AI discussions are often blurred by the fact that the term artificial intelligence is used to mean everything from nearly all modern software to chat-based services like ChatGPT. The definition in the EU AI Act is so broad that, taken literally, it covers a large share of today’s software, analytics, and automation technologies: if a system infers from inputs how to produce predictions, content, recommendations, or decisions, it already falls under AI. This is why we need more precise terminology.
I propose structuring AI along two axes. The first describes what the system does: does it generate, or does it analyse, predict, and optimise? This is the classical way of categorising AI applications. As generative solutions have become widespread, a second axis has emerged: the maturity of the generative application. Whether it relies solely on a foundation model, or whether it also incorporates search, external data sources, or agentic workflows. In practice, the same solution usually belongs to multiple categories at once.
Generative and agentic AI are typically built on a foundation model, a general-purpose model trained on large-scale data that can be adapted to many tasks. However, a foundation model alone is not yet a product. Only when it is combined with instructions, a user interface, retrieval, tools, memory, and safety mechanisms does it become an assistant or an agent.
By purpose
Generative AI
Creative Generation
Use case: Producing text, images, audio, video, and code.
Description: Generates new content by modelling patterns from existing data. The output can be useful, original, and convincing, but not automatically true. A particularly important application is software development, where AI performs strongly due to the logical structure of programming languages.
Interactive Generation
Use case: Conversation, ideation, guidance, drafting, and embedded assistants in user interfaces.
Description: Adapts content based on the user’s query, conversation history, and provided context. This is the form of generative AI most people recognise as a chatbot.
Simulating Generation
Use case: Scenarios, design alternatives, concepts, and situations that do not yet exist.
Description: Produces alternative futures or solution spaces for human evaluation. Rather than merely predicting the most likely continuation of the past, it helps explore new possibilities.
In this classification, algorithm families such as GANs, diffusion models, and transformers are best treated separately as model technologies. They describe how something is implemented, not what it is used for.
Operational AI
Analytical AI
Use case: Classification, anomaly detection, segmentation, and data analysis.
Description: Identifies patterns in data that have not been explicitly encoded as rules by humans.
Predictive AI
Use case: Demand forecasting, risk modelling, predictive maintenance, and other probabilistic estimates.
Description: Estimates what is likely to happen next based on historical observations.
Optimising AI
Use case: Routing, resource allocation, production control, energy optimisation, and scheduling.
Description: Finds better solutions to known objectives, such as lower cost, shorter time, or higher output.
Recommending AI
Use case: Decision support, personalisation, recommendation systems, and prioritisation of alternatives.
Description: Goes beyond prediction by suggesting what should be done to achieve a given goal. A long-standing application area is targeted digital advertising.
Autonomous AI
Use case: Robotics, self-directed digital processes, industrial automation, and vehicle applications.
Description: Makes and executes decisions partially or fully without continuous human control. As a result, safety and reliability requirements are exceptionally high in this category.
By implementation approach
Foundation model-based AI
Foundation model-based AI builds on general-purpose models. The defining feature is versatility: the same model can write, summarise, translate, code, analyse images, and answer questions. However, modern platforms rarely bind the user to a single model. Instead, they include an orchestration layer that selects the most suitable model or reasoning mode for each task. For example, ChatGPT partially exposes this through automatic mode switching, while different versions of Copilot, Claude, and Gemini implement similar principles through different research, search, and agent features, even if the actual model selection is often hidden behind the interface.
Retrieval-augmented AI
Retrieval-augmented AI is grounded in external sources. In practice, this includes web search, document reading, enterprise data access, or integrations with other systems. This category also includes grounding and the generative concept known as RAG (retrieval-augmented generation). The key idea is that the model does not rely solely on its training data but incorporates up-to-date or organisation-specific information.
This is now a core operating model for modern AI platforms. ChatGPT can automatically search the web and use files and connected applications. Microsoft 365 Copilot combines web grounding with work grounding, secured via Microsoft Graph. Claude retrieves information from both the web and connected tools. Gemini Deep Research uses Google Search by default and can also access Gmail, Drive, files, and other sources. Le Chat and Perplexity similarly provide multi-step, source-grounded research workflows.
Agentic AI
Agentic AI is an operating paradigm rather than a model type. An agent plans tasks, breaks them into steps, decides when to retrieve additional information, when to use tools, and when to ask the user for confirmation. As a result, it can browse websites, fill out forms, draft and send emails, modify spreadsheets, interact with enterprise systems, or trigger workflows.
At this point, the chatbot metaphor becomes too limited. ChatGPT Agent, Copilot agents, Claude’s research architecture, and Gemini Agent all represent a shift toward goal-oriented task execution. When an agent delegates subtasks to other agents, we speak of multi-agent AI, where one agent coordinates and others execute tasks in parallel.
Modern AI products such as ChatGPT, Microsoft 365 Copilot, Claude, Gemini, Le Chat, and Perplexity are therefore no longer best understood as single models. They are AI platforms with orchestration layers that interpret user prompts and select the appropriate processing mode: quick response, deeper reasoning, web search, file analysis, enterprise data grounding, or agentic execution. This shift is fundamental also from a terminology perspective.
Task-specific AI
The examples above focus on widely available AI products. However, most real-world business use cases are designed for specific processes. In these cases, different AI types are typically combined as needed, for example, to build a digital twin, optimise a manufacturing process, or deliver high-quality customer service.
