Before a new drug ever reaches you, an enormous amount of strategy decides whether it gets developed at all and how it's sold to your doctor. Jennifer Preston is a PhD biomedical engineer who now leads AI strategy for the consultants shaping those decisions. Her firm uses AI to create synthetic patients and synthetic physicians: digital stand-ins that get interviewed and surveyed to test drug messaging before any real human is involved.
Jeff opened with the uncomfortable question underneath all of it: what does it mean when healthcare starts making decisions based on people who don't exist? Her answers hold lessons well beyond pharma.
1. Synthetic research is phase zero, not the decision
Preston's first move is to deflate the scary version of the story. No decision gets made exclusively on synthetic users. Her team calls synthetic research "phase zero": a cheap, fast filter that kills bad ideas before expensive real-world research begins.
The economics are brutal. When a campaign or a new drug indication fails after full traditional market research, the cost lands on the pharma company and eventually on your prescription price. Vet the losers synthetically, then spend the real interviews on ideas worth testing. The humans still catch the outliers. The machines just clear the junk.
The transferable pattern: use AI to eliminate options, not to pick winners. That division of labor holds in marketing, product, and strategy work far outside life sciences.
2. In high-stakes domains, small beats large
Preston has watched her own doctor complain that AI keeps summarizing test results wrong, turning a time-saver into a time cost. Her diagnosis: general-purpose LLMs trained on Reddit and public sources, where accuracy is an opinion, have no business summarizing medical records.
"The doctor's office doesn't need to know housing prices. By bounding things a lot more, you're going to be able to make them more accurate."
Her prediction: the industry moves from large general models to small, purpose-built, tightly bounded ones. The LLMs don't go away. They become the base layer, not the tool. If your AI touches decisions where being wrong is expensive, the bounded model is the trust strategy.
3. AI can't fix a process you refuse to change
Jeff pointed at the absurdity: paper intake forms in 2026, siloed patient portals, patients repeating their history to every new doctor. Preston's response is the quote of the episode for anyone deploying AI anywhere:
"People think, oh, I'll just do this and it'll make my job easier. It will, but only a small amount. What you really need to do is change the process."
Her team redesigns workflows first: how do we do our work, where should AI fit, where can it fit, and where does the human stay in the loop. Bolting AI onto a broken manual process gets you a faster broken process. Sound familiar, revenue teams?
4. The engineer's edge: obsess over inputs
Asked what the engineer in her sees that strategists miss, Preston went straight to the black box. Engineers are trained that input and output are everything. So before she trusts any agent, she interrogates the input: what is it, where does it come from, what's the source, how clean is the data.
Her sharpest observation about her own industry: plenty of strategy consultants hand over the strategy and never have to implement it. The how isn't their problem. Engineers don't get that luxury, and neither does anyone accountable for AI output.
5. Respect the 10-year pipeline (and the 90% failure rate)
The episode's crash course in drug economics explains a lot about your pharmacy bill. A single molecule takes at least 10 years to reach market. 80 to 90% of phase one clinical trials fail. Go-to-market strategy starts before phase one, weighing speed, population size, rare disease pathways, and a 10-year competitive pipeline. Trial recruitment alone can add two years of delay, and it's one of the places AI is already delivering: finding the right patients faster.
The most consequential version of that: most medications were trialed on populations averaging a white male. FDA diversity mandates are pushing the other way, and Preston sees AI-assisted recruitment as the path to both faster trials and something bigger. A doctor who can say "here's how this drug performed in patients like you" rebuilds trust in the whole system.
What's coming
Preston's horizon list: a broad shift to small models to cut cost and energy, drugs reaching market faster, direct-to-patient distribution reshaping an industry built on selling to physicians, and personalized medicine, real individual drug cocktails, still 10 to 15 years out. Her caveat is the most honest line in the episode: "In the life science industry, you're generally 10 to 15 years away from everything."
And the one-pill-a-day dream for the 30-pill senior? Not soon. "When you start mixing those molecules together, they're not always friends."
FAQ
Who is Jennifer Preston? A PhD biomedical engineer who moved from the lab bench (including Nike-funded biomechanics research in grad school) through mass spectrometry and life sciences roles into commercial strategy and competitive intelligence, where she now leads AI strategy for consultants serving pharma and biotech.
What are synthetic patients and synthetic physicians? AI-generated digital respondents that get interviewed and surveyed to pressure-test drug messaging, delivery formats, and campaign concepts before real market research begins. Preston's team calls this "phase zero" research.
Does pharma make decisions based on synthetic research alone? No. Synthetic research filters out bad ideas cheaply so that traditional research with real physicians and patients focuses on viable concepts and still captures outliers.
Why do drugs cost so much? A 10+ year development pipeline, 80 to 90% phase one failure rates, and expensive late-stage trials. Every failed asset is paid for by the ones that succeed.
What's the main business takeaway? Use AI to eliminate bad options cheaply, bound your models to the domain, and redesign the workflow before deploying the tool. AI added to a broken process just breaks faster.