Consulting work includes telling clients when an ML build is the wrong spend. Low-volume decisions with high liability often need clearer rules and better interfaces, not a classifier.
We walk through volume forecasts, acceptable latency, and who absorbs the cost of mistakes. If a human already handles the case in under two minutes and volume stays under a few dozen a day, a well-written checklist may beat a model that needs constant labelling.
Saying no early is part of implementation consulting. It protects budget for the problems where prediction genuinely changes outcomes.