Every drug company depends on a handful of scientists whose judgment decides which programs go forward. They read the same literature as everyone else and notice different things in it. One sets aside a paper that looks persuasive because the method underneath cannot carry the conclusion. Another trusts a quieter, less-cited study because it fits what is already known about the mechanism.
Almost none of that reasoning is written down in a form another scientist could pick up and use, so when these people retire or leave for a competitor, the company finds itself rebuilding knowledge and expertise it once heavily relied on.
Organizations have responded by giving those scientists better AI tools, and at the level of the individual task it works. A toxicologist can draft a section of a safety assessment in two hours instead of two days, and a literature search that used to take an afternoon now takes minutes.
The assessment still arrives on the same date, because it moves through the same committee on the same schedule, and the time saved upstream is gone before it reaches anything a program leader measures. Task-level speed does not become program-level speed on its own.
A program’s pace is set by how an organization weighs the evidence and decides when it has enough confidence to commit, and that still lives in the experience of very few people.
Those few earn their standing by catching the errors a face-value reading misses. An experienced scientist might notice that what appears to be five independent papers actually traces back to a single dataset reanalyzed multiple times. The apparent consensus collapses into one experiment wearing five coats.
Knowing to test for that, and where to look, is the kind of reasoning no procedure records. It is what a general-purpose model glosses over when it counts the same finding, published in several places, as independent support.
What Drug Discovery Can’t Borrow From Software
Software engineering encountered a version of this and worked out an answer. A coding agent runs inside a loop that closes on its own. It writes a function and runs the test, and if the test fails it tries again, with no engineer in the room until the code already works. The teams that got real throughput out of those agents did it by changing what surrounded the agent, mainly how the output was reviewed and who stayed accountable for it.
Drug discovery has no loop that closes itself. That difference shapes the architecture of these systems. An agent can synthesize a vast body of published evidence and rank five candidates or 500 by how well the biology coheres, and no part of that ranking tells you which candidates are real. What settles it is an assay or a patient cohort, and the answer comes back in months or years, long after the budget is committed. A system that cannot rerun itself to check its work has to be checked by a person instead, which means building the work so a person can take it apart.
Every claim should carry the evidence under it and the path back to its source. A reviewer needs to open the ranking down to the studies and assumptions that produced it, including the contradictory finding and the reason it was put aside. A fluent answer that hides how it was reached is worth less to that reviewer than an awkward one that shows its work, because the scientist is the verifier the system does not have.
Encoding the Judgment, Not Just the Steps
An organization can take the evaluation logic itself, the ordered decisions its best scientist makes, and write it down as something a system executes. Before a step runs, it fixes what evidence to gather, how that evidence is weighed when sources disagree, and what the output must show before it counts as finished. Documenting the sequence of steps is the straightforward part.
Capturing the reasoning between those steps is much harder: getting a scientist who has made a call by reflex for 20 years to say what the reflex actually checks. That judgment was rarely spoken aloud, because it never had to be, and making it explicit enough for a system to run is unfamiliar even for the people best at making it.
The check for five papers resting on one dataset stops being something one reviewer happens to catch and becomes a step the process runs the same way every time, whoever is at the keyboard. The scientist keeps the decisions that genuinely need judgment, while the evidence gathering and provenance run underneath and outlast the day that scientist takes another job.
One large pharma team that built its indication work this way cut about a month out of every indication it assessed, and held each one to the same standard instead of letting every team rebuild it. The same constraint applies to computational output: a prediction model or a simulation produces a signal, and a signal is not a decision until prior knowledge and mechanistic reasoning have been brought to bear on it and every claim can be traced to its source.
The Limit That Doesn’t Move
None of this touches the biology. A cohort takes as long as it takes to show whether a drug works, and no model shortens that. What gets shorter is the work stacked around the experiment, the reading and the reconciling and the rebuilding that had quietly become the thing programs waited on.
Organizations have always known how to write down what their scientists do. What these systems ask them to capture is harder: how an experienced scientist decides the evidence is finally good enough to act on.
