I Almost Used AI to Classify User Input. Simple Rules Worked Better

The author recently rebuilt a small name tattoo tool and encountered a problem that seemed like a good use case for AI. They initially treated every input the same, but realized the inputs were different and needed a more tailored approach. At first, they considered using an LLM to classify the input, but after listing the cases they cared about, they realized the classification problem was small and predictable.

Instead of adding another model call, they used deterministic rules to classify the input. The rules distinguished between factors such as initials, initials + date, name + date, two names, hyphenated names, multi-word text, short single names, medium single names, and long single names. By using deterministic rules, the classification only impacted the UI, helping to choose the first few lettering options that made more sense for the input.

This approach proved to be a good decision, as the AI classifier was not necessary for this specific problem. The author now asks themselves if there's enough uncertainty in the input space to justify using AI, rather than asking if AI could handle the task. In the end, they prefer deterministic rules when the problem is not fuzzy enough to require AI, as they offer predictability, simplicity, and ease of reproduction.

The main takeaway is that for certain problems like classifying user input, rules can be a better product decision, while AI is more valuable when tackling genuinely open-ended tasks.