Menu
Close
.svg)
Many years ago, shortly after founding ThinkGen, one of our pharmaceutical industry clients gave me what became one of the most interesting and memorable assignments of my career. She and her team wanted to understand, in considerable depth, what really happened when physicians spoke with patients with early-stage Alzheimer’s disease, and their care partners, about treatment.
The operative words were really happened. Rather than asking physicians to reconstruct these conversations afterwards, as we often do, we embarked upon an ambitious longitudinal observational study. We observed and video-recorded, with all appropriate consents, encounters among physicians, patients and care partners in doctors’ offices. We subsequently visited patients in their homes on several occasions, exploring what they had heard, what they understood, what concerned them and what they intended to do next.
The research was exceptionally rich. It was also complicated and expensive.
What stayed with me most was the disparity that sometimes emerged among three ostensibly similar things: what the physician communicated, what the patient and care partner actually understood, and what they ultimately did. A physician could clearly provide medically accurate information yet fail to make the treatment decision intelligible to the people sitting across the desk. Indeed, part of our remit was to test design elements that might facilitate those conversations. Patients could leave an appointment with impressions quite different from those the physician thought had been conveyed. Care partners, meanwhile, could become translators, advocates, skeptics or de facto decision-makers.
For a researcher, there is something almost anthropological about having a front row seat and watching these dynamics unfold. Depth interviews are invaluable, but human beings are imperfect chroniclers of their own behavior. Observational research offers a different and, in some respects, more authentic vantage point. It allows us to examine what occurs rather than relying solely upon what people subsequently say, or remember, occurred.
There is, however, a catch. Reality is gloriously untidy. Every patient in our Alzheimer’s study was different. One was anxious, another sanguine. One care partner dominated the conversation, another barely spoke. Medical histories, family circumstances, understanding of the disease and willingness to divulge concerns all varied. That heterogeneity was precisely what made the research so human. It also made comparison challenging.
If two physicians conducted very different conversations, how much of the difference reflected the physicians and how much reflected the patients sitting opposite them? How could we distinguish an idiosyncratic encounter from a generalisable pattern?
I have recently found myself wondering how I would design that Alzheimer’s study if the same client RFP landed on my desk today.
First, I would not abandon ethnography with real patients. Quite the opposite. There was something irreplaceable about hearing from people living with the disease and entering the homes in which that disease was actually being lived. But today I would consider adding something we did not have at our disposal then: a synthetic patient.
At ThinkGen, we call them THINKMates. They are AI-enabled synthetic avatars that can represent patients, caregivers, physicians and other healthcare stakeholders in realistic conversations. Their value, at least to my mind, resides less in the technological wizardry that breathes life into them than in the research possibilities they create.
Imagine, for example, recreating elements of that Alzheimer’s study with several synthetic patient and care-partner avatars. One might represent a recently diagnosed dementia patient frightened by what the diagnosis means but determined to preserve independence. Another might minimize symptoms and resist the implications of disease progression, much as one individual in our original research did. A third scenario might involve a highly engaged care partner whose presence subtly alters the tenor of the conversation.
Then imagine a sample of physicians encountering the same situation, up to 15 or 20 of them. The possibilities become intriguing.
We could observe which physicians recognize important cues, which pursue them and which miss them altogether. We could see how treatment benefit is explained, whether clinical vernacular is translated into intelligible language, how physicians respond to misunderstanding or hesitation, and whether the conversation concludes with genuine comprehension or merely its appearance.
The patient, in effect, becomes the constant. Differences in physician behavior become more conspicuous, potentially revealing distinct behavioral archetypes and even providing the basis for a qualitatively derived segmentation.
There is precedent for this idea. Medicine has used standardised patients for decades to examine clinician behaviour under controlled conditions. A landmark JAMA study published in 2000 used standardized patients as a reference method for measuring actual physician practice, and simulated patients have continued to feature in health-services research. The underlying principle is therefore hardly futuristic. Standardization affords the researcher a consistent situation against which differences in professional behavior can be examined.
What has changed is the machinery.
Synthetic avatars potentially make this kind of interaction more practical, adaptable and scalable. A simulated stakeholder can participate repeatedly in a research program without the logistical complexity involved in recreating comparable human encounters again and again. More importantly, the consistency of the situation gives researchers a different lens through which to examine behaviors.
That does not make synthetic-avatar research synonymous with conventional observational research. A simulated interaction remains a simulation. But it creates an intriguing new territory between asking someone what they would do and observing what they actually do when confronted with a realistic clinical interaction.
That discrepancy matters enormously in pharmaceutical marketing research.
Among the most common questions we ask physicians and other respondents are questions of intent: what they think and what they plan to do. Yet commercial strategies ultimately have to influence behavior in the real world. A message that performs beautifully on a Powerpoint slide and/or that earns high ratings in a quant survey may prove cumbersome when a physician has to explain it to an anxious patient. An HCP who embraces shared decision-making may behave differently when a patient challenges the recommended treatment. An access barrier that appears manageable in the abstract may become considerably more vexing when the patient's actual out-of-pocket cost enters the conversation.
Synthetic avatars give us another way to investigate those moments. They offer a more realistic glimpse of the situation itself, including the frictions that can alter its course.
The principle travels surprisingly well across commercial questions. An oncologist might encounter a synthetic patient struggling to weigh survival benefit against toxicity and time. A rare-disease specialist might meet a parent carrying misinformation gathered online. A physician treating chronic disease might encounter an ostensibly stable patient who divulges an adherence problem only when the right question is asked. The data supporting a new drug may sound compelling until the doctor realizes she must persuade a needle-phobic patient to self-inject. A representative or MSL might interact with a synthetic physician harbouring realistic objections, entrenched habits or considerable scepticism.
In each instance, the avatar is not particularly interesting because it looks or sounds human. It is interesting because of what the human interacting with it reveals.
That is an important distinction. Much of the current fascination with AI avatars concerns verisimilitude: How realistic do they look? How natural do they sound? How convincingly can they sustain a conversation? Those are reasonable questions, but from a research standpoint they are not the destination. The more consequential question is whether a sufficiently credible synthetic interaction can elicit behavior that we could not readily have uncovered by asking another question on a discussion guide.
There is another reason I find this concept promising. Watching a respondent navigate a realistic, potentially difficult interaction is only the beginning of understanding it.
A skilled researcher can interrogate the reasoning behind what occurred. Why did a physician pursue one cue but overlook another? What prompted the conversation to shift in direction? What did the physician believe the patient was communicating? What was easy to explain, and what proved unexpectedly awkward?
This conjunction of behavior and subsequent reflection is potentially powerful. It allows researchers to examine not merely what people say they would do, but what they actually do in a particular situation and how they make sense of their behavior afterwards. Could this help narrow the perennial “say-do gap” in research? It may well be part of the answer.
AI does not obviate the need for good researchers. Arguably, it makes their role more exacting. Technology can create an encounter and make it repeatable. It cannot absolve us of the responsibility to discern which behaviors matter, which are epiphenomenal, and what recurring patterns might mean for a brand, patient-support program or commercial strategy.
There are, of course, boundaries. A synthetic patient is not a human patient. An AI-mediated conversation cannot reproduce all the emotional gravity, unpredictability or consequence of an actual clinical encounter. Nor should THINKMates be regarded as substitutes for physicians, patients or caregivers in research.
Traditional qualitative work, quantitative research, ethnography and real-world observation answer questions synthetic personas cannot. The more sensible future is methodological pluralism, not methodological conquest. THINKMates add another instrument to the research arsenal rather than rendering the existing ones obsolete.
That is why I keep returning to the Alzheimer’s study.
I would not want to sacrifice what we learned from entering those patients’ homes. The quotidian details mattered. So did the asymmetries amongst what physicians said, what patients heard and what families eventually did. But if I were conducting that research today, I would want to complement that human richness with an ability we did not possess then: allowing multiple physicians to encounter sufficiently consistent patient situations that differences in their behavior become easier to discern.
For decades, pharmaceutical marketing research has become increasingly adept at asking people what they think, believe, prefer and intend to do. We have also embedded behavioral science into our fundamental operating system, recognizing that what people intend to do and what they actually do often diverge. Those questions remain indispensable. But intention and behavior are not synonyms. A strategy can look excellent in a PowerPoint deck, perform beautifully in message testing and still falter when a physician has to explain it to a frightened human being.
This, to me, is where THINKMates become considerably more interesting than the technology that animates them. AI's contribution to marketing research may not reside solely in doing familiar things faster or more cheaply. In some circumstances, synthetic avatars may allow us to investigate questions previously too cumbersome, costly or difficult to standardize, and to repeat those encounters until individual idiosyncrasies begin to reveal discernible behavioral patterns.
The most interesting future for AI in research may therefore have relatively little to do with replacing people (as many fear it will). It may instead lie in creating new situations through which we can understand people better. Perhaps the next frontier is not asking more ingenious questions about behavior, but finding better opportunities to observe behavior itself.
And sometimes, to understand what changes, it helps to have a patient who does not.
