From Hype to Lab Reality: Defining AI Co-Scientists
AI co-scientists are software agents built on large language models that work alongside researchers to propose hypotheses, automate experimental workflows, interpret data, and connect digital designs to wet lab execution. Despite a rush of new life science AI products, the gap between this promise and daily bench work remains wide. According to the Pistoia Alliance, 54% of life science teams see AI delivering value in regulatory submissions and reporting, but only 1% report value in the wet lab. That divide explains why wet lab AI integration is now the central challenge in life science AI. Vendors are experimenting with different architectures to turn lab automation AI from a clever chat interface into a reliable co-worker whose suggestions can be traced through instruments, protocols, and results, rather than staying trapped in documents and slide decks.
Sapio and Potato: Agents, MCP, and the Data Plumbing Problem
One cluster of vendors is betting that AI co-scientists will earn trust by sitting directly on top of lab data systems and acting as cross-application agents. Sapio Sciences rebuilt its electronic lab notebook around an Anthropic-powered assistant accessed through the Model Context Protocol, turning what began as a chat box into an agent that can query records, pull files, and draft reports under a single instruction. Rob Brown says he no longer builds queries by hand because natural language and even voice now drive his workflows. Potato and similar players focus on the same problem from the infrastructure side: stitching together fragmented life science AI tools and proprietary data so models are grounded in consistent, well-governed information. Here, the architectural bet is that scientists will trust wet lab AI integration when the assistant can explain every recommendation in terms of specific samples, experiments, and source systems.
Google, OpenAI, and the Computational-Only Co-Scientist
Another group, including Google DeepMind and OpenAI, is building co-scientists that live mostly in computational space. Google’s Co-Scientist combines six specialized agents that debate and rate hypotheses in an Elo-style tournament, designed to sharpen reasoning on complex problems. OpenAI’s GPT-Rosalind is framed as a reasoning model that plugs into existing computational biology pipelines at companies like Amgen, Moderna, and Thermo Fisher. These tools sit firmly on the life science AI side, not lab automation AI, and they excel at document analysis, simulation planning, and data mining. However, they rarely control instruments or trigger experiments directly. As Christian Baber of the Pistoia Alliance notes, human review still separates transformer outputs from any external system that matters. Until these models connect to reproducible wet lab workflows, scientists may see them as powerful advisors rather than true AI co-scientists trusted to influence physical experiments.
Benchling’s Lab-First Strategy: Grounding AI in Physical Experiments
Benchling argues that an AI scientist only earns the title when it can run work in the physical lab. Co-founder Ashu Singhal emphasizes that the biggest bottlenecks are not ideas but the gruntwork of ordering reagents, setting up notebook entries, running assays, and cycling the resulting data back into design. Benchling’s response is an architecture that couples life science AI with lab automation AI and vendor ecosystems. Its one-click ordering with partners such as Twist Bioscience, Adaptyv, and Ginkgo Bioworks, along with a Model Hub and automation tools, is meant to shrink the distance between a prompt and a plate on a robot. Singhal divides experimentation into repetitive tasks worth full automation, work better sent to contract research organizations, and one-off assays that keep humans at the bench. In this view, AI co-scientists succeed only when they can orchestrate all three reliably.
Trust, Traceability, and the Path to Reproducible AI Workflows
Trust in AI co-scientists will not come from clever conversation alone; it depends on traceability from model suggestion to lab outcome. Today, many scientists are dissatisfied with traditional electronic lab notebooks, and a Sapio survey found that 45% of bench scientists use public generative AI tools through personal accounts to fill feature gaps. That shadow usage underlines demand but also highlights risk when outputs cannot be linked to controlled data or validated workflows. Vendors such as Ginkgo Bioworks, Parallel Bio, and Perceptic are building cloud labs, autonomous workcells, and connective software that tie prompts to priced protocols, instrument runs, and structured datasets. The emerging consensus is that wet lab AI integration must produce experiments that can be repeated and audited, not just drafts and suggestions. When an AI-generated protocol can be executed, tracked, and reproduced end to end, lab teams may finally treat these systems as dependable co-scientists.
