat in a photograph, a flooded field in a satellite image or a sarcastic sentence.

Most modern AI takes a different route. Instead of writing the rules, you show the system many examples, and it works out the rules for itself. Show it thousands of photographs labelled cat and not cat, and it learns which patterns of pixels tend to go with cats. The result is not a list of rules a person can read, but a model: a large collection of numbers that, together, turn an input into an output.

How a model learns, step by step

1. 01Collect examples. These are the data: photographs, documents, sensor readings, past decisions. The quality and variety of this data shapes everything that follows.
2. 02Prepare and label them. For many tasks, each example needs a correct answer, such as flooded or not flooded. This is often the slowest and most expensive step.
3. 03Make a first guess. The model begins with random settings, called parameters or weights, so its first answers are poor.
4. 04Measure the error. For each example, the system compares its answer with the correct one, and calculates how wrong it was.
5. 05Adjust the settings. A mathematical method nudges the parameters in the direction that reduces the error. This is the core of training, and gradient descent, described elsewhere in this library, is the best-known version.
6. 06Repeat many times. Millions of small adjustments gradually turn a useless model into a useful one.
7. 07Test on new examples. The model is checked on data it has never seen. If it performs well there, it has learned something general. If it only does well on its training data, it has memorised, which is called overfitting.
8. 08Use and monitor it. In use, the trained model produces answers, called inference. Over time, the world changes, so the model must be watched and sometimes retrained.

That is the whole loop. The remarkable abilities of modern systems come from doing it at enormous scale, with huge amounts of data, very large models and a great deal of computing power.

Inside a neural network

Most of today’s headline systems use neural networks. Imagine a stack of layers. The first layer receives the input, for example the pixels of an image. Each layer passes its results to the next, and each connection has a weight that says how much one unit should influence another. Early layers pick up simple features, such as edges. Later layers combine them into larger ideas, such as eyes, wheels or roofs. The final layer gives the answer.

Training adjusts millions or billions of weights until the network’s answers become reliable. Nobody sets these numbers by hand, and no single number means anything on its own. The knowledge lives in the pattern of all of them together, which is why these systems can be hard to explain.

What AI is not doing

• It is not thinking like a person. It finds statistical patterns, and does not understand the world as you do.
• It does not know what is true. A language model predicts plausible next words, so it can state falsehoods fluently.
• It does not learn from you in the moment, unless the system is specifically designed to. Most deployed models are fixed until their makers update them.
• It is not neutral. It reflects the data it was trained on, including gaps and biases.

The branches of AI

AI is better thought of as a family of fields that share a goal, which is to make machines do tasks that normally need human intelligence. The branches below overlap, and real products often combine several.

Machine learning

Machine learning is the branch that learns from data, and it underpins most of the others. It has three main styles.

• Supervised learning learns from labelled examples. It is used to classify, such as spam or not spam, and to predict numbers, such as next month’s demand.
• Unsupervised learning looks for structure without labels, for example grouping customers with similar behaviour or spotting unusual readings.
• Reinforcement learning learns by trial and reward. A system tries actions, receives a score and improves over time. It is used in game playing, robotics and some scheduling problems.

Deep learning is a powerful form of machine learning that uses large neural networks. It is behind most of the recent progress in vision, speech and language.

Natural language processing and language models

This branch deals with human language: reading, writing, translating, summarising and answering questions. Modern large language models are trained on vast amounts of text to predict what comes next, and as a result they can draft, explain, translate and write code. Common uses include assistants that answer questions from your documents, tools that summarise long reports and systems that sort and route messages.

Computer vision

Computer vision lets machines interpret images and video. It recognises objects, reads text, measures things, detects change and segments a scene into parts. In geospatial work, it finds buildings, roads, crops and floods in satellite and drone imagery. In industry, it inspects products for defects. In immersive technology, it helps headsets understand the room around you.

Speech and audio

This branch converts speech to text, text to speech and recognises sounds. It powers transcription, voice assistants and call-centre analysis. Combined with language models, it allows natural spoken conversations with a system.

Generative AI

Generative AI creates new content, such as text, images, audio, video, code and 3D models, based on patterns learned from existing examples. It is built on machine learning, and overlaps with language and vision. It is excellent for drafts, ideas and variations, and needs human review wherever accuracy matters.

Knowledge representation, reasoning and planning

Before machine learning dominated, much AI work focused on writing knowledge down explicitly, in rules, graphs and logic, and reasoning over it. This branch is still valuable where decisions must be explained, rules are known and errors are costly, such as eligibility checks, configuration and scheduling. Planning algorithms work out sequences of actions to reach a goal, and are used in logistics, routing and robotics. Many modern systems combine learned models with explicit rules, to get flexibility and reliability together.

Robotics and autonomous systems

Robotics combines sensing, decision-making and physical action. Warehouse robots, drones, farm machines and self-driving vehicles all draw on vision, planning and control. The difficulty lies in the real world, which is messy, unpredictable and unforgiving of mistakes.

AI agents

An agent is a system that pursues a goal over several steps. It decides what to do next, uses tools such as search or a database, checks its results and tries again if needed. Agents typically combine a language model with planning and tool use. They are promising for multi-step tasks, and need careful limits and human oversight because errors can compound.

Predictive analytics and recommendation

These are among the most widely used, least glamorous forms of AI. They forecast demand, flag likely fraud, estimate risk and suggest the next product, article or route. Often a simple model on well-prepared data beats a sophisticated one on poor data.

Choosing a branch for a problem

• Predicting a number or a category from past records points to machine learning on structured data.
• Reading or answering from documents points to language models, often with retrieval from your own knowledge base.
• Finding things in images or satellite data points to computer vision.
• Following known rules with a clear audit trail points to rules-based systems and knowledge representation.
• A multi-step task that crosses several systems may call for an agent, but a fixed workflow is often cheaper and safer.
• Creating drafts, images or variations points to generative AI, with a person reviewing the result.

Narrow AI and general AI

Everything in use today is narrow AI: systems that are strong at particular tasks and weak outside them. A model that reads scans well cannot draft a contract. A language model that writes fluently cannot reliably measure a field. Artificial general intelligence, a system with human-level ability across almost any task, remains a goal and a subject of debate, not a product you can buy. Plan for what exists now.

What this means for your organisation

• Data is the raw material. Without relevant, reasonably clean examples, a learning system has nothing to learn from.
• Different problems need different branches, so start with the problem.
• Models make mistakes in patterned ways, so design checks, human review and monitoring from the start.
• Systems age, because the world changes. Budget for upkeep, not only for building.
• A simple, explainable approach is often the right first step.

Frequently asked questions

Is AI the same as machine learning?

No. AI is the broad goal of making machines do intelligent tasks. Machine learning is the main method used today, in which systems learn from data. Some AI uses explicit rules and reasoning instead.

Does AI really learn like humans do?

Only loosely. It adjusts numbers to reduce errors on examples, which needs far more data than a person usually needs, and it has no understanding or common sense in the human sense.

Why do language models make things up?

Because they are built to produce plausible text, not to check facts. Connecting them to trusted documents and requiring sources reduces the problem, but does not remove it.

Which branch of AI is best for a small organisation to start with?

Usually practical, well-defined uses: a document question-answering assistant, a simple forecast on your own records or automating a repetitive, rule-based task. Start where the data already exists and the benefit is easy to measure.

Do we need huge amounts of data?

It depends on the approach. Training a model from scratch needs a lot. Using a pre-trained model, adapted with a modest amount of your own data or documents, needs far less.

AI works by finding patterns in examples, and it comes in many forms suited to different jobs. With that picture in your head, you can listen to vendors critically, match problems to the right branch and plan an AI project that is grounded in what the technology actually does.