Menu
Close
.svg)
Last Thursday I spent the day with two ambitious, fast-rising pharmaceutical-industry professionals at Lehigh Valley Sporting Clays, about an hour north of Philadelphia as the crow flies. Sporting clays, for the uninitiated, is essentially miniature golf with over-under shotguns. You wander through the countryside from station to station, except that instead of fretting over a seven-iron, you attempt to dispatch small orange mini frisbees hurtling unpredictably through the air.
It is also, as it turns out, a surprisingly good venue for discussing artificial intelligence.
As we circumnavigated the course, one of my companions, who works in market access in oncology, raised a problem that would have sounded peculiar only a few years ago. When he asks ChatGPT about the particular cancer in which his company competes, his drug does not surface as prominently as he thinks it should.
I mulled over that observation on my drive home. Twenty years ago, he might have fretted about where the brand appeared in a Google search. Today he is wondering where it appears in an AI-generated answer. There is an important distinction between the two.
Google largely helped us find information. Generative AI increasingly reviews it, weighs it, synthesizes it and presents an answer in a voice of authority. And according to fascinating new research from IQVIA, physicians around the world are making a habit of asking.
The algorithm, in other words, has become an audience.
For years, pharmaceutical companies have steadily expanded their conception of whom they must understand and influence. There was a time, not long ago, when the physician was overwhelmingly the customer. Then came a more complicated constellation of payers, health systems, patients, caregivers, nurses, pharmacists and other stakeholders.
We may now need to add a rather unusual member to the cast: the machine.
Last week, I attended a terrific IQVIA webcast entitled Launch Excellence in the Era of AI. The presenters shared results from a survey of 23,769 healthcare professionals across 34 countries. Fully 80% reported using AI tools, and half were doing so either daily or several times a week. Younger HCPs were somewhat more enthusiastic, but this is hardly an affliction of youth: even among those born before 1980, three-quarters reported using AI.
The remarkable finding is not simply adoption. It is habit formation. An earlier IQVIA study of oncologists and GPs in France, Germany and Britain found a similar pattern: 85% reported using generative AI professionally, with 41% using it daily and another 32% weekly.
Nor are physicians merely asking ChatGPT to polish an email or summarize meeting notes. IQVIA found HCPs using GenAI for disease information, literature review, treatment decisions, knowledge refresh and diagnostic support. Nearly half said they employed it to support treatment decisions.
This is no longer simply automation at the administrative periphery of medicine, where much of the early activity occurred. AI is sidling towards its cognitive center.
The most intriguing part of IQVIA's work may not be its enormous survey at all.
In a separate pilot, the company analyzed more than 2,000 anonymized prompts from more than 100 HCPs in Britain, France and Germany. Rather than asking doctors what they said they used AI for, IQVIA examined what they actually asked it. Nearly a third of the topics were patient-specific, encompassing treatment selection, diagnosis and test interpretation. Other questions concerned dosing, efficacy, adverse-event management and guidelines.
This introduces a potentially important new reservoir of healthcare insight. Professionals in our neighborhood of pharma marketing research have become extraordinarily sophisticated at asking doctors questions. We measure awareness, attitudes, message recall, treatment preferences and prescribing intentions. We conduct interviews, advisory boards, ATUs and conjoint studies. All remain valuable.
But a prompt captures something different. It captures the question an HCP chooses to ask at the moment uncertainty arises.
IQVIA makes a useful distinction between topics and tasks. An HCP asking AI to compare two medicines is not merely expressing interest in two brands. They may be trying to resolve a differentiation problem. A question about whether a treatment is appropriate for a frail patient may expose uncertainty about patient fit. Across its prompt analysis, IQVIA found recurring decision tasks involving evidence, differentiation, eligibility, positioning and safety.
For I&A professionals, this is tantalizing. Search data revealed what people wanted to find. Social listening revealed what they wanted to say. Prompt data may increasingly reveal what they are trying to decide.
None of this appeared ex nihilo with ChatGPT. Medicine has been inching towards algorithmically assisted decision-making for decades. A recent scoping review of rule-based clinical decision-support (CDS) systems examined 28 studies spanning applications from diabetes and cardiovascular disease to medication safety and breast cancer. The technology evolved from relatively simple SQL-based rules towards systems incorporating natural-language processing, machine learning, real-time monitoring and personalized decision support.
Another systematic review of reviews, encompassing 18 reviews and 669 underlying articles, found AI being applied across clinical, organization and shared decision-making. What generative AI changes is accessibility. A CDS system traditionally had to be designed, purchased, integrated and implemented. ChatGPT merely requires a browser and curiosity.
Institutions have consequently found themselves in the unusual position of governing a revolution that has already occurred. In IQVIA's international survey, only 31% of HCPs reported that their organization had approved or endorsed even one AI tool they were using. Only about 11% were using approved tools exclusively.
The doctors, it seems, arrived before the governance committee.
This creates a peculiar problem for pharmaceutical companies. A physician contemplating a treatment may increasingly query an AI system to compare therapeutic options, interpret a guideline or explain the evidence for a particular patient type. The resulting answer will be assembled from some combination of publications, guidelines, websites and other sources, filtered through a model whose synthesis the pharmaceutical company neither controls nor necessarily observes.
The old commercial question was: Have physicians seen our evidence?
The new one is: Can their AI find and correctly interpret it?
IQVIA calls the emerging contest “share of algorithmic trust”. Credible peer-reviewed publications, guideline inclusion, comparative evidence and current, well-structured scientific information may become more important because AI increasingly mediates how HCPs encounter the underlying evidence.
This has spawned another acronym, because no technological revolution is officially recognized until somebody supplies one: GEO, or Generative Engine Optimization.
The analogy with SEO is useful but imperfect. Search-engine optimization was largely about being found by a human searching for information. GEO is about making evidence discoverable and intelligible to a machine that will synthesize an answer for that human. IQVIA argues that this should not mean manipulating AI towards a favored product, but ensuring that authoritative, current evidence is structured so AI systems can identify and interpret it appropriately.
That distinction matters enormously. This is not “marketing to the algorithm”. It is ensuring that good evidence survives algorithmic digestion.
There is an even more consequential implication. Let's suppose physicians repeatedly ask AI whether a medicine is appropriate after a particular prior therapy. Or in patients with a certain comorbidity or who is taking a certain concomitant medication. Or with an unusual biomarker profile. Or how its safety compares with an alternative in a particular population.
Those prompts are not merely commercial intelligence. Collectively, they may constitute a map of where the evidence stops being self-explanatory.
IQVIA explicitly suggests that prompt analysis can identify places where clinical confidence begins to fray, potentially informing evidence generation, Medical Affairs and scientific communications.
That creates intriguing possibilities for R&D, HEOR, RWE and Phase IV research. Today, evidence gaps are identified through literature reviews, advisory boards, primary research and conversations with experts. Tomorrow, aggregated and appropriately anonymized prompt behavior could provide another signal: the questions HCPs repeatedly encounter for which the existing evidence does not provide an easy answer.
The prompt could become a new unit of insight.
Paradoxically, none of this necessarily diminishes the importance of pharmaceutical field teams. It may simply render one of their traditional functions increasingly otiose.
If an HCP can retrieve a trial endpoint, interrogate a guideline and compare published evidence in seconds, there is diminishing value in sending a highly trained human being into the clinic merely to recite information. The opportunity moves towards interpretation.
IQVIA's speakers described a shift from information delivery towards consultative problem-solving: assisting HCPs to understand patient fit, clinical trade-offs, risk and how evidence applies in less tidy real-world circumstances.
AI may therefore make mediocre human interactions less valuable and excellent ones more so. That is not the extinction of the representative or MSL. It is a considerably more demanding job description.
I have written previously about AI becoming another participant in the treatment journey, alongside physicians, patients and health systems. What IQVIA's research adds is something more concrete. We can now see the beginnings of a new behavior: HCPs routinely turning to AI at moments of curiosity, ambiguity and clinical uncertainty.
That should alter how pharma thinks about its audience.
For decades, success meant showing up in the physician's mind. Then it meant showing up with the payer, the patient and the health system. Increasingly, it will also mean showing up accurately in the corpus of evidence from which machines construct their answers.
Not first because somebody has gamed an algorithm. Not because a brand has produced more digital detritus than its competitors. But because its evidence is credible, accessible, well structured, current and sufficiently useful to earn its place there.
Which brings me back to a golf cart, a over-under and a conversation in the Lehigh Valley.
My companion was worried that his medicine did not appear prominently enough when he asked ChatGPT about his cancer market. I suspect many pharmaceutical executives will soon find themselves asking versions of the same question. They should. The algorithm is no longer merely carrying the message.
It has joined the audience.
