I Almost Shipped a Wrong Number Because I Trusted the Wrong Kind of AI Calculator

Was building a small pricing calculator feature into my side project and used an LLM to sanity-check a multi-step markup calculation. It gave me a confident, clean, wrong answer. Explanation looked completely correct. Final number was off. Caught it by luck, not by suspicion — that's the scary part.

What I learned about "AI calculators"

Two totally different things get called this. Wolfram Alpha-style tools actually compute — symbolic math engine, deterministic, essentially always correct for well-formed problems. LLMs like the one I used predict the most statistically likely next answer based on patterns. Great at explaining, not guaranteed correct on multi-step math.

The failure mode is the dangerous part: it doesn't fail obviously. It fails confidently, with clean reasoning wrapped around a wrong number, which is exactly the situation where you don't think to double check.

What I do now

Any actual computation that matters — pricing logic, unit conversions, anything a user-facing number depends on — goes through a deterministic tool (ended up scripting the math directly rather than trusting an LLM call for it). LLM stays in the loop for explaining things to users in plain language, not producing the number itself.

The broader takeaway for other solo builders

If you're using any AI assistant to generate or verify a calculation that ships to users, know which category it falls into first. "Sounds right" and "is right" are not the same thing with language models doing math, and the gap between them doesn't announce itself.

More detail

Longer version on my blog: AI Calculator Online — What Actually Works. Full breakdown on my site: AI Calculator Online. Also relevant if you're using AI around numbers generally: AI in Google Sheets.

Anyone else caught a quiet wrong-number bug from an LLM before shipping? How'd you catch it?