⌘K
ENES
GitHub
conceptLearning Foundations~2 min readUpdated 2026-06-07#foundations#scoping#judgment#decision

When not to use AI

Knowing when not to reach for AI is as valuable as knowing how to build it. Models add cost, latency, unpredictability, and a maintenance burden. If a deterministic solution works, it is almost always better — cheaper, faster, testable, and explainable.

Prefer something simpler when…

  • A rule or formula suffices. If the logic is known and stable ("flag orders over $10k"), write the rule. ML to relearn a known rule is wasted complexity.
  • You can't tolerate being wrong. ML is probabilistic; for tasks needing guaranteed correctness (accounting, safety interlocks), use deterministic code, with AI at most assisting a human.
  • There's no data (supervised ML) or no way to verify outputs (generative). No signal in, nothing reliable out.
  • The stakes are high and unmonitored. High blast radius + no human oversight is where AI failures become incidents.
  • Explainability is mandatory (some legal/medical/credit contexts) and the model can't provide it (transparency).
  • A heuristic gets you 90% at 1% of the cost and effort — ship that first.

Questions to ask before adding a model

  1. What decision does the output drive, and what does a wrong answer cost?
  2. Could a rule, lookup, or existing software do this acceptably?
  3. Is there data / a way to evaluate quality? (If you can't eval it, you can't trust it.)
  4. Can we tolerate variability and the occasional confident error?
  5. Who is accountable when it's wrong, and can a human catch it?

The nuance for LLMs

LLMs lowered the barrier — you can "solve" a task with a prompt and no training data. That makes it tempting to use them everywhere, including where a regex, a database query, or a form would be more reliable and far cheaper. Use the LLM for the genuinely fuzzy, language-shaped part; use plain software for the rest.

Pitfall

"AI" as a mandate rather than a tool — adding a model because it's expected, then inheriting hallucinations, cost, and latency to solve a problem deterministic code had already solved. Start from the problem, not the technology.

Connects to: mental models for AI systems · frame the problem · the smallest stack that works

Related notes

Learning FoundationsEvaluation metrics & what they hideAccuracy lies on imbalanced data. Precision, recall, F1, ROC-AUC — what each measures, and how to pick the metric that matches the cost of being wrong.Learning FoundationsFeatures, representations & the curse of dimensionalityModels don't see the world — they see features. How representation quality decides everything, and why high dimensions are weird.Learning FoundationsGeneralization, overfitting & the bias–variance tradeoffA model is only useful if it works on data it never saw. Overfitting, underfitting, and the tradeoff that governs both.Learning FoundationsHow learning works: loss, objective, and ERMEvery model that "learns" is minimizing a loss function over data. Understand that loop and most of ML stops being magic.