By: Ilya Letnik, Product Manager, Cenevo
While some organizations are already training and deploying specialized AI models for scientific discovery: protein design, molecular simulation, and in silico development, the adoption of LLMs for everyday laboratory work tells a different story.
Here, progress has been fragmented and haphazard: one tool has its own LLM integration for writing or data tasks, another remains a legacy system with no AI layer, and the two are expected to work together seamlessly, which they rarely do.
Labs operating at this level are not yet able to adopt agentic AI, which directly affects lab workflow itself. In agentic AI-enabled labs, AI goes from individual, one-off applications to comprehensive workflow orchestration. Agentic AI “does” scientific work: it acts independently rather than providing predictions.3,4 However, the approach varies by application and vendor. For example, scientists can provide a hypothesis, with suggested approaches, and then the agentic AI carries it out. In other cases, a PDF protocol can be transformed into a step-by-step form with automated error notification – e.g. if a two-digit result is entered as a three-digit result, the system will notify the scientist that the data or the data entry is incorrect. Despite the promise of agentic AI, the operational reality of most pharmaceutical labs remains fragmented.5,6
The Current State of AI in the Lab
Fragmented best describes many of today’s lab operations. Data is within laboratory information management systems (LIMS), electronic lab notebooks (ELNs), and a variety of instrument software platforms. While AI has accelerated specific areas of research: predictive modelling, image analyses, and compound innovation, the volumes of data AI creates has increased scientists’ data-management burden. They must manually transfer data from one system to another, interpret results without integrated context across systems, and manually run the experiments “suggested” by AI.1,2
Before LLMs can be adopted lab-wide, scientists need to implement FAIR (findable, accessible, interoperable, reusable) data infrastructure (if they haven’t already) and review new integration frameworks, such as model context protocol (MCP) and the emerging science context protocol (SCP).5,6
To truly benefit from AI, data management must be robust, orchestration protocols need to be standardized, and automation must be implemented. The combination must create a single source of truth, one data repository that allows AI to truly drive results.5,6
Step One: Managing Data
Digital transformation and digitization within labs add an extra layer of challenge that “traditional” industries don’t: much of lab data is stored within disparate instruments. Resulting measurements and software-based lab operations generate enormous volumes of data. Even fragmented AI itself generates enormous volumes of data. Without a comprehensive approach to data management, that data remains siloed, and results remain elusive.
Therefore, implementing agentic AI is a challenge. It can’t autonomously run a follow-up experiment if it doesn’t have access to all the previous results, which may be in a spreadsheet “somewhere,” or “trapped” within an instrument’s proprietary formatted data.5
Comprehensive data infrastructure must be built and align with the FAIR principles before any lab wide AI implementation is considered.
Many labs are leveraging cloud-based storage to create structured, semi structured, and unstructured databases or data lakes to house their data, making it easy for “data replatforming” converting all raw files into contextualized, AI-native data sets. This gives all data a single address for access by any researcher, agentic AI, or other key stakeholder in the drug discovery process.5,7,8
Different vendors in the space have seen this problem and are solving it. For example, TetraScience has developed scientific data and AI clouds to streamline data management. This ensures that data from multiple sources, with proprietary and non-proprietary formats, is “AI ready.”7,8 Benchling and Cenevo support data lakes, ensuring that all operational metrics, including everything from reagent inventory to instrument utilization and scientific outcomes leverage a common computational foundation.5,7-9
Step Two: Ensuring AI Compliance with Workflow-based Automation Agents
Often, lab scientists must either have become programmers as well or waste time translating their requirements to external support services who don’t always understand labs’ specific requirements.
That’s the challenges researchers have faced as they are trying to automate routine workflows, like design of experiments (DoE) or Design-Make-Test-Analyze (DMTA) cycles. Scientists either needed to learn how to code, find low-code workflow rule builders they could master quickly, or translate all their requirements into language that specialized automation could understand.
Now, specialized AI agents are “stepping up” for the job. While that’s all well and good, the AI agent may not easily “reveal” its secrets, acting autonomously in a black box and, therefore, creating a significant compliance risk.10,11
The vendors who understand this risk are developing AI-assisted, rule-based workflow builders. Instead of giving an LLM free run of the lab with complete autonomous control, workflow builders use natural language processing to translate what the scientist wants the agentic AI to do into easily auditable rules.10,11 Although this method allows for better clarity and transparency, this doesn’t solve another critical issue with AI: the possibility of it hallucinating or fabricating information. Effective guardrails and real-time monitoring mechanisms to catch such errors are still a challenge.11,14,15
Step Three: Your New Lab Partner: Embedded AI within Lab Platforms
Beyond regulatory compliance, integrating AI directly into lab operations using external chatbots creates significant data security risks. If a researcher forgets to turn off “chat history and training,” any data may become publicly accessible, easily discoverable by anyone querying the chatbot, including competitors.16,17
To counter this challenge, AI assistants are being developed that “live” within the lab’s data systems. Of course, these assistants are only as good as the data included within the system. If the data is comprehensive and complete and generated using orchestrated hardware and software in the lab, then the embedded AI assistants can provide true value.16,17
Sapio Sciences is now offering the Sapio ELaiN (electronic laboratory artificially intelligent notebook), which has been trained on the platform’s LIMS and ELN. ELaiN makes it simple for chemists to use conversational prompts to synthesize compounds or design complex plate layouts without exposing any of their source data to external parties.17,19
Benchling’s Benchling AI uses native agents that can verify notebook entries, generate SQL queries, and execute R or Python scripts directly alongside the experimental data.9,18
Cenevo is already offering three separate AI agents as part of its agentic lab platform, including the Labguru Assistant, which is integrated directly into the Labguru platform and allows scientists to use simple natural language to troubleshoot experiments, optimize protocols, compare results across studies, and analyze complex data.6,12,13 The AI Protocol Conversion agent automatically transforms legacy, “paper-based,” static documents and protocols into fully structured, reusable, and compliant protocols.12,15 Meanwhile, the AI Automation agent allows scientists to describe what they want to automate, without needing programming expertise.12-14
These models are grounded within the proprietary data layer of each platform. That means their outputs are not only context-aware, but they are also traceable back to specific, validated laboratory records, which is mandatory for patent documentation and QA/QC.
Step Four: Scale Connectivity Using the Model Context Protocol (MCP)
Integration and orchestration remain a consistent challenge, with or without AI in the mix. With AI in the mix, it gets even more complex. AI agents need to be deployed across different R&D and manufacturing environments and departments within the pharmaceutical company. Implementing point-to-point API integrations across software and hardware platforms is very time consuming and very expensive. Updates require time and money as well. When equipment must be replaced, the integrations start all over again.
To reduce the necessity for point-to-point integrations (some may be necessary due to the age of the equipment or other internal factors), the model context protocol (MCP) has been developed as a universal connector – a “USB port” – to connect AI with workflows and data. The MCP provides a standardized path to allow AI assistants to securely discover and access external software environments, databases, and tools.20-22
MCP connectivity enables cross-vendor communications, smoothing the path to lab-wide connectivity. If a high-performance liquid chromatography (HPLC) instrument has a built-in MCP server and it’s connected to an ELN that also contains an MCP server, the agentic AI can query both simultaneously, as long as the human researcher has authorized access to both.21-23
After connectivity comes orchestration.
Step Five: Leverage AI to Orchestrate Lab Processes
Once everything is connected, orchestration is the next logical step. Drug discovery requires many players across multiple laboratories, all trying to work together to develop the next life-saving pharmaceutical. The larger the pharma, though, the more labs need to coordinate efforts and activities. Researchers in computational dry labs are designing molecules in-silico; they are then synthesized within wet labs using closed-loop platforms. The results are then distributed to external CROs so they can conduct characterization.9,24-26
To address this issue within the agentic AI environment, the science context protocol (SCP) has recently been introduced. It serves as the domain-specific orchestration layer above the MCP, acting as “essential infrastructure for scalable, multi-institution, agent-driven science”.25,27
SCP is a centralized lab hub that manages the entire scientific lifecycle. Its capabilities include:25,27
Persistent state – Maintains the experiment’s context across multiple days, tracking the individual software platforms and the specialized AI agents.26,27
Multi-agent coordination – The SCP orchestrates complex tasks throughout the entire path: the computational agent designs a compound. The parameters are passed to the synthesis agent in control of the lab’s robotics, and then it routes the data to an analytical agent so it can be reviewed.26,27
Wet and dry lab integration – The MCP directly addresses the challenges of managing the virtual and physical lab, all while maintaining strict provenance and chain of custody. It connects digital simulations to physical laboratory execution via standardized lab device drivers.25
While SCP may be overkill in a single-lab environment – the MCPs should be sufficient – it is absolutely necessary when managing multi-system experiments. Using SCP ensures that these experiments remain auditable, compliant, and reproducible.23,26,28
Moving Forward with Your Agentic AI Lab
AI is already integral to the entire R&D process. While embedded AI is a given across most software and hardware platforms, agentic AI, natural-language workflow platforms, embedded AI assistants, and other AI applications are going to take drug discovery to levels that used to take decades to achieve.
Platforms like Benchling, Biosero, Cenevo, Sapio, and Synthace are just a few vendors ensuring that their clients can start to fully leverage the promise of agentic AI.9,11,13,17,32 Many more have their agentic AI in progress.
However, unless the most critical part of infrastructure is addressed – eliminating non-FAIR compliant data silos across individual labs and the organization itself – AI adoption will be catch-as-catch-can. It will remain superficial and actually cause problems because it will only be able to analyze individual data sets, delivering incomplete results.5,7,9-11,13,19
To prepare for comprehensive agentic AI in the labs, R&D, compliance, QA/QC leaders need to take the following steps:
Manage data – Create a single source of truth via a data lake, on a scientific data cloud or on-prem, that is easily accessible, secure, in compliance with FAIR, and containing data from all computational tools, software, hardware, AI platforms, instrumentation, etc.
Ensure AI compliance with workflow-based automation agents – Streamline operations and automation by using workflow-focused AI tools that “understand” the scientist, all while ensuring human-in-the-loop overview and full regulatory compliance.
Leverage embedded AI within lab platforms – Maximize the value of embedded AI within the lab while ensuring full traceability and validation.
Scale connectivity using the model context protocol (MCP) – Choose vendors that support interoperability standards to ensure your modular lab tech stack is AI ready.
Leverage AI to orchestrate lab processes – Ensure that SCP or something similar is in place to coordinate AI activities across the organization.
All these steps need to be taken with an eye toward regulatory compliance, traceability, FAIR principles, data integrity, and human in the loop. If they are done properly, they can allow scientists to focus on the more creative aspects of their role, streamline operations, and reduce costs, all while significantly accelerating drug discovery.
References
- https://pmc.ncbi.nlm.nih.gov/articles/PMC12426084/
- https://pure.johnshopkins.edu/en/publications/ai-agentic-models-and-lab-automation-for-scientific-discovery-the/
- https://arxiv.org/abs/2503.08979
- https://www.mckinsey.de/industries/life-sciences/our-insights/reimagining-life-science-enterprises-with-agentic-ai
- https://www.cenevo.com/blog/digital-lab-transformation-fair-data-cenevo
- https://www.cenevo.com
- https://eliteai.tools/tool/tetrascience
- https://www.tetrascience.com/platform/tetra-scientific-data-ai-cloud
- https://www.benchling.com/news/benchling-reveals-the-next-chapter-of-rd-at-benchtalk-2025
- https://www.businesswire.com/news/home/20230524005075/en/Synthace-combines-ChatGPT-and-digital-experiments-in-first-step-toward-a-true-AI-scientist
- https://biosero.com/news-and-media/biosero-advances-laboratory-automation-with-the-introduction-of-assistive-ai-solution/
- https://www.rdworldonline.com/cenevo-launches-two-ai-agents-for-lab-protocol-conversion-and-workflow-automation/
- https://clpmag.com/lab-essentials/lab-automation/cenevo-ai-agents-automate-lab-workflows/
- https://help.labguru.com/en/articles/13225365-introducing-the-automation-agent
- https://www.selectscience.net/article/cenevo-advances-toward-agentic-labs-with-two-new-ai-agents
- https://help.benchling.com/hc/en-us/articles/42092744616461-Release-Notes-Volume-11-2025
- https://www.sapiosciences.com/blog/sapio-elain-your-science-aware-ai-lab-assistant/
- https://www.benchling.com/webinars/whats-new-webinar-august-2025
- https://clpmag.com/lab-management/sapio-sciences-enhances-ai-powered-lab-assistant/
- https://mcptoolbox.app
- https://cloud.google.com/discover/what-is-model-context-protocol
- https://modelcontextprotocol.io/docs/learn/server-concepts
- https://intuitionlabs.ai/pdfs/model-context-protocol-mcp-in-pharma.pdf
- https://biosero.com/solutions/speed-up-dmta-cycle/
- https://the-decoder.com/science-context-protocol-aims-to-let-ai-agents-collaborate-across-labs-and-institutions-worldwide/
- https://www.linkedin.com/pulse/028-science-context-protocol-scp-can-next-mcp-drug-rd-nagesh-nama-bh3ze
- https://arxiv.org/html/2512.24189v1
- https://www.linkedin.com/posts/stawil_should-your-lab-adopt-science-context-protocol-activity-7417592935925518336-zjtG
- https://www.benchling.com/news/introducing-benchling-ai
- https://www.prnewswire.com/news-releases/benchling-reveals-the-next-chapter-of-rd-at-benchtalk-2025-302583467.html
- https://www.youtube.com/watch?v=chmARjwttPk
- https://www.synthace.com
- https://www.bioanalysis-zone.com/synthace-collaborates-with-charles-river-to-offer-the-global-pharma-industry-more-robust-faster-and-cost-effective-assays/
- https://www.youtube.com/c/Synthace/videos
- https://www.drugtargetreview.com/news/193003/charles-river-and-synthace-announce-partnership-for-assay-development/
- https://www.synthace.com/blog/tag/lab-automation
About the Author
Ilya Letnik is a Product Manager at Cenevo, which specializes in lab management systems, automation, orchestration, data management and AI technology for life sciences. He holds a PhD in biotechnology from Hebrew University, where he researched biotechnological applications of encapsulated microbial cells.