Types of Artificial Intelligence: A Practical Business Guide

Types of artificial intelligence become easier to understand when we separate the technology, task, capabilities and autonomy of a system. Machine learning, generative AI, language models and AI agents are not four mutually exclusive options. One application can combine all of them.

Imagine an assistant that reads an enquiry, finds the customer's details and prepares a quotation. Analysing the message involves natural language processing; drafting can use a generative model; looking up the CRM requires an integration; sending the quotation depends on the application's permissions. The label “AI” alone describes neither the architecture nor the level of control.

This guide shows how the main categories fit together and how they inform the choice of custom applications with AI integration. The examples explain possible solutions, not results attributed to customers or performance tests carried out by WebHat.

Why are there several classifications of artificial intelligence?

In NIST definitions, an AI system produces outputs such as predictions, recommendations or decisions for a set of objectives, influencing a physical or digital environment. This covers much more than chatbots, from image analysis to planning actions.

Confusion starts when different questions are mixed together. “What data does it process?” is not the same as “How was it trained?” or “Is it allowed to change a record?”. A useful comparison therefore keeps the criteria separate.

Example: an assistant for company documents can be based on deep learning, specialised in language, generative, connected to a knowledge base and limited to proposals requiring approval. These descriptions are compatible, not competing. None replaces checks on access, data and the final answer.

How do AI, machine learning and deep learning relate?

Artificial intelligence is the broad field; machine learning is a family of methods within it; deep learning is a branch of machine learning based on neural networks with multiple layers. AI systems can also use search, planning or explicit knowledge representation. However, an automated rule is not, by itself, necessarily AI.

In machine learning, an algorithm adjusts a model using data or experience. Instead of manually specifying every combination that could indicate an urgent enquiry, a classifier can be trained on labelled examples. It must then be evaluated on examples that were not used to adjust it.

Deep learning uses successive layers to learn representations of data. CNNs are a family used in image analysis; RNNs handle sequences through recurrent state; Transformers use attention mechanisms. Architectures serve different purposes and can be combined. The number of layers alone does not determine whether an application meets a business requirement.

Supervised, unsupervised and reinforcement learning describe learning approaches, not three commercial sectors. There is also self-supervised learning, where training targets are constructed from the data's structure, and semi-supervised approaches. A project can involve several stages: pretraining on a broad corpus, adaptation and evaluation on the specific task.

Types of AI by the data and problems they handle

Natural language: text, search and conversations

Natural language processing, or NLP, covers methods for analysing and generating language. A business can classify enquiries, extract product names, identify the intent of a message or produce summaries. An LLM is a large language model; not every NLP solution uses an LLM.

To answer questions using internal information, a model can receive passages retrieved from approved sources. This is the approach explained in our guide to an AI chatbot connected to company documents through RAG. Answer quality also depends on the selected documents, their versions and the user's right to access them.

Computer vision: information from images and video

Computer vision can classify an image, locate objects, delineate regions or track movement. In a warehouse, tasks might include counting parcels or checking whether a label is present. Detecting an object, reading text and recognising a person's identity are different problems.

Results must be tested with the lighting, angles, products and cameras used in the real setting. A demonstration on selected photographs does not automatically describe behaviour across every working shift.

Audio: transcription, speech synthesis and sound analysis

A speech-to-text system converts speech into text; text-to-speech generates speech from text. A voice assistant also needs conversation management, data rules and connections to the available operations. Recognising speech and confirming a booking are separate components.

Structured data: estimates, classifications and anomalies

Models can use tables of sales, stock or service times to estimate values, group behaviour or flag unusual patterns. Classical methods are useful here too, not only generative models. The choice starts with the target being measured and the historical data available.

Robotics and embodied AI

In a robotic system, perception is connected to actions in the physical environment. Sensors, planning, motion control and safety mechanisms work together. “Embodied AI” describes research and systems where interaction with the environment matters to behaviour; it means more than adding a chatbot to a robot.

Multimodal AI

Multimodality means working with several forms of information, such as text and images. A workflow can analyse a product photograph alongside its description and code. The inputs and outputs supported by the specific model must be stated: image support does not automatically imply support for every file format, audio or video.

ANI, AGI and ASI: classification by capabilities

These concepts describe the breadth and level of abilities, not the infrastructure running an application. In Levels of AGI, researchers propose assessing generality, performance and autonomy separately. The framework supports explicit comparisons without treating a label as a measurement.

ANI, artificial narrow intelligence, targets defined tasks or domains. A model can transfer knowledge between tasks or languages while remaining limited in other respects. Narrow AI should therefore not be defined through an assumption that transfer is impossible.

AGI refers to general capabilities for learning and problem-solving, compared with human performance across a broad range of activities. The criteria must be named: which tasks, which level, under what conditions and with how much assistance. A high score on one test does not answer all these questions.

ASI is the concept of superintelligence substantially exceeding human capabilities across a broad range of activities. In this guide it is a conceptual category, not a feature selected from a software subscription. Choosing a business application depends on measurable tasks, not a forecast date for AGI or ASI.

Reactive AI, limited memory, theory of mind and self-awareness

Another taxonomy found in introductory material concerns how a system uses state and interaction. It is a conceptual explanation, not a universal set of four exclusive technical classes or four versions every model must pass through.

• Reactive systems: respond to currently available information without maintaining a session history. “No memory” in this description does not mean that code, rules or learned parameters are absent.
• Limited-memory systems: use some history or relevant state for the current decision. A conversation context, working state and persistent database are different mechanisms, even though each can supply information about the past.
• Theory of mind: concerns representing other participants' beliefs, intentions and perspectives. Studies involving LLMs evaluate specific tasks and protocols; their results need to be interpreted in those terms.
• Self-awareness: concerns questions about subjective experience and representing one's own existence. Fluent statements about emotions or the self do not, on their own, establish these properties.

For a business application, the practical questions are where state is stored, who can read it and how long it is needed. Saving a conversation does not automatically mean that the model updates its parameters through training. We inspect the actual mechanism rather than assuming “continuous learning” from every message.

Generative AI, discriminative models, LLMs and SLMs

Generative AI produces content: text, images, sound, code or other outputs, depending on the model. In a company it might prepare a summary or a draft reply. Exact data, calculations and actions must be connected to the application's sources and rules.

Discriminative models learn the relationship between an input and a target, such as an enquiry category. The statistical distinction between generative and discriminative is not the same as “creates versus predicts”: generative classifiers also make predictions. The scikit-learn documentation on discriminant analysis illustrates classification using the distribution of data within each class.

A foundation model is trained on broad data and can be adapted to downstream tasks. The definition systematised by Stanford CRFM describes its role as a base for other applications. “Foundation” does not automatically mean linguistic, multimodal or authorised to act on company systems.

LLM and SLM describe language models of different sizes. Selection depends on the task, quality in the required language, memory requirements, response time and hosting arrangements. We do not apply one universal parameter threshold to every model described as “small”.

An SLM can be trained directly or produced through other approaches; distillation is one technique, not a mandatory definition. Quantisation reduces the numerical precision of representations and is not synonymous with reducing the parameter count. These differences affect implementation and must be assessed for the chosen model.

What does an AI agent add to a model or chatbot?

An agent perceives information from its environment and selects actions. The classical treatment in Artificial Intelligence: A Modern Approach includes reflex, model-based, goal-based and utility-based agents; learning can be added to these architectures.

In applications using LLMs, an agent can use tools to search, calculate or request operations through APIs. The model proposes steps, but the application must check identity, permissions, arguments and the outcome of each operation. An available API does not imply unrestricted access.

A chatbot can explain the contents of a quotation. An agent connected to systems can prepare it using approved data if the necessary tools and permissions are available. Sending it to the customer can remain subject to explicit approval. These distinctions are developed in our guide to chatbots versus AI agents for business.

Autonomy is a design decision. An application can allow reading only, accept proposals or execute clearly bounded operations. Increased autonomy needs accompanying tests, logs and stopping mechanisms. It does not establish that the model is AGI.

Three examples: how types of AI combine in a company

1. Documents become validated records

A PDF might require OCR to obtain text, NLP to extract fields and deterministic rules to verify formats or totals. A generative model can explain discrepancies, while the user confirms fields before they are saved. This is a hybrid system, not a single “read documents” feature.

2. A CRM prepares the next action

An application can group enquiries, summarise history and propose a reply using customer information accessible to the user. A classical model can estimate a probability while an LLM drafts the text. In a custom CRM for a company, statuses, responsibilities and approvals remain explicit product rules.

3. Visual checks within a warehouse workflow

A camera can supply images to an object detector; the result is matched to an order and referred to an operator when a discrepancy appears. Counting objects does not automatically replace checks on codes or contractual quantities. Projects involving business automation with AI integration connect perception to rules, exceptions and records.

The context photograph shows a human checking stock in a warehouse; the workflow described is an explanatory architecture example. In all three cases, AI contributes to a stage while the application organises the complete process.

How do you choose the right type of AI for a business?

1. Define the outcome: classification, estimation, extraction, generation or execution of an operation. Express it in a form that can be tested.
2. Check the data: format, volume, quality, access, freshness and usage rights. More data does not automatically mean suitable data.
3. Compare against a simple solution: clear rules or a classical model can solve a bounded task. Compare alternatives on the same set of cases.
4. Define control: what the system may read, propose and change, who approves and how errors are corrected.
5. Test operation: response time, cost per accepted task, integration, monitoring and the ability to change the model.

These choices form part of software development for company processes. Model selection follows process definition and is checked through concrete tests. Internal information also requires the controls described in our guide to data protection when integrating AI.

How do you measure whether the solution works well?

The metric must match the task. In classification, we examine which categories are confused; in extraction, we check correct, incorrect and missing fields; in search, we assess whether the necessary source is retrieved. For generation, we evaluate compliance with the request, support for claims and usefulness, not fluency alone.

Teaching example using illustrative data: we check 200 expected document fields. If 190 are extracted exactly and 10 are incorrect or missing, the proportion of correct fields is 190 ÷ 200 = 95%. This does not show how many documents are entirely correct or whether an error concerns a critical total or a secondary note.

Report error types, human verification time and cost per accepted outcome separately. Maintain a test set and repeat evaluation when the model, instructions, documents or integration change. For the full budget, see our guide to AI application costs and maintenance.

Mini-glossary: model, training, inference and memory

• Model: a component that transforms particular inputs into outputs according to its parameters and mechanism.
• Training: adjusting parameters using data and a learning objective.
• Inference: using a model to obtain an outcome for a new input.
• Context: information supplied to the model for the current request.
• Application memory: information retained and retrieved by the system under access and retention rules.
• Agent: a system selecting actions in relation to its information and objective; it can use one or more models.

Frequently asked questions about types of artificial intelligence

How many types of artificial intelligence are there?

The number depends on the criterion. ANI, AGI and ASI classify capabilities; NLP and computer vision describe processing domains; generative and discriminative describe model families. A system can carry several labels at once. A useful list keeps the criterion visible.

Are generative AI and predictive AI opposites?

No. “Predictive” describes use for estimation or classification, while “generative” relates to modelling distributions and producing content. Generative models can be used for classification, and an application can combine generation with predictions and deterministic calculations.

Is a language model automatically an AI agent?

No. A model is a component. An agent includes mechanisms for selecting actions and interacting with an environment; the application defines its tools, state and permissions. A model drafting a reply does not automatically gain permission to modify a CRM.

Does a chatbot learn from every conversation?

Behaviour depends on implementation. Retaining history or retrieving information does not, by itself, constitute training. Updating parameters requires a separate process; conversation usage, retention and access must be checked in the specific system's configuration and documentation.

Is a larger model always the right choice?

Selection is based on task evaluation, not size alone. A smaller model may cover a bounded requirement using suitable resources; another requirement may need different capabilities. Compare quality, time, memory, hosting and cost per accepted outcome.

Where does an AI project for a company begin?

With a clear process, an accountable person and representative examples. Define the input, expected output and who confirms the result. Then choose the technology, check the data and measure performance on a set of cases before extending its use.