AI Navigator #11: Five stages of AI use in software development

Contents

Welcome to the eleventh edition of the DOAG AI Community's AI Navigator column!

Bastian Weinlich

is a computer scientist, software architect, project manager and AI enthusiast. He has been developing software in a wide variety of domains for over ten years. Since the advent of GitHub Copilot and Cursor, he has preferred to program with AI support and actively promotes the concept of Vise Coding. He regularly shares his experience at conferences and gives training courses on AI and software architecture.

Semjon Mössinger

is a software developer at WPS - Workplace Solutions and has been developing applications in various languages and frameworks for over ten years. His focus is on test automation, requirements engineering and the productive use of AI-supported development tools such as GitHub Copilot and JetBrains AI Assistant. He shares his practical experience in training sessions, meetups and at various conferences.

In the software development environment, different types of AI use can be identified, which we at WPS GmbH divide into five types:

1. Non-users of AI
2. ChatGPT users
3. Copilot Coder
4. Chat First Coder
5. Vibe Coder

When used correctly, AI usually leads to greater efficiency and better code quality. Nevertheless, many developers have not yet reached their personal optimum when using AI. The extent to which AI supports the development process depends on several factors: the level of experience, the AI tools available and the technologies used.

In the following, we present the five types in more detail – how they work, why they work the way they do and our assessment of each.

The non-users of AI

Those who do not use AI at all work as they have done for the past 15 or 20 years: They write all the program code themselves. They get an overview of existing code without AI support, research via Google, documentation, forums and discussions such as on Stack Overflow or GitHub.

There may be various reasons why they work in this way. The most obvious and probably most common is that their company does not (yet) allow the use of AI, for example, due to security concerns. Others may not yet have tried out AI due to fundamental reservations. Some have tested AI and were disappointed with the results.

Disappointment may be due to a lack of experience in prompting or a lack of awareness of the necessary context – both of which can be quickly remedied. The effect of the so-called Jagged Technological Frontier may also have been responsible for this. The principle describes how AI solves seemingly similar tasks, sometimes surprisingly well, sometimes surprisingly badly.

Whether you benefit from AI is also a question of type. If you have been working with a certain technology for a long time, can research quickly and adapt efficiently, you may not (yet) notice any real efficiency gains from AI. However, a study has shown that top performers in particular tend to have reservations about AI, although paradoxically, according to research, they benefit more.

Finally, the question arises: how recent are the negative experiences with AI support? Our observations show: What did not work satisfactorily three to six months ago may already be robustly operational today.

The ChatGPT users

According to our definition, ChatGPT users only use AI occasionally – about a few times a day. This usage behavior is often observed when ChatGPT is the only AI tool used. The intensity of use depends heavily on the tooling available: If AI is integrated directly into the development tool, as is the case with GitHub Copilot, it is usually used much more frequently.

ChatGPT users primarily use AI for research and thus partially replace googling, reading documentation and searching forums. They also occasionally use ChatGPT to generate code snippets or examples, explain code or analyze error messages.

What is the motivation to (only) use pure ChatGPT or comparable chatbots such as Claude or Gemini? On the one hand, its use is similarly low-threshold as Google search. Secondly, developers retain full control over which of their own code they make available to the AI.

ChatGPT is frequently the preferred tool when a company has not defined clear or official rules for the use of AI tools and their use is in a gray area. To counteract this undesirable situation, we have developed an AI guideline in our company that provides developers with clear guidance.

The co-pilot coders

Copilot coders have their AI programming assistant in constant use. It is directly integrated into the development environment and can therefore be used easily and without any hurdles. Probably the biggest advantage over using ChatGPT is the autocomplete function: While coding, the AI automatically provides suggestions for further code – often with surprisingly good results.

Many such AI tools now exist. The best-known representative is GitHub Copilot, but JetBrains AI Assistant or the AI development environment Cursor also offer impressive functions. Examples include fixing programming errors with a click, automatically generating documentation and seamlessly integrating an AI chatbot with a selection of the best and most up-to-date models.

Copilot coders usually remain strongly code-centric. They deliberately limit the use of AI to a small, self-defined context within their projects, for example when generating a filter function for a list. Inserting a button that triggers the filter or saving the result would then take place in separate, AI-supported steps.

In our experience, this way of working establishes itself quite naturally with dedicated developers who work with a Copilot-like tool. Typically, the decisive factor is that the company provides a business license. According to our observations, this approach marked the maximum feasible with the tools available at the time, until around the end of last year.

The chat-first coders

This approach has only been practicable since the turn of the year 2024/2025. The decisive factor was the introduction of AI agents in common programming assistants, initially in Cursor and since April 2025 also by default in GitHub Copilot and Junie from Netbrains. Apart from a few teething problems and longer computing times, we can recommend AI agents for most use cases – even for those who still work more in the style of Copilot coders.

Chat-first coders treat the source code and the chat with the AI assistant as equivalent elements. You describe a feature in full or in part in the chat – and the assistant then independently develops a plan, adapts the program code, tests changes and makes corrections if necessary. For the above example request to filter a list, the AI would add the necessary button and saving in one step. Our approach here is that we develop a clear expectation before the chat with the AI, which we then use to check the generated result.

For chat-first coders in particular, it is essential to never accept the generated source code unseen or misunderstood. Furthermore, a clear idea of the structure of the program code, both large and small, and careful testing are essential. This is where the responsibility lies with the developers, and our experience shows that this is precisely where the first developments are already taking place: This is precisely where the first quality problems are currently already arising because care is sometimes lacking.

The controlled, tightly managed form of code generation by AI agents is sometimes referred to as vision coding. In contrast to this is the final stage of AI use: vibe programming.

The vibe coders

Put simply, vibe coding means developing software without even looking at the resulting program code: You merely formulate a prompt, observe the result and then write the next – the work is done according to the motto: "The main thing is that it works." Andrej Karpathy coined the term at the beginning of 2025.

This is the direction in which tools such as Devin have been pushing for some time – so far with rather moderate success or a very limited range of applications. Since the introduction of agent mode, this approach can now also be implemented with tools such as GitHub Copilot, Cursor, Codex and Google Firebase Studio.

As things currently stand, this method definitely cannot be used to implement larger, robust software projects – but it can be helpful for small prototypes. Good results can already be achieved, particularly with short scripts, for example in an evaluation code in Python.

Tips for developers and companies

It is foreseeable that AI assistance will become increasingly powerful in the future – and that workflows based on it will become increasingly established among developers. In our view, Lars Röwekamp provides a helpful forecast in his column article "Is AI heralding the end of the software developer species?" If you would like to look at the stages of AI adaptation from an alternative perspective, we recommend an entertaining article on Zef+.

Finally, we would like to make two recommendations to drive forward the use of AI programming assistants:

Developers should specifically consider the areas in which AI can help them with their coding – there are almost always suitable use cases. And they should repeat this evaluation regularly, as the possibilities change in very short cycles. Especially at present, a re-evaluation every three to six months can be worthwhile.

Companies should enable the use of AI programming assistants. Where it makes sense, they should give their developers freedom of choice when it comes to tools, but provide them with clear guidelines on how the tools are to be used.

Anyone who would like to find out more about AI programming assistants and the use of AI in general will have the opportunity to do so at the AI Navigator conference on November 19 and 20 in Nuremberg.

(dahe)