By Jason Smith, CTO Within3
There is a version of “AI for Medical Affairs” that is essentially glorified transcription. You run your advisory board, a language model generates a summary, and a moderator spends the next two weeks cleaning it up and distributing a PDF that gets read once and filed. This version exists. It is being sold. It is not the point.
The version worth building is different. It treats every structured scientific engagement as a node in a continuously learning system, where expert input is captured in governed, structured form, synthesized against prior context, and routed to the teams that can act on it fast enough to influence the decisions in front of them. The technology to do this exists. The design discipline to do it in a compliant, governed way is the harder part, and it is the part that ultimately determines whether AI creates strategic value in Medical Affairs or just reduces transcription costs.
Why General-Purpose GenAI Misses the Point
The market has moved toward embedding generative AI productivity tools inside general workflow systems and meeting platforms. The argument is that AI-assisted summarization, available wherever the meeting happens, is sufficient for medical insights capture.
This argument is architectural, not preferential, and it is wrong.
Medical Affairs engagements are not generic meetings. They involve expert opinion on scientific evidence, off-label discussion that requires documented handling under company SOPs, potential adverse event mentions that trigger regulatory obligations, and strategic content that is competitively sensitive. A general-purpose meeting AI has no context for any of this. It processes audio or transcript and produces a summary. It cannot distinguish between a routine physician comment and a spontaneous AE report. It cannot apply company-specific guardrails for off-label content. It cannot route an insight about unmet patient need to the right Medical Excellence owner versus an insight about competitive mechanism to the right brand team.
The failure mode is not that general-purpose AI produces bad summaries. It is that it produces plausible summaries that look complete and are not, and that the gaps are precisely in the categories where accuracy matters most.
The Insight Yield Problem
Here is the operational problem that most Medical Affairs teams face and that most AI solutions do not address: insight yield is low not because engagements are unproductive, but because the translation from engagement to usable intelligence is slow, lossy, and disconnected from the systems where decisions are made.
A top-tier advisory board might generate three to five high-quality strategic insights that are genuinely decision-relevant for a launch or lifecycle management question. Identifying those insights, structuring them against the knowledge gap they address, and routing them to the right cross-functional owners typically takes weeks when done through manual reporting processes. By the time the insight reaches a brand team or a field medical leader, the window for acting on it may have passed.
This is not a novel observation. The chronic failure to convert scientific engagement into timely organizational intelligence is well understood inside Medical Affairs functions. What is less understood is that solving it requires a different architecture, not a faster version of the same process.
The EMA’s published guidance on LLMs in medicines regulation is explicit about the risk on the other end: overreliance on AI-generated summaries, hallucinations in complex scientific content, and inadequate evaluation of model outputs.² Peer-reviewed research has documented that evaluation approaches for healthcare language models are inconsistent in ways that create material reliability risks.¹⁵ The answer to both problems converges on the same design requirement: retrieval-grounded summarization, mandatory human review gates, and draft-mode labeling that prevents any AI-generated summary from being treated as authoritative until a qualified reviewer has signed off.
A well-designed scientific exchange architecture threads this needle. AI accelerates synthesis through draft generation grounded in what was actually said in the engagement and traceable to specific participant statements. Human review validates before distribution. And the routing layer ensures that what gets delivered reaches the owner with both the context and the authority to act.
The Consent and Governance Architecture
Deploying AI in scientific exchange requires consent architecture that is precise, not implied. Participants need to know that AI is assisting with summarization, what data is being processed, and how the outputs will be used. This is not a legal formality. It is the foundation of trust with the external experts whose willingness to engage candidly is the entire premise of an advisory program.
Beyond consent, the governance requirements for this workflow are specific and non-negotiable. Role-based access controls need to distinguish between those who can see raw transcript-level content and those who receive only the structured insight summary. AI-generated drafts must be clearly labeled throughout their lifecycle. Human sign-off must be documented with the reviewer’s identity and timestamp. Where multilingual translation is involved, as it is in any global advisory program, the translation governance layer needs to be explicit about which language was treated as authoritative and how translation quality was assessed.
HIPAA’s minimum necessary standard governs any Protected Health Information that flows through these processes.²⁵ GDPR’s lawful basis requirements apply to European participant data.¹⁴ These are standard requirements for any global Medical Affairs program. They need to be designed into the platform and process rather than added as configuration by the customer team.
What “Year-Round Insight Engine” Actually Requires
The aspiration to turn advisory programs into a year-round insight engine is real and achievable. The requirements to make it real are worth naming precisely.
Structured capture across multiple engagement types is the first requirement. Advisory boards are one source. MSL field interactions, congress symposia, publication steering committee meetings, and investigator meetings all produce scientific exchange content. A year-round system needs to draw from all of them under a consistent governance architecture, not treat each engagement type as a separate data silo.
Insight accumulation with deduplication is the second. If your fourth advisory board of the year produces the same three insights as the first two, that confirmation is signal about something: consistency of expert perspective, perhaps, or persistence of an unmet need. But it is different signal than genuine novelty. A system that cannot distinguish between confirmation and new information generates noise at scale, and noise at scale is worse than no system at all.
Routing to documented action is the third. An insight that reaches the right owner and triggers a documented response closes the loop. An insight that disappears into a report is expensive notes. The KPI that validates the architecture is not how many summaries were generated. It is the percentage of insights that were routed to a cross-functional owner and resulted in a documented action or decision.
The Medical Affairs function has the most direct relationship with the scientific experts whose perspective shapes launch strategy, prescribing behavior, and clinical practice. Converting that relationship from a series of discrete engagements into a governed, continuously learning intelligence asset is one of the highest-leverage transformations available to a pharmaceutical organization. Platforms like Within3’s Virtual Advisory Board and Moderator Assistant are explicitly designed around this architecture.⁴⁶ ⁴⁸ The design philosophy, structured engagement, AI-accelerated synthesis, mandatory human review, and insight routing to action, is the pattern the industry should be building toward regardless of which platform executes it.
References
European Medicines Agency / Heads of Medicines Agencies. “Harnessing AI in Medicines Regulation: Use of Large Language Models (LLMs).” https://www.ema.europa.eu/en/news/harnessing-ai-medicines-regulation-use-large-language-models-llms
European Parliament and Council. General Data Protection Regulation (GDPR). https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng
Published evidence on evaluation inconsistency in healthcare LLM deployments. JAMA. https://jamanetwork.com/journals/jama/fullarticle/2825147
U.S. Department of Health and Human Services. “HIPAA Privacy Rule.” https://www.hhs.gov/hipaa/for-professionals/privacy/laws-regulations/index.html
Within3. Virtual Advisory Boards. https://within3.com/application/virtual-advisory-boards
Within3. Moderator Assistant. https://resources.within3.com/moderator-assistant-video
