Define the ML problem

Prediction.
Classification.
Ranking.
Recommendation.
Computer vision.
Natural language processing.
Forecasting.
Anomaly detection.

Evaluate the data reality

Model performance depends heavily on the availability and quality of training and evaluation data.

Ask about baseline performance

A good engagement should compare the proposed system against a meaningful existing baseline rather than assuming machine learning is automatically better.

Model complexity is not the goal.The useful system is the one that performs well enough, can be evaluated and can be operated.

Plan production ownership

Clarify who manages deployment, retraining, monitoring and infrastructure after the initial model is delivered.