Could We Have Predicted the Study Before We Ran It?

A Live Test of AI Prediction Against Real-World PMR in a Competitive Oncology Launch

A new TPP needs to be tested before real investment goes behind it, and the standard path is a full qualitative-then-quantitative primary market research (PMR) study. That’s the right call when the stakes are high. But it’s also the default, run the same way, at the same scale, regardless of how much you already know going in. In a launch environment where timelines have compressed and pipelines have multiplied, that default is worth questioning: how much of a PMR study’s findings could you predict before spending a dollar on it?

We tested that question directly with a global oncology client preparing early-line PMR for a new asset in a competitive, fast-moving tumor category. Before commissioning primary research, we asked Stella, our pharma-native AI agent, to predict what the study would find. Three questions framed the test: Can AI predict HCP reactions to a new early-line asset before PMR runs? Where do AI predictions hold up, and where do they break down? And what’s the right role for AI in how we plan research going forward?

How the test worked

Stella synthesized what the client already had: existing PMR and opportunity-assessment data on the asset landscape, the digital context around it (HCP discussion, ad board commentary, conference presentations, published articles), and competitive signals (ongoing early-line trials, competitor announcements, emerging regimen activity). From that, she produced a full slate of predictions: anticipated treatment priorities, expected strengths and weaknesses of each TPP under consideration, sequencing dynamics, source-of-business shifts, projected HCP responses to the client’s and competitors’ profiles, appropriate patient types for each, and directional first-line share allocations.

Then the real PMR ran — both qualitative interviews and a quantitative TPP test — and we compared the actual findings against what Stella had anticipated.

Where the prediction held

The qualitative story held up well. Stella’s read on treatment decision logic matched how physicians actually sorted patients in the real interviews. She correctly flagged the newest asset in the set as the strongest new option, and the broader asset hierarchy she predicted closely aligned with what PMR confirmed. Directionally, she also got source of business right — physicians in the real study were still shifting patients away from an existing standard-of-care therapy, just faster than she projected.

Where the gap showed up

The quantitative share numbers are where the model’s limits appeared. Stella’s directional predictions on first-line share were in the right neighborhood for most assets. Two things moved the real numbers. First, switching off the existing standard of care ran stronger in the quant data than Stella predicted. Second, one TPP was refined between the AI exercise and the quant fielding, shifting its resulting share on its own. Those gaps are important because they define the boundary:

What the qualitative research still gave us

The obvious next question: could the qualitative round have been skipped? No, but the sample size could have been reduced. These were high-profile assets, and understanding how physicians actually reasoned through the decision, in their own words, mattered. The qualitative interviews captured something the AI prediction couldn’t: the genuine excitement and perceived potential physicians attached to these assets. That’s a real-world, in-the-room signal, and it’s not one you get from analyzing available data, however good that evidence is.

A framework for deciding how much PMR you still need

The test produced a practical framework we now use before scoping PMR for a new TPP. Three questions determine how much you can lean on AI before you commission primary research:

The answers sort into three paths:

The point isn’t AI instead of PMR

TechManity™, our operating model, is built on the idea that AI amplifies human expertise rather than replaces it. This test puts that idea into practice. Stella didn’t replace the PMR study — she helped identify, in advance, how much of the answer we already knew and where primary research could add the most value.

The future isn’t AI versus PMR. It’s using AI to determine what research is actually worth doing — and focusing human expertise on what we still need to learn.

Teams can now use Stella directly to generate this kind of predictive read before scoping their next PMR study.