AI in the place to be!

Using AI Only Where It Adds Value in Code Automation

When we talk about AI-assisted software development, the conversation often starts with agents: which model should generate the code, how much autonomy should it have, and how many tools should we give it?

I want to start somewhere else.

What if the goal is not to make AI do as much as possible, but to make it do as little as possible?

That is the idea behind this new series of videos.

For tasks based on well-defined patterns, traditional automation is still faster, cheaper, more predictable, and easier to validate. AI becomes interesting only at those particular points where interpretation or judgement is required and the decision cannot be conveniently formalized.

In the first video, I use a deliberately simple example.

I start with a C# WinForms designer containing literal UI texts in English and automate the process of converting them into multilingual resources.

Most of the work does not require AI at all.

The application can:

• locate the relevant text properties,
• generate standardized resource names,
• replace literals with resource references,
• modify the .resx files,
• update their generated designer code,
• collect the modified files,
• and present all changes for review before they are accepted.

The language model only enters the pipeline for the part where language understanding is actually useful: adapting the texts into Spanish while taking their context into account and preserving class and control names.

After that, the deterministic process takes control again.

This is intentionally a very small example. The purpose is to establish the basic approach before applying it to increasingly complex code-generation tasks.

The application becomes the assembly line.

AI is not the assembly line. It is the specialist that is called in only when the assembly line reaches a problem that cannot be solved mechanically.

That leads to the question I want to explore throughout the series:

How much of a code automation task actually needs AI?

Video

Watch in YouTube

In the next examples, I’ll gradually increase the complexity while trying to keep the same rule: deterministic code should do everything it can, and AI should only deal with what remains.