AI is incredible for tasks with judgment, language, and context. It's terrible for tasks with judgment, regulation, and consequences. Knowing which is which is 80% of the work.
1. When determinism matters
If the output has to be exactly the same every time (invoicing, tax calculation, ID validation), an LLM is the worst possible friend. Even if you nail it 99% of the time, that 1% can cost you the monthly close.
2. When per-execution cost adds up
An LLM parsing every inbound email at €0.01 per call seems like nothing. Multiply by 50,000 emails per month and that's €500 for a task a regex does in milliseconds and for free.
3. When audit is non-negotiable
In processes where you must explain exactly why a decision was made (HR, legal, financial), "the model decided" is not an answer. You need written rules an auditor can read.
4. When the system already works
The most expensive mistake: dropping AI into a workflow that already works because "it's time." The cost isn't just development, it's the technical debt of maintaining a more complex system with no business reason.
AI is not a goal. It's a tool. If you don't know what specific problem it solves, it's not time.
