AI is being operationalized into biopharma CI / BD teams at different rates, but it's interesting observing the problems those furthest along focus on vs. those just getting started. It comes down to producing work vs. verifying it.
1. The cost to produce a credible-looking output is collapsing
2. So we see an explosion of knowledge work output
3. But there's always a system bottleneck, and now it's
(a) validating these outputs
(b) deciding what actually matters for the question at hand
(c) taking ownership
This is clearly a problem with knowledge work broadly, but Pharma CI is an interesting extreme case:
• The work is full of classifications that require domain expertise / judgement
• There's rarely a clean "ground truth" to benchmark against
• Sources are scattered, incomplete and frequently contradictory
• The feedback loop on whether you're right can be years long. A persuasive output (AI is great at this) and the correct output can look pretty indistinguishable for a long time.
The old days of taking days / weeks to aggregate data & draft an analysis isn't coming back, but it's clear now that one hidden function was that time spent building the output was in itself a verification layer. The analyst had to engage directly with the messy, raw data, make many inclusion decisions and ultimately understand how the answer came together. As a byproduct, they also got training on how to do the task better.
This process didn't mean the answer was correct or the output was high quality, but it did create a chain of custody. AI inverts this by separating the final output from the context required to defend it:
• Your agent returns 50 assets for a landscape. How do you know that was comprehensive?
• Your team fixes many edge cases. How do those judgements get inherited by the next 100 analyses across teams, rather than trapped in one chat?
The last mile to bridge output --> decision is everything now. Intelligence ≠ auditability.
To be clear, the answer is not to return to 100% manual research or require experts to check every AI-generated cell, essentially duplicating the work. It's to build systems that make verification scalable:
• Construct a tailored dataset for every decision
• Preserve the evidence trail behind every claim and classification
• Surface conflicts, uncertainty and material exceptions for expert review
• Capture the expert’s corrections and reasoning so they improve future work
• Keep the analysis current as new evidence changes the market
The model is an important part of that system. But the model alone is not the system.
