Pharma insights · · 8 min read
Primary research vs. synthetic data: where pharma insights budgets should go
Acumen Editorial · Updated November 20, 2026
Pharma insights teams are being asked to do more with AI and less with fieldwork. In a November 16, 2026 Pharmaceutical Executive feature, Spherix's Dan Barton describes anecdotal signs that some spend is moving from primary research toward internal AI builds, synthetic data and advanced models. He puts healthcare AI spending at an estimated $480 million in 2024 and $1.4 billion in 2025.
His warning is short: "AI without fresh evidence amplifies inference, not truth."
We agree with the direction and disagree with the framing that one method must win. The real decision is which question goes to which method, and what keeps the synthetic layer honest. If you lead primary market research, forecasting or insights procurement, here is how we would make that call.
What is the difference between primary research and synthetic data?
Primary research collects new evidence from real people. A cardiologist explains why she switched a patient. A payer describes how a new launch will be managed. A patient describes why he stopped a therapy. Nothing in that answer existed before the interview.
Synthetic data generates modeled answers from evidence that already exists. ZS, in a July 23, 2026 piece, splits it into two very different products:
- AI personas are in-character respondents you talk to one at a time. A language model writes each answer. ZS calls a persona an archetype, not a calibrated sample.
- Simulated panels are synthetic populations of hundreds to thousands of respondents. Their choices come from patterns learned from real respondent data and behavioral-science constructs. They produce rating distributions and segment-level reads.
The distinction matters for budget. ZS reports that persona-style approaches correlate with real market research at roughly 50% to 65%. Its simulated panels reach roughly 85% to 95% in blinded validation studies. Those are vendor-reported figures, but the gap shows that "synthetic" covers tools of very different quality.
How do they compare side by side?
| Factor | Primary research (HCPs, patients, payers) | AI personas | Simulated panels |
|---|---|---|---|
| Source of answers | Real respondents, collected now | Language model, prompted in character | Models trained on real respondent data |
| Best for | Novel questions, lived experience, sizing, calibration | Early exploration of the "why" | Repeat message and concept testing |
| Statistical output | Yes, with a proper sample | No | Yes, distributions and segments |
| Reported fit to real research | The reference point | ~50–65% (ZS) | ~85–95% blinded (ZS) |
| Speed | Six to twelve weeks per cycle (ZoomRx) | Minutes | Minutes, after training |
| Main risk | Cost and time | Plausible but uncalibrated answers | Drift as the market changes |
Can synthetic respondents replace primary research?
No, and the vendors building them say so. ZS lists the cases where real research remains essential: first-in-class questions with no historical analog, deep emotional and lived-experience work, absolute measures of size or share, nonverbal cues and final validation of the panel itself.
Barton makes the launch case directly. At launch there is no historical data to synthesize. A model can summarize what the market said last year. It cannot tell you how specialists in a narrow indication are reasoning about a drug approved last month.
The cost case is weaker than it looks, too. ZoomRx's May 2026 position paper, Synthetic Data Is Not Enough, estimates that sample costs are 15% to 30% of a typical pharma market research budget. Replacing respondents saves that line item. It does not remove design, analysis or interpretation.
Where do synthetic panels earn their place?
They earn it between fieldwork waves. Once a panel is trained on recent, high-quality respondent data, it can screen dozens of message variants, rank concepts and test sequences before you field the best few with real HCPs. ZS reports that using a panel ahead of a major study cut research costs by roughly 30% to 50% and timelines by a third to a half.
Note the condition: ahead of a major study, not instead of one. The panel narrows the options. Real respondents confirm the choice.
Three conditions make this work:
1. Recent training data. ZS says about 50 to 60 recent transcripts can be enough when combined with a strong proprietary decision dataset.
2. A refresh cadence. ZS suggests many panels should be retrained once to a few times a year.
3. A blinded check. Each panel should be compared against a fresh wave of real research before its outputs drive a brand decision.
All three depend on primary research. The synthetic layer is only as current as the last real conversation it learned from.
What does a sound budget split look like?
We would not set a fixed percentage. We would route each question by type. Use this checklist when you plan the year:
- Is the question new? First-in-class, new indication, new competitor or new access rules. Send it to primary research.
- Do you need a number? Market size, share, prevalence or switching rates. Primary research.
- Is it emotional or lived experience? Patient journeys, caregiver burden, physician fear of an adverse event. Primary research.
- Is it iterative? Ten versions of a message, five concept framings, sequence order. Simulated panel, then confirm the winners with real respondents.
- Is it exploratory? Early hypotheses before a study design. An AI persona is acceptable, with outputs labeled as unvalidated.
- When was the panel last trained? If the market has moved since then, fund a refresh wave before trusting it.
- Who checks the output? Name the expert who reviews every AI-assisted finding before it reaches a brand team.
The last item is not optional. Barton notes that at Spherix every AI-assisted output passes through a human expert before it reaches a client. We hold our own work to the same standard.
What do regulators expect from AI in drug development?
There is no rule on synthetic market research. The broader direction is clear, though. On January 14, 2026, FDA and EMA jointly issued 10 guiding principles of good AI practice in drug development, as reported by RAPS. They include a risk-based approach tied to a clearly defined context of use, multidisciplinary expertise, good data governance and documentation, and risk-based performance assessment across the lifecycle.
Insights work sits upstream of regulated decisions, but it feeds forecasts, launch plans and payer strategy. Applying the same logic is sensible: define what each synthetic tool is used for, document what data trained it and test it against real evidence on a schedule.
What should procurement ask before shifting budget?
Ask these before moving money from fieldwork to a synthetic tool:
1. Is this a persona or a simulated panel, and what statistical output does it produce?
2. What real respondent data trained it, from which specialties, and how recent is it?
3. What blinded validation was run, against which study, and what was the result?
4. How often is it retrained, and who pays for the refresh research?
5. Which questions does the vendor say it should not answer?
6. Who reviews outputs before they reach a brand team?
ZoomRx frames the core test well: ask where the behavioral foundation came from and what happens to it as the market changes. For a fuller list, see our guide on what to ask a synthetic persona vendor.
How Acumen helps
We supply the real evidence underneath. Our market research data for AI service delivers verified HCP, patient and payer interviews, structured transcripts and expert-graded validation sets. Insights teams use them to train simulated panels, refresh them on a schedule and run blinded checks against fresh responses.
Every respondent is verified by specialty and every dataset is reviewed and adjudicated. You keep a primary research layer that is current, documented and built for AI from the start.
FAQ
What is the difference between primary research and synthetic data in pharma?
Primary research collects new evidence from real HCPs, patients and payers through interviews and surveys. Synthetic data generates modeled answers from evidence that already exists, either through AI personas written by a language model or through simulated panels trained on real respondent data. Primary research finds what is new. Synthetic data reuses what is known.
Can synthetic respondents replace primary research?
No. Even vendors that build them say real research stays essential for first-in-class questions, lived experience, absolute sizing, nonverbal cues and validating the synthetic tool itself. At launch there is often no historical data to model. Synthetic respondents work best between fieldwork waves, narrowing options that real respondents then confirm.
What is the difference between AI personas and simulated panels?
ZS describes AI personas as in-character respondents generated by a language model, useful for exploration but not statistically calibrated. Simulated panels are synthetic populations of hundreds to thousands of respondents whose choices come from patterns in real respondent data. ZS reports roughly 50 to 65 percent fit to real research for personas and 85 to 95 percent for panels.
How often should a synthetic panel be retrained?
ZS suggests many panels should be retrained once to a few times a year, and sooner when the market changes. A new competitor, label update, guideline change or access shift can make a panel stale. Each retrain needs fresh primary research from the right specialties, followed by a blinded check against real responses.
How much can synthetic data save on pharma market research?
Less than the headline suggests. ZoomRx estimates sample costs at 15 to 30 percent of a typical project budget, so replacing respondents saves only that share. ZS reports 30 to 50 percent cost savings when a simulated panel is used ahead of a major study to narrow options, not instead of the study.
What data do you need to build a reliable simulated panel?
You need recent, specialty-matched respondent data with clear decision signals. ZS says about 50 to 60 recent transcripts can be enough when paired with a strong proprietary decision dataset. The data should come from verified HCPs, patients or payers, be documented for governance and be refreshed on a schedule.