🤖 #AI in Manufacturing: What AI Can—and Cannot—Do for Industrial #Innovation – #Harvard Business.
Artificial intelligence is rapidly changing how companies approach engineering, manufacturing and industrial innovation.
AI systems can analyze enormous datasets, identify patterns, generate design alternatives, assist with troubleshooting and help engineers evaluate ideas faster than traditional methods. In modern manufacturing, AI is increasingly being connected with sensors, digital twins, robotics, process-control systems and industrial data platforms.
But there is an important distinction between accelerating innovation and creating innovation.
Recent research highlighted by Harvard Business Review raises an important question: if companies increasingly have access to similar generative AI models, why do some organizations achieve significantly better innovation outcomes than others?
Part of the answer may be that AI does not automatically eliminate the human limitations within an innovation process. In some circumstances, it can reinforce them.
For industrial companies, this leads to a more practical question:
Where should we trust AI—and where do experienced engineers, scientists and operators remain indispensable?
AI Is Becoming Part of Modern Manufacturing
Artificial intelligence is no longer limited to chatbots and office productivity.
Industrial AI is increasingly being applied to areas such as:
- Predictive maintenance
- Process optimization
- Automated quality inspection
- Production scheduling
- Supply-chain optimization
- Robotics and autonomous systems
- Digital twins
- Advanced sensing
- Engineering data analysis
- Generative design
- Energy and resource optimization
NIST’s 2026 roadmap for artificial intelligence and machine learning in smart manufacturing identifies industrial data analytics, sensing, autonomous systems, digital twins, robotics, supply chains, generative AI and large language models among important areas of development.
The potential is substantial.
But the same NIST roadmap identifies continuing challenges involving industrial data, integration, explainability, reliability, availability, maintainability and safety.
That is particularly important in heavy industry, chemical processing, metallurgy and advanced materials manufacturing.
AI Can Find Patterns—but It Does Not Automatically Understand the Process
Consider a metallurgical plant.
Thousands of operating variables may influence production:
Temperature → Pressure → Gas composition → Reaction kinetics → Particle characteristics → Product quality
An AI system can potentially analyze historical relationships among all these variables much faster than a person.
For example, it might discover that a particular combination of reactor temperature, pressure and feed composition frequently precedes an off-specification product.
That information can be extremely valuable.
However, correlation is not necessarily causation.
An experienced process engineer may recognize that the apparent relationship is actually caused by another variable that wasn’t adequately represented in the dataset.
This is where domain expertise becomes critical.
AI can tell an engineer:
“Something unusual is happening here.”
The engineer still needs to determine:
“Why is it happening?”
The Quality of Industrial AI Depends on the Quality of Industrial Data
Industrial AI has another fundamental limitation: its conclusions depend heavily on its data.
NIST has specifically highlighted the importance of understanding the data, assumptions and rules feeding industrial AI systems. It notes potential problems including incomplete data, inadequate variation and gaps in datasets.
Imagine training an AI system using five years of plant operating data.
That sounds impressive.
But suppose the plant has never operated under a particular combination of feed composition, temperature and pressure.
The historical database may contain no reliable information about that operating condition.
The AI model can still produce an answer.
That does not necessarily mean the answer is physically correct.
For industrial applications, therefore:
A confident AI prediction should never be confused with a validated engineering result.
AI Cannot Replace Physical Experimentation
Engineering ultimately operates in the physical world.
A proposed process must actually work.
A material must actually possess the required properties.
A reactor must actually remain stable.
A product must actually meet specification.
This is why laboratories, pilot plants and industrial trials remain essential.
AI might predict that changing a process variable will improve yield.
But the hypothesis still needs to be tested.
The real process may reveal effects that were absent from the model:
- Unexpected reaction kinetics
- Contamination
- Corrosion
- Equipment limitations
- Heat-transfer constraints
- Mass-transfer limitations
- Particle agglomeration
- Instrumentation errors
- Previously unidentified side reactions
Industrial innovation has always progressed through the interaction between theory, experimentation and experience.
AI adds a powerful new tool to that process.
It does not eliminate the process.
AI Can Generate Ideas—but Novelty Is More Complicated
Generative AI is exceptionally good at producing ideas quickly.
Ask an AI system for 50 possible solutions to an engineering problem and it may produce them within seconds.
That represents a significant productivity improvement.
But generating more ideas is not necessarily the same as generating better ideas.
The recent Harvard Business Review research highlights this problem. When innovation teams use similar AI systems, AI can sometimes steer ideation toward familiar concepts rather than truly differentiated solutions.
This creates an interesting paradox.
AI dramatically increases the speed of idea generation while potentially making it easier for organizations to converge on similar ideas.
For companies pursuing genuine technological differentiation, human creativity therefore becomes more important, not less important.
Experienced Engineers Possess Something Difficult to Digitize
An engineer who has spent 20 years operating a process possesses knowledge that may never appear in a database.
They may know that:
- A certain sound indicates a mechanical problem.
- A small pressure fluctuation precedes an unstable operating condition.
- A laboratory result looks technically acceptable but is inconsistent with experience.
- A particular raw material behaves differently despite meeting specification.
- An instrument reading is technically possible but probably incorrect.
- A proposed modification works theoretically but will create maintenance problems.
Much of this is tacit knowledge.
It develops through observation, mistakes, troubleshooting and years of interaction with physical equipment.
AI can help capture and organize some of this knowledge.
But replacing it entirely is far more difficult.
The Future Is Human + AI
The most productive question may therefore not be:
“Will AI replace engineers?”
A better question is:
“How can engineers equipped with AI outperform engineers without it?”
NIST is actively researching human-AI teaming in manufacturing, including applications involving digital twins and production scheduling. Its work reflects a broader shift toward combining computational capabilities with human expertise rather than treating them as competitors.
The division of responsibilities could increasingly look like this:
| AI Strengths | Human Strengths |
| Processing enormous datasets | Engineering judgment |
| Pattern recognition | Understanding physical context |
| Rapid calculations | Evaluating causality |
| Generating alternatives | Challenging assumptions |
| Monitoring thousands of variables | Managing unusual situations |
| Searching technical information | Experimental validation |
| Detecting anomalies | Safety responsibility |
| Repetitive optimization | Creative problem solving |
The strongest industrial organizations will likely combine both.
AI Could Make Experienced Engineers More Valuable
There is another consequence that receives less attention.
If AI automates routine engineering work, the value of experienced technical judgment may actually increase.
Junior engineers traditionally develop expertise partly by performing calculations, analyzing failures, reviewing drawings, examining operating data and troubleshooting equipment.
If AI performs increasingly large portions of those tasks, companies will need to think carefully about how the next generation develops deep engineering judgment.
Human oversight only works when the person overseeing the technology understands the underlying process.
Industrial companies therefore need to invest simultaneously in AI capability and technical capability.
Five Principles for Using AI in Industrial Innovation
Companies implementing AI in engineering and manufacturing should consider five basic principles.
1. Use AI to augment expertise—not blindly replace it.
AI should help engineers analyze information faster while leaving critical technical decisions subject to qualified review.
2. Validate AI recommendations against physical reality.
Simulation, laboratory testing, pilot trials and operating data remain essential.
3. Protect engineering knowledge.
Companies should capture the experience of senior engineers, operators and scientists rather than assuming an AI model already contains that knowledge.
4. Understand the data.
Before trusting an AI prediction, engineers should understand where the training and operating data came from and whether it adequately represents the situation being analyzed.
5. Maintain human accountability.
For safety-critical processes, responsibility cannot simply be transferred to an algorithm.
AI Is a Powerful Engineering Tool—not a Substitute for Engineering
Artificial intelligence may become one of the most important industrial technologies of this generation.
It can help manufacturers analyze more information, identify problems sooner, optimize complex systems and explore potential solutions faster.
But industrial innovation ultimately has to survive contact with physical reality.
Reactors, furnaces, pumps, materials, chemical reactions and production lines do not respond to persuasive language. They respond to physics and chemistry.
That is why the future of industrial innovation is unlikely to be AI versus engineers.
It will be AI combined with engineers, scientists and operators who know how to question its conclusions, validate its recommendations and turn computational insights into technologies that actually work.
And that combination may prove far more powerful than either one alone.
