1. The model is not the product. Everyone can call the same models. What makes a product defensible is everything around the call: your data, the workflow it sits in, the evaluation that keeps it good, and the accumulated handling of cases where it is wrong. If a pitch is entirely about which model is used, there is no moat being described.
2. "It works" is not a measurement. Ask how quality is measured. The right answer involves a set of test cases with known-correct answers and a score that gets tracked over time. "We tried it and it seemed good" means nobody will notice when it stops being good.
3. Inference costs money per use, forever. Unlike normal software, every interaction has a marginal cost. Ask what a single user interaction costs today and what it becomes at ten thousand users. If nobody has calculated it, that is the finding.
4. Hallucination is a property, not a bug to be fixed. Any vendor promising it has been eliminated is either misinformed or hoping you are. The real question is how the system makes wrongness visible — citations you can check, confidence thresholds, human review on the consequential paths.
5. Your data is the asset. Whatever proprietary data you have — support conversations, documents, transaction history — is the part competitors cannot copy. An AI strategy that does not use it is a strategy anyone could execute.
