AI workflows - TetraScience: When AI meets the lab (lessons from scientific data workflows)

From artisan shop to assembly line

Laboratory data workflows have undergone a transformation that mirrors the evolution of software development itself. For decades, each data integration was a bespoke creation, hundreds of handcrafted point-to-point connections between instruments, databases and analysis tools. Sound familiar?

This artisanal approach created beautiful, highly specialised and bespoke solutions. It also created technical debt, brittleness and an inability to scale. The Covid-19 pandemic exposed these limitations when labs needed to collaborate at unprecedented speed and scale.

The response wasn’t to hire more integration specialists. Instead, leading organisations adopted what they call “industrial-scale data factories” – standardised, reusable components that could be composed into complex workflows without starting from scratch each time. Life sciences organisations want to move beyond asking “how do I move my data?” to asking “how do I use AI to reduce clone selection time by 70% or predict equipment failures?”

Here’s where it gets interesting for enterprise developers…

These “data factories” embody several patterns that apply far beyond scientific computing:

Standardised Schemas as APIs:

Instead of negotiating data formats between systems, successful labs established common data schemas—essentially contracts that any system could implement. Think of it as API-first development, but for data structures.

Composable Pipeline Components:

Rather than building monolithic data processing applications, they created libraries of small, focused components that could be chained together. Each component did one thing well and could be reused across different workflows.

Self-Service Infrastructure:

Perhaps most importantly, they built platforms that allowed domain experts (scientists) to create and modify workflows without waiting for IT teams. The infrastructure handled the complexity; the interface remained simple.

The self-service revolution

When AI entered these environments, something unexpected happened. Instead of replacing human decision-making, AI became a force multiplier for workflow automation. AI assistants now generate data transformation code, suggest optimal pipeline configurations and automatically handle edge cases that previously required manual intervention. But the key insight is that AI works best when operating on standardised, well-structured workflows – not chaotic, artisanal integrations.

This suggests a different approach to AI adoption in enterprise development: instead of trying to apply AI directly to legacy systems, focus first on standardising and automating core workflows. AI can then accelerate what’s already working well.

The most striking innovation in scientific workflows is the democratisation of complex data operations. Labs now increasingly have bench scientists—not developers—creating sophisticated data pipelines using visual interfaces and AI assistants. This is going to become even more routine as the tools get better fast.

This isn’t dumbing down the technology; it’s making the technology more accessible by hiding complexity behind well-designed abstractions. The underlying systems are more sophisticated than ever, but the cognitive load on users has decreased dramatically.

For enterprise developers, this suggests rethinking who our “users” really are. Instead of building tools for other developers, what if we built platforms that domain experts could use directly? Because, while the industry often focuses on data workflows and the plumbing that moves scientific data around, what most practitioners really need are AI workflows that deliver transformational outcomes. Scientists, for example, aren’t just asking “how do I move my data?” but “how do I use AI to reduce clone selection time by 70% or predict equipment failures?”

What they often find is that AI workflows are only as good as the data foundation beneath them.

Lessons for enterprise architecture

Three principles emerge from successful scientific workflow transformations:

• Design for Composition, Not Customisation: Build small, focused services that can be combined in unexpected ways rather than large, configurable applications.
• Standardise Interfaces, Not Implementations: Focus on common data formats and APIs rather than forcing everyone to use the same tools.
• Optimise for Self-Service: The best workflow platforms make complex operations feel simple to end users while handling the complexity behind the scenes.

Beyond the lab

These patterns are already appearing in other domains. DevOps platforms increasingly offer visual pipeline builders. Data teams use notebook-based environments that hide infrastructure complexity. Modern APIs prioritise developer experience over feature completeness.

But scientific workflows push these concepts further because they operate under constraints most enterprise systems don’t face: regulatory compliance, zero-error tolerance and the need to integrate with thousands of proprietary systems. The solutions they’ve developed – standardised schemas, composable pipelines, AI-assisted automation and true self-service platforms – represent a mature approach to workflow orchestration that other industries are just beginning to discover.

The question isn’t whether these patterns will spread to enterprise development. The question is which organisations will adopt them first.

Kondapalli oversees TetraScience‘s product management, engineering, scientific data architecture and forward-deployed teams that build TetraScience’s platform and customer solutions and ensure their successful implementation worldwide. He was previously the vice President of Engineering at AppDynamics, an observability company acquired by Cisco.