Transforming Lab Compound and Sample Management From Automation to AI Reasoning

Across the industry, the role of compound and sample management is expanding. Teams need to manage much more than small molecules; they must manage across modalities, workflows, instruments, solvents, and reagents. Agentic AI will help this significantly.

Most labs have already embraced AI in some form, but usually that AI is built directly into the latest version of “traditional” lab software or is being used for in silico analysis. The next logical step in AI evolution—agentic AI—operates at the intersection of lab hardware and software. In a broad sense, agentic AI drives the transition from in silico analysis, which generates multiple compound suggestions, by reducing that list to a smaller volume of highly promising compounds ready for real-world experimentation.

Generally, lab technologies are highly specialized. However, agentic AI is flexible across industries. For example, an agent trained on manufacturing principles, such as the theory of constraints, when applied to a lab setting, can also analyze plumbing issues in data flows in ways you wouldn't normally bring in cross-domain expertise for.

Moving beyond scripts, towards reasoning

AI-based compound and sample management goes beyond inventory management and registration. It’s at the crossroads of automation and integration, where AI tracks the physical and virtual environments, moving data, accelerating analysis, improving the traditional Design-Make-Test-Analyze (DMTA) cycle, and reducing lab bottlenecks.

Existing lab automation processes follow specific scripts; if they fail, the entire process stops—the “if, then” logic fails. Agentic AI moves lab automation to the next level; it analyzes the failure points and takes steps to mitigate the issues.

With the DMTA cycle built into its reasoning, agentic AI can:

• Design: Select compounds and doses and propose experiments
• Make: Generate protocols and plate maps, and book equipment slots
• Test: Monitor runs, capture raw readings, and note anomalies
• Analyze: Fit curves, score hits, and suggest improvements for the next cycle

For compound and sample management, AI needs to be trained on typical sample usage patterns across automated systems, so it can determine the best way to run sample preparation and and experimental workflows.

Improving AI-to-AI communications and data

While Application Programming Interface (APIs) and command line interfaces (CLIs) allow direct access to hardware and equipment, data-to-data connectivity is still evolving. New pathways include model context protocol (MCP) and agent-to-agent (A2A) communications.

MCP and A2A speed and streamline data exchange, smoothing data connectivity. MCP allows AI agents to “talk” to existing lab software platforms, such as electronic lab notebooks (ELNs) and laboratory information management systems (LIMS). A2A also allows for cross-company communication, such as between a pharma and a contract research organization (CRO).

Automated data transfer among equipment, software, and AI can speed-up analysis timelines, accelerating drug discovery and other life science lab processes.

However, before AI is “allowed” to start feeding databases directly, the databases themselves need to be standardized. Databases must have the ability to manage structured, unstructured, and semi-structured data under consistent schemas, in line with the FAIR principles (findable, accessible, interoperable, and reusable) for scientific data management. In addition, access management is just as critical for AI as it is for human team members. AI can be permitted read/write access to only relevant database entries.

Agentic AI in pharma, CRO, and CDMO applications

Pharma R&D benefits from agents that can prepare samples, manage documentation, and suggest and prepare experiments. The same applies to CRO and Contract Development and Manufacturing Organization (CDMO) operations, but with an added layer of reporting and guidelines.

Agentic AI can greatly speed up the reporting process for CROs and CDMOs; agents automatically pull relevant data, eliminating the need for researchers to “find” and compile it. Instead, researchers become the “human-in-the-loop” needed to review and approve the information before it is sent on to the pharma industry client.

Compliance, regulation, and governance

In an agentic AI-enabled lab, agents write the experiments while scientists provide a comprehensive review of all AI decision-making and guide the overall goals and strategies of the lab. Agents can reserve the equipment, select the compounds or samples, decide doses, and then perform the experiments. Then, open-source large language learning (LLM) engineering platforms—such as Langfuse, which can monitor, debug, and evaluate agentic AI activities—score the outputs to ensure reliability. Lab scientists can then review those results with a critical eye.

But when it comes to compliance and documentation, how do you prove that the AI suggestions were right? While most existing metadata systems record who did what, where, when, and on what equipment, this is naturally not possible when such tasks have been assigned to machines. How do you ensure compliance and governance in these cases? The audit trails’ metadata sections must now include who approved the agentic AIs’ activities.

Humans must also look out for AI usage with respect to the “laws of unintended consequences”—to borrow a term from economics and social sciences. AI may not understand plate size and might split 50 experiments across plates instead of consolidating, generate a plate that uses too much compound or the entire sample, repeat past experiments, leave stores empty of rare ingredients, or “hog” the equipment.

Unless taught in advance, AI doesn’t understand the difference between common, inexpensive compounds and rare, expensive ones. When developing new compounds, AI must “understand” reasonable restraints, such as dose size. Yes, a two-pound dose of a compound might wipe out a bacterium instantly, but what patient would be able to comply with that prescription? The ideal situation is to wrap agents with input and output validations—gates that verify the agent's proposed output (e.g., a dose within a defined ceiling)—before it proceeds.

Advancing AAV Characterization Through Mass Photometry

This infographic shows how single-particle mass photometry delivers AAV purity analytics in less than five minutes while supporting development and GMP manufacturing workflows.
View Infographic

Advertisement

Confidential projects pose another area of concern. When a lab is working on very specific research that should not be shared with other parts of the organization, AI must be taught not to “learn” from that specific data and include it in broader contexts.

Relevant standards and the need for human oversight

The International Organization for Standardization (ISO) has already released relevant AI standards, including:

• ISO/TC 276 (biotechnology): standardization relating to large biological datasets used to train AI, with an emphasis on integration and validation.
• ISO/IEC 5259 (data quality for all): a framework for assessing and enhancing data quality for reliable analytics and machine learning (ML).

While FAIR data standards are usually integrated within generative AI and agentic AI for labs, formal regulations, such as those from the US Food and Drug Administration (FDA), still lag behind the state of technology.

Implementation and the future of lab operations

From day one of agentic AI, humans are in the loop or in the center, depending on on your perspective. At the beginning, the sample management team is responsible for implementing AI across its processes, handling the minutiae of training the agents, and setting appropriate constraints. Upstream and downstream researchers will coordinate their efforts, working across the entire research organization to ensure all AI activities are in sync.

Once everything is ready to go, lab workers will be able to spend less time pipetting and more time guiding and planning research, while the agentic AI—via robots and automated stores—connects the physical lab and software environments. The more efficient AI becomes at compound and sample management, the more requests the team will receive for its services.

In an effective lab that is utilizing agentic AI, the people overseeing processes will be able to operate at much higher levels. Autonomous mobile robots will be pulling samples and inventory, running analyses 24/7. Each scientist will simultaneously be a lab manager and researcher, focusing on process design and making critical decisions about the research results—not how it’s done “on the ground.”

Of course, agentic AI is still in its early stages. Not everyone is sure how it will all play out. It will play out, though.