The Factory Data That Management Could See but Could Not Use

A two-site garment manufacturer employing approximately 350 people had built its quality system around twelve QC staff, most of whom worked as inline inspectors close to the sewing process. Their role was to identify defects as garments moved through production, record what they found and support the correction of faults before finished goods reached the customer. The system generated a substantial volume of information, although much of it remained on handwritten forms completed in Vietnamese, using several layouts, a mixture of printed categories and free-text annotations, and numbers that were sometimes overwritten or added at the end of the shift.

The company’s newly appointed production engineering manager was Chinese and worked primarily in English. He had been recruited to improve production performance and began by asking a straightforward question: what did the QC reports actually say about the way the factory was operating?

At that stage, his immediate concern was access rather than analysis. The documents were familiar to the Vietnamese QC team, yet their meaning was difficult to transfer to an English-speaking manager who wanted to understand how process failures were being recorded across the company. The initial review began with three reports from one production line. Translation alone would have required someone to interpret handwritten Vietnamese, recognise garment-production terminology, understand the layout of several different forms and explain how the individual observations related to the sewing process.

AI provided a practical way through that barrier. The first task involved reading the documents, identifying the printed defect categories, extracting the handwritten quantities and explaining the contents in English. What began as a translation and document-reading exercise gradually became something more valuable, because once the information had been extracted, it could be compared, reorganised and analysed.

From Quality Records to Capacity

One set of reports covered a sewing line of 27 operators working an eight-hour shift. The line produced 242 garments during the day, while the QC records identified 43 defective pieces requiring some form of correction. The defect rate approached 18 percent, although that percentage alone did little to explain the production manager’s concern about capacity.

The connection became clearer when the defects were converted into production time. The line had used 12,960 labour minutes to produce 242 garments, giving an average labour content of approximately 53.5 minutes per garment. The production manager estimated that repairing a defective garment consumed around 30 percent of the original sewing time. On that basis, the 43 defective pieces represented approximately 691 minutes of rework, equivalent to 11.5 labour hours or 1.4 people working for an entire shift without producing an additional garment.

This calculation changed the management discussion. The production engineering manager had entered the business expecting to examine line speed, staffing levels, work allocation and production planning. The QC documents showed that a material part of the apparent capacity shortage was already inside the existing process, absorbed by defects and the effort required to correct them.

The direct rework represented approximately 691 minutes, or 12.9 garment-equivalents, each day. Where defects are returned to the original sewing operators, interruptions, machine resetting and disrupted line balance can increase this loss to between 17 and 21 garments. In this factory, however, defective garments were generally passed to a specialist repair group. This protected the immediate flow of the sewing line, while creating a different problem: the people and processes responsible for the original defect received weaker feedback about the consequences of their work.

The factory had possessed the information needed to explain part of its capacity problem throughout the period in which the reports were being completed. The information remained difficult to use because it was dispersed across handwritten forms, expressed in another language and organised for daily inspection rather than management analysis. AI converted those records into a form that could support a different kind of question, moving the discussion from how many garments had failed inspection to how much productive capacity the failures were consuming.

The Same Documents Revealed More as AI Improved

The first analysis depended mainly on quantities and predefined defect categories. The AI could identify the number of defective garments, recognise categories such as broken stitches, incorrect label attachment, measurement failure and pressing defects, and calculate their effect on output. That level of capability was sufficient to establish the cost of poor quality and to identify the largest defect groups.

The handwritten notes offered a deeper layer of information. Inspectors had added short Vietnamese descriptions explaining the precise nature of individual faults, where they appeared and, in some cases, the production circumstances surrounding them. The earlier model could reliably read handwritten quantities and predefined defect categories, although the free-text Vietnamese annotations remained too uncertain to support analysis. Six months later, improved recognition made those annotations usable, while stronger reasoning connected related evidence across several reports.

Several months later, the same documents were reviewed again using more capable AI. The improvement came from two linked developments. Recognition had become better, allowing the model to read handwritten Vietnamese annotations with greater accuracy, while reasoning had also improved, allowing evidence from several reports and production stages to be connected.

A broad category such as distorted shape could now be associated with a note describing a skewed zipper tail. A broken-stitch entry could be connected to a comment suggesting fabric tension before failure. A measurement defect could be linked to an armhole cross-point offset that had appeared earlier in the process under a different classification. The terminology used by inspectors varied according to the stage at which the fault was observed, but the underlying production issue could now be followed across documents.

The pressing reports offered a particularly useful example. Defects appeared during sewing, resurfaced at final inspection as panel mismatch and were then recorded after ironing as pressing failures. Viewed separately, the final form suggested a problem in the ironing area. Viewed across the full sequence, the ironing team appeared to be receiving and exposing a fault that had originated earlier.

This distinction changed the value of the analysis. The first use of AI answered the question of how much quality failure was costing the factory. The later use came closer to answering where the failure began. The documents remained the same, while the information obtainable from them became progressively richer as AI capability developed.

For management, this has an important implication. Historical records should not be treated as having a fixed analytical value. A document that supported only basic extraction six months earlier may later support translation, classification, process tracing and causal investigation. The growing capability of AI can increase the value of information that the organisation already owns.

Repairing the Garment While Preserving the Defect

A subsequent review of the workflow revealed that defective garments were being transferred to a specialist repair group, which corrected the fault and returned the garment to production. This arrangement helped protect the immediate flow of the sewing line because experienced repair operators could deal with difficult faults while the main operators continued working.

The same arrangement weakened the feedback loop between the defect and the process that created it. The repair team received the garment and solved the visible problem, while the original operator, supervisor, technician or production engineer received limited information about the recurring cause. A garment could therefore be repaired successfully while the method, machine setting, jig, measurement instruction or work sequence remained unchanged.

Over time, the repair group became a buffer between process failure and process learning. Its effectiveness at correcting garments reduced the visibility of recurring defects at source. Production continued, the repaired garments passed inspection, and the organisation absorbed the cost through lower effective capacity.

The required response involved more than reducing the number of repairs. The workflow needed to preserve the advantages of specialist repair while restoring immediate feedback to the point of origin. Each defect had to generate two connected actions: the garment required timely correction, and the source operation required information capable of preventing recurrence.

This meant linking the repair record to the original operation, defect type, responsible supervisor and technical owner. Repeated faults could then trigger a review of machine settings, work instructions, operator training, tooling, pattern accuracy or production engineering. The repair team remained part of the production system, while its information became an input to improvement rather than the final destination of the defect.

From Document Reading to Operational Redesign

The case began with an English-speaking production engineering manager trying to understand Vietnamese QC forms. AI initially served as a bridge between languages and document formats, extracting information that had previously been difficult for management to access.

Once extracted, the same information changed the understanding of production capacity. Defects that had been viewed mainly as quality events became visible as lost labour time, disrupted flow and delayed output. When more capable AI became available, the analysis moved deeper into the handwritten annotations and connected related faults across production stages. This made it possible to distinguish the department that discovered or repaired a defect from the process that had created it.

The final insight concerned workflow. The factory had developed an effective mechanism for repairing defective garments, while the transfer of defects to a specialist group had weakened the feedback required to prevent their recurrence. AI helped expose that separation and provided the evidence needed to redesign the flow of information alongside the flow of garments.

The broader lesson extends beyond garment manufacturing. Many companies continue to operate with handwritten records, spreadsheets, emails and locally designed forms. These systems often contain detailed operational knowledge, although their structure makes analysis difficult and their language may limit access to a small group of employees. AI can release that information in stages. Early capability may support reading and translation, later capability may support measurement and comparison, and further improvement may allow patterns, relationships and process causes to emerge from the same underlying records.

The value lies in the progression. Information that was previously inaccessible becomes readable, readable information becomes analysable, analysis changes management understanding, and deeper analysis creates the basis for operational redesign.