As AI shifts from novelty to necessity in life sciences R&D, many scientists are already using powerful models in their daily work—sometimes without sanctioned tools, policies, or audit trails. That reality makes governance, reproducibility, and traceability urgent, not optional. At Sapio, this takes the form of embedding AI agents directly within the electronic laboratory notebook (ELN) environment, allowing contextual reasoning without leaving the experimental record.
Few observers have a wider vantage point on this transition than Rob Brown, head of the scientific office at Sapio Sciences. With extensive leadership experience in the life sciences technology sector and a background in pharmaceutical research—Brown combines a hands‑on bench perspective with decades of informatics strategy.
In this article, he discusses why shadow AI is emerging, how AI-enabled ELNs and laboratory information management system (LIMS) can bring scientific reasoning into the notebook, and why strong guardrails, clear audit trails, and “trust, but verify” workflows are essential if AI is to be embedded responsibly into everyday research practice.
Shadow AI is a signal that urgency now outpaces provisioning
“There’s an element of, ‘well, if I’m not going to get given something, I’m going to do it for myself anyway.’” — Rob Brown
Brown cautions that slow review cycles, reminiscent of early cloud adoption, will not suffice—the window is months, not years—or ungoverned usage could proliferate.
The core risk is unvalidated science, or outright hallucination of scientific methods. Without clear records of what the scientist did versus what the AI contributed, organizations may face future IP, legal, or regulatory uncertainty.
Figure 1: The risks associated with Shadow AI and capabilities of governed AI in scientific workflows. Credit: AI-generated image created using Microsoft Copilot (2026).
What this means:
- AI adoption is now an operational imperative.
- Slow approval cycles increase the likelihood of shadow AI.
- Governance, training, and auditability must accompany access from the start.
From passive records to reasoning workspaces
Historically, ELNs captured experiments and made them searchable. Scientists reviewed prior work, then stepped outside the ELN to consult the project team and computational experts before returning to document the next experiment.
Brown argues that AI changes that pattern. At Sapio, AI agents are embedded directly into the ELN, known as an AILN, so the system can understand experimental data and context, and provide analysis without researchers leaving the notebook. He adds that scientists can control integrated lab instruments with natural language consistent with the rest of the AILN.
Crucially, this is not unconstrained large language model (LLM) reasoning. The architecture pairs LLMs with AI agents that invoke validated tools—cheminformatics, bioinformatics, structure-based design, and other methods already trusted within organizations. Brown adds that Sapio is building a partner network of these trusted vendors so that, when a scientist asks a question like “Calculate the ADMET profile of my proposed compounds?”, the ELN calls the exact package the organization’s computational team has already validated, returning results inside the notebook without altering the underlying algorithm.
“It’s like having a bioinformatician or cheminformatician right over a scientist’s shoulder” — Rob Brown
Because the underlying validated algorithms remain unchanged, accessibility improves without compromising scientific rigor. Brown notes that many established computational vendors are actively integrating into AI-driven ELN ecosystems, extending specialist tools across broader teams.
He likens the impact to giving every junior scientist direct access to the most experienced expert in the organization—amplifying capability rather than replacing expertise.
“With AI-driven ELNs, you can get the best answer your experts could have provided for you —without having to jump through all the hurdles.” — Rob Brown
How the environment changes for scientists:
- Experimental reasoning happens inside the ELN rather than outside it.
- LLMs orchestrate calls to validated computational tools.
- Specialist capabilities become accessible through prompts.
- Researchers can access public or commercial models from directly within the ELN.
Trust, accountability, and governed intelligence
Building trust in AI-driven ELNs, Brown argues, begins with clarity: scientists must be able to trace exactly how decisions were made. The ELN should record what the scientist did and what the AI contributed so decisions can be reconstructed later. That auditability supports regulatory and IP compliance and reduces uncertainty in downstream review.
Transparency should also extend to analytical methods. When the AI proposes to analyze a dataset, it should first outline the intended methods and ask permission to proceed. Where the AI calls pre-approved computational agents, it inherits existing verification; where it reasons more independently, method visibility becomes even more important.
Even with these safeguards, responsibility remains human. AI can remove time-consuming tasks, but scientists must interpret results and exercise judgment.
“It absolves you of the grunt work… but it doesn’t absolve you of the validating the results .” — Rob Brown
What matters for governed use:
- Clearly show what the scientist did versus what the AI did.
- Use validated tools for important analyses, with the AI showing its planned method first.
- Keep the scientist responsible for interpreting the results.
Two emerging operating models: AI in the loop and lab in the loop
Brown describes two emerging paradigms.
Today’s systems are largely AI in the loop: scientists lead experiments and consult AI assistants embedded within the ELN.
Looking ahead, he anticipates scenarios where AI conducts extended virtual research cycles—designing candidates, refining models, and iterating computationally—before handing off to the laboratory for physical validation. In that case, the scientist and the lab become in the loop at key checkpoints.
He expects both models to coexist.
Emerging patterns:
- Scientist-led workflows supported by AI assistance.
- AI-orchestrated research cycles requiring lab validation.
- Human oversight remains central in both models.
AI’s growth curve
Brown describes the past two years of embedded AI progress as moving from “toddler” (frequent failure, scant knowledge) to “teenager” (selective cooperation, limited knowledge) to something approaching the level of an expert researcher.
Because the platform can switch foundation models rapidly, improvements are incorporated as soon as new versions outperform older ones. Agility, rather than long-term lock-in, allows organizations to benefit from the accelerating model arms race.
Long-range predictions are difficult at this pace—but normalization is near.
“In less than a couple of years… my AI assistant and I go into the lab, and we do science together.” — Rob Brown
As governed systems mature, Brown expects shadow AI to recede—not because AI use declines, but because it becomes standard, sanctioned, and embedded within core informatics platforms.
What to watch:
- Rapid iteration of foundation models.
- Decline of shadow AI as governed systems mature.
- ELNs and LIMS are becoming the operational surface for AI-driven research.
“What today feels experimental will soon feel inevitable — a world where scientists and AI work side by side, not as replacements, but as teammates accelerating discovery.” —Rob Brown
AI is becoming inseparable from everyday scientific work. The question is no longer whether to adopt it, but how to embed it responsibly.
For Brown, the path forward is clear:
- Integrate AI directly into ELNs and LIMS.
- Pair LLMs with validated scientific agents.
- Maintain auditability and provenance.
- Keep scientists accountable for conclusions.
The future of AI in life sciences is not autonomous replacement, but governed augmentation—deeply embedded, transparent, and accountable.
This content includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks’ AI policy can be found here.