Why the same nonconformities keep coming back, and what AI changes
Why the same nonconformities keep coming back despite 8D and CAPA, and how AI surfaces the recurrences that case-by-case handling never sees.
Open your nonconformity database for the last three years. You will find, almost certainly, the same defect described ten times, in ten different wordings, handled ten times as if it were new.
This is not a lack of rigour. It is a problem of vision: when you handle nonconformities one by one, the repetition stays invisible.
This is exactly the terrain of recurring nonconformities, the ones that each record closes out cleanly but that nobody ever connects to one another.
The essentials
An isolated nonconformity is handled well. A recurrence, though, cannot be seen case by case, because nobody is tasked with rereading a thousand NCs together. It is the shared root cause, never looked at as a whole, that makes the defect come back. AI makes that recurrence visible in a few hours; the root cause and the action remain a human decision.
Why do the same nonconformities really keep coming back?
Three reasons stack up, and none of them is fixed by more goodwill. They stem from the way a nonconformity is born, handled and closed.
Free text. A NC is described in words, and two people do not describe the same defect the same way.
"Excessive play in the bearing", "gearbox vibration", "abnormal noise at the outlet" may point to the same origin. Sorting by category does not bring them together: they fall into three different boxes, and the recurrence gets lost between the labels.
Case-by-case handling. Each NC has its own cycle, its owner, its deadline.
That is necessary in order to resolve it, but the exercise of rereading the whole year exists in no job description. There is no time for it, and the task is thankless. Nobody is paid to do the nonconformity analysis as a whole.
The 8Ds left dormant. The corrective action is described, but its effectiveness check, the step always pushed back, is rarely closed out.
We correct, we do not verify that it held. Six months later, nobody knows whether the action served any purpose. The 8D of a nonconformity is opened, documented, then abandoned before its last step.
- Believing that "every NC handled" equals "problem solved".Closing a thousand records cleanly says nothing about the common cause feeding them. The recurrence of quality defects lives precisely in the aggregate, not in the individual record.
- Sorting by category and stopping there.A category groups labels, not meanings. Two descriptions of the same defect, worded differently, will stay separated forever.
- Closing without checking effectiveness.The CAPA of a nonconformity is worth something only if its last step, the proof that the defect does not return, is genuinely closed out.
What does AI see that you do not?
This is precisely the kind of task where artificial intelligence brings something real, provided you understand what it does.
A language model turns each nonconformity, free text included, into a numerical representation of its meaning. It can then group those that talk about the same thing, even written differently, across the full depth of your history.
Where you saw a thousand tickets, it brings out forty families, and shows you that three of them account for half the volume. It can also reread the 8Ds and flag those whose corrective action was never verified.
The contrast with the classic approach fits in one table:
| What changes | Case-by-case handling | AI analysis at scale |
|---|---|---|
| Unit of work | One NC at a time | The whole history at once |
| Grouping | By predefined category | By meaning, free text included |
| What you see | A thousand separate tickets | Forty families, three of which account for half the volume |
| Recurrence | Diluted, invisible | Quantified and ranked |
| Time | Days of manual rereading | A few hours |
Table scrolls horizontally on small screens.
What used to take days of manual rereading, when someone found the courage to take it on, is done in a few hours.
The gain is not handling each NC faster. It is finally seeing, at scale, what repeats. That is the whole value of AI in quality control: it does not replace the act of handling, it makes the pattern visible.
A diluted recurrence, not a hidden one
A site logs around eight hundred nonconformities a year, spread across four workshops and two quality teams. Each is handled properly, on time. Yet the same sealing defects come back every quarter.
Rereading the history by grouping on meaning rather than by category, it emerges that a third of one workshop's NCs all trace back to the same gasket-and-torque-procedure pairing, described each time with different words. No corrective action had treated it as a single problem, because none had seen it as such. The recurrence was not hidden: it was diluted.
What must AI absolutely not do here?
One point is non-negotiable. AI groups, it does not conclude.
It shows you that these forty NCs form a family; it does not tell you what their root cause is, and it closes nothing. The root cause of a recurrence is found with someone who knows the process, not with a system that knows the words. The grouping is a hypothesis to investigate, not a verdict.
This is where responsible AI is decided, and it is not a moral afterthought, it is the mechanism itself. The dividing line is clear:
| AI prepares the decision | The human makes the decision |
|---|---|
| Group the NCs by meaning | Name the root cause |
| Quantify the recurrences | Decide the corrective action |
| Flag the 8Ds never verified | Validate effectiveness and close out |
| Propose families to investigate | Rule on conformity |
Table scrolls horizontally on small screens.
AI prepares the decision: here are the recurrences, here are the actions never verified. The human makes it: here is the cause, here is what we do.
An AI that closed out nonconformities or ruled on conformity would cross exactly the line a quality system cannot allow to be crossed. Accountability, regulatory as much as moral, stays attached to a person.
Where to start, without launching a big project?
Good news: you already have the raw material. It is your nonconformity export, just as it comes out of your tool.
No new software to deploy, no project to budget first. The first step plays out on what you already own.
Start from the existing export
Take your NC database as it comes out of your quality tool, free text included. That is the starting point, not a prerequisite to build.
Test on an already-known defect
Pick a defect you already know recurs, and check that the analysis finds it and quantifies it correctly. If it sees what you knew, you can trust it for what you did not.
Work locally
Do it on your own machine, without your nonquality data leaving your walls. For quality data, this is the only reasonable approach.
The right first step is modest and verifiable. It calls for no budget, no rollout, no promise kept in advance: just a check on a case you already master.
Do you need new technology, or a building block you already know?
Grouping texts by their meaning is the same capability that lets you read a report and extract what matters, a subject already covered in turning technical documents into usable data.
Nothing quality-specific in that: a nonconformity is one text among others, and industrial AI knows how to handle it. What makes the quality cluster credible is precisely that it does not call for new technology, but for a skill already installed elsewhere.
What stays specific, on the other hand, is the weight of accountability.
The same requirement set out for the use of AI on inspection and maintenance data applies here, in a stricter form: a nonconformity wrongly closed by an automatism costs more than ten nonconformities handled slowly by hand.
That is why the boundary sits not at the level of what AI can do, but of what we allow it to decide. Grouping, yes. Closing out, never.
The same mechanism recurs elsewhere in quality:
- on the visual inspection that depends on the operator,
- on the cost of poor quality that goes unmeasured,
- on the preparation of an audit that takes two weeks.
Each time, AI surfaces what scale made invisible, and the human keeps the decision.
What to take away?
Recurring poor quality is not a fate born of insufficient rigour. It is a blind spot of scale: too many nonconformities, handled too separately, for a human eye to see the pattern.
AI makes the pattern visible, in hours rather than days. What you do with it, and the accountability for the decision, stay entirely on your side.
Can AI find the root cause of a nonconformity?
No, and you should not ask it to. It groups the nonconformities that describe the same problem and brings out the recurrences. The root cause is determined with someone who knows the process. The grouping is a starting hypothesis, not a conclusion.
Do you need a big piece of software to analyse your nonconformities with AI?
No. The starting point is your existing NC export. You can begin on a raw history, locally, and check the result on an already-known defect before going further.
Our nonconformities are described in free text, is that a problem?
On the contrary, that is where AI helps most. A language model groups by meaning, not by exact keywords, so it brings together different descriptions of the same defect, something a sort by category never does.
Does AI replace the quality manager?
No. It gives back the time spent rereading hundreds of records so it can go into deciding and acting. The conformity decision, the closing out of a NC and the validation of an action stay human.
Written by Adama CamaraAI Consultant · Industry · view profile
Published on July 31, 2026
Software and data
Industrial AI: What It Really Is, and What It Changes on the Plant Floor
Industrial AI without the jargon: use cases by function and by sector, limits, costs, financing and a roadmap to get started in your plant.
AI and maintenance
Can AI Replace the Inspector? The Line Between Preparing and Deciding
What AI can genuinely do with inspection and maintenance data, what it must never decide on its own, and why the sign-off stays human.
Software and data
OCR or AI for reading your technical documents: three things people confuse before they buy
OCR, extraction, AI: three notions people confuse before buying. What each one really does to a technical report, and which one costs you.
AI and maintenance
Why responsible industrial AI is local, bespoke and off the cloud
Responsible AI in industry is not a label but an architecture: protecting sensitive data calls for bespoke, on-premise tools that stay off the cloud.