beepm’s approach
With 3 weeks to go, my immediate priority is to decouple the Launch Success from Data Perfection.
First, I would conduct a Sensitivity Analysis. I’d ask my data scientists to simulate the 30% loss on historical sets to see the 'Delta' in our North Star metric (e.g., Click-Through Rate). If the model is robust enough that the delta is marginal, we proceed with a 'Soft Launch.'
Second, I’d implement a Heuristic Fallback. For any user where the data is missing or confidence is low, the engine will revert to a high-performing 'Popularity' logic. This ensures a 0% 'Broken Experience' rate, even with 30% 'Data Loss.'
Third, I’d manage the dependency by defining an MVD (Minimum Viable Data) with the upstream team. Rather than demanding a full fix, I’d identify the 20% of data fields that drive 80% of the model's value and ask them to prioritize those.
Finally, I would present a 'Confidence Scorecard' to stakeholders. I wouldn't just say 'we have a problem'; I’d present a recommendation: 'We launch on time with a Hybrid Model, which still outperforms our current baseline by X%, while we phase in the full AI over the following 3 weeks.