Bridging the Expectation Gap: Building a Science-Smart AI Strategy for R&D

How lab leaders can close the gap between investment and impact

Written bySarah BauderPresented byAndrea Jacobs
| 6 min read
AI-driven laboratory research and automation in action
Register for free to listen to this article
Listen with Speechify
0:00
6:00

Most R&D organizations have already placed their bets on AI—adoption is widespread—yet the results rarely match the promise. That disconnect was the central theme of Andrea Jacobs' session at the 2026 AI in Lab Automation Digital Summit. Jacobs, senior director, AI at CAS, leads a team of data scientists building AI-driven solutions to accelerate R&D workflows. Drawing on surveys and CAS's own conversations with research organizations, she opened with a sobering observation: the vast majority of labs have launched AI in R&D initiatives in the last two years, but satisfaction remains surprisingly low. The pattern holds industry-wide—Deloitte's 2026 Life Sciences Outlook found that just 22 percent of life sciences leaders had successfully scaled AI, and only 9 percent reported significant returns. "The gap between AI investment and meaningful impact is exactly what we're here to discuss today," Jacobs told the audience. Her prescription is what she calls "science-smart AI."

Female tech leader presenting on laboratory automation

Andrea Jacobs, senior director, AI at CAS.

Why is scientific AI so hard to get right?

Jacobs identified five obstacles that make R&D a fundamentally harder environment for AI than  finance or HR, which keep AI adoption from translating into results:

  • Diverse, complex scientific data where minute details drive fundamental meaning. 

  • Rapidly evolving science that keeps shifting the underlying data foundation.

  • Missing core competencies in data science, AI, and their intersection with science.

  • Skeptical scientists trained to disprove rather than trust.

  • Unreasonable expectations imported from unrelated business functions.

The first is the nature of scientific data itself. Leaders, Jacobs cautioned, sometimes assume complexity can be managed with an 80-20 rule. In science, the details are the meaning. "A wedge versus a dash versus a line is a meaningless formatting difference in most fields," she explained, "but that tiny difference has a dramatic impact on the chemical application." A transcription slip (mistaking a numerical value) can yield an entirely different compound. As she put it, "our challenge in science goes beyond just, there's a lot of formats, there's a lot to wrangle. It's a deeper problem because the details drive true scientific meaning."

The second obstacle is the pace of change. Labs must contend not only with the explosion of published literature but with the terabytes of data their instruments and researchers generate daily, much of it exported without the contextual metadata that makes it usable for AI downstream. Furthermore, science itself is iterative. "What's true today can be refined or overturned tomorrow," Jacobs noted, comparing the task to "trying to build your AI house, but the foundation is a raging river."

Third, successful AI implementation demands core competencies most R&D organizations were never built to house. Data modeling, data governance, chemical structure embeddings for machine learning—these are distinct disciplines, and few research companies have them woven into their organizational fabric. The gap is widening: in the Pistoia Alliance's 2025 Lab of the Future survey, 34 percent of respondents cited a lack of skills as a barrier to AI adoption, up from 23 percent in 2024. Worse, Jacobs pointed out, the skills required "aren't just data or AI skills, but they're data and AI skills plus science skills."

Fourth is a cultural reality every lab leader will recognize: scientists are trained skeptics. "The entire foundation of our scientific education is disproving the hypothesis," Jacobs said. When a researcher spots a flaw in an AI-generated answer, their instinct is to discard it and move on. Because errors are inherent in any AI system, that low tolerance for flaws "undermines buy-in, adoption, and trust" she noted, and trust remains the central barrier to AI adoption in scientific settings.

The fifth obstacle is unreasonable expectations, often imported from outside the lab. A senior leader reads a few case studies promising 50 percent cost reductions or tenfold revenue gains—drawn from finance, HR, or cybersecurity, "areas where the complexity of science just doesn't play a role." Set that expectation, Jacobs warned, and "you're bound to underwhelm." The frustration is measurable: a Roland Berger survey of R&D decision-makers found 67 percent dissatisfied with the speed of AI implementation in their departments.

Lab manager academy logo

Lab Quality Management Certificate

The Lab Quality Management certificate is more than training—it’s a professional advantage.

Gain critical skills and IACET-approved CEUs that make a measurable difference.

Defining "science-smart AI" 

Jacobs' answer to these obstacles is AI built with scientific reality embedded from the outset. "This is AI with a lab coat," she said. "AI that recognizes and integrates the unique challenges of science from the ground up. In the data, in the problem framing, in the user engagement, in the definition of goals." The reframing matters: AI success isn't a matter of buying the right tool, but of building the right foundation that integrates the unique challenges of science from the ground up, beginning long before a model is ever trained. 

Why does data foundation matter more than the model?

Science-smart AI starts with data, and Jacobs was careful to separate two distinct sets of problems that leaders often conflate.

The first is non-scientific: data accessibility, security, and technical infrastructure. Organizations with longer histories, especially those built through acquisitions and mergers, tend to carry a tangle of data silos that don't communicate with each other. "The data fabric is really a mess," Jacobs acknowledged, a real and persistent barrier, but not a scientific one.

Interested in lab leadership?

Register for a FREE Lab Manager account to subscribe to our Lab Leadership Digest Newsletter.
Subscribe for Free

The second problem is scientific, and it's where most efforts fall short. True scientific data readiness requires rigorous quality control, curation, harmonization, and adherence to the widely adopted FAIR data principles (Findable, Accessible, Interoperable, Reusable). "Good science in, yields good science out from an AI perspective," Jacobs explained. Additionally, she was emphatic that this is not a project with a finish line. "There's no such thing as done. It's an ongoing commitment. And don't believe anyone who tells you otherwise." It's a point that echoes across the field: as other Lab Manager coverage has noted, AI amplifies whatever is already in the data, scaling clean inputs into insight and gaps into confident-looking errors.

Jacobs also cautioned against letting large language models (LLMs) and generative AI paper over these gaps. Their capabilities are genuine, offering genuine utility for translation or summarization. Yet, "large language model expertise alone isn't enough. It can only get you so far." Without scientific context and a strategy for harmonizing data, an LLM cannot deliver the robust foundation a lab needs. Better, she argued, to treat these generative AI models as one component that works in concert with scientific experts, both in building the data foundation and in generating insight on top of it.

Defining the right problem

Even with clean data, Jacobs argued, AI in the lab fails when teams rush to solutions before defining the problem. She invoked the guidance widely attributed to Einstein: "If I had an hour to solve a problem, I'd spend 55 minutes thinking about the problem and five minutes thinking about solutions."

The difference shows up in how a use case is framed. A directive like "use AI to make our lab more efficient" is both vague and fixated on the solution rather than the problem. Compare that, Jacobs suggested, to "reduce total synthesis time while using in-stock starting materials and ensuring yield stays above 80 percent." The second version lets models optimize for the right metrics and produces solutions that map to real problems on the bench. Defining the right question, with realistic success metrics connected to it, is itself an act of scientific expertise.

The triangle for success

To pull these threads together, Jacobs described what she calls the "triangle for success," three types of expertise required for any serious AI effort:

  • Domain experts: Scientists with deep understanding of the workflows, who bring real-world lab data to the solution.

  • Tech and algorithm experts: Specialists who engineer the inner workings of AI systems

  • Content experts: Data management partners who bring curation, harmonization, and modeling expertise. 

"To get to science-smart AI that can deliver real results and be trusted and adopted by our skeptical users, all three of these experts are needed," Jacobs said—a recurring warning for lab leaders, who are cautioned against treating AI as something to buy from a vendor without deep science skills in-house.

The benfit of this curated, expert-driven approach is measurable. Jacobs shared two examples where scientist-curated data outperformed noisier public or machine-curated datasets. In drug-protein binding prediction, models trained on smaller, higher-quality CAS-curated data achieved twice the prediction accuracy with 50 percent fewer data points.  "Half the data, double the results," she summarized. "Data quality really beats data quantity every time." Similarly, a partnership with Molecule One using scientist-curated atom mappings produced a 30% increase in pathway accuracy, winning the trust of working lab scientists. 

How can lab leaders win over the skeptics?

Because adoption ultimately hinges on trust, leaders must proactively manage the cultural rollout of AI tools. Jacobs offered practical advice on building scientist confidence:

  • Education: Provide transparent  information about an application's purpose, its role in the workflow, and its historical evaluation results

  • Explainability: Utilize explainable AI that ties predictions directly back to source data, showing how a generative model reasoned through a question. 

  • Involvement: Bring bench scientists into the development and deployment phases to build buy-in and surface critical workflow feedback.

Researchers consistently state that transparent sourcing and explained reasoning are are prerequisites for reliance. It also increasingly carries regulatory weight: the FDA's 2025 draft guidance on artificial intelligence in drug development proposes a risk-based credibility framework for AI models used to support decisions on drug safety, effectiveness, and quality.

The takeaway for lab leaders

Bridging the expectation gap, Jacobs stressed, "isn't about lowering our ambitions for AI in R&D. It's about building smarter strategies that account for the unique challenges of scientific applications." The organizations that succeed share a few common traits:

  • They combine computational power with deep scientific expertise.

  • They treat data quality as a continuous operation.

  • They define the scientific question before the solution, with realistic success metrics.

  • They set expectations grounded in scientific reality, not case studies from other industries.

For a lab manager launching an AI project, Jacobs’ first-step advice was concrete: assemble the expert triangle, and define a clear, well-scoped use case before anything else. The Lab Manager guide to evaluating and implementing intelligent workflows reinforces the same sequence: prove value on one workflow, then scale with a roadmap the entire team understands. Ultimately, the goal, in Jacobs' words, is "empowering scientists with science-smart AI tools that amplify their expertise and accelerate discovery."

Frequently Asked Questions (FAQs)

  • What is science-smart AI?

    Science-smart AI is AI built with the unique challenges of scientific data embedded from the outset—in the data, problem framing, and goals. Andrea Jacobs of CAS calls it "AI with a lab coat," designed to handle scientific complexity rather than treat it as an afterthought.

  • Why do AI initiatives in R&D often fail?

    Scientific data is complex and detail-driven, science evolves faster than data foundations can keep up, and most labs lack the blend of data, AI, and science expertise required. On top of that, skeptical scientists and unrealistic expectations undermine trust and adoption.

  • Why don't scientists trust AI tools?

    Scientists are trained to disprove hypotheses, so they instinctively discard AI outputs the moment they spot a flaw. Building confidence takes education, explainability, and involving researchers directly in developing the models.

  • What is the first step to launching an AI project in the lab?

    Assemble Jacobs' "triangle for success" (domain, technical, and content experts) before anything else. Then define a clear, well-scoped use case with realistic success metrics rather than starting from the solution.

Add Lab Manager as a preferred source on Google

Add Lab Manager as a preferred Google source to see more of our trusted coverage.

About the Author

  • Placeholder image

    Sarah Bauder is the senior editor at Lab Manager. She possesses a diverse background spanning editorial, digital marketing and film and television production. She brings over 15 years of experience in editorial writing, B2C and B2B content creation. A student of history, she graduated from York University in Toronto, Ontario, Canada. She can be reached at sbauder@labmanger.com.

    View Full Profile

Speaker

  • Headshot

    Andrea Jacobs

    Director of Data AnalyticsCAS

Related Topics

Loading Next Article...
Loading Next Article...
Current Magazine Issue Background Image

CURRENT ISSUE - October 2026

Why Lab Teams Resist Change—And How Leaders Can Respond

Addressing the human side of automation, AI, and operational change

October 2026 Lab Manager Cover Image