# AI quality inspection report triage.

URL: https://levystreet.com/industries/manufacturing/quality-inspection-report-triage/
Updated: 2026-10-19

How AI quality inspection report triage works for NZ manufacturers: sorting check sheets and complaints by defect and severity, with QA keeping dispositions.

## Short answer

AI quality inspection report triage is a workflow where a model reads inspection records, check sheets, photos and customer complaints, classifies each defect against your taxonomy and specification, suggests a severity, spots repeats across lots or tools, and routes the record to the right quality queue. Quality staff decide every disposition, nonconformance and recall. Fixed keyword rules escalate safety terms regardless of what the model suggests.

- The model classifies and routes; QA decides use-as-is, rework, scrap or return.
- Safety keywords escalate by rule, and the model can raise severity but never lower it.
- Repeat-defect linking across lots, tools and suppliers is often the biggest gain.
- A voluntary recall for injury risk must be notified within 2 working days under the Fair Trading Act.

## What triage does and does not do

AI quality inspection report triage reads the paperwork that surrounds a defect and decides where it goes next. Pass or fail against a measured tolerance is already a rule in your gauges or inspection software. What slows quality teams down is everything around it: free-text notes on check sheets, photos from the line, third-party lab reports and customer complaint emails, all arriving in different formats and all needing a person to read them before anything happens.

The model classifies each record against your defect taxonomy, links it to the drawing or specification, suggests a severity using your written criteria, and looks for the same defect on the same tool, cavity, shift or supplier lot. Then it puts the record in the right queue. That is a [document processing](/services/ai-document-processing/) job with a strict escalation rule.

## How severity routing works

| Severity | Typical triggers | Route |
|---|---|---|
| Critical | Injury risk, contamination, electrical or pressure failure, regulatory nonconformance | Quality manager within one hour, by rule |
| Major | Out of tolerance on a functional dimension, customer line stoppage | QA engineer within the shift |
| Minor | Cosmetic outside the agreed standard | Daily batch review |
| Observation | Within tolerance but trending | Weekly trend report only |

Two rules sit above the model. First, a keyword and defect-code list (burn, sharp, glass, metal fragment, shock, allergen, leak) routes straight to Critical whatever the model suggests. Second, the model may raise a severity but never lower one set by a rule or an inspector. Errors in the cautious direction cost minutes; errors the other way can cost a recall.

## How the workflow runs


- **Inputs**: Incoming, in-process and final inspection records; check sheets (digital or photographed); defect photos; lab and test reports; customer complaints; the defect taxonomy, specifications and severity criteria.
- **Systems**: The quality management system or NCR register, the ERP for lot, tool and supplier data, a shared inbox for complaints, a review queue, and a model API.
- **Steps**: Extract fields and notes; classify defect type and location; link the specification clause; apply rule-based escalation; suggest severity; search for repeats across lots, tools and suppliers; route; record the suggestion.
- **Human review**: QA staff clear every queue and decide disposition (use as is, rework, scrap, return to supplier), whether to open a nonconformance, and containment. The quality manager decides customer notification and recalls.
- **Outputs**: Routed records with suggested codes, repeat-defect alerts linking related records, and a weekly trend summary by line, tool and supplier.
- **Measures**: Time from inspection to decision for Critical and Major records, how often people raise the model's severity, repeats caught before reaching a customer, and NCR closure time.
- **Risks**: Under-called severity, misread handwriting, taxonomy drift, and staff approving suggestions without reading the record.


## Illustrative example: an Auckland plastics moulder

Illustrative example, synthetic numbers: an injection moulder producing parts for appliance and building product customers logs about 250 inspection records a week across three shifts, plus 15 customer complaints a month. Assumptions: a QA engineer spends 6 hours a week reading and sorting records; with triage, clearing pre-sorted queues takes 2 hours.

- **Time:** about 4 hours a week back for the QA engineer, roughly 17 hours a month.
- **Repeat detection:** linking records by mould and cavity number surfaces a recurring short shot on cavity 3 of one tool that had been logged as four unrelated minor defects.
- **Model fees:** about 1,100 records a month at 4,000 input tokens (form, one photo, taxonomy) and 300 output tokens on OpenAI's gpt-6-luna, at US$0.10 and US$0.50 per million tokens ([OpenAI](https://developers.openai.com/api/docs/pricing), as listed on 19 October 2026), is about US$0.60 a month. Use a larger model for complaint letters if the small one misreads them; the cost difference is still trivial.

## NZ rules that touch quality triage

When a defect means goods may cause injury, the [Fair Trading Act 1986](https://www.legislation.govt.nz/act/public/1986/121/en/latest/) matters. Section 31A says a supplier that voluntarily recalls goods because they will or may cause injury, or do not comply with a product safety standard, must notify the chief executive within 2 working days, where no other law already sets a reporting process. The notice must be in plain language and describe the goods, the danger and what consumers should do. Your triage queue is where the clock often starts, so Critical routing has to be fast and logged.

Food manufacturers also operate under the [Food Act 2014](https://www.legislation.govt.nz/act/public/2014/32/en/latest/), which imposes documentation, traceability and recall requirements on food businesses. MBIE's [quickstart for businesses using AI tools](https://www.mbie.govt.nz/assets/businesses-using-ai-tools.pdf) says a human in the loop is needed where important decisions are informed by AI, which is exactly where disposition sits.

## Where triage goes wrong

- **Rubber stamping.** When suggestions are usually right, people stop reading. Hide the model's severity on a random 10% of records and compare independent calls.
- **Taxonomy drift.** New defect types get forced into old codes. Review "other" classifications monthly and update the list.
- **Bad inputs.** Photos taken in poor light or check sheets photographed at an angle fail quietly. Return unreadable records to the line the same shift.
- **Over-trusting trends.** A trend summary can make noise look like a pattern. Treat it as a prompt for an engineer to look, not a finding.

Queue design is covered in our guide to [human review queues](/engineering/human-review-queues/), and the [AI pilot scorecard](/resources/ai-pilot-scorecard/) helps decide when a pilot is ready to widen. Related workflows are [maintenance knowledge search](/industries/manufacturing/maintenance-knowledge-search/) and [supplier document extraction](/industries/manufacturing/supplier-document-extraction/); the [manufacturing overview](/industries/manufacturing/) covers the sector.

## Questions people ask

### Can AI decide whether a part passes inspection?

Not in this workflow. Measured pass or fail against a tolerance is a rule your gauges or inspection software already apply. Triage handles what happens after: reading the free-text notes, photos and complaints that come with a failure, classifying them, and putting them in front of the right person quickly. Disposition remains a documented human decision in your quality system.

### Will AI triage work with handwritten check sheets?

Partly. Modern vision models read tidy handwriting and ticked boxes reasonably well, but smudged, abbreviated or photographed-at-an-angle sheets fail more often. Measure extraction accuracy on 100 real sheets first. Many plants find the triage project is the push they needed to move check sheets onto tablets, which removes the reading problem entirely.

### How does this fit with ISO 9001 or our existing quality system?

Treat the triage step as a documented process inside your quality management system, with its own work instruction, owner and records. The model's classification is an input to the nonconformance process, not a replacement for it. Keep the original record, the model's suggestion and the person's decision together so an auditor can see who decided what.

### What happens if the AI misses a serious defect?

Design for it. Safety-related words and defect codes route by fixed rule straight to the quality manager, whatever the model says. Every record still lands in a queue a person clears, and a weekly sample of low-severity calls is re-checked. Track how often people raise the model's severity; if that rate climbs, stop and recalibrate.


## Sources

- [Fair Trading Act 1986](https://www.legislation.govt.nz/act/public/1986/121/en/latest/)
- [Food Act 2014](https://www.legislation.govt.nz/act/public/2014/32/en/latest/)
- [Quickstart: Businesses using AI tools](https://www.mbie.govt.nz/assets/businesses-using-ai-tools.pdf)
- [API pricing](https://developers.openai.com/api/docs/pricing)

Published by Levy Street (https://levystreet.com), Wellington, New Zealand.