Your Data Doesn't Need to Be Perfect. It Needs to Be Relevant
McKinsey’s 2025 State of AI report finds that 88% of organizations use AI in at least one business function. Yet only about a third have deployed it across multiple workflows. IDC reports that 88% of AI proof-of-concept projects never reach production.
Scale is where things break.
In highly regulated and industrial environments, the breakdown often traces back to data readiness.
Secondarily, teams face constraints around talent, budget, and change management. Those shape outcomes. But when AI efforts stall, it usually comes down to something more immediate: whether teams can access, trust, and use the data required to support decisions in practice.
Many organizations respond by focusing on data cleanup efforts — centralizing systems, standardizing formats, and trying to create a unified foundation before moving forward.
That approach slows things down more than it helps.
The Myth of the Perfect Dataset
A common internal assumption is that AI work begins once data is fully organized.
That leads to large-scale efforts to centralize, clean, and standardize everything before meaningful work starts.
Industrial R&D doesn’t behave that way.
In chemical, materials, energy, and manufacturing domains, data is expensive to generate and tightly tied to experimental context. Each data point often carries meaning that goes beyond its numeric value.
A 2023 review in npj Computational Materials describes this as the “dilemma of small data,” where traditional machine learning methods struggle because they rely on large, uniform datasets that rarely exist in scientific environments.
Instead, teams work with limited but high-value experimental observations.
Progress depends less on completeness and more on whether the available data can support a meaningful decision.
That shifts the core question: Do we have enough relevant data and scientific context to answer a specific question with confidence?
Once that question becomes central, the path forward changes.
Work tends to move away from large remediation programs and toward focused decisions. Teams identify a problem worth solving, examine what data already exists, and then determine what additional information would actually improve outcomes.
What “Ready” Actually Means in Practice
Data readiness shows up when available information is sufficient to support modeling, interpretation, and decision-making at an acceptable level of confidence.
At NobleAI, we use a Rubric we call:
DECIDE
D – Decision to improve
E – Evidence available (structured & unstructured data)
C – Context required to preserve scientific meaning
I – Information gaps that must be addressed
D – Domain experts and stakeholders required
E – Execution readiness (organizational adoption)
- Decision. What decision needs to be made? AI should be tied to something that matters: which formulation to test, which ingredient to replace, or which operating condition to change. Without a clear decision, the assessment becomes abstract.
- Evidence. Where does the relevant data live? Structured sources like ELNs, LIMS, databases, and simulation outputs. Unstructured sources like PDFs, patents, lab reports, and literature. Both count.
- Context. What context must be preserved? Units, test methods, processing history, batch information, and environmental conditions. Data stripped of scientific context isn't ready. It's just numbers.
- Information gaps. What gaps matter most? Not all missing data is equally important. Some gaps block modeling. Some reduce confidence. Some can be addressed through targeted experiments or external literature sources.
- Domain experts. Who needs to be in the room? Recent research from MIT's NANDA initiative highlights that many enterprise AI pilots struggle to scale because of organizational and workflow integration challenges rather than model performance alone.
- Execution readiness. Is the organization actually set up to act on this? A model can be technically sound and still go nowhere if there's no owner, no workflow it plugs into, and no plan for adoption beyond the pilot team.
Why Industrial AI Needs a Different Starting Point
Most AI systems assume abundant, structured, and consistent data.
Industrial R&D operates differently.
Consider a materials team with 35 experiments testing alloy compositions against strength and corrosion resistance. A general-purpose ML model needs hundreds or thousands of examples to find reliable patterns. With 35 data points, it overfits, or it simply won't commit to a confident prediction. The team is told to run more experiments before AI can help, which defeats the purpose.
Science-Based AI handles the same 35 data points differently. Because the model already incorporates known metallurgical relationships, it doesn't have to rediscover basic chemistry from scratch. It uses the experimental data to refine what's already known, rather than learning everything from zero. That's a usable prediction from a dataset that would leave a generic model guessing.
Learning chemistry from data versus applying data to known chemistry: that's the distinction, and it's why SBAI works in industrial R&D, where data is inherently expensive and limited.
We use SBAI because it reflects that reality by combining domain knowledge with limited, imperfect, and context-rich datasets. The emphasis shifts from data volume to decision relevance.
What matters is whether the available signal is strong enough to support a targeted application, not whether the dataset is complete in a general sense.
That changes how teams begin.
What an AI Acceleration Assessment Actually Delivers
The assessment focuses on where data, domain context, and business value intersect in a way that supports action.
At NobleAI, this process is used to identify where AI can directly improve R&D decisions rather than simply improving data visibility.
In selected deployments, outcomes have included up to a 20× increase in R&D productivity, a 30% reduction in experimental workload, and approximately $40M in revenue impact alongside $5M in operating cost savings. These results vary based on use case and data maturity, but they reflect what becomes possible when work begins with the right problem framing.
In one energy-sector deployment, downstream testing cycles were reduced by 75%, and large sets of candidate formulations were evaluated in minutes instead of months.
These outcomes are less about model sophistication and more about whether the underlying data carries enough scientific signal to support prediction and exploration.
The organizations that capture the greatest value from AI won't be the ones waiting for perfect data. They'll be the ones that identify where they already have enough signal to begin.
Move beyond AI readiness. Start delivering faster discoveries, smarter decisions, and measurable business impact with NobleAI’s AI Acceleration, available through the Microsoft Azure Marketplace. Qualifying organizations can begin today.
References
- McKinsey & Company. (2025). The State of AI in 2025.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai - IDC, as reported by CIO.com. (2025). 88% of AI pilots fail to reach production, but that's not all on IT.
https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html - Xu, H. et al. (2023). Dilemma of Small Data in Machine Learning for Materials Science. npj Computational Materials.
https://www.nature.com/articles/s41524-023-00999-7 - MIT NANDA Initiative findings, as reported by Fortune. (2025). MIT report: 95% of generative AI pilots at companies are failing. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- McKinsey & Company. (2025). The Next Frontier for AI in R&D.
https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-for-ai-in-r-and-d
Frequently Asked Questions
McKinsey’s 2025 State of AI report finds that 88% of organizations use AI in at least one business function. Yet only about a third have deployed it across multiple workflows. IDC reports that 88% of AI proof-of-concept projects never reach production.
Scale is where things break.
In highly regulated and industrial environments, the breakdown often traces back to data readiness.
Secondarily, teams face constraints around talent, budget, and change management. Those shape outcomes. But when AI efforts stall, it usually comes down to something more immediate: whether teams can access, trust, and use the data required to support decisions in practice.
Many organizations respond by focusing on data cleanup efforts — centralizing systems, standardizing formats, and trying to create a unified foundation before moving forward.
That approach slows things down more than it helps.
The Myth of the Perfect Dataset
A common internal assumption is that AI work begins once data is fully organized.
That leads to large-scale efforts to centralize, clean, and standardize everything before meaningful work starts.
Industrial R&D doesn’t behave that way.
In chemical, materials, energy, and manufacturing domains, data is expensive to generate and tightly tied to experimental context. Each data point often carries meaning that goes beyond its numeric value.
A 2023 review in npj Computational Materials describes this as the “dilemma of small data,” where traditional machine learning methods struggle because they rely on large, uniform datasets that rarely exist in scientific environments.
Instead, teams work with limited but high-value experimental observations.
Progress depends less on completeness and more on whether the available data can support a meaningful decision.
That shifts the core question: Do we have enough relevant data and scientific context to answer a specific question with confidence?
Once that question becomes central, the path forward changes.
Work tends to move away from large remediation programs and toward focused decisions. Teams identify a problem worth solving, examine what data already exists, and then determine what additional information would actually improve outcomes.
What “Ready” Actually Means in Practice
Data readiness shows up when available information is sufficient to support modeling, interpretation, and decision-making at an acceptable level of confidence.
At NobleAI, we use a Rubric we call:
DECIDE
D – Decision to improve
E – Evidence available (structured & unstructured data)
C – Context required to preserve scientific meaning
I – Information gaps that must be addressed
D – Domain experts and stakeholders required
E – Execution readiness (organizational adoption)
- Decision. What decision needs to be made? AI should be tied to something that matters: which formulation to test, which ingredient to replace, or which operating condition to change. Without a clear decision, the assessment becomes abstract.
- Evidence. Where does the relevant data live? Structured sources like ELNs, LIMS, databases, and simulation outputs. Unstructured sources like PDFs, patents, lab reports, and literature. Both count.
- Context. What context must be preserved? Units, test methods, processing history, batch information, and environmental conditions. Data stripped of scientific context isn't ready. It's just numbers.
- Information gaps. What gaps matter most? Not all missing data is equally important. Some gaps block modeling. Some reduce confidence. Some can be addressed through targeted experiments or external literature sources.
- Domain experts. Who needs to be in the room? Recent research from MIT's NANDA initiative highlights that many enterprise AI pilots struggle to scale because of organizational and workflow integration challenges rather than model performance alone.
- Execution readiness. Is the organization actually set up to act on this? A model can be technically sound and still go nowhere if there's no owner, no workflow it plugs into, and no plan for adoption beyond the pilot team.
Why Industrial AI Needs a Different Starting Point
Most AI systems assume abundant, structured, and consistent data.
Industrial R&D operates differently.
Consider a materials team with 35 experiments testing alloy compositions against strength and corrosion resistance. A general-purpose ML model needs hundreds or thousands of examples to find reliable patterns. With 35 data points, it overfits, or it simply won't commit to a confident prediction. The team is told to run more experiments before AI can help, which defeats the purpose.
Science-Based AI handles the same 35 data points differently. Because the model already incorporates known metallurgical relationships, it doesn't have to rediscover basic chemistry from scratch. It uses the experimental data to refine what's already known, rather than learning everything from zero. That's a usable prediction from a dataset that would leave a generic model guessing.
Learning chemistry from data versus applying data to known chemistry: that's the distinction, and it's why SBAI works in industrial R&D, where data is inherently expensive and limited.
We use SBAI because it reflects that reality by combining domain knowledge with limited, imperfect, and context-rich datasets. The emphasis shifts from data volume to decision relevance.
What matters is whether the available signal is strong enough to support a targeted application, not whether the dataset is complete in a general sense.
That changes how teams begin.
What an AI Acceleration Assessment Actually Delivers
The assessment focuses on where data, domain context, and business value intersect in a way that supports action.
At NobleAI, this process is used to identify where AI can directly improve R&D decisions rather than simply improving data visibility.
In selected deployments, outcomes have included up to a 20× increase in R&D productivity, a 30% reduction in experimental workload, and approximately $40M in revenue impact alongside $5M in operating cost savings. These results vary based on use case and data maturity, but they reflect what becomes possible when work begins with the right problem framing.
In one energy-sector deployment, downstream testing cycles were reduced by 75%, and large sets of candidate formulations were evaluated in minutes instead of months.
These outcomes are less about model sophistication and more about whether the underlying data carries enough scientific signal to support prediction and exploration.
The organizations that capture the greatest value from AI won't be the ones waiting for perfect data. They'll be the ones that identify where they already have enough signal to begin.
Move beyond AI readiness. Start delivering faster discoveries, smarter decisions, and measurable business impact with NobleAI’s AI Acceleration, available through the Microsoft Azure Marketplace. Qualifying organizations can begin today.
References
- McKinsey & Company. (2025). The State of AI in 2025.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai - IDC, as reported by CIO.com. (2025). 88% of AI pilots fail to reach production, but that's not all on IT.
https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html - Xu, H. et al. (2023). Dilemma of Small Data in Machine Learning for Materials Science. npj Computational Materials.
https://www.nature.com/articles/s41524-023-00999-7 - MIT NANDA Initiative findings, as reported by Fortune. (2025). MIT report: 95% of generative AI pilots at companies are failing. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- McKinsey & Company. (2025). The Next Frontier for AI in R&D.
https://www.mckinsey.com/capabilities/operations/our-insights/the-next-frontier-for-ai-in-r-and-d
