When NOT to Use AI

The question nobody asks

Every AI product roadmap starts with "what can we build with AI?" The better question is "what should we leave alone?"

The AI hype cycle pushes teams to bolt LLMs onto everything. Most of those features would be better as a form, a filter, a SQL query, or a human process. Knowing when to skip AI is more valuable than knowing how to prompt.

The decision checklist

Before adding AI to any feature, run through this:

1. Can I eval it? — Do you have 30+ examples with known correct answers? If not, you can't verify it works. Don't ship.
2. Does the answer already exist? — If it's in a database, API, or spreadsheet, query it directly. LLMs guess; databases don't.
3. What's the cost of a wrong answer? — Low (suggested emoji)? Ship with confidence. High (medical, legal, financial)? Human review or no AI.
4. Is the input structured? — Dropdowns, checkboxes, and date pickers beat chatbots for structured data collection. Every time.
5. Will users verify the output? — If users blindly trust AI output without checking, you're building a liability.

If you fail checks 1, 2, or 3 — stop. Use a different tool.

Where AI actively makes things worse

Exact answers from structured data

"What's our Q3 revenue by region?" is a SQL query, not an LLM prompt. The model will hallucinate a plausible number. Your CFO will make decisions on it.

Use instead: SQL, BI tools, APIs, spreadsheets.

This applies to anything where the ground truth lives in a system of record: inventory counts, account balances, employee records, compliance deadlines.

Zero-error-tolerance calculations

Dosage calculations, tax computations, structural engineering loads, aviation weight-and-balance — these need deterministic math, not probabilistic text generation.

LLMs can do arithmetic sometimes. "Sometimes" is not acceptable when lives or money are at stake.

Use instead: Calculators, validated formulas, rule engines.

Real-time or rapidly changing data

Stock prices, weather, sports scores, breaking news — anything that changes minute-to-minute. LLMs are frozen at training time. RAG helps only if your retrieval pipeline is faster than the data changes, which is harder than it sounds.

Use instead: Live APIs, webhooks, traditional search with freshness ranking.

Accountability and audit trails

When someone needs to answer "why did the system make this decision?" in a courtroom, regulatory audit, or post-incident review — LLM outputs don't cut it. You can't explain a probability distribution to a judge.

Use instead: Rule-based systems with explicit decision logs, human decision-makers with documented rationale.

Workflows where the human is the product

Therapy, executive coaching, legal counsel, creative direction — the value is the human relationship, judgment, and accountability. Wrapping an LLM around these doesn't scale the human; it replaces them with something worse.

AI can assist (draft notes, suggest questions), but the human must own the outcome.

Where AI is fine — with caveats

Not everything is black and white. These are the gray zones:

The pattern: AI proposes, human or system disposes. The moment you remove the verification step, you move from the right column to the "don't use AI" list.

The "we added AI" trap

Three common failure modes when teams add AI because they "should," not because they "need to":

1. The chatbot that should be a form

"We replaced our 5-field intake form with an AI chatbot." Now users type free text, the model extracts fields (sometimes wrong), and support tickets increased because the chatbot misunderstood.

Fix: Keep the form. Add AI behind the scenes if you need NLP on free-text fields.

2. The dashboard that should be a query

"We built an AI copilot for our analytics platform." Users ask "what were sales last quarter?" and the model hallucinates a chart.

Fix: Natural language → SQL translation can work, but show the generated SQL and let users verify before running. Better yet, give them the query builder.

3. The automation that should have a human

"We automated compliance review with an LLM." It misses edge cases. Regulators find out. Fines follow.

Fix: AI flags suspicious items. Humans make final determinations. Log everything.

How to say no (without sounding like a Luddite)

When stakeholders push for AI everywhere:

1. Ask for the eval set. "Show me 50 examples where we know the right answer." If they can't, the feature isn't ready — AI or not.
2. Calculate the cost of wrong. "If this is wrong 5% of the time, what happens?" Quantify it.
3. Propose the simpler alternative. "A filtered dropdown solves this in a week. The AI version takes a month and we can't verify it." Let the tradeoff speak.
4. Offer AI where it fits. "I'll add AI to the draft-generation step, but the approval workflow stays human." Compromise without compromise on safety.

The meta-point

The best AI engineers aren't the ones who add LLMs to everything. They're the ones who know when a regex, a SQL query, or a well-designed UI solves the problem better.

AI is a tool, not a strategy. Use it where probabilistic generation adds value. Skip it where determinism, accountability, or simplicity wins.

If you can't eval it, don't ship it. If the answer is in a database, query it. If a wrong answer is expensive, keep a human in the loop. Everything else is negotiable.