Adding AI can make a roadmap sound modern while making the product slower, less predictable, and more expensive. The first question should not be “Which model?” It should be “Why does this need a model at all?”
Use AI for Ambiguity
Generative models are strongest when inputs are messy and several valid outputs may exist: summarizing documents, interpreting natural language, extracting meaning from images, drafting content, or choosing among tools based on context.
Traditional code is stronger when the rule is exact. Permissions, prices, calculations, state transitions, and schema validation should remain deterministic whenever possible.
Score the Opportunity
I evaluate a candidate feature across five questions:
- Value: does AI remove meaningful effort or unlock something impossible before?
- Tolerance: can the experience survive occasional uncertainty or error?
- Evidence: can we define and measure a good result?
- Economics: does the value exceed inference, review, and support cost?
- Fallback: what happens when the model is wrong or unavailable?
If the feature is high-risk, difficult to evaluate, and has no safe fallback, excitement is not enough.
Prefer Hybrid Systems
The best answer is often not AI or code. It is both.
Let the model interpret an unstructured request, then validate its proposed action with deterministic rules. Let it generate structured fields, then reject invalid values. Let it recommend, while the user confirms consequential actions.
AI should handle the ambiguity. Software should enforce the contract.
A worthwhile AI feature makes the product easier for the user without making its core promise less trustworthy. If a form, search filter, or rules engine solves the problem better, use it proudly.