Artificial intelligence is a useful toy

Some years ago OpenAI released the first version of ChatGPT to the public. They made bold claims about their artificial intelligence and many people were impressed. For me, the product was quite handy and impressive from a technical perspective, but mostly a useful toy. Small coding tasks, text corrections or summaries of given information were most of the tasks I handed over. Whether it was ChatGPT, Google Gemini, Perplexity or Kagi assistant, the interaction with each of them was limited to an input field and attachments. More complex tasks I had to cut down into small chunks and then process them step by step. AI wasn’t able to get the big picture of a project.

Don’t get me wrong, even this was useful and made it possible for me to tackle projects, I would’ve needed far more time otherwise. But this approach on AI wasn’t intelligent at all. Just a big algorithm nobody was able to understand anymore, trained with most content of the world wide web, slicing a question in pieces and calculated word for word the mathematically most fitting response. The answer of an LLM is just the statistically most likely combination of words.

Considering, that correct facts and data are the majority of content in the WWW it works rather well, especially for commonplace stuff. But for unknown topics, current news and more complex contexts, this approach wasn’t only useless, but even dangerous. LLMs don’t know, if their response is correct. It’s just the best response the LLM can give and it will give it, always. In the meantime, there are ways to stop them giving an answer just to answer. The confidence threshold needed for an answer to be valid was adjusted higher.

But even then I wasn’t able to follow the hype around LLMs. I saw the potential, but I wasn’t as enthusiastic about the state of AI until mid 2025 as many people were. And even today, I assume around 95% of people only use the simple LLMs in the browser, celebrating the progress and thinking this is artificial intelligence.

Claude AI shows me the way

While these models and services were just a rock to hammer in nails for me - better than no tool - Claude AI has surprised me the last weeks. It’s not a nail gun, but a solid hammer. A hammer that I use shockingly regularly and that supports my personal and work tasks a lot. I’m not only more productive, I also can do stuff I couldn’t do before.

Complex coding projects? No problem. Claude isn’t limited to edit just one file, it could work through whole project folders. Multi-level analysis? Easy. File sorting and modification? Claude can do this.

To be honest, I didn’t expect how fast Claude Code and Claude desktop convinced me. The agentic approach, that allows for multi-level and multi-step actions, is my game-changer to make LLMs useful. This doesn’t make them real artificial intelligences, but really helpful tools. Like an untrained assistant that does the monkey work that you always should verify. But in the end it saves more time than needed for verifying.

Actual examples

As Product Owner I have access to our code and I have a basic understanding of our code. But the ERP code is complex, nested, divided and full of dependencies. To understand an “easy” function in a workflow I have to crawl for hours through the code. Or I just let Claude Code crawl through it, connected with our ticket system Jira and knowledge base Confluence. In just a few minutes, Claude checked the code, stories and documentation, then created a Confluence page with the answers to my questions. With basic IT knowledge the answers can be understood, but there are also enough details to go into discussion with a real developer.

People who know Confluence will also know, that the built-in search is no masterpiece. Even with Rovo, the newly implemented AI assistant by Atlassian, the search results couldn’t satisfy me. Claude, on the other hand, searches through Confluence, follows links to Jira and external sites and provides a great summary to topics. Not perfect, not without mistakes, but mostly solid and always reasonable and verifiable.

But also in private I use Claude Code a lot nowadays. I have implemented it as a sidebar in Obsidian. When you check my older blogposts you will find, that I had some troubles with nearly all notebook applications. My thinking is too fast and erratic to write more than some bullet points. So my notes were short, chaotic and therefore not helpful afterwards. But how would it be with a tool, that could take my chaos and convert it into structured, readable notes? Send “Rework the structure and wording of this note” to Claude and 30 seconds later I have a structured note that will be handy in future.

With this blogpost Claude has supported me too. Not for the writing itself, but for the structuring of the post. I write a note with the topics and ideas for a blogpost and let Claude bring structure to it. If you only knew how often I had to rewrite or rework blogposts in the past just because their structure was chaotic. Additionally I let Claude proofread my blogposts to improve spelling, syntax and forgotten words. But I typed this whole blogpost myself.

Hopes and worries

My experiences with Claude in the last weeks were an emotional rollercoaster. On one hand I’m extremely impressed by the capabilities and independence of it, especially with complex topics and tasks. On the other hand, this AI has reached a level that really could make some jobs redundant. Especially jobs that are based solely on data processing and presentation seem to be in danger. Consultation in the local electronics store may already be better done by AI than by a bored student. But also everything around research and summarisation. Why do I need a Controller who creates and updates reports into an Excel template? Let AI query and visualise the data itself. Always the newest numbers, always the numbers I need.

With creative jobs, I don’t see a real problem. Current AI isn’t creating anything and they don’t understand anything. A good developer will be supported by agentic AI and most likely more efficient, but current AI will have troubles creating sustainable complex software. Also my work as Product Owner won’t be taken away by AI, but it will take over the monkey work. At least for now.

But maybe we are soon at a point where brain work will be done fully by AI, just a small group of people will manage them and give them the tasks. What will be left over? Manual labour and fundamental science. For the first, I’m too left-handed, for the latter, too dumb. We’re living in thrilling times.