Many organizations see AI and machine learning as the next major technological development – something that could unlock greater insights and accelerate drug discovery and development. But those are very broad ambitions. AI often seems to be the answer, but organizations do not necessarily know what the question is. They want to use it, but they do not know how or where to start.
Organizations may have some initial ideas, but the foundational data are often not ready. Almost all organizations have some form of digitalization in place: instrument data are stored in databases, and systems such as electronic laboratory notebooks (ELNs), laboratory information management systems (LIMS), and chromatography data systems (CDS) are widely used. However, just because information is digital does not mean that it is ready for AI. In many organizations, the data still exist in isolated systems: an ELN from one vendor, a LIMS from another, and instruments from multiple manufacturers, all producing different file formats, which are not interoperable.
AI requires more than a collection of digital files. The outputs generated by individual analytical instruments represent only a portion of a chemical study. To make them useful for AI, organizations must assemble those files with data from ELNs, LIMS, and other relevant systems, preserving the context needed to understand data provenance – how the experiment was designed, and how the results relate to one another.
This remains a key bottleneck, difficult for many organizations. Workflows may still involve manual steps, and the necessary information is often distributed across disconnected systems. Even when it can be brought together, it must be normalized, cleaned, and accompanied by complete metadata.
Terminology and vocabulary must also be consistent. A field such as “sample ID,” for example, might appear elsewhere as “SAMID,” “SampleID,” or simply “sample.” Although those fields represent the same definition, an AI system may not interpret them consistently if the organization has not established a common vocabulary and ontology.
There has been some discussion about whether generative and agentic AI reduce the need to bring everything into one place. These systems could connect to different systems and repositories, retrieve information from multiple files, and perform some of the data assembly. However, this is likely to be most effective for numerical and tabular data. Without appropriate applications for visualizing and manipulating analytical data, the ability to leverage more complex outputs, such as spectra and chromatograms, will remain limited. AI therefore needs to work alongside context-specific analytical systems to produce the desired results.
Even where AI can retrieve and assemble the relevant data, a further challenge remains: the underlying data may not be directly comparable. If data come from three different LC-MS systems, for example, each platform may process or interpret them slightly differently. AI could pull those results together into a table, but that does not mean they have been generated or processed consistently. Differences between instruments, vendors, and processing workflows could affect the conclusions drawn from them. A common platform can help organizations standardize data processing and resulting spectra/chromatograms. These standardized datasets can then be interpreted and leveraged as consistent, normalized data.
Where AI can help today
Building a robust, sufficiently structured data foundation may take time, but organizations should think about how they will leverage AI once the data is ready.
AI can examine historical data, identify patterns, and flag anomalies for a scientist to investigate. It might identify something that appears unusual or falls outside the expected range, allowing the scientist to focus their review accordingly.
AI is also useful for working with documents. Large language models can scan multiple sources, extract relevant information, and bring it together. If the assembled data and context from a stability or method development study are available, for example, an LLM could produce a draft report following a defined specified format or set of guidelines. The scientist would still need to review that report, but the initial draft could be generated very quickly.
Another opportunity lies in workflows that generate so much data that it becomes difficult for scientists to extract useful insights. Scientists may already have access to business intelligence and statistical tools such as Tableau, but extracting insights from very large datasets can still be time-consuming. AI may help them reach those insights more quickly by processing larger volumes of data efficiently.
The risks of adopting AI without a clear strategy
There is a risk that organizations select AI projects according to whichever data happen to be readily available, rather than the applications that could deliver the greatest scientific or operational value. A better approach is to identify the desired use case and then determine what data foundation it requires.
My advice is to start small. Examine your existing workflows for focused applications that could deliver tangible value, then assess whether the necessary data are sufficiently digitalized, assembled, and contextualized. Workflows that already have some of those foundations in place may offer sensible opportunities for early progress.
You should also look for workflows in which data pass through many hands during review. Release or batch testing, for example, could benefit from AI to flag anomalous results for further investigation.
The key questions are: where could faster analysis and interpretation reduce manual data curation or review? And crucially, do we have the data foundation in place to make use of AI?
It is an exciting time. As people become more familiar with AI in their everyday lives, they will also become more comfortable using it in the workplace. But to be certain, when AI produces information, humans will still play a critical role in deciding what to do with it. AI will highlight issues that require attention and identify trends that scientists would have eventually found themselves, but much more quickly. Scientists will then apply their knowledge and experience to interpret those insights and decide how to act.
Organizations may not know which AI enabled workflows will prove most valuable until further down the road. The important thing is to be ready when they emerge. That means treating analytical data as a scientific asset: assembled, contextualized, and ready to support the next generation of AI tools.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
