The simple explanation
Adding an LLM introduces uncertainty, cost, and extra failure modes. If the right answer follows a clear, testable rule — arithmetic, an exact filter, a fixed routing rule — ordinary code is the better starting point. It’s deterministic, testable, cheap, and it fails loudly instead of smoothly. An LLM earns its keep where language is ambiguous and judgment is needed: drafting, summarizing, translating intent. Even there, its output needs validation. AI proposes; rules dispose.
There’s also a cost-and-speed argument hiding here. A rule answers in microseconds for fractions of a cent; a model call takes seconds and costs real money per request. At scale — millions of decisions a day — the LLM path isn’t just riskier, it’s slower and dramatically more expensive. And when a rule breaks, it breaks visibly: a failed test, an exception, a wrong-but-obvious output. When a model breaks, it breaks politely, handing you a confident wrong answer that slides past review.
See this idea move.
The AI or Plain Logic? experiment walks through this concept step by step — press run and watch it happen. Everything is simulated in your browser; no real AI runs.
A concrete example
Routing invoices over $500 to a manager is one comparison — code does it perfectly, forever, for free. Summarizing fifty varied customer messages needs language understanding — AI may genuinely help. The strong pattern is a hybrid: AI suggests a category, plain rules control what happens next — who gets emailed, what gets paid, what gets filed.
The same hybrid thinking applies inside AI products: let the model draft the customer reply, but let a rule decide whether it sends. Let the model extract the invoice total, but let code compare it against the purchase order. Put the model where judgment helps and the rules where correctness matters — and never the other way around.
Where people get misled
“AI makes it smarter.” Adding an LLM to a solved problem makes it worse: slower, pricier, and occasionally wrong in ways your tests can’t catch, because the wrongness is fluent. The reverse mistake exists too: refusing AI for genuinely ambiguous work and hand-writing a thousand fragile regexes instead. Match the tool to the shape of the problem.
The honest limits
Never let a plausible sentence substitute for a verified payment amount, a permission check, or a safety decision. And before reaching for a model, ask the plain-logic question: does a simpler solution already meet the requirement? Usually, it does.