The pharmaceutical business is generating more clinical information than ever before, but it is finding it difficult to translate that evidence into timely strategic decisions.
The National Library of Medicine reports that PubMed added over 1.5 million citations in fiscal year 2023, or around 4,300 every day. In addition to that literature, medical affairs teams have to take in data from Congress, empirical findings, and field inputs from healthcare professionals (HCPs). This volume makes it more difficult for pharmaceutical companies to respond to clinical evidence as soon as it becomes available.
AI has its own failure modes, yet it can reduce synthesis time. In a 2023 study published in npj Digital Medicine, researchers testing large language models (LLMs) on medical evidence summarization discovered that the models could generate factually inaccurate and unduly convincing summaries, with higher errors on lengthier texts. Since then, the U.S. Food and Drug Administration has developed a risk-based credibility approach for AI models used to assist medication regulatory decisions.
Taken together, these results point executives to a more focused question: which choice should AI enhance, and how will leaders tell?
On the AI in Business podcast, Senior Editor Yolandi de Weerdt and Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals, discuss what practical AI adoption looks like in evidence-based, regulated pharma processes.
This essay looks at the fundamental knowledge pharmaceutical executives need to make quicker, more convincing clinical and strategic decisions using AI-accelerated evidence:
- Decision-anchored AI assessment to demonstrate evidence-to-action impact: Identify the clinical or business decision that AI should expedite, then assess success using a downstream outcome that the company already monitors, such as clean trial data or site activation.
- AI synthesis with source verification to avoid compounding evidence errors: Before AI-generated summaries make it to field reports or strategy decks, where a single confused finding is multiplied across all subsequent decisions, compare them to the original abstracts, posters, and publications.
- Funded learning curves for the deployment of AI in regulated work that complies: AI business cases should include a training quarter and a supervised-use quarter so that teams can prompt, evaluate, and validate results before gains are tallied.
The entire episode can be heard below:
Episode: Using Practical AI Adoption to Accelerate Evidence to Action in Pharma with Pfizer’s Nabil Khan
Visitor: Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals
Medical affairs, clinical development, clinical trial operations, and scientific engagement are among the areas of expertise.
Brief Recognition: As Pfizer’s Medical Director for Internal Medicine Antivirals, Nabil Khan collects field ideas from healthcare professionals and communicates them to clinical development teams. Prior to joining Pfizer in 2024, he worked in top clinical research positions at PPD, ICON, and IQVIA, monitoring clinical trial sites and supporting studies in neurology, cardiometabolic illness, ophthalmology, and vaccines. He was Impact Physician Group’s Director of Clinical Operations. He graduated from Xavier University School of Medicine and is a doctor with over ten years of experience in clinical research and operations.
Measuring Decision-Anchored AI to Show Impact of Evidence-to-Action
In the pharmaceutical industry, the delay in making choices is more often the issue than the evidence itself. According to Nabil Khan, the difference is one of the most significant bottlenecks in the sector.
Access to publications, congressional presentations, real-world facts, and field insights is rarely a barrier for medical affairs teams. He claims that because the capacity to synthesize, evaluate, and act upon clinical and scientific data has not grown at the same rate as the volume being produced, pressure is mounting.
According to Khan, each seat in the process has a different bottleneck. Medical teams must determine which HCP conversations provide clinically meaningful information in the field. These findings come secondhand and must be translated into strategy inside clinical development or medical affairs. He claims that the primary challenge on both sides is separating noise from real signal and proof.
Khan lists three ways that AI modifies the process. It expedites synthesis by allowing teams to process enormous numbers of publications, congress abstracts, field insights, and real-world data much more quickly, freeing up time for interpretation and decision-making. It finds new signals and patterns that are hard to find in a manual review. Additionally, it improves alignment between clinical development, medical affairs, and commercial teams by creating a more uniform body of knowledge, which he refers to as a stabilizing characteristic. “Allowing the experts to spend less time gathering the information and more time acting on it” is the aim, according to him.
When asked what a senior leader should do when a team reports that AI is accelerating its progress, Khan likens adopting AI without a clear problem to having a toolbox full of tools and not knowing which one to use. The choice itself is where he begins:
“A well-defined business problem, not just the technology itself, must be the first step in the successful implementation of AI. Therefore, leaders should question themselves and their teams, “What decision are we trying to improve on?” Which choice are we attempting to expedite? What choice are we attempting to scale?
— Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals
His response highlights four queries that executives should address in any suggested workflow for AI evidence:
- Which choice are we attempting to make better? According to Khan, a team may determine whether or not AI is guiding them in that direction if the business aim is clearly stated up front.
- What proof does the system rely on? Data quality is crucial since AI is only as trustworthy as the data it is fed, trained on, and analyzed.
- What role do expert oversight and governance play? Khan advocates for the early establishment of governance in a regulated business to guarantee openness, compliance, and supervision, as well as for the close involvement of independent physicians and scientific experts to ensure that outputs remain actionable and clinically relevant.
- Which downstream result will demonstrate its effectiveness? According to Khan, success doesn’t become apparent until the AI-supported choice is put into action.
- How far downstream that final response can sit is indicated by the trial site selection. AI can assist in locating potentially pertinent HCPs in an area, but only after field medical teams screen the investigators and clinical development puts the sites through protocol training, regulatory approval, compliance, and activation—are those sites generating clean data? Only then can a team determine whether the effort upstream has been successful, according to Khan.
AI Synthesis with Source Verification to Avoid Compounding Evidence Errors
AI is appealing for evidence synthesis because of its speed and consistency, but its mistakes are also expensive. Khan considers consistency while assessing the risk:
It’s not always a good thing to be consistent in the wrong direction. You consider consistency to be a good thing, but it’s obviously not a good thing if you’re constant in the wrong way. The dangers I’ve observed are that any errors or inconsistencies will intensify the hostility.
— Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals
At medical congresses, his team makes extensive use of AI to summarize the vast amount of content presented as well as to arrange the limited time on site. He claims that since several presentations can occasionally be combined into the final product, what they receive needs to be verified twice:
Because their titles are so similar, presentations and posters can be combined in some way. Additionally, their subjects may be similar in addition to their titles. The speaker may even be the same, but they are giving two different presentations. AI takes that data and occasionally compiles it when it shouldn’t. When you are building on that information after it has been assembled, you are building upon inaccurate information.
— Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals
He reiterates his operating principle, which is to treat AI as “a tool rather than a source,” at the end of the discussion. If AI is given inaccurate or poor information, it amplifies the misinformation, and once it spreads, it is extremely difficult to contain in the healthcare industry, as it is on social media.
When the approach is applied to his Congress example, it boils down to three checks before an AI summary is sent to other teams:
- Traceability to primary material: Each assertion in Congress summaries, literature reviews, and evidence digests has a clear path back to the publication, abstract, poster, or presentation that supported it.
- Closer examination for look-alike inputs: The synthesis is subject to more scrutiny due to similar titles, overlapping subject matter, and shared presenters.
- Named accountability: Before additional teams build upon an AI-assisted synthesis, it is validated by a designated human expert.
Funded Learning Curves for Adoption of Compliant AI in Regulated Work
Access to an AI tool and being prepared to utilize it on regulated tasks are two distinct things in a regulated industry. When asked which stage businesses most frequently omit when using AI, Khan mentions training. However, he is unsure if businesses completely ignore it or just don’t take it seriously enough.
In addition to mastering the tool itself, professionals in the pharmaceutical industry also need to understand how AI use interacts with external communications, copyright, regulatory restrictions, and compliance obligations. He describes how that exposure is produced using a general-purpose technology:
“AI wasn’t developed for a particular field. It wasn’t created only for the pharmaceutical industry, accounting, or healthcare. Since it’s a general product, there are several rules, compliance concerns, and copyright difficulties that people who aren’t accustomed to handling them won’t be able to handle. They run the risk of being non-compliant and violating federal standards if they merely utilize AI to provide them with responses and to elaborate on subjects.
— Nabil Khan, Pfizer’s Medical Director for Internal Medicine Antivirals
According to Khan, when an employee posts AI-generated content that turns out to be non-compliant, both the corporation and the individual are at risk. Additionally, the tools themselves are different. He points out that utilizing AI on a personal phone is different from using it on a business computer for a regulated job, and teams must learn how to utilize ChatGPT, Claude, Copilot, and the many other accessible systems in their respective context.
Khan adds a disclaimer to any timeline: big businesses are still figuring out where AI works best through trial and error. In light of this, his learning curve sequence consists of three stages:
- Teams receive training for roughly a quarter (three to four months) during which they learn how the tools operate, how to trigger them, and where regulatory and compliance constraints apply.
- Another quarter or so of supervised use: Workers apply AI to actual tasks, observe what works and what doesn’t, and develop the ability to determine whether an output is trustworthy or requires verification.
- Regular integration: Once teams are able to confidently interpret AI’s output, it is included in daily workflows.
The use case also affects the speed. He points out that summarizing a clinical or journal article or creating a presentation based on obtained medical evidence involves more preparation than summarizing or writing emails. Khan’s minimum estimate for that evidence-based work is five to six months, and he anticipates that real adoption will take longer.