AI agent · Manufacturing · Read & extract
Non-conformity reports
The software reads non-conformity reports as operators write them, checks the part and batch against production records, codes the defect against your list and creates the record once. Quality engineers get clean data and early warning of repeats.
Typical volumes for this process, not a client figure.
Written on paper, typed up later, analysed rarely.
Captured, categorised, and trended without a second entry.
Where the time goes today
When a part fails inspection or an operator spots a defect, someone fills in a non-conformity report: line, workstation, part number, batch, quantity affected, what is wrong, a suspected cause and the containment action taken. On many sites this is still a paper form, sometimes a tablet form with free-text boxes, occasionally a photo sent by message. Paper forms go into a tray and reach the quality office at the end of the shift.
A quality technician then types them into the quality management system or a spreadsheet, often days later. Handwriting, shop-floor abbreviations and local names for defects slow this down. The same defect is described five ways ('burr', 'sharp edge', 'flash on edge'), so it lands under different codes, or under 'other'. Part numbers are mistyped and batch numbers left blank, which makes containment and traceability harder when a customer complaint arrives.
The result is data nobody trusts. A defect repeating on the night shift is visible only at the weekly meeting, if at all. Pareto charts reflect how technicians coded the forms rather than what went wrong, and the monthly review becomes a count of forms instead of an analysis.
How the agent works
- Collect the reportScanned forms, photos from the line, tablet submissions and emails are collected from one inbox or folder. The source image stays attached to the record for its whole life.
- Extract the fieldsThe agent reads line, workstation, part number, batch, quantity, defect description, suspected cause, containment action, reporter and time. Each field carries a confidence level, and a field read with low confidence is marked for confirmation, not guessed.
- Check against production dataPart numbers are checked against the item master and batches against the production orders for that line and shift. A part that was not produced on that line that day is flagged.
- Code the defectThe free-text description is mapped to your defect and cause codes, and the operator's original words are kept beside the code. If no code fits, the report is left unclassified for a quality engineer instead of forced into the nearest one.
- Create and trendThe record is created in your quality system and linked to the batch. Counts by code, line, part and shift are updated, and the quality engineer is alerted when a code repeats more often than the threshold you set.
What stays with a person
Disposition stays with quality: use as is, rework, scrap or return to the supplier. So does root-cause investigation, the decision to open a corrective action, stopping a line, and any notification to a customer. The agent records the operator's suspected cause; it does not confirm it.
Reports touching a safety-critical characteristic, or a customer-owned part, go to a quality engineer as soon as they are read, whatever the extraction confidence. The quality team also owns the defect code list. The agent can show which descriptions keep landing in 'unclassified', but adding a code is a person's decision.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Paper report forms, scanned or photographed | A structured record in your quality system, with the source image attached |
| Tablet and email reports from the line | A short list of fields waiting for confirmation |
| The item master and production orders or batch records | Defect counts by code, line, part and shift |
| Your defect and cause code lists | Repeat-defect alerts and duplicate candidates |
| Open non-conformity records, to spot duplicates |
Controls that come with it
- Below the confidence threshold a field stays blank and goes to a technician; the agent never fills a gap with a plausible value.
- Part and batch are validated against production data before the record is created.
- The form image and the extracted fields sit side by side, so a check takes seconds.
- Safety-critical characteristics route to a person regardless of confidence.
- The agent creates and updates records but never closes one; every change is logged and can be edited or voided in the quality system.
How you know it works
- Time from defect found to record in the quality system
- Field accuracy against a monthly sample re-keyed by a technician
- Share of records with a valid part and batch
- Share of reports left unclassified or re-coded by an engineer
- Time from the start of a repeat defect to the first alert
Is your process ready?
- Written rules: a defect and cause code list exists, even if it is applied loosely today.
- Systems: the quality system accepts records through an interface or import, and item and batch data can be read.
- Cheap check: a technician compares each record with the form image in a few seconds.
- Volume: hundreds of reports a month across several lines repay the build; a site with twenty a month may be better served by a structured tablet form.
- Same description: supervisors on every shift use the same form and mean the same thing by 'containment'.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Forms photographed at an angle, in poor light or with oil stains; agree a simple capture routine first.
- Each line has its own vocabulary for defects, so the mapping must be taught with past reports from every line, not one.
- A code list that is too coarse, with 'other' used heavily; the agent reproduces that vagueness faithfully.
- The same defect reported by the operator and by the inspector, doubling the trend unless duplicates are checked.
Questions we get
Can it read our operators' handwriting?
That depends on your forms, so it is tested rather than assumed. Before go-live the agent is run on a few hundred past forms that technicians have already typed up, and you compare field by field. Fields it cannot read with confidence are left for a technician, who sees the image next to the record. The rate of those fields is one of the numbers you track.
Should we move to tablets instead?
If your reports can become a structured digital form with drop-down codes and validated part numbers, do that first: it is cheaper than an agent. The agent earns its place when reports keep arriving in mixed forms, when operators need free text to describe what they saw, or when paper remains the practical option on part of the floor.
Does it decide what happens to the parts?
No. It records what the operator wrote about containment and makes the report visible quickly. Disposition, rework instructions, supplier returns and corrective actions stay with your quality engineers, as does anything involving customer notification. The agent's job is to make sure they are deciding on a complete, correctly coded record rather than on a form in a tray.
How does it learn our defect codes?
It is configured with your code list and its definitions, then tested on past reports that your engineers coded. Where the agent and the engineers disagree, you review the cases and either correct the mapping or clarify the code definition. New codes are added by the quality team; the agent starts using them from the next version, which is tested on the same past reports first.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.