Protecting industrial knowledge, connecting factory systems and turning AI into reliable operational intelligence.
By Dr. Rajendra Shekatkar | Expert Contribution to MFG Magazine
Consider a situation that many manufacturing leaders would recognise.
It is the beginning of a busy production week. The factory has its schedules in place, materials have been planned, machines are allocated and customer deliveries are already committed.
The sales team is pursuing new opportunities while production is focused on meeting existing deadlines.
Then an important customer calls with an urgent requirement. They need an additional 5,000 units within the next ten days. It is a valuable order, and the company would certainly not want to lose the opportunity.
The sales manager needs to respond quickly. Normally, this would involve checking with production planning, inventory, procurement and maintenance before making a delivery commitment.
Depending on the complexity, it could take several hours or longer to establish what the factory can realistically deliver.
But things are changing.
The company has recently started using AI tools. Someone suggests asking ChatGPT or Claude whether the additional order can be accommodated.
The available order details are entered into the assistant, with a request to prepare a production and delivery plan.
Within seconds, a response appears. It estimates the production capacity required, suggests a schedule and recommends how the additional order could be managed.
The response looks logical, detailed and convincing.
But there is something the AI does not know.
The factory is already operating at 90% capacity. A critical machine is scheduled for maintenance, a key raw material has a four-week replenishment lead time, and several existing customer orders are competing for the same production resources.
None of these details were provided to the AI assistant.
These constraints do not necessarily mean that the order cannot be delivered. There may be available inventory, alternative production capacity or opportunities to reschedule existing work.
But these possibilities must be verified before making a commitment.
The AI has produced an answer. But has it helped the company make the right decision?
This is a hypothetical example, but it illustrates an important distinction that manufacturing leaders need to understand.
The Difference Between Artificial Intelligence and Operational Intelligence
ChatGPT, Claude, Gemini and similar general-purpose AI assistants have made artificial intelligence accessible to business users.
They can interpret documents, summarise information, prepare reports and help people explore possible solutions.
Behind these applications are Large Language Models (LLMs), such as the GPT and Claude model families, which can also become components of broader enterprise solutions.
Their usefulness is not in question.
The concern arises when organisations assume that a convincing AI-generated response is equivalent to a validated operational decision.
Manufacturing does not operate on information alone. It operates on the relationship between machines, materials, people, processes, costs and customer commitments.
An AI assistant working only with a prompt cannot automatically establish these relationships or verify the factory’s current operating conditions.
Manufacturing therefore needs more than access to an LLM. It needs intelligence designed around the realities of its operations.
The Factory Is Not One Connected System
Most manufacturers have invested in technology over several decades. Yet the reality inside an established factory is often very different from the integrated environment shown in a technology demonstration.
A typical manufacturing environment may contain:
- Enterprise Resource Planning (ERP) systems managing sales orders, procurement and inventory.
- Programmable Logic Controllers (PLCs) and Computer Numerical Control (CNC) machines.
- Supervisory Control and Data Acquisition (SCADA) systems.
- Human-Machine Interfaces (HMIs).
- Standalone quality and inspection databases.
- Maintenance management applications.
- Excel spreadsheets containing critical production information.
- Manual registers maintained by operators.
This is particularly common in brownfield plants, where infrastructure has evolved over many years.
A company may have modern enterprise applications alongside machines installed twenty years ago. Those machines may still perform well, but they may not offer the connectivity expected by newer digital systems.
Does this mean the manufacturer must replace them before introducing AI?
Certainly not.
A manufacturing AI solution can be designed to connect information progressively, beginning with the systems relevant to the problem being addressed.
ERP information may be accessed through approved application interfaces. Machine operating data may be obtained from existing equipment historians, SCADA systems or industrial gateways.
Quality records may be retrieved from databases, while selected information from spreadsheets or legacy applications can be incorporated through controlled file ingestion.
Where direct connectivity is unavailable, validated manual inputs may still be necessary.
These connections must respect the existing technology environment.
Industrial control networks should not be exposed directly to an AI application simply to make integration easier. Approved gateways, controlled interfaces and appropriate network separation are essential.
Data Quality Is Equally Important
A material may have one identifier in ERP and another in production.
A machine may be known by different asset numbers in maintenance and operations. Batch records may be incomplete, and timestamps may not be consistent across systems.
If the solution cannot reliably connect an inspection result to the relevant production batch, machine and material, its conclusions may be misleading.
Before AI can interpret manufacturing information, the underlying records must be reconciled, validated and placed in the correct operational context.
Where information is missing, the solution should identify the limitation rather than make an unsupported assumption.
The objective is not to connect every machine or replace every legacy system. It is to bring together reliable information from the systems that matter for a particular decision.
Manufacturing Knowledge Is Not Just Another Data Source
There is another reason why AI adoption in manufacturing requires careful consideration: intellectual property.
Imagine an automotive component supplier developing a specialised part for an international Original Equipment Manufacturer (OEM).
The product may have taken months or even years to develop, test and qualify.
Behind the finished component are:
- Engineering drawings and product designs.
- Tooling designs.
- Material formulations.
- Process recipes and tolerance limits.
- Inspection procedures.
- Proprietary costing models.
- Manufacturing methods and process adjustments.
- Considerable accumulated industrial experience.
This information represents far more than a collection of documents.
It can be the supplier’s competitive advantage.
Now consider an engineer facing a recurring production issue.
To obtain a quick explanation, the engineer uploads the customer’s drawing, process parameters and inspection records into an external AI service.
The intention is reasonable. The engineer wants to solve the problem faster.
But several important questions arise.
Has the customer permitted those documents to be shared with an external provider? Does the supplier own the engineering information, or merely have the right to use it for manufacturing?
Where will the uploaded data be processed and stored? Who can access it? How long will copies remain in logs or backups? Can any of it be used to train or improve models?
These are separate issues, and manufacturers need to understand the difference.
| Area of concern | What manufacturers need to establish |
|---|---|
| Intellectual property ownership | Who holds the rights to the information? |
| Confidentiality | Who may access, use or disclose the information? |
| Data retention | How long will the information remain stored? |
| Model training | Can the information be used to develop or improve AI models? |
| Contractual restrictions | What uses or transfers have been permitted by the customer? |
Even where model training is contractually excluded, confidentiality and data-handling responsibilities remain.
The risks are particularly relevant when dealing with customer-owned designs, defence-related technical information or manufacturing processes developed under restrictive agreements.
This does not mean that hosted AI services are inherently unsafe.
Enterprise offerings may provide contractual protections, access controls and configurable data-handling policies. These vary by provider, product, subscription and configuration, and must be verified before use.
Equally, an AI model running within the manufacturer’s premises is not automatically secure.
The real question is whether the manufacturer has sufficient control over what information the AI solution can access, how it processes that information and where the information ultimately goes.
Where Should Manufacturing AI Run?

This brings us to a technology decision that is often overlooked during early AI discussions.
Should manufacturing AI operate through a hosted service, a private cloud, within the factory itself or through a combination of these environments?
There is no single answer.
1. Hosted AI Services
A hosted AI service can provide access to advanced models without requiring the manufacturer to maintain the underlying model infrastructure.
This may be practical for approved business information, document processing or selected enterprise applications.
However, the organisation must understand the service’s contractual protections, data locations, retention policies, external dependencies and operating costs.
2. Private Cloud AI
A private-cloud deployment can provide a more controlled or dedicated environment, depending on its design.
It may suit organisations requiring particular infrastructure, access and network arrangements.
However, private-cloud deployment alone does not establish where every external service call or stored record will reside.
3. On-Premises Manufacturing AI
An on-premises deployment allows a manufacturer to run suitable AI models and associated applications on infrastructure within its own controlled environment.
With local inference, approved information can be processed by a locally hosted model without sending prompts or underlying documents to an external model service.
For plants with restricted internet connectivity, sensitive process information or customer-imposed data-transfer limits, this may be essential.
However, the solution must genuinely support this arrangement.
A system described as on-premises may still depend on external services for authentication, monitoring, updates or other functions.
Manufacturers must establish whether these connections can be disabled or controlled without affecting the required functionality.
Local deployment also brings responsibilities, including:
- Suitable computing infrastructure.
- Model maintenance and software patching.
- Access management and audit logging.
- Secure backups.
- Ongoing performance and security monitoring.
- Network segmentation between business and industrial systems.
4. Hybrid AI
A hybrid arrangement may be appropriate for some organisations.
Restricted engineering data, plant records and sensitive calculations can remain within a customer-controlled environment, while approved non-sensitive activities use hosted AI services.
The key is to define which information may cross that boundary and enforce the restrictions technically, rather than relying only on employee instructions.
The choice should follow the manufacturer’s risk profile, customer agreements, infrastructure capabilities, performance requirements and total cost of ownership.
On-premises does not automatically mean secure, and cloud does not automatically mean unsafe.
Security depends on how the complete solution is designed, configured, operated and governed.
Manufacturing Needs Process Intelligence, Not Just Document Intelligence

Consider another hypothetical situation, this time in a foundry.
The quality manager notices an increase in porosity and shrinkage-related defects in a particular casting.
Rejections are rising, and the production team wants to understand what has changed.
A general-purpose AI assistant can explain possible causes. It may discuss gas entrapment, melting practices, pouring temperatures, feeding arrangements, mould characteristics and cooling behaviour.
That information can help engineers structure an investigation.
But it does not explain why defects are increasing in this particular foundry.
For that, the solution needs to understand the manufacturing process through actual plant information.
The investigation may require batch chemistry, melting and pouring temperatures, moulding conditions, cooling records, inspection findings and previous corrective actions.
These records must be matched to the correct product, production batch, equipment and time period.
Step 1: Classify the Defects Correctly
Porosity and shrinkage-related defects can arise through different mechanisms, and treating every rejection as the same problem could lead the investigation in the wrong direction.
Reliable analysis therefore begins with accurate defect classification.
Step 2: Identify Recurring Patterns
The system can then look for patterns across production batches.
For example, a suitable analytical approach might examine whether certain defect categories occur more frequently under particular combinations of material composition, pouring conditions and cooling behaviour.
Step 3: Perform Multivariate Analysis
Manufacturing outcomes are rarely influenced by only one parameter.
Several variables may interact, and examining them independently can overlook important relationships.
Depending on the available data, statistical methods or appropriately validated machine-learning models could help identify combinations of conditions associated with increased defects.
Step 4: Investigate Potential Root Causes
An association is not proof of causation.
The process engineer must examine alternative explanations, consider engineering knowledge and determine whether the suspected factors are credible.
The findings may narrow the investigation, but confirming a root cause requires supporting evidence and appropriate validation.
Step 5: Recommend Next-Batch Adjustments
If the analysis supports a change, the solution could help propose adjustments to selected next-batch parameters, such as permissible pouring conditions or other approved process settings.
These would remain recommendations until reviewed by the responsible engineer.
Any adjustment must stay within established tolerances, customer specifications, equipment limits and approved procedures.
Step 6: Monitor the Results
The change should be documented, and the subsequent batch inspected to assess whether the expected improvement occurred.
Results from later batches could then be incorporated into future analysis, subject to data-quality checks and model validation, allowing alerts and recommendations to be refined over time.
The complete process would follow this sequence:
Batch and Process Records → Defect Classification → Pattern Detection → Multivariate Analysis → Engineering Investigation → Approved Next-Batch Adjustment → Outcome Monitoring
A purpose-built solution may still use an LLM to interpret questions, retrieve supporting records and explain analytical findings.
But the underlying intelligence must come from plant-specific information, relevant analytical methods and validated engineering knowledge.
Knowing what can cause a defect is different from understanding what may be causing it in your own manufacturing process.
An Intelligent Recommendation Is Not Yet a Business Decision

Let us return to the additional 5,000-unit order.
A standalone assistant, working with incomplete information, may produce a plausible production plan.
A manufacturing AI ecosystem should be capable of assessing the request against actual operating conditions and business constraints, provided the required information and analytical capabilities are available.
It would first establish whether the required material is available.
This involves more than checking inventory quantities. Some stock may already be reserved for existing orders, may not meet the required specification or may still be awaiting quality clearance.
The solution would then evaluate usable machine capacity, considering current loading, planned maintenance, tooling availability, labour requirements and whether alternative production lines are qualified for the product.
It must also assess existing customer commitments, since accommodating one urgent order should not silently create delays elsewhere.
The commercial consequences are equally important.
Additional shifts or overtime may increase production costs. Expedited material procurement may affect margins. Higher inventory or earlier purchases may increase working-capital requirements, while a revised schedule may involve contractual delivery obligations.
A properly designed solution could bring these factors together and present alternatives with their assumptions, constraints and estimated consequences.
| Business function | Questions the AI ecosystem should help evaluate |
|---|---|
| Sales | Can the customer’s requirement be fulfilled? |
| Production | Is sufficient qualified production capacity available? |
| Procurement | Can the required materials be sourced on time? |
| Maintenance | Will planned maintenance affect production availability? |
| Quality | Are materials, processes and alternative lines approved? |
| Finance | What is the impact on margins and working capital? |
| Planning | Will existing customer commitments be affected? |
The manufacturer might be able to deliver part of the order within ten days and the balance later.
Alternatively, the full quantity might be feasible through approved overtime or rescheduling, provided materials, qualified capacity and other commitments permit it.
A revised delivery date may ultimately prove commercially preferable.
The system should also identify situations where information is incomplete and a reliable recommendation cannot yet be made.
The responsible production, planning and commercial managers would evaluate the alternatives and approve an appropriate course of action.
Production planning may then revise the schedule, procurement may expedite materials, and customer service may communicate the confirmed delivery arrangement.
Each activity should have an assigned owner, an authorised workflow and a means of monitoring completion.
The process should not stop when AI generates a recommendation.
Instead, it should support the complete decision cycle:
Customer Requirement → Verified Operational Data → Material, Capacity and Financial Validation → Feasible Alternatives → Management Approval → Assigned Actions → Controlled Execution → Outcome Monitoring
If operating conditions change, the system should identify the exception and bring it back for review rather than continuing to rely on outdated assumptions.
Operational Safety Cannot Be Compromised
There is also a boundary that must be respected.
An AI solution may recommend operational changes, but it should not directly control safety-critical equipment through unrestricted LLM instructions.
Industrial equipment requires validated control systems, appropriate safety interlocks and authorised operating procedures.
Any AI-assisted control functionality must be engineered, tested and governed according to the risks involved.
The National Institute of Standards and Technology (NIST), in its Guide to Operational Technology Security, highlights the need to protect industrial systems while recognising their requirements for performance, reliability and safety.
Getting an answer is one thing. Making an informed decision is another.
Executing that decision safely and responsibly is where the real value and accountability lie.
General-Purpose AI Versus a Manufacturing AI Ecosystem
For manufacturing leadership, the distinction between using an AI assistant and investing in a manufacturing AI ecosystem is important.
Both have value, but they are designed to address different needs.
A general-purpose AI assistant can improve how people access and work with information.
A manufacturing AI ecosystem brings together the data, analytical capabilities and operational controls required to support manufacturing decisions.
| Dimension | General-Purpose AI Assistant | Manufacturing AI Ecosystem |
|---|---|---|
| Business context | Works with supplied information and configured tools | Connects relevant enterprise and shop-floor information |
| Process intelligence | Explains problems and interprets available information | Evaluates plant-specific processes and operational constraints |
| Technology | Primarily LLMs and associated tools | LLMs, business rules, predictive analytics, optimisation and industrial integration |
| Data protection | Depends on product, subscription and configuration | Designed around industrial IP, customer restrictions and security requirements |
| Deployment | Commonly accessed through hosted services | On-premises, private cloud, hosted or hybrid deployment |
| Decision-making | Suggests possible approaches | Evaluates feasible options against validated operational and commercial information |
| Execution | Depends on available tools and permissions | Authorised workflows, assigned responsibilities and controlled execution |
| Business value | Faster reporting, information retrieval and administrative tasks | Reduced downtime, lower rejections, better delivery reliability and improved operational performance |
The distinction is in how the complete solution is designed, not simply which AI model is used.
GPT, Claude and other LLMs can be important components of a manufacturing AI ecosystem.
This does not mean every manufacturer needs to invest in a large AI platform or develop its own language model.
A smaller manufacturer may begin with a focused solution connecting existing ERP information, production records and established business rules.
A larger operation may require specialised analytical models, plant-level integration and additional controls.
What matters is not the size or sophistication of the technology, but whether it is designed to solve the problem reliably and deliver measurable business value.
What Should Manufacturing Leaders Ask Before Investing in AI?
Deloitte’s 2026 Manufacturing Industry Outlook highlights cost, talent, data, technology, governance and workflow transformation as considerations for manufacturers moving towards wider AI deployment.
Its 2025 Smart Manufacturing and Operations Survey also reports concerns about unauthorised access, intellectual property theft and operational disruption in industrial technology environments.
These findings reinforce why management needs to look beyond an impressive technology demonstration.
Before approving an AI investment, manufacturing leaders should evaluate the following areas.
1. Intellectual Property and Data Protection
Where will engineering information, prompts, generated responses, activity logs and backups reside?
Who may access them, how long will they be retained and what contractual restrictions apply?
Can external model calls, monitoring connections and data transfers be disabled or restricted when necessary?
2. Compatibility With Existing Infrastructure
Can the solution integrate with the manufacturer’s existing ERP applications, production systems, quality databases and legacy equipment?
How will missing or inconsistent information be handled?
Does implementation require major infrastructure replacement or can the organisation adopt the technology progressively?
3. Reliability of Recommendations
Which calculations, business rules, engineering constraints and analytical methods support the findings?
How does the system respond when information is incomplete or contradictory?
Can the recommendation be traced to the records and assumptions behind it?
4. Accountability and Governance
Who has the authority to approve a recommendation, initiate a transaction or change an operating procedure?
How will actions be recorded, exceptions escalated and results monitored?
5. Measurable Business Value
What measurable improvement should the organisation expect?
This might include:
- Faster operational decisions.
- Reduced machine downtime.
- Lower rejection rates.
- Better delivery performance.
- Reduced inventory.
- Improved profitability.
These outcomes should be agreed before implementation and compared with an established baseline.
A convincing demonstration is not evidence of a successful implementation.
How Manufacturers Should Begin Their AI Journey
Manufacturers should begin with one well-defined operational problem, establish how the process performs today and identify the information and controls required to improve it.
A pilot should then be tested using representative operating conditions, including exceptions, missing information and situations where human intervention is required.
Only when the results demonstrate acceptable reliability, security and business value should the implementation be expanded.
The approach is straightforward:
- Define the business problem.
- Establish the existing performance baseline.
- Assess data, intellectual property and operational risks.
- Select the appropriate technology and deployment architecture.
- Conduct a controlled pilot.
- Validate operational and commercial outcomes.
- Scale the implementation responsibly.
This applies to both large manufacturers and Micro, Small and Medium Enterprises (MSMEs).
Their scale and investment may differ, but the discipline should remain the same.
Technology must be selected because it addresses a business requirement, not because a particular AI model is attracting attention.
Beyond the LLM: The Future of AI in Manufacturing
GPT, Claude, Gemini and other LLMs have changed how people interact with technology.
They have made advanced AI capabilities more accessible and opened possibilities for improving the way businesses work.
Manufacturing can benefit significantly from these developments.
But a factory is not simply a collection of documents or transactions.
It is a network of machines, materials, people, processes, engineering knowledge, costs and customer commitments.
A general-purpose AI assistant can help employees access information, explore possibilities and work more efficiently.
A dependable manufacturing AI solution must go further.
It must work with verified operational information, respect industrial and commercial constraints, protect proprietary knowledge and support decisions that can be validated and acted upon responsibly.
That does not mean every manufacturer needs to develop its own LLM or invest in an expensive AI platform.
Nor does every manufacturing problem require AI.
Sometimes better data integration, a validated business rule or conventional optimisation may provide the most appropriate answer.
The real opportunity lies in bringing the right capabilities together around the problem the manufacturer needs to solve.
Manufacturing does not need just another chatbot or an AI model that produces impressive answers.
It needs intelligence designed around its machines, materials, people, processes and intellectual property, with the ability to support reliable decisions and responsible action.
Ultimately, the measure of successful AI adoption is not how intelligently the system responds, but how effectively the manufacturing business performs.
References and Further Reading
- Deloitte Insights — 2026 Manufacturing Industry Outlook
https://www.deloitte.com/global/en/insights/industry/manufacturing-industrial-products/manufacturing-industry-outlook/2026.html - Deloitte Insights — 2025 Smart Manufacturing and Operations Survey
https://www.deloitte.com/us/en/insights/industry/manufacturing-industrial-products/2025-smart-manufacturing-survey.html - National Institute of Standards and Technology (NIST) — Guide to Operational Technology (OT) Security, SP 800-82 Revision 3, 2023.
https://csrc.nist.gov/pubs/sp/800/82/r3/final
About the Author
Dr. Rajendra Shekatkar is a contributing author to MFG Magazine, and Executive Director at Prodware Solutions, with over four decades of industry experience spanning manufacturing, supply chain management, enterprise technology and digital transformation. Having started his career in manufacturing and supply chain operations, he brings a practical business perspective to the adoption of emerging technologies. He holds a Doctor of Business Administration (DBA) in Generative AI from Golden Gate University, USA. His work focuses on applying AI within enterprise systems and business processes to improve operational efficiency, decision-making and business outcomes.
Connect With MFG Magazine
Are you a manufacturing business evaluating artificial intelligence for your operations? Are you facing challenges related to legacy system integration, data confidentiality, on-premises deployment or identifying the right AI use cases?
MFG Magazine welcomes industry experiences, practical questions and discussions around manufacturing technology adoption.
For editorial contributions, industry interactions and networking enquiries, contact our team at ask@mfgmagazine.com.
MFG Magazine — Manufacturing Intelligence for Industry Leaders.
Editorial disclosure: The industrial scenarios presented in this article are illustrative examples, not independently documented implementation case studies.



