This article was originally published by Anna Forsythe in Forbes on 03 March 2026.
I spend a lot of time working with artificial intelligence, and I am constantly struck by a contradiction: On one hand, AI has become remarkably fluent. It can summarize dense material, surface insights from massive datasets and produce text that looks confident enough to pass for expertise. On the other hand, in some of the places where accuracy matters most, AI remains almost entirely clueless about real decision making. Pharmaceutical evidence generation is one of those places.
A Systemic Blind Spot
Despite years of excitement about AI in healthcare, there are still no widely used systems that autonomously maintain living, submission-ready evidence for regulatory or reimbursement purposes. This is often framed as a technology gap. In my experience, it is much more than a gap; it is a systemic blind spot not easily remedied. It exposes a misunderstanding of what regulated evidence actually is and what today’s AI systems are built to do.
Systematic literature reviews are not summaries. Nor are they answers to questions posed after the fact. They are strictly governed by carefully regulated processes. They begin with predefined protocols, require transparent inclusion and exclusion logic, and must withstand scrutiny long after the original analysis is complete. Regulators and payers do not just ask what conclusions were reached. They ask how those conclusions were formed, what was excluded along the way and whether the same decisions would be made again if the review were repeated. They need to understand how those decisions were made.
This is where conversational AI systems struggle in ways that cannot be fixed with better prompts or larger models. Large language models are optimized for plausibility, not traceability. They are designed to produce likely responses, not to preserve the reasoning trail behind each decision. While they can describe evidence convincingly, they cannot reliably explain why one study mattered more than another, or why a marginal trial was excluded at a particular point in time. When the literature updates—as it constantly does in oncology—those inconsistencies compound.
I have seen teams experiment with using chatbots to “speed up” evidence reviews, only to discover that what looks efficient at first quickly becomes indefensible under scrutiny. The problem is not that the chatbot models are unsophisticated. It is that the task itself—the special sauce behind the systematic literature review—is not generative in nature. That special sauce needs to be procedural, auditable and accountable.
At the same time, the pressure to maintain living, constantly updated evidence has never been higher. Clinical data no longer arrives in neat cycles. New trials appear between guideline updates. Regulatory decisions shift comparators. Payers ask questions that did not exist when the original review was written. Static evidence simply cannot keep up.
This has created a strange stalemate. Fully manual processes are too slow and resource-intensive. Fully automated ones are not trustworthy. Many organizations quietly accept the friction, even as they invest heavily in AI elsewhere.
A Change In Design
What ultimately breaks that stalemate is not better AI, but a different way of designing work.
In regulated environments, the only approach that scales is human-in-the-loop intelligence. Machines do what they are good at—continuous surveillance, structured extraction, pattern detection—while humans retain ownership of judgment, interpretation and accountability. When designed properly, this does not slow teams down. But it does change where expertise is applied.
What surprises many leaders is that this challenge is not unique to pharma. Several years ago, I had a conversation with an executive in commercial aviation who described a similar tension. Modern aircraft are astonishingly automated. They can take off, navigate complex airspace and land with minimal human input. Yet aviation has never tried to remove pilots from the cockpit. In fact, as automation has increased, pilot training has become more rigorous, not less.
The reason is trust. When something goes wrong at 35,000 feet, no one accepts “the system thought it was likely” as an explanation. Automation is expected to assist, not absolve. Human oversight is not a fallback; it is part of the system’s credibility.
Evidence generation works the same way. Regulators do not reject AI because it is new. They reject opacity. They expect to see where judgment was applied, and by whom. Systems that blur that boundary undermine trust, even if their outputs look impressive.
What’s Holding Organizations Back?
What ultimately holds organizations back from building these hybrid systems is rarely technology. It’s culture. Most companies are still organized around projects with defined endpoints, not living assets that require continuous stewardship. Evidence is treated as a document to be delivered, not an infrastructure to be maintained. AI initiatives are evaluated on novelty and visibility, not on whether they quietly reduce friction year after year.
Changing that requires leadership restraint. It means resisting the temptation to deploy tools that demo well but cannot be defended later. It also means investing in governance, workflow redesign and cross-functional ownership—none of which make headlines, but all of which determine whether AI creates real value.
The most important lesson I have learned working at the intersection of AI and regulated decision making is this: Fluency is not the same as reliability. The organizations that succeed with AI will not be the ones that generate the fastest answers, but the ones that can explain and stand behind those answers when it matters.
In pharma, as in aviation, intelligence is only as valuable as the trust it earns. And that means human beings in the chain of command.
