Artificial intelligence is beginning to reshape how scientific organizations approach research, decision-making, and operational strategy, particularly in environments where complex datasets, specialized instrumentation, and multidisciplinary teams intersect. For many laboratories, the question is no longer whether AI has value, but how it can be deployed responsibly to accelerate discovery, enhance productivity, and support more efficient technical workflows.
In this Q&A, Michael Connell, chief operating officer at Enthought, discusses how purpose-built AI is emerging in scientific settings, what lab leaders should understand about its capabilities and limitations, and where he sees the greatest opportunities for meaningful impact across research and industry.
Q: Can you define agentic AI and explain what value it offers to labs today?
A: AI, generally speaking, is the capability of a computer system to perform tasks that typically require human intelligence, such as hypothesis generation, reasoning, performing complex data analysis, and synthesizing experimental data into a summary of findings. There are different kinds of AI systems. Some are agentic, and some are not. While a non-agentic system might execute a specific task when prompted, such as drafting an email or answering a question, an agentic AI system is distinguished by the fact that it can operate autonomously and make decisions in pursuit of a goal without constant human intervention. It can autonomously create a plan, execute the necessary steps to carry out the plan, and adapt its strategy based on real-time information and feedback.
This capability is particularly valuable in scientific R&D, which is uniquely path-dependent and highly variable, making it difficult to manage with rigid, rule-based “classical” AI. Unlike these earlier systems, large language model (LLM)–based agents can handle novel inputs adaptively—for example, interpreting anomalies in the data and adjusting experimental conditions on the fly. In a fully automated laboratory, an agentic AI system could conduct a series of experiments, learn from each one, and adapt the design of subsequent experiments as it searches for a formulation that meets specific target criteria. This flexible, closed-loop behavior enables a level of automation previously seen only in highly predictable and standardized environments such as factory manufacturing.
Q: What is the relationship between Design of Experiments and AI?
A: The relationship between Design of Experiments (DoE) and AI in R&D is a synergistic feedback loop. Together they transform R&D into a dynamic, closed-loop cycle often called "active learning."
The core shift is that traditional AI/DoE optimizes based on numbers, while LLMs optimize based on meaning. This changes a number of things:
- Parameter definition: While Bayesian Optimization already enables dynamic experimentation, LLMs automate the setup. Instead of humans manually guessing constraints, LLMs can mine thousands of academic papers to extract and construct valid parameter ranges (e.g., temperature) automatically.
- Hypothesis generation: Traditional AI optimizes existing parameters, whereas LLMs can use reasoning to suggest new parameters to try. For example, in the old way, the goal might be to optimize the amount of catalyst A being used; in the new way, the LLM might observe that the literature suggests Catalyst A degrades in the presence of moisture and suggest switching to Catalyst B or adding a drying agent.
- Use of natural language: Before LLMs, connecting a DoE algorithm to a lab robot required writing complex, rigid custom code. LLMs can write the necessary control code on the fly, adapting to what is learned in previous iterations, allowing the experimental design loop to execute itself without a human having to code the interface.
Q: For lab managers interested in fine-tuning LLMs on data in their lab’s domain, where should they start with collecting that data and finding an expert who can lead that process?
A: Start with a clear understanding of your lab’s knowledge landscape, not just its data assets. The most valuable sources are often unstructured, including protocols, analysis reports, lab notebooks, and decision logs that capture human reasoning. Curating and structuring this information creates a foundation for fine-tuning models that truly reflect your scientific language and workflows. For most traditional R&D labs, a key mindset and operational shift is required. In particular, data collection, storage, and access infrastructure should be designed with secondary analysis in mind, because by and large, data collected for one purpose (e.g., to complete a specific development project) will not automatically be useful for different applications (such as fine-tuning LLMs). You’ll then need experts who understand both the science and the data infrastructure, who can bridge lab knowledge with machine learning workflows.
Lab Quality Management Certificate
The Lab Quality Management certificate is more than training—it’s a professional advantage.
Gain critical skills and IACET-approved CEUs that make a measurable difference.
Q: What are the essential components of a successful business case for agentic AI?
A: A compelling business case for agentic AI recognizes that the investment is not just in a tool or technology rollout, but in a strategic capability that compounds over time. While near-term ROI may be difficult to quantify, the true value lies in significantly accelerating product development cycles and de-risking R&D decisions. A strong proposal must align with corporate priorities such as speed to market, sustainability, or scientific competitiveness. It should define a phased approach that begins with "quick wins" to engage stakeholders and provide proof points, while building foundational infrastructure for data readiness and workflow redesign, followed by measurable impact through reduced experimentation, more reliable predictions, and faster discovery. This progression helps executives see how early investments evolve into lasting competitive advantage and a more innovative, resilient R&D organization. In a landscape where competitors are already using AI to compress discovery timelines and expand their footprint, the cost of waiting can quickly outweigh the cost of adoption.
Q: What types of data would be most valuable to feed into AI to speed materials R&D?
A: To accelerate R&D, AI needs comprehensive information, not just empirical data. There are multiple sources of knowledge and intelligence that are invaluable for materials R&D specifically. First is experimental and synthetic data, or combining high-fidelity lab measurements (e.g., XRD) with high-volume simulation data (e.g., DFT) to fill coverage gaps. You also should include physics-based constraints that reflect fundamental laws (e.g., thermodynamics) to prevent physically impossible predictions and drastically reduce data needs. Unstructured knowledge is also essential, including text from patents and papers to capture synthesis recipes ("how-to"), often missing from structured datasets. You also need to integrate constraints like cost, toxicity, and supply chain to ensure solutions are commercially viable, not just theoretically possible.
Q: Do you think that AI can predict physical properties of novel materials?
A: Absolutely, and now with increased precision. AI models can now learn property–structure relationships using next-gen hybrid approaches that blend data-driven methods with physics-based constraints. The next frontier, however, is more trustworthy prediction: optimizing the quantified uncertainty in those outputs so that scientists know where the AI/ML model is solid and where it’s extrapolating beyond its domain. This is putting Materials by Design (MbD)—a methodology that inverts the traditional R&D process by starting with the desired end-use properties and applying computational tools to design the precise structure needed to achieve them—within reach for more companies. MbD creates a massive acceleration of the innovation pipeline and the ability to create entirely novel, proprietary materials that are custom-engineered for property, performance, and cost targets.
Q: What does the future hold for AI in materials R&D?
A: AI is beginning to redefine the scientific method itself—from a linear, forward process of hypothesis, experiment, and analysis into a continuous, intelligent inverse loop. Experiments, simulations, and models will operate together within adaptive systems that learn, reason, and improve with every iteration. For researchers, AI will serve as both collaborator and catalyst, expanding the boundaries of what can be explored while accelerating the path from idea to validated material. As this transformation unfolds, entirely new roles will emerge in R&D: AI workflow designers who translate human reasoning into automated processes, and heads of AI operations who oversee intelligent systems across the organization. For enterprise R&D organizations, success will depend on how quickly and strategically they build AI fluency and infrastructure to be ready for the next era of innovation.
