In a low-wage garment economy, computer vision pays back by saving fabric and preventing rework rather than replacing inspectors

Institutional uniforms and repeat domestic orders can make quality failures commercially serious, but the base case should not invent a chargeback schedule. A rejected lot, an emergency replacement, a return or a lost tender has a different contractual cause and cost. None belongs in annual benefits until the factory has buyer terms, an observed baseline and a defensible estimate of how the inspection change alters the event probability.

The Ethiopian trouser study reports image processing of 0.45 seconds a sample and a substantial throughput improvement over manual inspection. That result supports testing faster classification. It does not demonstrate elapsed garment handling, complete surface coverage or prevention of buyer-lot rejection. The economic station record must add presentation, review, diversion, re-presentation and queue time before claiming a production-rate benefit.[S19]

Final acceptance and process control also answer different questions. A sampled buyer audit can accept or reject a lot after production, while inline vision can flag a changing process before every item is complete. A camera may reduce the chance of defective work reaching the final audit, but the saving arises only if the factory avoids a real downstream disposition. Better inspection coverage alone is not proof of fewer claims.

For that reason, the central model assigns zero value to avoided AQL failure, air freight, returns, reputational damage and future tender access. Zero does not mean those effects are unimportant. It means their frequency and unit costs are unknown here. When evidence becomes available, the external-failure calculation can use lots exposed, the measured change in verified rejection probability and the actual incremental cost per event, with legal and commercial review of the contract terms.

A similar rule applies to quality premiums. A buyer might pay more for documented inspection, or a factory might win work it would otherwise lose, but those are commercial scenarios rather than automatic vision benefits. Counting both a higher price and the full margin on newly won volume would overstate the inspection effect if product design, delivery or relationship management also contributed. The attribution should be agreed before the sale is included.

GFT recommends an external-failure register linked to lot, buyer requirement, defect class, verified cause, disposition and incremental cost. The baseline should include accepted lots as well as failures so a rate can be calculated. Until that register supports a probability, management should treat avoided external loss as upside outside the payback decision. This makes the investment harder to exaggerate and easier to reconcile after commissioning.

The Ethiopian trouser study estimates a 12 to 18 month return for medium-sized factories. The central illustrative result of 15.6 months sits inside that interval, but this is arithmetic coincidence rather than independent validation. The opened abstract does not disclose a capital and operating-cost bridge, and its image-classification station is not the same as this chapter's two-station allowance. Its estimate should be used to frame questions about scope and payback, not to mark the Cameroonian model as observed.[S19]

That distinction suggests a disciplined evidence ladder. First, establish the manual counterfactual with verified defect and disposition records. Next, demonstrate local capture and false-alarm performance on retained production examples. Then measure the physical outcomes that create value: recovered cloth, avoided repair minutes and changed downgrade destinations. Only after those measurements should management annualise benefits and compare them with quoted ownership costs.

The low-cost edge study reinforces the same gate from the operating side. Its triage design can reduce repetitive work, but its reported false-alarm forwarding shows why review and calibration remain cost centres. A vendor demonstration with high recall is not yet a labour budget. The factory should time the complete review workflow and include that workload in annual operating cost before accepting a payback claim.[S07]

GFT's procurement gate is consequently conditional. Scale only when the verified net benefit meets the sponsor's threshold under actual utilisation, the recovery fraction remains positive after all cutting consequences, and manual review can absorb the alarm load. If those conditions fail, change the optical setup, defect scope or station count before adding capital. Do not repair a weak operating case by adding speculative buyer penalties.

The illustrative sensitivity model retains every central assumption except utilisation and fabric-origin defect share. Annual volume changes with utilisation, fabric recovery changes with both volume and defect share, and the direct rework and seconds buckets scale with volume. Inspector reallocation and annual operating cost remain fixed. Payback is XAF 50 million divided by positive net annual benefit, multiplied by 12. This is a constructed comparative model, not a statistical confidence interval.

Exhibit 49 shows the interaction. At 30 percent utilisation and a 2 percent incoming defect share, the fixed cost base leaves only XAF 1.310 million net annual benefit and an illustrative payback of 458 months. At 85 percent utilisation and an 8 percent defect share, the same assumptions produce 8.0 months. The wide range is precisely why a payback statement without volume, prevalence and disposition assumptions is not decision-ready.

At the central 5 percent defect share, changing only utilisation gives illustrative paybacks of 57.6 months at 30 percent, 34.5 months at 40 percent, 15.6 months at 70 percent and 12.3 months at 85 percent. A ramping factory should therefore avoid installing every station against nameplate volume on day one. A smaller initial scope can limit fixed operating cost while the factory learns which dispositions genuinely recover value.

Fabric price changes alter the same investment gate. Holding the other central inputs fixed, illustrative fabric cost of XAF 2,000 a garment equivalent produces XAF 26.586 million net annual benefit and 22.6-month payback. XAF 3,000 gives 15.6 months, and XAF 4,500 gives XAF 56.117 million net benefit and 10.7 months. The correct input is the avoidable cost of the actual material mix, including only those logistics and processing elements that change under the counterfactual.

The defect-rate axis needs equal caution. A higher observed defect share can increase addressable value, but only if the additional detections are real and recoverable. A difficult print that creates both more true faults and more false alarms may increase review, recutting and downtime faster than gross benefit. The recovery fraction r must therefore be remeasured by fabric family rather than held at the illustrative 45 percent because the central spreadsheet requires it.

The sensitivity results support three practical responses. At low utilisation, reduce fixed scope and prove one high-value route. At high material cost, prioritise incoming inspection and cutting coordination. At high sewing-rework prevalence, assign corrective action to the verified station and class before expecting camera detection to become saved minutes. These are different operating interventions, and the station configuration should follow the dominant loss rather than a generic vision package.

The final investment record should preserve the assumptions beside the realised values each month. Replace nameplate volume with reconciled production, estimated capture with audited capture, assumed recovery with net disposition value, and allowance costs with invoices and measured review time. Report simple payback only after net annual benefit is positive and externally supportable. If the realised case diverges, change the operating decision rather than hiding the divergence inside a blended quality score.