Non-conformance triage: a worked manufacturing example

A familiar quality problem

Consider a manufacturer that discovers a dimensional discrepancy during inspection. A batch of components has been produced, and one measured feature falls outside the specified tolerance. The immediate question is what should happen next. Depending on the circumstances, the material might need to be quarantined, reworked, scrapped or considered for a formally authorised concession. The decision could affect delivery commitments, production capacity, cost and customer confidence. It is the sort of situation in which an experienced quality engineer may need to gather information from several places before reaching a view.

This is an illustrative example rather than a report of a completed implementation. I am using it to examine where AI might make a practical contribution and where the limits of that contribution should be drawn. The aim is not to design a system that decides whether non-conforming material can be released. It is to explore whether the evidence required for a sound decision can be assembled more efficiently, and whether doing so would help the people who remain responsible for the outcome.

Following the incident through the business

The inspector records the measured deviation and identifies the affected batch. A quality engineer then needs to establish which drawing and revision apply, whether the measurement is reliable, how much material may be affected and whether the same issue has occurred before. Relevant information might be held in inspection records, production documentation, previous non-conformance reports, emails, customer specifications and a quality management system. Some of it may be structured, while other details exist only in documents or notes written during earlier investigations.

The engineer is unlikely to treat these sources as equally reliable. The current customer specification may carry more authority than an old internal note. A previous investigation may describe a similar symptom without establishing that the cause was the same. A record of corrective action may show what was proposed but not whether it was completed or effective. These distinctions are central to the work. An apparently relevant document is not necessarily sufficient evidence for the decision now being considered.

There may also be practical delays. The engineer could be waiting for a revised measurement, clarification from production or confirmation of the customer’s requirements. The non-conformance may therefore spend considerably longer in the system than the time required to read and assess the available evidence. If we want to improve the workflow, we need to understand both the information-gathering effort and the delays created by handovers and unresolved questions.

What an AI assistant might do

A limited AI-assisted triage tool could begin by helping the engineer assemble a case file. Given a non-conformance identifier and access to approved information sources, it might retrieve the relevant inspection record, identify the applicable specification, locate previous incidents involving the same component or characteristic and produce a structured summary of what the records show. It could flag apparent conflicts, missing measurements or references to earlier corrective actions whose outcomes are not recorded. Each significant observation should be traceable to the source from which it was derived.

The value would be in reducing the effort required to establish the background to the problem. An engineer might otherwise search several systems, open numerous files and reconstruct the history from fragments of information. The AI could help with that preparation, particularly where useful evidence is buried in unstructured text. It might also prompt a useful question, such as whether the measurement was taken using the current inspection method or whether a previous incident involved a different material supplier.

Asking the system for a single confident conclusion about disposition would be a mistake. The same dimensional deviation could have very different implications depending on the component’s function, the assembly into which it fits, the customer’s requirements and the potential consequences of failure. A superficially similar previous case might be a poor precedent. The system could identify relevant comparisons and help make the differences visible, but the decision about what those comparisons mean belongs with the authorised people in the business.

Where the judgement remains

The engineer must still establish the technical significance of the deviation. They may need to consult design authority, production, purchasing or the customer. If a concession is possible, it may require formal approval under the manufacturer’s quality arrangements and contractual obligations. The AI’s role would be to support this process with accessible evidence, not to bypass the people or controls that give the decision legitimacy.

A well-designed triage output would therefore distinguish between established facts, possible interpretations and unresolved questions. It might say that three earlier reports mention the same component and characteristic, but only one records a confirmed root cause. It might identify that a drawing revision appears inconsistent between two records and recommend verification before proceeding. Such statements are less dramatic than an automated disposition recommendation, but they are more useful if they help the engineer avoid a mistaken assumption.

The organisation should also be able to see what the system did not examine. If historical reports are held on a network drive that the tool cannot access, or if an older quality database is outside the connected information sources, the output should not imply that a comprehensive search has been completed. An apparently empty search result is ambiguous unless the user understands the scope of the search. This is one reason why information architecture and system integration matter as much as the language model chosen for the task.

Testing the idea on real cases

Before investing in a substantial system, the sensible test is a small set of completed non-conformances whose records experienced staff can examine. The purpose would be to see whether the assistant can locate the right evidence, represent it accurately, recognise gaps and avoid implying conclusions that the records do not support. The cases should include straightforward incidents as well as awkward ones involving conflicting documents, incomplete investigations or misleading similarities to earlier problems.

We could then compare the AI-assisted preparation with the existing way of working. How long does it take to assemble the case information? How many relevant records are missed? Does the output help the engineer identify the questions that still need answering? Are source references reliable enough to inspect? Does the assistance reduce administrative effort without encouraging the reviewer to accept an unsupported recommendation? These questions are more informative than whether the model produces a fluent summary.

There is no measured performance result to report for this illustrative design. Whether it saves time or improves consistency would depend on the quality and accessibility of the manufacturer’s records, the range of incidents encountered and the way engineers use the output. It might turn out that the greatest improvement comes from standardising how incidents are recorded or making existing quality information easier to find, with AI playing a relatively small role. That would still be a worthwhile result from the investigation.

The wider lesson

Non-conformance triage is a useful example because it contains several different forms of work within one familiar activity. Some information can be captured and checked through straightforward rules. Some evidence may be difficult to locate and compare without assistance. Other questions require engineering judgement, authority and an understanding of consequences. Treating the entire process as a single automation opportunity would obscure these differences.

For me, success would mean helping the responsible people understand the case more quickly and with greater confidence, while preserving a clear account of the evidence and the decision. AI may have a useful part to play in that arrangement, but it should earn its place by improving the work. The starting point is to understand how a non-conformance is actually investigated, what makes some cases difficult and where better access to evidence would change the outcome.