The Control Gap: The Prediction Question
- Jonathan Olsen
- Jul 28
- 2 min read
"AI predicts launch success" is the line everyone is selling right now. It also happens to be pointed at the easier half of the problem.
The short version:
Predictive launch tracking is real, and the people running pharma's biggest AI programs are still careful about the claim.
Prediction works when a problem holds still, and a launch doesn't.
The ROI lives in response time, not in the forecast, so buy the sensor and build the reflexes.
IQVIA describes active launch tracking against AI-generated success predictors, with real-time course-correction, and ZS frames every launch as a learning system that updates its own assumptions. The technology is impressive and worth having, so let's take it at its strongest rather than knock it down.
Here's the tell, though. The people running pharma's largest AI programs are notably careful about this exact claim. Lilly's Diogo Rau, who is about as pro-AI as senior leaders get, is direct that AI doesn't magically compress what biology and large organizations make slow, and the places where Lilly sees AI clearly outperform people are narrow and contained, like demand forecasting deep in the supply chain or defect detection on a production line. Prediction shines when the problem holds still.
A launch does not hold still, and prediction was never the hard part.
The hard part is the distance between the signal and the act. Picture a payer's rejection rate climbing in week three. Everyone can see it, and the dashboard has never looked better. Then the questions that actually matter start stacking up: whose call is it to intervene, does the field have new guidance or last quarter's, and how many meetings pass before someone owns the response. By the time the organization has digested the insight, the window has already moved, the model was right, and the launch drifted anyway.
A system that learns faster than the company can act just gives you a more precise record of how you fell behind. Buy the sensor if you want, but understand that the reflexes are the actual work.
Where's your real bottleneck: seeing the signal, or acting on it before it goes stale?
Comments