When I started writing this post (beginning of December 2025), I was on a sick leave and decided to write one post to make the most out of that time. Here it is, after a month of delaying it, finishing the last edits at 3am and publishing it finally. The content of this post is more of a generic and semi-unstructured collection of thoughts on my 2025 experience building AI-powered products and a few tangent thoughts.

Random fact: All my posts in 2025 was about AI.

AI Unlocks Orders of Magnitude

We have been experimenting a lot with AI, and there was a clear pattern. Any AI feature we developed increased the user productivity. It is the same efficiency increase when we got calculators for complex calculations, or the way Quantum Computers arguably will change the speed of computation radically. AI is not just a new computer, new phone, or a new car. It is the first computer, the first phone or the very first car ever developed.

Intelligence and generative capability transforms how we approach the problems. Image generation is becoming so advanced that one does not need any photoshop skill, or someone with that skill. Code-writing is another example. My sales colleagues develop prototypes in a matter of an hour and showcase how they’d envision to see the product working. No one could’ve envisioned it few years ago.

Coding Agents can develop applications incredibly quickly given the ambigious requirements. With the help of MCPs, Skills, Plugins, memory extensions these tools bring your product from idea to reality in very short time. One cool thing I have been using is opening multiple tabs of CLI and running multiple agent sessions in parallel using Git worktrees that allows to check out multiple branches and develop each feature in isolation and then merge them at the end. I am simply amazed.

The AI Excitement Asymmetry

On the one hand, there is a high visibility of AI everywhere. All platforms, news, websites, blogposts like this, products bombard people with AI marketing. They show as if “their AI” can solve everything and people buy it. The reality is that those ads do not turn into an actual value generation.

On the other hand, we are trying to build it in-house, but it does not bring the same level of excitement. “But ChatGPT is faster” is what we got as a feedback. It is difficult to bring people’s perception away from the abstract ads world back to what’s feasible and viable.

AI as a Feature

In an enterprise setting, the most optimal usage of the AI product that we’d developed was not a standalone product. It was rather an integrated feature within existing processes. In general my main product hack is to have it as part of my usual routine. Instead of opening another link to do the job, I’d rather want to have it as a step within my current workflow. Our users also have this expectation. Someone asked me one day, cannot we integrate the AI into the excel? then I can give my prompt there and it’d do the job. Someone else mentioned it would be great if AI can record my meetings (regardles online or on site) and provide a summary to me, directly on my laptop.

Another project I built for my own was an intelligent file manager. I developed a standalone user interface to talk to my files, organize them, and do some actions on them. I know that it’d be even better if Apple would’ve had it natively integrated as part of their “Apple Intelligence”, then I’d not need an extra application on Docker to manage my files. Here my main takeaway is that AI would work better as a feature within the product, and not as standalone tool.

AI development at a lightning speed

It is crazy how fast the industry is moving. Every week there is a new model, new method, new tool, new idea made public that changes how we use AI. It signals how broad and robust the AI ecosystem will become. At the same time, many of the developments are half-baked, incomplete, dangerous to use, and before they can be matured the next hype begins and gets all the attention.

I do not consider myself to be the most up-to-date person, neither do I want to be one. I adopted Claude Code very lately, and I am happy about it because the product feels very mature in comparison to what people were writing about it at the beginning. Now there are skills to learn. It can become very overwhelming if one digs deeper and faster. I take it steady and easy.

Even more AI in 2026

I am excited about the next year. As AI becomes a part of our daily tool stack, we will see even more applications for it. With the maturity of coding assistants (like Claude Code), we will be able to turn our ideas into actual products (check the ClaudeAI subreddit for the apps built with Claude Code).