Insight: Why data context, not algorithms, is the real key to industrial AI

Artificial intelligence is generating understandable excitement across manufacturing. Its potential applications include predictive maintenance, automated quality inspection, production scheduling, supply-chain optimisation and the identification of patterns that would be difficult for people to detect manually. However, operational AI adoption remains relatively limited. Nick Hallam from ELF Productivity looks at why industrial AI does not begin with an algorithm.

Research referenced in the UK manufacturing sector indicates that only a small proportion of manufacturers have embedded AI widely across their operations. Many businesses describing themselves as AI users are applying it to activities such as research, document summarisation and correspondence rather than production, maintenance, quality or operational planning.

There is an important difference between using an AI tool and operating an AI-enabled factory.

For manufacturers, the greatest obstacle may not be access to AI technology. It may be the condition of the operational information being supplied to it.

AI can process data, but it cannot create missing context

A typical manufacturing business may hold information across:

• ERP and MRP systems
• Time and attendance applications
• Machinery and PLCs
• Quality systems
• Maintenance platforms
• Financial applications
• Warehouse systems
• Spreadsheets
• Paper records
• Employee and supervisor knowledge

Each source contributes something valuable. The challenge is that those sources frequently describe the operation in different ways.

A job may have different identifiers across estimating, production and finance. A machine may record that it stopped, but not why. Labour may be recorded against a shift without being attributed accurately to a job or activity. Scrap may be declared without identifying the process condition that caused it.

An AI system may therefore receive a large volume of data while still lacking the information required to reach a reliable conclusion.

Consider a production order that finishes 15% above its expected labour cost.

AI might identify a correlation between that product type and poor performance. But the true cause could have been:

• A late material delivery
• A longer-than-normal machine set-up
• An experienced operator being reassigned
• Unrecorded rework
• Employees remaining booked to the job during downtime
• An unrealistic standard time
• Incorrect production quantities
• A quality problem attributed to the wrong order

The cost variance is visible, but the operational truth behind it may not be.

Manufacturers need four connected dimensions

Before AI can provide meaningful operational insight, manufacturers need to connect four areas of information.

What was planned
This includes the production order, routing, expected quantity, standard time, planned labour, machine allocation, target cost and anticipated margin.

What actually happened
This includes who was present, the hours worked, jobs and activities completed, machines used, quantities produced, downtime, waste, rework and actual completion time.

Why performance varied
The reason may have been a breakdown, material shortage, changeover delay, labour constraint, quality issue, planning change or customer amendment.

What the variation meant commercially
This includes additional labour cost, lost production, delivery risk, increased Work in Progress, cost variance and margin erosion.

Without these connections, AI may identify patterns.

With them, it can begin to support important operational decisions.

The first AI question should be about trust

Manufacturers should not begin by asking: What AI application should we buy? They should begin by asking: Can we trust the data describing what is happening inside our operation?

That means reviewing whether:

• Jobs are identified consistently
• Labour is attributed accurately
• Machine events have operational context
• Downtime reasons are meaningful
• Standards remain current
• Waste and rework are connected to their causes
• Planned and actual performance can be compared
• Operational variance can be translated into financial impact
• Industrial AI does not remove the need for operational discipline.

It increases it. A sophisticated model trained on incomplete or inconsistent information may provide an answer quickly, confidently and incorrectly. The manufacturers most likely to benefit from AI will not necessarily be those that acquire the technology first.

They will be those that first create a reliable understanding of what is happening now. Because industrial AI does not begin with the algorithm.