When not to use AI

By Dylan Bailey, Certified Claude Expert · Updated 4 August 2026 · 6 min read

Most writing about AI is about what it can do. Rather less is about where reaching for it actively makes things worse — which is a shame, because knowing the boundary is what lets you move quickly everywhere inside it.

1. When a deterministic tool already does it

A spreadsheet formula gives the same right answer every time. An AI estimate gives a plausible answer that's usually right. For anything calculable, the formula wins — not because AI is bad at arithmetic, but because reliability is the entire requirement.

2. When you can't verify the answer

If you couldn't tell a correct answer from a convincing wrong one, you have no way to catch the failure. Using AI in a domain nobody at your business understands isn't leverage — it's outsourcing to something that can't tell you when it's out of its depth.

3. When being wrong is expensive and quiet

Loud failures are survivable — someone notices. The dangerous combination is high cost and low visibility: a misstated figure in a tender, a wrong clause in a contract, a compliance date that's out by a month. AI is confident in exactly the same way whether it's right or not.

4. When the work is the relationship

A condolence message. An apology to a client you've let down. A reference for someone who worked for you for ten years. The value of these isn't the prose — it's that a person sat down and thought about someone. Generating them produces the right words and none of the point.

5. When the process is the product

Sometimes the thinking is the deliverable. Writing a strategy forces you to confront what you actually believe. A generated strategy gives you a document without the argument you needed to have with yourself — and it reads fine, which is what makes it hard to notice.

6. When it decides something about a person

Screening, ranking, assessing suitability, scoring risk. The bias risk is real and the accountability is entirely yours. Summarising applications against stated criteria is defensible; producing the shortlist is not.

7. When you'd have to explain how you got there

If a regulator, a client or a court might ask why you reached a conclusion, "the AI suggested it" is not an answer. Anywhere you need a defensible reasoning trail, the reasoning has to be yours.

8. When the input can't leave your systems

Some material simply shouldn't be sent to a third party — under an NDA, under a regulatory obligation, or because a client would reasonably object. That's a boundary decision, not a prompting problem.

9. When it's genuinely faster to just do it

A two-line email. A quick note to a colleague. Watch someone spend ninety seconds prompting for something they'd have typed in twenty, and you're watching a habit that's stopped being useful. The tool became the point.

An Australian small-business example

A Brisbane consultancy uses AI to draft a scope-of-works for a government client. The document is well written and the client signs it.

Two months in, a dispute surfaces over a deliverable that reads clearly but doesn't match what was verbally agreed. Nobody on the team can explain why the clause was worded that way — the draft was reviewed for tone, not intent.

Situations 3 and 7 together: quiet, expensive, and no reasoning trail. AI was fine for drafting the covering email. It shouldn't have been near the scope.

The question that covers all nine

Before reaching for AI, ask: if this output were confidently wrong, what would it cost, and would we notice?

Cheap and obvious — go ahead, that's most work. Expensive and hard to spot — do it yourself, or build in checking that genuinely catches it. That single question generalises further than any list of banned tasks.