インサイトに戻る

AI enables faster exploration

Agentic AI platforms offer a solution by orchestrating data, tools, and workflows directly. Engineers and scientists can drive complex processes that previously required multiple experts. This reduces overheads of exploration, shortens iteration cycles, and makes previously impractical investigations feasible. At the same time, a wide range of computational tools can be used to generate insight before committing to often costly physical experiments. Together, this significantly increases the speed and scale at which ideas can be explored in R&D environments. By making these capabilities directly accessible at the point of need, agentic systems reduce the reliance on organizational coordination.

Applicable across R&D domains

This approach is applicable across a wide range of R&D domains where workflows span multiple tools, datasets, and areas of expertise. For example, in pharmaceutical research, agentic systems can support drug discovery by orchestrating literature review, target identification, molecular modeling, and candidate evaluation. This allows researchers to explore therapeutic hypotheses more rapidly before committing to experimental validation.

In cosmetics and formulation science, developing new products often requires balancing multiple parameters such as stability, efficacy, and sensory properties. Agentic AI can accelerate formulation development by exploring combinations of ingredients, predicting performance, and iterating across simulation and experimental data to identify promising candidates more efficiently.

Similarly, in materials research, agentic systems can support challenges such as the discovery of more sustainable chemicals. In Microsoft’s demonstration of PFAS-free coolant discovery, agentic workflows were used to explore large chemical spaces by combining simulation, data, and iterative evaluation. Rather than relying on sequential experimentation, the system was able to propose and assess candidate molecules against multiple criteria. This enables faster identification of viable alternatives while reducing the need for extensive manual exploration of the search space.

Across these domains, the common theme is the ability to combine data, computational tools, and domain knowledge into a unified exploratory process. This reduces coordination overhead and increases the speed and breadth of innovation.

Our collaboration on Microsoft Discovery

With capabilities across a wide range of disciplines including electronics, wireless technologies, signal processing, chemistry, AI and synthetic biology and the facilities to support them, CC was well positioned to collaborate on Microsoft Discovery. We were thrilled to explore how agentic systems can be applied to real-world R&D challenges.

Our previous work in AI-driven protein engineering, for example, has shown how computational methods can significantly accelerate the scientific discovery process. But it required expertise across multiple domains. We used this as the arena to build a demonstration of how agentic AI can reduce coordination overhead and accelerate workflows such as ours.

Our Microsoft Discovery life sciences demo

As a starting point, we built on our previous work in protein optimization, focusing on plastic-degrading enzymes. This provided us with a dataset that could both support exploration and serve as a benchmark to evaluate the system’s performance.

Rather than prescribing a fixed workflow, we defined the overall approach, model evaluation followed by optimization, and allowed Microsoft Discovery to determine the best execution. We equipped it with access to relevant agents, like protein foundation models, some of which we asked it to create, and configured access to compute on Azure.

Given this setup, Microsoft Discovery began by exploring the available tools and resources. It analyzed the dataset, consulted relevant literature via a PubMed agent, and iteratively built and evaluated models using protein representations and basic biophysical features. Through this process, it converged on a strong modeling approach.

This is where it genuinely impressed us. The ability to plan an investigation, break down work, and then orchestrate all the necessary steps of the process on compute available while adapting to errors within hours and validating on our benchmark is something that would take us weeks.

Another highlight was the ability to automatically generate reports that clearly communicate findings, including interactive elements such as 3D visualizations of the optimized enzyme. Together, this demonstrates how agentic AI can turn complex, multi-step R&D workflows into fast, iterative exploration processes.

How R&D challenges evolve

With the ability to explore ideas more rapidly, the limiting factor in R&D now changes. The key challenge is no longer primarily how to execute complex workflows, but how to enable and govern them effectively.

The first constraint is enablement. Regardless of how powerful the platform is, it needs access to tools and data to convert ideas into insights. In our case, that means access to data from previous experiments on plastic degrading enzymes that we have carried out as well as protein foundation models and connecting with our experiment management system.

This requires that datasets are accessible by Microsoft Discovery, which in turn requires building tools and interfaces that can be used by agents. With expanding amounts of information available to the agentic system, good management of it becomes a necessity. With its Science Bookshelf, Microsoft Discovery offers a solution targeted at knowledge spheres with a lot of interdependencies. This is an effective solution where relevant concepts are relational, but knowledge can as well flexibly be pulled in through tool calls (e.g. we used a PubMed agent to search for relevant literature) and stored in classic or hybrid RAG systems.

A second constraint becomes assurance. As exploration is amplified and delegated to agentic systems, trust in the outputs becomes critical. Understanding how results were produced, validating them against known benchmarks, and ensuring reproducibility are essential. Particularly where decisions and large follow-on costs depend on these outputs. These concerns can be addressed by establishing or reusing strong benchmarks, expert oversight and encapsulating workflows into repeatable chunks through agents. Together, these shifts redefine the challenge: from executing individual analyses to building systems where capabilities are both accessible and trustworthy at scale.

From platform to impact

To conclude, it’s clear that realizing the potential of agentic platforms requires more than the platform itself. Success depends on making tools and data accessible in the right way, designing workflows that can be reliably orchestrated, and ensuring that outputs can be trusted in real‑world R&D environments.

This is where many organizations will need support to unlock the true value of agentic AI in R&D. We see a significant opportunity to help teams identify and connect their existing capabilities, structure them for agentic use, and build the assurance mechanisms required to make accelerated exploration both effective and dependable.

It’s a space where Cambridge Consultants and our parent Capgemini’s combined strengths – deep tech innovation, systems engineering and enterprise‑scale delivery – can help organizations move from experimentation to impact with confidence. Do please reach out to me by email if this topic resonates and you’d like to discuss your R&D challenges.