A lab automation agent made a small dosing error. Nobody caught it for three weeks.
An internal review at a mid-size drugmaker shows how a rounding mistake in an automated protocol spread through 40 experiments before a technician noticed.
By Lena FischerPublished · 1 min read
The mistake was a single digit. An AI agent responsible for translating written protocols into instructions for liquid-handling robots rounded a dilution factor from 2.45 to 2.5. Every plate prepared afterward carried a compound concentration about 2% higher than intended.
For most assays that would be noise. For the potency study Kestrel Ridge was running to support a regulatory filing, it was enough to push several results outside their acceptance range.
According to a review the company shared with Vitality Journal, the agent then did what it was built to do: it flagged the out-of-range results, proposed an explanation and reran the experiments using the same faulty instruction. It repeated that loop for 22 days.
The system was very good at noticing that something was wrong and very bad at suspecting itself.
A technician finally traced the problem by preparing one plate by hand. Kestrel Ridge said no data from the affected runs was submitted to regulators, and that the delay to its filing will be "weeks, not months."
The company has since added a rule that any protocol change generated by software must be countersigned by a person before it reaches the robots. It is a modest fix, and one that several larger drugmakers told us they already had in place.
TrackerPatent cliff
Patents protect pharma profits. Track when the biggest ones expire.
Updated Sept. 10, 2026Pharma
