Manufacturing AI Needs to Understand More Than the Data

By Harri Pihlajavesi, CEO | Predisys

When someone asks a question about manufacturing, the words alone don't always tell us what they need to know.

Suppose a process engineer, a factory manager and a VP of Quality all ask the same question:

What changed?

The process engineer may be looking at a shift in a particular characteristic and trying to determine which process parameters moved with it. The factory manager may be trying to understand whether production is at risk and what needs attention. The VP of Quality may be concerned that the same problem could be developing across several products, plants or suppliers.

They are working from the same manufacturing reality, but the context of the question is different.

People handle this naturally. We know something about who is asking, what they are responsible for and what is happening around them. If Manufacturing AI is going to become useful across an organization, it needs some of that context as well.

Manufacturing questions have context

A process engineer rarely starts an investigation by describing everything they already know.

They may ask why a dimension shifted on Line 3 because they are already looking at the control chart. They know which product is running, which characteristic changed and approximately when it happened.

Another engineer hearing the question may know much of this too.

An AI agent cannot assume that context unless the manufacturing environment provides it.

If the engineer is looking at a particular process, the system should be able to understand what product is running, which machine and tool are involved, what characteristics and parameters belong to the process, what has changed recently and what the process was doing before the current issue appeared.

That context narrows the question considerably.

Instead of searching everything the company knows about dimensional variation, the AI can begin with the manufacturing situation the engineer is actually investigating.

The same principle applies at other levels of the organization.

A factory manager asking, "What should I be worried about today?" is unlikely to want a list of every alarm generated during the last 24 hours. The useful answer depends on which processes are behaving unusually, whether production or quality is at risk, how significant the issues are and whether someone is already responding to them.

A quality executive asking about a supplier may be interested in something broader still. Are there unusual patterns across incoming materials, production processes or test results associated with that supplier? Are the same problems appearing at more than one plant? Is the situation getting better or worse?

The underlying information overlaps, but the useful answer depends on why the person is asking.

The manufacturing structure provides part of that context

Manufacturing data already contains many of the relationships needed to understand a question properly, provided those relationships are maintained.

A measurement belongs to a characteristic. The characteristic belongs to a product and process. The process runs on equipment and may involve a particular tool or cavity. Materials and components come from lots and suppliers. Process parameters describe conditions under which production occurred.

Time provides another connection because it lets us see what was happening around an event.

If a characteristic begins shifting at 10:42, we can look at which product was running, which equipment was involved, what material was being used and what happened to related parameters around 10:42.

We can also move in the other direction.

If a supplier lot is suspected, we can identify where that material was used, which products and processes were affected and whether any unusual statistical behavior appeared afterward.

The usefulness comes from being able to move through those relationships according to the question being investigated.

Statistical context helps determine what deserves attention

A manufacturing operation can generate an enormous number of events every day. Measurements change continuously, parameters move, alarms occur and production conditions change.

Most of those changes do not require an investigation.

Statistical understanding helps provide another layer of context by telling us whether what we are seeing is consistent with normal process behavior.

If a parameter changes slightly, the value itself may not be particularly interesting. If the change occurs at the same time a normally stable quality characteristic begins showing unusual behavior, it may deserve a closer look.

Likewise, a measurement can remain comfortably within specification while the process that produces it begins behaving differently.

This is one reason Manufacturing Intelligence needs both process context and statistical context. Knowing where a piece of data came from tells us what it belongs to. Understanding process behavior helps us determine whether what is happening is unusual enough to matter.

For an AI agent helping with an investigation, both are useful.

Understanding what changed is only part of the work

Once a meaningful change has been identified, manufacturing becomes operational very quickly.

Someone needs to know about it. Someone may need to investigate it. An action may need to be assigned, production may need to be held or a supplier may need to be contacted.

The details depend on the process and the organization, but the analysis usually leads somewhere.

Suppose SPC identifies an unusual shift in a process characteristic. An engineer investigates and finds that the shift began shortly after a tool change. Maintenance inspects the tool, an adjustment is made and the process returns to its previous behavior.

There is useful information throughout that sequence.

We have the original statistical change and the conditions around it. We have the investigation and what the engineer found. We know who became involved and what action was taken. Then we have the process behavior afterward, which helps us determine whether the action actually worked.

If those pieces remain connected, the organization has learned something that can be used the next time a similar situation occurs.

Analysis and operations inform each other

Manufacturing systems often separate analytical information from the operational activity surrounding it.

The analysis may happen in an SPC or analytics system. An alarm may be handled somewhere else. Corrective actions may live in the quality system, maintenance work in a CMMS and supplier communication in another application.

There are good reasons for specialized systems, but an investigation can cross several of them.

An engineer trying to understand a recurring process problem may need the statistical history, the previous alarm, notes from an investigation, a maintenance event and the action that was taken afterward.

The relationships between those records can be more valuable than any one record on its own.

They also continue to develop as people work on the problem.

An investigation adds knowledge. An action changes the process. The resulting measurements tell us whether the action had the expected effect. That result then becomes useful context for the next investigation.

This is how Manufacturing Intelligence can continue to develop during normal manufacturing operations.

AI agents will need to work within manufacturing operations

As AI agents become more capable, they will likely participate in more than analysis.

An agent may recognize an unusual statistical condition and bring together the process information most relevant to it. It may find similar events from previous production and show what was discovered at the time. It may help determine who normally owns that type of issue or whether an existing investigation is already open.

The people involved still need to know why the information was surfaced and where it came from.

This becomes especially important in regulated manufacturing and other environments where decisions and changes need to be traceable. If an AI agent contributes to an investigation, the evidence it used, the actions that followed and the resulting process behavior may all become part of the history.

Over time, this creates an interesting relationship between AI and Manufacturing Intelligence.

The agent uses the manufacturing knowledge that already exists to help with the current situation. The investigation and actions that follow create new manufacturing knowledge. Future investigations can then use what was learned.

That can happen without creating a separate AI knowledge base. The useful knowledge is being created through the work people are already doing.

Different roles can work from the same Manufacturing Intelligence

One of the more useful possibilities is allowing people at different levels of the organization to work with the same underlying manufacturing knowledge without requiring everyone to navigate it in the same way.

A process engineer might ask:

What changed when this characteristic shifted?

The answer may involve detailed measurements, process parameters, tooling information and previous statistical behavior.

A factory manager might ask:

Which processes need attention today?

Now the system needs to look across processes, identify meaningful conditions, understand their production impact and know whether someone is already responding.

A VP of Quality might ask:

Are we seeing this problem anywhere else?

The scope expands again. The relevant information may cross products, plants, equipment and suppliers.

Each question starts from a different point, but the answers depend on many of the same relationships underneath.

This becomes particularly useful when someone wants to continue asking questions.

If the VP sees an unusual pattern associated with three plants, the next question might be whether they share a supplier. From there, the question might move to which material lots were involved, which products used them and whether statistical changes appeared before any failures were reported.

A useful Manufacturing AI environment needs to be able to follow that investigation as the question changes.

Context makes AI more useful in manufacturing

Natural-language interfaces make manufacturing systems easier to question, but the quality of the answer will depend heavily on what the AI can understand about the environment behind the question.

It needs to know what the data belongs to, how the process normally behaves, what was happening when something changed and what the organization has learned from similar situations.

It also helps to understand who is asking and the level at which they are trying to make a decision.

When an investigation leads to action, the information created during that work becomes part of the manufacturing history as well. The alarm, ownership, investigation, action and resulting process behavior provide context that may matter months or years later.

Manufacturers already create much of this knowledge during normal operations. Connecting it gives an AI agent a better basis for helping an engineer investigate a process, helping a manager understand what needs attention or helping a quality leader recognize a pattern across the organization.

The question may begin with one person and one process, but the answer can draw on what the organization has learned across manufacturing.