Non-Conformance Report With AI: Cause, Containment, Action

A good non-conformance report states what was wrong against what the specification required, what was done at once to contain it, what was decided about the affected parts, what the verified cause is, and what action will stop it returning. AI can tidy the facts and ask the five-whys questions. It must never supply a cause.

That last rule is the whole difficulty. An assistant is very good at producing a fluent root cause, and a fluent root cause that nobody checked is worse than a blank field, because it gets signed off. Use the assistant for structure, wording and gaps. Keep the evidence, the decisions and the signatures with people.

Follow me on Instagram@sagnikteaches

The structure of an NCR, section by section

Different firms use different forms, and any standard or customer you work to may require particular fields. The structure below covers what most small manufacturers need. Each section answers one question.

Connect on LinkedInSagnik Bhattacharya

Notice that containment comes before cause. When a defect is found, the first job is to stop more of it leaving the building. The investigation can wait a day; the quarantine cannot.

Subscribe on YouTube@sagnikteaches

A filled-in example: bracket holes out of position

Consider an illustrative six-person metal fabrication shop making a steel mounting bracket for a customer. During final inspection, an operator measures the distance between two holes on a sample of 50 from a batch of 400. Tolerance on the drawing is 62.0 mm, plus or minus 0.3. Ten of the 50 measure 62.8 mm.

The rough note left by the inspector reads: "holes wrong on the brackets from job 2217, about 10 of 50, too far apart, think it might be the new fixture". Here is how that note becomes a report.

Identification and description

NCR no.: NCR-2026-041
Date found: [date]    Found by: final inspector
Part: Mounting bracket, drawing rev C    Job: 2217    Batch size: 400
Requirement: hole-to-hole distance 62.0 mm +/- 0.3 mm (drawing, note 4)
Finding: 10 of 50 sampled parts measured 62.8 mm (0.5 mm outside the upper limit).
Method: calibrated digital calliper, ID [number], checked against gauge block before use.
Evidence: measurement sheet and photos attached.

The report states the requirement, the measurement and the method, and it keeps opinion out. "Think it might be the new fixture" is not in the description; it goes into the cause section as a hypothesis to test.

Containment: stopping the spread

Containment answers: where else could these parts be? List every location and what was done. For the brackets, the shop checks three places.

Those figures need to add up. The batch is 400: 160 + 120 + 120 = 400, which accounts for every part. The sample of 50 came from the finished goods group, where the full inspection later found 31 bad in 160, about 19 per cent, against 20 per cent in the sample (10 of 50). The full inspection is the number to rely on, and the sample only told you to look. The 30 earlier parts sit outside the 400, as a check on whether the fixture had been wrong before. An assistant can check these sums when you paste the table in, and it will flag if the locations do not add up to the batch size. That is the kind of check it does well.

Disposition: what happens to the parts

The common choices are use as is, rework, scrap, return to supplier, or regrade. For parts outside tolerance on a customer's drawing, "use as is" needs the customer's written agreement, not your own judgement. In this example the bracket holes can be re-reamed, so the decision is rework for the 69 bad parts found so far (31 + 38) if the customer agrees, with a full re-inspection afterwards. The quality owner signs the decision, and the customer's reply goes in the file.

A quick sum shows why the note matters. Rework takes 12 minutes a part, so 69 parts is 828 minutes, or 13.8 hours. At a loaded labour cost of $40 an hour, that is $552, plus the cost of the customer's inspection time and any delay. Recording it makes the cost of the fault visible, which helps when you justify the corrective action.

Using AI to tidy the report without changing the facts

The first helpful job is turning the inspector's rough note into the fields above. The second is to check the report for gaps before the quality owner sees it. Both prompts carry the same rule: no facts that are not in the input.

Below is a rough inspector's note and measurement data from a non-conformance. Fill the NCR fields: identification, description (requirement, finding, method), and evidence list. Use ONLY facts in the input. Where something is missing, write "NOT PROVIDED" and list it at the end as a question for the inspector. Do not state or imply any cause. Remove opinions such as "think it might be" from the description and list them separately under "Inspector's suggestion (unverified)".

Input:
[PASTE NOTE AND MEASUREMENTS]

A realistic output gives you the clean description, a short list of questions such as "Which fixture was used for the second operation?" and "Which measuring instrument ID?", and a separate line for "New fixture (unverified)". That separation is the safeguard: the suggestion is recorded, but not as a cause.

Five-whys prompts that ask, not answer

The five-whys method asks "why" repeatedly until you reach a cause you can fix. It goes wrong when each answer is a guess. The assistant should act as the facilitator, asking the next question and asking for the evidence behind each answer, while your team supplies the answers.

Act as a facilitator for a five-whys analysis of a manufacturing non-conformance. I will give you the problem statement and then answer one question at a time.
Rules:
1. Ask ONE "why" question at a time and wait for my answer.
2. After each answer, ask "What evidence shows that?" and record it. If I say there is none, mark that answer "UNVERIFIED" and suggest a quick check that would test it (a measurement, a record, an inspection).
3. Never supply an answer yourself, never suggest a cause, and never skip to a conclusion.
4. Stop when I say we have reached a cause that is within our control, then summarise the chain with each link's evidence status.

Problem statement: 10 of 50 mounting brackets from job 2217 measured 62.8 mm hole-to-hole against 62.0 mm +/- 0.3 mm.

An illustrative session might run like this.

1. Why were the holes 0.8 mm further apart than the nominal 62.0 mm? Team: the second drilling position was offset. Evidence? The fixture stop was measured 0.8 mm further out than the setup sheet. Verified.
2. Why was the stop 0.8 mm out? Team: it was re-seated after the fixture was cleaned and no one checked it. Evidence? The operator says so; there is no record. UNVERIFIED, check the maintenance log and ask the cleaner.
3. Why was there no check after re-seating? Team: the setup sheet has no step for first-off inspection after cleaning. Evidence? The sheet was read and has no such step. Verified.

The chain stops at a cause the shop controls: the setup sheet lacks a first-off check after the fixture is cleaned. The unverified link is flagged, and a person goes to find the record. That is better than a smooth five-line story where every link looks equally sound.

For more prompts of this kind, the root cause analysis prompts for recurring problems cover alternative methods, and the guide to spotting patterns in complaints and defects helps when you suspect the same fault has appeared before.

Corrective and preventive action

The fix should match the cause, not the symptom. "Retrain the operator" is the weakest corrective action because it relies on memory, and it is the one that tends to appear when no real cause was found. Stronger fixes change the process so the error is hard to make.

For the bracket, one illustrative set of actions looks like this.

The third row is easy to forget. If a cause is in a process, ask where else the same process exists. The assistant can draft the list of other fixtures from your equipment register, but you still check each one.

Verification and closure: did it work?

An NCR is not closed when the action is done. It closes when evidence shows the problem has not returned. Agree the check at the time the action is set. For the brackets: the next three setups show first-off measurements within tolerance, and the next two job batches show no hole-position defects at final inspection. Put the date of each check on the report, and have someone other than the action owner confirm it.

Ask the assistant to produce a closure checklist from your own report, and to list anything open.

Here is a completed NCR. List, as a checklist, every section that is empty, every action without an owner or date, every cause marked UNVERIFIED, and every statement in the description that is not backed by an item in the evidence list. Do not fix anything or suggest new causes.

That is an audit by a careful reader. It will not understand your process, but it will see the box that was left blank. It is also worth using before an external audit, where an unfinished NCR is a weakness.

When the fault came from a supplier

If the defect arrived in goods from outside, the NCR still starts the same way: describe, contain, decide. The cause section then points outward. Do not write the supplier's cause for them. Record what you found, ask them to investigate, and attach their reply when it comes. The guide to drafting a supplier corrective action request shows how to word that request, and the supplier audit checklist helps you decide whether a visit is warranted when problems recur.

An assistant is useful here for one more job: comparing the supplier's reply with your findings. Paste both in and ask it to list claims in the reply that your evidence does not support, such as "the parts left our factory within tolerance" when your measurements say otherwise. Treat that list as questions to put back to the supplier, not conclusions.

A second report: a process fault with no bad parts

Not every NCR is about a measurement. Take an illustrative case where a packer ships an order of 24 bracket kits with the wrong revision of the instruction sheet. Nothing is wrong with the brackets, and no one is hurt, but the shop has not met its own requirement that every kit carries the current sheet.

NCR no.: NCR-2026-044
Requirement: each kit ships with instruction sheet rev D (document control register).
Finding: 24 kits on order 3190 shipped with rev C.
Containment: customer contacted the same day; rev D emailed; sheets replaced on the 6 kits still in their goods-in area; remaining 18 in use.
Disposition: no rework of product; replacement sheets posted.
Cause: [to be established by the team: evidence needed]
Corrective action: [after cause]

The five-whys questions then look at the paperwork, not the product: where do packers get sheets, who removes the old ones when a revision changes, and what proves the shelf is clear. An assistant can draft those questions, but the answers come from walking to the packing bench and looking in the drawer. Often the finding is a stack of old sheets sitting under the new ones. The action is a change to document control, such as removing superseded sheets on the day of issue and signing a log.

Grading severity so the response fits

Treating every NCR as a crisis wastes effort, and treating all of them as paperwork hides the serious ones. Agree three grades and write down what each triggers. An illustrative scheme follows; adjust it to your own product and to any customer or standard you work to.

The brackets would be major: they fail a customer requirement and some had shipped. The sheet error might be minor or major depending on whether the instructions affect how the product is installed safely. That judgement belongs to the quality owner. An assistant can suggest a grade from the description if you give it your scheme, but treat it as a second opinion.

The message to the customer

If affected parts have shipped, the customer needs to hear from you before they find out. Draft the message while containment is under way. Give the assistant the facts and forbid it from adding any.

Draft a short, plain email to a customer about a non-conformance in parts we shipped. Use only these facts: part (mounting bracket, drawing rev C), job 2217, 120 parts shipped on [date], hole-to-hole distance measured at 62.8 mm against 62.0 mm +/- 0.3 mm on 10 of 50 sampled parts, we ask them to hold the shipment and not use it, we will rework or replace at our cost, and we will send a corrective action report by [date].
Do not state a cause. Do not say the parts are safe or unsafe. Do not offer compensation. Leave [your name] and [phone] as placeholders. Under 150 words.

The restriction list is what makes this safe to use. Without it, assistants often add reassurance such as "this poses no risk", which only a qualified person can say. The quality owner reads and edits the email; a person sends it.

A cost-of-quality sum worth doing once

Put numbers on a few NCRs and the case for fixing causes becomes clear. For the bracket job: rework of 69 parts at 12 minutes each is 13.8 hours, or $552 at $40 an hour. Add 3 hours of inspection of the full batch ($120), 2 hours on the investigation and customer contact by the quality owner (say $100 at $50 an hour) and the extra carriage for reshipment, perhaps $60. The total is $832, before the harder-to-count cost of a worried customer. Against that, a pin in the fixture that costs a few hours of toolroom time is a bargain.

The arithmetic: 552 + 120 + 100 + 60 is $832. Ask the assistant to produce this table for each closed NCR in a quarter, and to show the total for each cause type. If one cause type accounts for most of the cost, it tells you where to spend your attention. Check each figure against your timesheets; the assistant sums the numbers you give it and has no idea what your hours really cost.

A mistake: a plausible cause that nobody tested

A hypothetical shop had a recurring surface defect on painted panels. The first NCR named "contaminated paint" as the cause, because the assistant, asked for a probable root cause, offered it with confidence, and the paint was replaced. Defects returned in three weeks. A second look found that the defect appeared only on panels from one rack position, close to a door that opened onto a dusty yard. The cause was airflow, not paint.

What went wrong was the request: "what is the probable root cause?" invited a guess. The better prompt asks for questions, evidence and a test for each possibility. In this case, a test of "does the defect follow the paint or the rack position?" would have taken one afternoon. Keep that rule written on the NCR form itself: a cause is not a cause until something has been checked.

Trend review: are the reports telling you something?

Once you have ten or twenty NCRs, the pile itself is data. Ask the assistant to group them by product, process step, supplier and cause type, and to say which groups hold more than one report. Check its counts against the log, and treat small numbers cautiously. A list of "similar causes" is worth a human read, because two reports with the same wording may have different roots, and two with different wording may share one.

Quarterly, review three questions: which actions are overdue, which causes were never verified, and which NCRs reopened. A reopened NCR means the fix did not work or the cause was wrong. Both are worth knowing, and both are more likely to be spotted when the report records its evidence honestly.

A short routine to keep the paperwork light

1. The finder writes a rough note, with photos and measurements, on the day.
2. The assistant tidies it into the form, leaving gaps visible.
3. The production lead completes containment and the quality owner records disposition, the same day.
4. The team meets within a week for the five-whys, with the assistant as facilitator and a person writing the answers.
5. Actions are set with owners and dates, and the verification check is agreed.
6. The quality owner closes the report only after the verification evidence is attached.

Done this way, the assistant saves the quality owner the typing and the proofreading, and nobody can mistake it for the person who decided. The signatures are still human, and so is the responsibility for what went to the customer.

Non-conformance report questions from small manufacturers

What is the difference between containment and corrective action?

Containment stops the problem reaching a customer now: quarantine, sorting, a hold on shipments. Corrective action removes the cause so it does not return. Containment is quick and temporary. Corrective action takes longer and must be verified. A report that has only containment has not finished the job.

Can AI find the root cause?

It can structure the questioning and point out gaps, but the answers must come from evidence your people collected. An assistant asked for a cause will write a plausible one. Treat anything it proposes as a hypothesis to test on the shop floor, not as a finding.

Who should sign an NCR?

The person who raised it, the person who owns the disposition decision, and whoever verifies the action worked. In a small firm one person may hold two roles, but the verification should not be done by the person who made the fix if you can avoid it.

How long should an NCR stay open?

Set a target by severity, for instance containment the same day, cause within a week and verification within a month. The right numbers depend on your product and on any customer or standard you work to. What matters is that every open report has a date and an owner.

Do I need a separate NCR for a supplier fault?

Raise the NCR internally to record the finding and the containment, then send a supplier corrective action request so the supplier investigates their own cause. Link the two by reference number so the file shows the full story.

Further reads

• Near Miss Report Template: AI Logs, Categorises and Spots Trends — Log safety near misses with the same discipline.
• AI Supplier Management: Track Prices, Lead Times, and Risk — Track supplier lead times, prices and risk.
• How to Compare Supplier Quotes Side by Side With AI — Compare replacement supplier quotes side by side.
• AI Systems for Manufacturing: MRP Add-On or Chatbot on Your Own Data? — Decide between an MRP add-on and a chatbot.
• AI Quality Control for Small Manufacturers: Where to Start — Decide where AI helps quality work first.
• How Much Does AI Visual Inspection Cost a Small Manufacturer? — See what automated visual inspection costs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the report structure follows common quality-management practice; every figure and part in the example is illustrative. Check any standard or customer requirement that applies to you.