The question of whether AI belongs in research has been answered. It does.
The more interesting question is which operating model fits which category of work. Two distinct models are taking shape: one scientist-led, the other agent-led. They are not competing versions of the same process. They suit different kinds of research activity, place human control at different points and depend on the same underlying foundations: governed data, validated tools and a connected ecosystem.
The hurdle isn’t the model’s intelligence; it’s the scientist’s trust.
The scientist-directed model
Every scientist recognises the Design, Make, Test, Analyse (DMTA) cycle, with make and test happening in the ELN and design and analysis running in dedicated computational software. Running those tools means either learning new skills and UIs or engaging computational teams and waiting, sometimes a week, for results that should be available in minutes.
For two decades, the industry tried to solve this with custom interfaces of computational tools that scientists might actually use. They never worked. Every new interface was still another UI to learn, another context switch, another barrier between the scientist and the result.
Natural language finally removes that barrier entirely.
A single series of experiments might draw on 10 different computational packages from six vendors, each with its own UI and file format. An AI co-scientist embedded in the ELN calls the appropriate validated tool at the scientist’s direction and returns the output directly into the experiment record. Scientists don’t need to learn how the software works. They just say what they want.
The agent-directed model
The other approach is the agent-directed model. Here the scientist defines the research question or hypothesis; the platform orchestrates the tools, data sources and computational steps needed to test it, returning evidence or asking for scientists’ input at the points where human review matters most.
For functions where scale and complexity make manual coordination impractical, such as target assessment, molecular design and lead generation, agent-led execution is the more realistic way to move from question to tested hypothesis at the pace the work demands.
What makes this model credible, rather than just fast, is transparency. Agent-led operations require high data quality, well-defined review thresholds and full audit trails across the pipeline. Without that, autonomy creates exposure instead of value. With it, the coordination burden that would otherwise consume scientific capacity gets absorbed by the system and the scientist stays accountable for the decisions that matter.
Data is the glue
Neither model works without the same underlying discipline: accurate, structured, governed data.
Scientific data accumulates across instruments, LIMS records, legacy documents and disconnected systems that were never designed to speak to one another. Inconsistent naming conventions, absent metadata and siloed departmental standards mean the same entity can mean different things depending on where in the discovery pipeline it sits. AI doesn’t interrogate the data it’s given; it just acts on it. So bad data doesn’t slow the system down; it results in faster chaos.
The organisations making real progress treat data readiness as a prerequisite, not a parallel track. The AI readiness gap closes from the foundation up, not from the tool down.
Two companies, one ecosystem
That is the logic behind what we have built across Sapio Sciences and Sigmatic Sciences.
In the scientist-directed model, Sapio’s AI co-scientist works within the ELN, helping scientists to move work forward, access validated tools and keep every step inside a governed scientific record.
In the agent-directed model, Sigmatic Sciences, a Sapio Sciences company, extends that same environment into lab-connected orchestration: a scientist’s question is translated into testable hypotheses and evaluated across validated computational tools and live experimental data.
Sapio Elain and Sigmatic Scout are two points on the same journey: one making the scientist-led DMTA cycle faster and more capable, the other taking entire research cycles virtual before the wet lab is ever involved.
Science does not operate in one mode. The infrastructure supporting it should not either.
Rob Brown is Head of the Scientific Office at Sapio Sciences. He previously served as head of global presales, product marketing, and product management at Dotmatics and earlier led product marketing teams at Accelrys, SciTegic, and MSI.
