A customer-support pilot shows higher satisfaction after launch, but the period also introduced longer opening hours and a new staffing model. What would provide stronger evidence of AI impact?

1. Exclude every dissatisfied customer from the pilot report.
2. Attribute the entire improvement to AI because it launched in the same month.
3. Ignore customer satisfaction and measure only token count.
4. Use a suitable controlled comparison or staged experiment that accounts for the other changes.