Unpopular opinion: most SaaS AI features should have been a cron job

I review AI products for a living, including my own Tagwise. A depressing number of "AI-powered" features are: run on schedule, apply a threshold, send an email. The model adds latency, cost, and a new way to be wrong, over a 40-line script.

My rule now: ship the cron job first. Add the model only when you can name the specific input variation the rules cannot handle. For Tagwise that moment was genuinely ambiguous product titles ("navy boat shoes" — color or vessel?). Everything before that was regex wearing a trench coat.

Where did the model actually earn its place in your product?

Drafting free-text replies. Rules cannot write the paragraph, they can only pick one. Everything else in our pipeline (routing, priority, language detect) is regex and classifiers. The expensive part does the only thing rules cannot do.

Same split. Retrieval + rules decide what is true, the model only writes the sentence. The day we stopped asking the model to decide things, our error rate fell off a cliff.

Log pipelines agree. Parsing, grouping, thresholds: deterministic. The model earns its place exactly once: explaining the weird one at 3am in human language. Everything else is cron.

Flashcards from notes: the model earns it in distractor quality for quizzes. Rules make fine flashcards, terrible wrong answers. Students notice immediately.

Finally, an AI thread I agree with. My cold emails: segmentation is SQL, timing is cron, and the model writes exactly one sentence of personalization. One. The rest is rules, and open rates went up when I stopped letting it freestyle.