Skip to content
Every leader wants to talk about AI. Fewer want to talk about what has to happen before AI can be applied. Most labs aren’t ready to “do AI” in any meaningful way – not because the models aren’t good enough, but because the data those models would need is scattered, siloed, poorly contextualised, and locked inside proprietary formats. If you want to play the AI end game, you first have to build the foundation underneath it: universal data accessibility.
That foundation rests on three principles: connecting instruments to leading systems, giving data a proper home, and exploiting that data for real value. Skip the first two and the AI conversation becomes an impossibility.
Think of scientific data as moving through a life cycle. First, generation: experiment execution, time-series data collection from sensors, and the “last mile” connection between an instrument and whatever system requested the measurement. Then data management: once data exists, it needs a home and a consistent way to be captured across every format a lab produces. Ideally, data is normalized to a common format at that point. Only following that comes application – where decisions are made and analytics and AI can provide additional insights. It’s tempting to skip straight to the exciting part, but an AI strategy that skips connectivity and data management is building on sand.
Pillar One: Vertical Integration Without the Custom Code Graveyard
Anyone who’s integrated instruments with a LIMS or ELN knows the pain. It’s hard to do once, harder at scale, even harder to maintain as the landscape keeps changing. The answer isn’t another bespoke driver for every pairing. It’s a vendor-neutral connectivity layer, implemented through a single interface to the leading system, so an integration pattern built once can be reused for every future instrument.
The second half of the problem is context. An instrument can tell you what it measured, but not what a sample ID actually means – which batch, which formulation, which recipe. That context lives in the leading system: the LIMS, ELN, SAP, or MES, each governing its own domain. The power comes from enriching instrument-level detail with system-level context, bidirectionally – matching data to sample records as it arrives, or pushing worklists down to the bench and tracking execution across connected equipment.
In practice, this looks like a pH meter showing a scannable worklist next to the probe, a balance displaying the exact job it needs to weigh, and a digital pipette sending dispensed volume into a CDS for the dilution factor to be automatically calculated. Weigh in a sample, and that weight flows straight into a chromatography sequence table, with the same viewing components embeddable directly inside common LIMS platforms, so operators don’t need to leave the system that governs the work.
Underneath it all sit two answers to messy, proprietary formats: push standardisation where the industry is converging, and build normalisation adapters for everything else. Both matter, as unfortunately it’s unrealistic to rely purely on standardisation.
Standards Matter, Precisely Because They’re Invisible
Standards have an odd property: missing, everyone complains; present, they become invisible. The 96-well SBS microplate format quietly transformed life sciences by simply working, and shipping containers did a similar thing for global trade.
In lab informatics, standards like SiLA address connectivity – a stable core with extensible layers, not unlike USB giving a stable connector while what runs over it continues to evolve. AnIML, an XML-based data format, tackles data structure itself, capturing materials, instructions, and results in a form that supports proper audit and quality control. Both are community-driven and vendor-neutral, with working groups now extending into robotics, cybersecurity, and AI.
The pragmatic approach isn’t insisting on one standard everywhere – it’s using the right tool for the job. Some integrations suit a standards-based approach; others suit a direct, proprietary connection. What matters is that everything talks to everything else reliably.
Pillar Two: Give Your Data a Home Without Forcing a Migration
Once data exists, it needs somewhere to be accessible from, not necessarily to physically live. A better approach sits on top of existing infrastructure where it exists, and provides storage where it doesn’t. Modern lab data is genuinely multimodal: files, time-series data, readings from simple instruments, relational data, metadata, chemical structures – all needing to be understood in context, whether it’s from a manufacturing batch or an R&D experiment. Accumulating so much accessible data creates a new problem: a haystack full of needles. Solving it starts with knowing what you have – achieved through proper data cataloguing and contextual metadata enrichment that treats existing systems as systems of record rather than replacing them. That enables real discovery, not just search: search works when you already know what you’re looking for; discovery requires a catalogue you can browse. By layering identification, access control, and governance on top, you have content that’s genuinely navigable.
Pillar Three: Turning Accessible Data Into Value
With connectivity and data management in place, data value creation becomes dramatically easier. Liberating data from a chromatography system and normalising it into an open data model means chromatograms, spectra, and audit trails all become viewable through consistent web renderers, regardless of the source system. The same tooling extends naturally to particle sizing, bioreactor data, Raman spectra, and plate readers — normalise once, visualise everywhere. Crucially, these components embed directly inside a LIMS, so an analyst can see chromatogram detail without a system switch, far faster than a from-scratch integration.
Where AI Actually Fits
With data federated, contextualised, and governed by proper identity and permissions, exposing it to large language models becomes a matter of building domain-specific tools on top of the existing data access layer, using standard interfaces like MCP for tool access and A2A for agent-to-agent communication.
In practice, that means asking an AI assistant, in plain language, which instruments are connected or what retention time drift looks like across a sequence — with the system finding the right files, understanding the metadata, and answering within the querying user’s own permissions. Data isn’t used to train a shared model; it’s served on demand according to who’s asking, keeping governance intact.
Getting a model to be genuinely useful requires us to build domain-aware tools so it understands a chromatogram or a qPCR run, and only expose the model to specific data subsets so it isn’t flooded with noise. None of this locks users into one interface or model – an open approach lets organisations bring their own model and combine it with agentic capabilities from other systems.
Closing the Loop
The natural extension is the closed loop, or “self-driving lab” — using accumulated data and AI to automatically determine what experiment to run next. “Loop” does not have to mean “unattended”: humans can sit inside the cycle in several roles — as operators carrying out the steps the equipment can’t, as scientists guiding the direction of the campaign and adjusting its constraints as evidence accumulates, and as supervisors reviewing each iteration’s results, approving or overriding the recommendation, and adjusting the parameters for the next round. Because connectivity, data management, and AI access are already solved, adding a processor that reads results at the end of each cycle and reprioritises a queue of upcoming experiments becomes a natural extension, not a separate project.
A Practical Checklist
The path to AI readiness isn’t mysterious, but it requires sequencing: connect existing systems rather than replacing them; federate data to support a genuinely FAIR approach; make content discoverable through a proper catalogue; and only then layer AI on top with carefully scoped tools. Universal data accessibility isn’t a nice-to-have alongside an AI strategy – it’s the foundation the strategy relies upon.
For more information, visit Splashlake.
