For the last couple of years, we’ve all been hearing the same thing over and over again.
AI Agents will run businesses autonomously.
Honestly, I think there is a problem with this idea.
LLMs are probabilistic by nature, while business systems usually expect deterministic outcomes. In real business life, mostly correct is sometimes still a problem. Invoices cannot randomly fail. Approval flows cannot suddenly behave differently. Customer data cannot unpredictably change.
That’s why I’ve never fully believed in the idea that autonomous AI agents will suddenly run entire companies without boundaries or human involvement.
But at the same time, I’m also seeing something very real.
AI coding tools are becoming incredibly powerful.
Recently, one of our customers asked whether CloudOffix had a social media management product similar to Hootsuite or Zoho Social.
At that moment, we did not. I told them that if this was a deal breaker for using CloudOffix, we could probably implement it.
Instead of forwarding the request to our Product and Development teams, I wanted to personally test how far today’s LLMs could actually go on the coding side. So I started building it together with Claude.
The experience was honestly impressive. Not because the process was smooth though. It definitely was not. There were broken flows, errors, strange outputs, missing logic and many moments where I thought “okay, this will probably not work.” But then something interesting happened.
Every time we tested, saw the logs, refined the prompts, regenerated the code and tested again, the product became better.
The AI was not magically building a perfect system from scratch. It was improving through iteration.
After around 3 or 4 hours, we had a surprisingly usable social media management product running inside CloudOffix. That part honestly surprised me a lot.
Then I started thinking about why this worked so well.
I think there were two main reasons.
The first one was iteration itself.
Coding environments naturally create feedback loops. You write code, test it, see errors, fix things and repeat the cycle again and again. Every iteration pushes the probabilistic system closer to a more reliable outcome.
The second reason was the platform foundation itself.
CloudOffix already provides structured workflows, forms, UI patterns, permissions and components. The AI did not need to redesign the entire architecture every time. It was operating inside a structured environment with boundaries already defined.
I think this part is extremely important and often misunderstood.
Many people think AI becomes successful simply because models are becoming smarter and smarter. I think AI becomes successful when the environment around it becomes more structured.
When boundaries exist, iteration becomes much more effective.
And honestly, this experience brought me back once again to the importance of foundation and platformization.
If we understand both the strengths and weaknesses of these technologies and create the right environment for them to operate in, the results can become incredibly powerful.
Otherwise, we mostly stay at the hype level.