AI can analyse evidence faster than ever, but primary research with customers still creates the evidence pharmaceutical companies require for strategic decision making. By Dan Barton, Spherix.
In the 2000s I participated in the pharmaceutical industry rush to offshore talent, chasing exactly the logic now driving parts of the AI revolution.
The efficiency case was real, and the savings were real. However, as firms leaned into a capability they didn’t fully understand, the costs boomeranged years later in people and work that had to be hired and rebuilt.
The lesson wasn’t that offshoring was wrong; it was that efficiency pursued without regard for what’s being hollowed out creates problems that surface long after the savings are booked.
Pharmaceutical firms are under more pressure than ever to expand R&D productivity, increase sales, and cut costs. At the same time, everyone is chasing opportunities to optimise big data, advanced analytics, and AI.
The pressures are justified, and at Spherix we feel the same pressures and at the same time excited by its potential.
We use AI every day ourselves to improve quality control, interrogate datasets, surface patterns across years of proprietary research, connect those findings with what’s publicly available, and provide insights to clients faster.
The goal isn’t to generate evidence artificially; it’s to extract more value from the evidence we’ve already invested in creating.
But there’s a quiet risk in the rush. Overreliance on AI alone can create problems, and so can stripping human oversight out of critical processes in the name of speed.
The debate worth having isn’t whether to use AI; it’s whether we’re using it for sustainable long-term value or for short-term efficiency gains that create systemic problems later.
A model isn’t neutral intelligence that improves every input equally. It reflects what it’s been trained on and what it can reach.
In the world of pharmaceuticals where decisions are expensive to reverse, the inputs are everything. As AI makes analysis faster and more accessible, competitive advantage increasingly shifts upstream to the quality and uniqueness of data feeding the model.
The models become commoditized. The inputs will not. Several ambitious pharma marketing ideas landed with a thud not because the underlying premise was wrong, but because the technology or the organizational execution wasn’t ready.
A few stand out: e-detailing, omnichannel marketing, closed loop marketing, and programmatic digital targeting. Each promised transformation drew capital fast, promised exponential growth or cost reduction without more understanding of why prescribers behave as they do.
The error was never the technology. It was the assumption that a new model, or a cheaper way of working, was a costless substitute for a capability that took years to build.
AI may ultimately prove more transformative than any of those shifts. That’s precisely why getting the operating model right matters.
The current wave carries a familiar risk, only faster. Healthcare AI spending nearly tripled in a single year, from an estimated $480 million in 2024 to $1.4 billion in 2025.
Anecdotal signals suggest a meaningful share of that spend is shifting from proven primary research towards internal builds, synthetic data, and increasingly sophisticated models on the argument that AI can serve as a substitute for physician, patient, and payer research.
There are some camps that believe synthetic data can completely replace valuable insights from primary research. That argument has no empirical support, yet.
It’s spend first, measure later, and, unfortunately, discover the gap when it shows up in brand performance.
What actually drives decisions
Consider the questions that brand teams need answered to make a launch call, respond to a competitive entry, or defend a resource allocation: How is a specialist reasoning through a narrow indication right now?
How does a biologic-experienced patient weigh a treatment switch? Where does a payer stand on a newly approved agent this quarter not last year?
That information isn’t published anywhere, and it isn’t sitting in a claims feed waiting to be scraped.
A general-purpose model can synthesise what’s already public, but it can’t manufacture behavioural evidence that doesn’t yet exist. It can tell you what a market looked like last year, and not why prescribing has shifted over the last six months among the specialists who matter to your brand.
That kind of evidence has to be gathered directly from the field, kept current, and generated independently from the public and manufacturer sponsored information that increasingly forms the common knowledge base available to company sponsored AI systems.
And one part of this is structural, not a matter of degree: at launch, there’s no historical data to synthesise. A more powerful model operating without evidence doesn’t create evidence, it creates a more sophisticated inference.
That distinction matters. AI is extraordinarily good at finding, connecting, synthesising, and interrogating information. But it is not inherently an evidence-generation system. When the question is what physicians, patients or payers actually think and do today, somebody still has to create the evidence.
Where synthetic data helps
Synthetic data has real uses. Synthetic data can extend existing evidence, simulate scenarios, fill gaps and stress test hypotheses. What it cannot do on its own is establish new empirical evidence about how real physicians, patients, or payers are behaving today.
Synthetic data can help generate hypotheses; primary research tests whether those hypotheses survive contact with reality. A synthetic respondent can generate a surprising answer.
What it cannot establish is whether that answer reflects a real physician, patient, or payer in the market.
In our research, the surprise is often the whole point: the physician whose reasoning breaks the pattern, the patient barrier the existing evidence didn’t predict, or the market shift that becomes visible only when new data are collected.
Ask synthetic data to carry a launch decision alone, and you’re still reasoning from what was already known.
There’s a broader risk here as well. AI systems largely reason from what has already been observed, documented, or generated.
If organisations increasingly replace primary evidence with synthetic outputs derived from existing evidence, the system risks becoming recursive: increasingly sophisticated analysis of an increasingly closed information set.
Primary research does something fundamentally different: it introduces new observations into the system.
The standard to which we hold ourselves
In January 2026, the FDA and EMA jointly issued Guiding Principles on Good AI Practice in Drug Development, emphasising data quality, validation rigour, and human oversight.
We’re not in drug development ourselves, but the direction is the same: even in the highest-stakes corner of pharma AI, regulators are converging on the standard I’d hold our own team to, regardless of what they require of us specifically: I want to know what the model is reasoning from, who reviewed the output, and who’s accountable if it’s wrong.
We establish documented workflows for how AI is used in client-facing processes, including where expert review and accountability sit. Every AI-assisted output at Spherix passes through a human expert before it reaches a client.
Not because technology can’t be trusted, but because human expertise isn’t simply the last quality-control step. It determines which questions are worth asking, which signals deserve scrutiny, what the evidence can and cannot support, and what a finding actually means for a brand. Judgement matters most when the stakes and ambiguity are highest.
The case, plainly
This isn’t a case against AI. It’s a case for using it towards long-term value instead of simply for efficiency gains.
Automate the mechanical work – data wrangling, first-pass synthesis, pattern-surfacing across a decade of studies – and let it run fast, then reinvest the time it buys back into deeper reflection, richer inputs, and sharper judgement on the calls that carry real consequence.
Get this right, and the next five to ten years look genuinely different: AI clears the mechanical burden, so experts spend their time on interpretation and decision quality, and insights move at the speed of the market. AI can find patterns, connect evidence, generate hypotheses, and accelerate interpretation.
Primary research with real physicians, patients and payers establishes what is actually happening in the market and continually gives those systems something new to learn from.
The competitive advantage won’t come from AI alone, proprietary data alone, or human expertise alone.
It will come from combining all three: differentiated evidence that others can’t simply prompt their way into, technology capable of extracting more value from it, and experts who know what the evidence means.
The models will keep improving. The inputs, and the judgement, won’t build themselves. That’s the work. That’s where the advantage lives.
