Five Types of AI Systems – And What They Do for Us

Why this article?

Anyone dealing with AI today – whether as a UX professional, product manager or strategist – quickly realizes that there is no such thing as the artificial intelligence. Instead, we encounter different systems with completely different capabilities. And more importantly, with completely different benefits for us as users.

Terms such as GenAI, RAG and agents are only of limited help here. This is because they usually describe technical concepts rather than user experiences. What matters is:

• What do I get as a user?

• What can I hand over – and what do I still have to do myself?

• Where does the system help me – and where are the risks?

  
  

In this article, I present five basic types of AI systems. All five are found – explicitly or implicitly – in modern tools. They differ in function, benefits and typical UX scenarios. And they all have their limitations. It's time to bring some order to the chaos.

1. GenAI – When creativity happens at the touch of a button

 The classic among AI systems: Generative AI (GenAI for short). You enter a prompt and get text, images, code or ideas in return. Usually without any waiting time, often quite impressive.

What is it?

 A trained language model that generates new content based on probabilities. It doesn't remember – it invents. And it does it damn well.

What do you get?

🟢 Ideas, first drafts, creative suggestions, visual variations – all fast, all with a prompt.

UX examples:

• Generate invitation texts for usability tests

• Write down 10 ideas for user feedback questions

• Sketch initial personas based on a few keywords

• Output wireframe description as HTML

  
  

Benefits:

• Starting point for creative or conceptual work

• Time savings for repeatable tasks (e.g., text drafts)

• Idea generator in stagnating processes

  
  

Limitations:

• Hallucinations: GenAI invents facts – even if it sounds plausible, it is often wrong

• Lack of context: Without a precise prompt, there is little useful output

• Style risk: Many texts sound the same – ‘AI speak’ rears its head

Conclusion: Great for creative rough drafts – dangerous if used without checking

2. RAG – When AI knows what we know

Retrieval-Augmented Generation (RAG for short) is an attempt to make GenAI smarter: The AI first searches specifically for relevant information – and then formulates an answer based on that. Sounds simple, but it's powerful.

What is it?

A combination of vector search (e.g. in PDFs, databases or Miro boards) and an LLM that uses this content instead of inventing it.

What do you get?

🟢 Context-based, well-founded answers – with direct references to your data, documents and reports.

UX examples:

• ‘What UX criteria apply to our MedTech products?’ → Quote from internal standard

• ‘What was said about the onboarding flow in the last interviews?’ → Text excerpts with sources

• ‘Which KPIs do we use for NPS tracking?’ → Extracted from workshop documentation

  
  

Benefits:

• Saves time searching for information

• Leads to more reliable answers

• Supports internal knowledge sharing

Limitations:

• If the data is poor, even RAG won't help

• Relevance logic often remains opaque

• High data maintenance effort (structure, timeliness, rights)

  
  

Conclusion: Finally makes large amounts of data usable – as long as it is well-prepared.

3. MCP – When AI can think in multiple stages

Modular Conversational Pipelines (MCP) are orchestrated systems in which a task is broken down into sub-steps and processed sequentially. This enables AI to suddenly analyse, structure and weigh – not just respond.

What is it?

 A chain of specialized AI modules or prompts – e.g. ‘Summarize → Cluster → Prioritize → Visualize’.

What do you get?

🟢 Logically structured results for complex, multi-layered tasks – including intermediate steps

UX examples:

• Automatically code and cluster open interview responses

• Harmonize multilingual user feedback

• Systematically evaluate use cases according to impact and feasibility

• Create benchmark comparisons based on UX metrics

  
  

Benefits:

• Saves time on recurring evaluations

• Reduces errors through standardization

• Makes complex data sets tangible

Limitations:

• Errors in the first step can carry through

• Often unclear why a particular cluster was named that way

• Requires good process definition and prompt strategy

Conclusion: Ideal for structured thought processes – provided you control the intermediate steps.

4. Function calling – when AI uses specific tools

An often overlooked but extremely practical category: tool-enhanced prompts or function calling. Here, the AI calls up specific predefined functions or APIs – and combines human language with concrete logic.

What is it?

 The AI passes tasks to an external module – e.g. a calculator, a weather API, an Excel plugin or a calendar system.

What do you get?

🟢 Answers based on current, accurate data or functions – no guessing, no hallucinations

UX examples:

• ‘How many test subjects did we have in March?’ → API call to CRM

• ‘Calculate the average of these UX scores’ → Excel function in the background

• ‘Show me open appointment slots next week’ → Access to calendar data

• ‘Give me live user numbers from Google Analytics’ → Direct integration

  
  

Benefits:

• Connects voice AI with real functionality

• Provides precise, data-based answers

• Enables automation of standard processes

  
  

Limitations:

• Must be explicitly connected and defined – not a self-runner

• Only works if permissions, formats and APIs are cleanly configured

• Little leeway – no creative input from the AI

  
  

Conclusion: Perfect when accuracy is key – but no substitute for thinking or interpreting.

5. Agents – When AI really does something for us 

The supreme discipline: AI agents. They act independently, execute steps, use tools, ask questions – and often deliver a result, not just an answer.

What is it?

A (semi-)autonomous assistant that plans tasks, makes decisions and uses tools – without every action having to be prompted individually.

What do you get?

🟢 Results, not suggestions. The AI takes over – and gets back to you when it needs something.

UX examples:

• Creation of a complete slide deck summary based on a study

• Consolidation of user feedback + metrics + logs into a journey map

• Better focus on analysis & strategy

• Planning a UX workshop with suggested methods, duration, agenda, participant logic

• Deriving recommendations for action from qualitative and quantitative data

  
  

Benefits:

• Relief from routines and multistep tasks

• Interaction feels more like a project partner

• 
  

Limitations:

• High loss of control – what is actually happening?

• Decisions are often not transparent or explainable

• Potential for errors due to ‘incorrect’ assumptions or poor tool execution

  
  

Conclusion: Extremely powerful – but to be used with caution. Responsibility remains with humans.

Conclusion: What fits when?

It's not about which system is ‘best.’ It's about what fits which problem.

• Need ideas? → GenAI

• Looking for internally distributed knowledge? → RAG

• Want to evaluate data systematically? → MCP

• Need exact figures or real-time data? → Function Calling

• Want to hand over an entire task? → Agent

💡 The real magic often happens when you combine things:

A RAG system retrieves content, an MCP structures it, an agent prepares it – and function calls ensure that the data is correct.

And that's exactly why it's worth taking a differentiated view. Not only because there are technological differences – but because it enables us as users to make better decisions:

→ Which AI do I use for what?

→ And where am I better off staying in the driver's seat?