When Not to Use AI

In one line: AI is a hammer; not everything is a nail — when a regex, lookup table, or small ML model would do, AI is just an expensive, slow, non-deterministic version of the right answer.

In plain English

The hype loop pushes everyone to use LLMs for everything. That's wrong. Many tasks are cheaper, faster, more reliable, and more private if you use a regex, a SQL query, or a tiny ML model trained for the job. Reach for AI when the problem genuinely needs language understanding or generation — not when a 20-line script would handle it.

AI is a hammer; not everything is a nail.

Don't use AI when

• A regex would work.
• A simple lookup is sufficient.
• The behavior must be 100% deterministic.
• The latency budget is < 100ms.
• The cost per request can't be justified.
• The privacy implications are unacceptable.
• The failure modes are too dangerous.

Often-misused cases

• Using AI to format data when a function would do.
• Using AI for classification when a small ML model would be cheaper.
• Using AI for math (it's bad at this; use code).
• Using AI for translation when Google Translate API exists.

The right test: "Does AI add something a simpler tool can't provide?" If not, use the simpler tool.

Worked example: a $40 bug caused by using AI for math

A team builds an internal tool that summarizes spreadsheets. They ship a feature that "uses AI to compute the totals."

Two weeks later, a finance lead notices the totals are off by a few cents on a few rows. Cause: the LLM is occasionally rounding floats incorrectly when adding 100+ numbers. The behavior is almost right, which is worse than being clearly wrong — nobody caught it for two weeks.

The fix: do the math in regular code (array.reduce((a,b) => a + b.amount, 0)). The AI still writes the narrative summary ("revenue up 12% week-over-week"), but the numbers come from arithmetic. Cost: zero per request. Correctness: 100%.

Rule of thumb: use AI for language; use code for math.

Highlight: the substitution test

Before you reach for an LLM, ask the substitution question: "If I replaced this with a regex, a SQL query, or a 50-line script, would it work?"

• "Extract email addresses from text" → regex.
• "Classify into 5 fixed buckets with high accuracy" → small fine-tuned classifier or even keyword rules.
• "Sum a list of numbers" → sum().
• "Translate UI strings" → professional translation or a translation API.
• "Reformat a date" → date-fns.
• "Open-ended Q&A about your docs" → ✅ legitimately AI.
• "Summarize a long meeting transcript" → ✅ legitimately AI.
• "Decide which tone to use in a reply email" → ✅ legitimately AI.

If a simpler tool exists, default to it. AI's strengths are language, ambiguity, and open-endedness — not deterministic transformations.

Common mistakes

Where people commonly trip up

• Using an LLM as a deterministic formatter. "Make this date look like MM/DD/YYYY" or "trim this to 200 chars" through a model call introduces non-determinism, latency, and cost for a problem date-fns and String.slice solve for free. If the rule fits on a sticky note, write the function.
• AI for arithmetic. Models are great at describing numbers and terrible at computing them, especially across many rows or with currency. Always compute totals, percentages, taxes, and aggregations in code; let the model write the narrative around the numbers, never produce them.
• Reaching for AI because it's faster to prototype, then never going back. "I'll replace this with real code later" rarely happens — the AI version ships, the costs accumulate, and nobody owns the migration. Set the bar upfront: if a simple tool would work, build the simple tool, even if the AI version is 30 minutes faster to demo.
• Ignoring the latency floor. Even a fast frontier model adds 300-800ms minimum. For autocomplete, keystroke validation, or anything inside a tight render budget, that's disqualifying. Use a local heuristic or a small classifier; reserve the model call for the slower, higher-value paths.
• Underrating privacy implications. "We'll just send the row to the LLM and have it decide" can quietly route customer data, internal documents, or regulated content through a third party. Even with enterprise no-training tiers, every AI call is a data-egress decision — apply the same review you'd apply to adding any new vendor.

Page checkpoint

Checkpoint Quiz

Did 'when not to use AI' stick?

Required

What's next

→ Continue to The 2026 AI Stack Summary — a one-page reference for the dominant tools.