Practice · 4 MIN READ
When not to use AI
The fastest way to lose credibility with an AI project is to apply it where a deterministic system would have been better, cheaper and correct.
AB19 Innovations
In short
AI suits problems where errors are cheap to detect and correct and the rules resist enumeration. Deterministic logic remains better where rules are known, correctness is binary, or wrong answers are expensive and invisible.
We turn down more AI work than we take. Not out of scepticism — we build these systems — but because the failure mode of applying a probabilistic tool to a deterministic problem is expensive and entirely avoidable.
The test: what does a wrong answer cost?
If a wrong answer is immediately visible and cheap to correct — a suggested tag, a draft, a ranked list a human scans — AI is often an excellent fit. If a wrong answer is invisible, expensive, or flows unchecked into a downstream system, the bar rises enormously.
Signals that you want ordinary software
- The rules are already written down somewhere and fit on a page.
- Correctness is binary and auditable, not a matter of judgement.
- The same input must always produce the same output.
- A regulator, an auditor or a court may ask why a specific decision was made.
Signals that AI genuinely earns its place
- The input is unstructured — language, images, audio, messy documents.
- The rules resist enumeration; experts recognise the answer but cannot fully specify it.
- Volume makes human handling impractical and sampling is acceptable.
- A human reviews the output before it has consequences.
The question is never "can AI do this?" It is "what happens when it is wrong, and who finds out?"
Saying so early costs one uncomfortable conversation. Discovering it after two quarters of build costs considerably more.