AI & Machine Learning
Before You Build an AI Feature, Check It Needs a Model
A surprising share of requested AI features are better served by a query, a rule, or a well-placed index. Knowing which is the useful skill.
We are asked for AI features constantly, and a meaningful share of them do not need a model. Not because the request is unreasonable, but because the underlying need turns out to be deterministic once it is stated precisely.
When a rule is enough
"Predict which orders are at risk of delay" often decomposes into three conditions the operations team already applies by eye. Written as rules, they are explainable, auditable, instant, and free to run. A model would approximate the same thing less transparently and cost more to operate.
The cases that genuinely need a model
The genuine cases share a trait: the input is unstructured or the pattern is genuinely learned rather than specified. Classifying free-text complaints, extracting fields from scanned documents, forecasting demand from years of seasonal data, routing a chatbot conversation — these do not reduce to a rule table without losing what makes them work.
The costs nobody estimates
The cost that gets underestimated is not training but everything around it: labelled data, an evaluation set that reflects reality, monitoring for drift, and a plan for what happens when the model is confidently wrong in front of a customer. A rule engine has none of those obligations.
The question is not whether a model could do this. It is whether anything simpler could do it well enough.
Start from the decision the feature is meant to support and work backwards. If the logic can be written down, write it down. Save the model for the problems that genuinely resist being written down.
