Seven data foundations manufacturers need before scaling AI

Manufacturers do not need a perfect data environment before they begin exploring artificial intelligence. They do, however, need a sufficiently reliable operational foundation. Without it, AI may amplify existing inconsistencies rather than improve decision-making. Nick Hallam from ELF Productivity shares seven practical foundations manufacturers should address before moving from experimentation to operational deployment.

1. Begin with an operational question

The starting point should not be: We need to use AI. It should be a defined business or production question.

For example:

• Can we predict which machines are most likely to fail?
• Can we identify jobs likely to exceed their labour budget?
• Can we forecast which production orders may miss their delivery date?
• Can we identify recurring causes of lost output?
• Can we improve labour allocation across lines and shifts?
• Can we predict the true cost of completing Work in Progress?

A clear question determines the data required and provides a measurable route to value. Without one, businesses risk collecting information and testing technology without improving a meaningful decision.

2. Create consistent master data

AI needs consistent definitions. The same employee, machine, product, job or activity should not be represented differently across multiple systems.

Manufacturers should establish common identifiers for areas including:

• Employees
• Jobs and production orders
• Products
• Operations
• Machines
• Departments
• Customers
• Cost centres
• Downtime reasons
• Waste and rework categories

This does not necessarily require every application to be replaced. It requires a reliable method of matching and connecting records between them. AI should not need to guess whether “Machine 12”, “CNC-12” and “Asset 0047” refer to the same equipment.

3. Attribute time and activity accurately

Manufacturing performance is closely connected to time.

AI needs to understand:

• Who was present
• When work started and stopped
• Which job consumed the time
• What activity was taking place
• Which machine or resource was used
• What was completed

Attendance information alone cannot explain how an employee’s working day was divided between production, set-up, maintenance, waiting, rework, training and indirect activity. Similarly, machine data may show that equipment was operating without identifying whether it was producing the correct item, achieving the required quality or running against the expected standard. Time, employee, activity, job and machine information need to be connected.

4. Capture reliable reasons behind events

Automated systems are generally effective at identifying that an event occurred. They may be less effective at explaining why.

A PLC can identify when a machine stopped. It may not know whether the operator was waiting for material, a quality check, a tool, the previous operation or a production decision. Human context remains essential and capturing that context should not involve lengthy administration. It should be quick, structured and relevant.

The machine provides the event, and the operational environment provides the explanation.

5. Connect business and factory systems

Industrial AI cannot operate effectively if information technology and operational technology remain separate. ERP, financial and workforce information needs to be connected with machinery, employee activity and production processes.

A connected operational layer should be capable of:

• Receiving information from existing systems
• Matching records through common identifiers
• Presenting relevant information to employees
• Capturing additional operational detail
• Applying consistent calculations
• Returning enriched information to approved systems
• Providing trusted data for analytics and AI

AI should not sit outside the business as an isolated experiment. It needs to be connected to the systems and processes delivering the work.

6. Establish ownership and governance

Operational data needs clear owners. Manufacturers should define:

• Who controls job and activity codes
• Who validates standards
• Who investigates missing information
• Who defines each metric
• Who approves calculation changes
• Who controls access
• Who determines which AI outputs require review
• Who monitors model performance over time

This becomes particularly important when AI-supported decisions influence quality, maintenance, customer delivery, workforce deployment, production priorities or safety. The greater the consequence, the stronger the governance required.

7. Design human oversight into the process

Industrial AI should support experienced employees, engineers, supervisors and managers rather than exclude them. AI may identify a statistical relationship without understanding:

• A temporary production change
• A new material supplier
• Engineering modifications
• Customer-specific requirements
• Planned maintenance
• A safety restriction
• An unusual but legitimate event
• Experienced people can distinguish between a meaningful signal and an understandable exception.

The question is not whether AI or a person should make the decision; it is how AI can provide better evidence to the person responsible for making it.

Build the foundation before increasing the ambition

Poor data does not simply reduce AI performance. It can create false optimisation, unreliable job profitability, misleading productivity comparisons and incorrect maintenance predictions. Manufacturers should therefore treat data readiness as part of the AI programme—not as a separate project to be considered later.

AI can process information more quickly than any management team, but the organisation must first ensure that the information reflects operational reality.