AI co-scientists meet the wet lab problem
AI co-scientists in life sciences are software agents that assist researchers by generating hypotheses, designing experiments, and interpreting results, but their impact stays limited until they connect to the physical tools, data structures, and constraints of real wet lab environments where experiments are planned, executed, and recorded. Today, that gap is stark. According to a Pistoia Alliance survey, only 1% of life science professionals say AI delivers measurable value in the wet lab, even though more than half see gains in regulatory and reporting work. The problem is not a shortage of models: Google’s Co-Scientist, OpenAI’s GPT-Rosalind, and various startup offerings all promise “AI scientists.” What they lack is grounding in day-to-day bench workflows—ordering reagents, scheduling assays, capturing instrument data, and looping results back into design decisions.
Why life science AI adoption lags behind code-heavy fields
In software engineering, AI agents already plug into code repositories, test suites, and deployment pipelines. In life science labs, workflows remain far more fragmented and physical. Bench scientists still shift between notebooks, inventory systems, instruments, and external CROs, often with manual handoffs at every step. Benchling’s Ashu Singhal points out that the hard part is not forming hypotheses but handling the grunt work of ordering materials, setting up methods, and processing results into usable data. Without a clear path from AI-generated plan to executable protocol on real equipment, wet lab AI integration stalls. Many experiments are repetitive and well defined, yet only 13% of surveyed teams report value from automating scientific workflows and experiments. AI co-scientists stay confined to the design phase, while human scientists remain stuck bridging the digital and physical worlds by hand.
Benchling’s bet: connect AI to instruments and CROs
Benchling’s AI lab automation strategy starts from the assumption that an AI scientist means little unless it can trigger and track real experiments. The company has been wiring together the pieces between design and execution: one-click ordering with partners like Twist Bioscience, Adaptyv, and Ginkgo Bioworks; a Model Hub that lets researchers run structure prediction and generative models within existing workflows; and AI Connectors to link Benchling records to external AI tools. Benchling Automation, its latest release, connects workcells and bench instruments so that when a scientist or AI agent designs an experiment, the system sends the run to the relevant hardware and routes data back into the correct notebook entries. Singhal argues that high-quality, structured scientific records—“not just capturing paper-on-glass data”—are the fuel that lets AI co-scientists act on real samples, not only abstract datasets.
Grounded AI co-scientists: from design to execution loops
Once AI agents are grounded in lab automation, they can participate in closed loops that look more like coding workflows. Benchling’s AI Scientist, used by Prime Medicine on its PM359 program, pulled together years of experimental data, mapped evidence to FDA validation requirements, and proposed targeted follow-up studies, all within Benchling’s structured environment. The loop is gated: the AI drafts methods and notebook entries, pauses while a human runs the assay in the wet lab, then resumes once instrument data flows back through Benchling Automation. Singhal compares it to a coding agent waiting for tests to finish. Merck’s bioanalytical group shows what mature AI lab automation can unlock, reporting a tenfold efficiency gain and a 25% reduction in rote tasks after rebuilding vaccine trial bioanalytics around automated Benchling workflows.
The future: human judgment, automated hands
A growing field of players is pushing toward more automated science, from Lila Sciences’ “AI Science Factories” to Medra’s robotic systems and Amazon’s Bio Discovery effort. Benchling positions itself as the connective tissue rather than the robot builder: a data model rich enough for AI co-scientists, links to CROs and instruments, and interfaces that working scientists will actually use. Singhal divides lab work into repetitive tasks ideal for workcells, outsourced projects suited to CROs, and one-off experiments that keep human scientists at the bench. That last third is why he does not expect labs to empty out. Instead, AI co-scientists grounded through AI lab automation will handle the structured, automatable parts of experimentation, while humans focus on judgment, edge cases, and novel assays. The promise of wet lab AI integration is less about replacing scientists and more about giving them automated hands that can keep pace with digital minds.