Put AI on the searching and keep the scientist on the anomaly

An AI assistant for wet lab scientists should take over the searching, the drafting and the number work, and leave the odd result that breaks the pattern to the scientist, because that is where the field's biggest discoveries came from. Whatever work a lab hands to software, its people stop practising, so choosing what to automate is also choosing which skills the lab keeps.

What a central tool does to the habits around it

Give a tool a task people used to do by hand and the organisation changes shape around the tool, and that is the risk worth planning for. A driver who has followed the satnav on every journey for a year has stopped learning the roads it picks, and on the day it sends them into a road closure the alternative has to come from somewhere, because the practice that would have built it stopped a year ago. A phone that corrects each typed word keeps spelling accurate on the screen, and plenty of people who write readable messages all day hesitate over a word on a handwritten list, because the correction moved the practice off the person and onto the device. In one sentence, the tool takes the work, and the capacity the work used to build goes with it, so deciding what to automate is also deciding what a team stops practising.

The case behind the question

One of the world's largest pharmaceutical companies cut 489 positions in 2025, with a further round in 2026, whole research teams among them and senior research leadership reduced, while the same company described AI as increasingly central to its research strategy and expanded its digital and automation capability. The company said the changes aligned investment with its priority therapeutic areas, touched a small part of the organisation, and left it still hiring, including for research. Both accounts hold. So a major research organisation reduced its experimental science while it invested in AI, and it leaves open whether one moved because of the other. That leaves the question every lab adopting AI now faces: what do its people stop practising once software starts doing the work?

Why the big discoveries came from odd results

The field's biggest discoveries began as results nobody was looking for. An attempt to deepen the colour of petunia flowers by adding an extra copy of a pigment gene produced white flowers, because the added gene silenced the pigment genes the plant already carried, and that unexplained result, followed for years, became RNA interference, a Nobel Prize in 2006 and approved medicines a decade later. The repeating stretches of DNA in bacterial genomes, noticed in a sequencing paper in the late 1980s and unexplained for most of two decades, became the guide-RNA system that genome editing runs on. Checkpoint inhibitors, reprogrammed stem cells and the ubiquitin system that tags damaged proteins for destruction share the same shape: a scientist, with bench expertise and the freedom to keep going, worked on a result the science of the day had no explanation for.

AI's achievements are real: predicting a protein's structure from its sequence, classifying medical images at specialist level, prioritising experiments. Structure prediction alone changed the questions a working lab can ask, because a structure that used to demand years of crystallography now arrives in minutes, and predictions exist for a far larger set of proteins than the experimental record covers. Those are wins at prediction, drawn from what the field already knows. Prediction starts from the record as it stands. The discoveries above started where the record stopped matching the bench, and the move towards that mismatch was made by a person.

Where the agent belongs in a working lab

A scientist reads a plate readout and one condition came back inverted against what the design expected, the kind of result that used to cost an afternoon of searching: has this assay behaved like this before, does the lab's own record hold anything, has anyone published the pattern. Asked in full, Elnora searches the lab's knowledge base and the published literature together, and what comes back is the matching passages from the lab's own files, each with the file it came out of named beside it, and papers from PubMed and the open literature, each with its identifier. A skill for generating hypotheses turns what it found into candidate explanations ordered by the evidence behind each, and the scientist reads the ranked list and picks the one worth chasing. Elnora writes the protocol that tests it, the run waits until the scientist has reviewed the protocol and approved it, and the readout goes back into the knowledge base alongside the protocol that produced it, failed runs included.

This division of work protects what a lab stands to lose. Institutional memory, the protocols, results and failed runs a lab has already paid for, sits in a knowledge base Elnora reads before it writes, so what one scientist learned is findable by the next one on the day it matters. Experimental expertise stays practised, because the bench work and the review of every plan remain the scientist's, and what Elnora produced can be checked step by step. The freedom to follow the odd result survives too, because the searching that used to cost the afternoon now costs minutes, and the afternoon goes back to the bench.

The agent belongs on the searching, the drafting and the number work, and the scientist belongs on the result that breaks the pattern. A lab that splits the work that way, in the software it adopts as much as in the strategy its leadership writes, gets the speed and keeps the discovery.