Does This Feature Actually Need AI?

Does This Feature Actually Need AI?

When building a new feature, it's becoming natural to ask:

"Can AI do this?"

But I think there's a more important question we should ask first:

"Does this feature actually need AI?"

I'm not talking about using AI to help write the code. I'm talking about making AI part of how the feature works in production.

I've noticed, including in my own work, that sometimes we use AI simply because we know it can solve the problem.

A feature needs to classify something, transform some data, or make a decision. Sending it to an LLM feels like an easy solution.

But after understanding the problem better, we might discover that the same thing could have been handled by a simple function, a few rules, or even a shell script.

And that's an important difference.

Instead of:

Input → LLM → Response → Validation → Result

We could have:

Input → Code → Result

The first solution introduces API costs, latency, nondeterministic results, external dependencies, and additional failure cases.

Sometimes all of that is worth it. A chatbot, for example, needs to understand unpredictable natural language. Other features deal with ambiguity or unstructured information where writing deterministic rules isn't practical.

But AI being capable of solving a problem doesn't mean AI is necessary to solve it.

Sometimes we choose AI because we haven't yet discovered the simpler programmatic solution.

So before adding AI as a dependency of a feature, I think it's worth asking:

Does this problem actually require AI, or do I just not know the deterministic solution yet?