GeoBusinessIQGeoBusinessIQ

Visual inspection: what a person looking at a part can and cannot decide

What this answers

Would two of our inspectors reach the same verdict on this part, and would they reach it again tomorrow?

Most acceptance decisions in a factory are made by somebody looking at something and forming a judgement. It is the lowest-cost inspection available and the least reliable, and the gap between how reliable people assume it is and how reliable it actually is causes an enormous amount of escaped defect. Improving it starts with accepting that two competent inspectors will disagree, and designing around that rather than exhorting them not to.

Written for: quality engineers, inspectors and line operators, production supervisors.

The written criterion is never enough on its own

Words like scratch, blemish, flash, orange peel and acceptable finish carry different meanings to different people, and a specification built from them will be applied differently by each shift. What resolves it is physical reference: retained samples showing the acceptable limit and the first unacceptable condition, ideally for each defect type and each surface class, held at the point of inspection rather than in an office. They need control like any other reference — identified, dated, reviewed for degradation, and reissued when the criterion changes, since a yellowed limit sample teaches the wrong boundary.

Lighting, distance, angle and time are part of the specification

The same part passes under one light and fails under another. If the criterion does not state the illumination level and type, the viewing distance, the angle, whether the surface may be tilted into the light, and how long the inspector has, then the plant is running an undefined test and will disagree with its customer about the result. Fixing the conditions physically — a defined station with consistent lighting and a marked viewing position — removes more variation than any amount of training, and makes the eventual argument with a customer a comparison of conditions rather than of opinions.

Measuring whether inspectors actually agree

The way to find out is deliberately: assemble a set of parts spanning clearly good, clearly bad and borderline, have several inspectors judge them blind and more than once without knowing the answers, and compare. The results are usually humbling, particularly for borderline items, and they distinguish two different problems. Inspectors who disagree with each other need a clearer criterion and reference samples. Inspectors who disagree with themselves on a repeat need better conditions, less time pressure or a different method. Running this exercise periodically also keeps the criterion from drifting quietly over the years.

The conditions that destroy detection

Human detection falls away sharply with monotony, with defect rarity, with speed, and towards the end of a long period of concentration. Someone checking a stream of good parts will miss the rare bad one no matter how conscientious they are, because attention is not sustainable at that duty. The mitigations are practical: rotate people between tasks, keep inspection periods short, break the visual field into a defined sequence rather than a general look, avoid pairing inspection with a demanding physical task, and never treat a hundred percent visual check as a screen that can be relied upon.

Deciding what should not be judged by eye at all

Some characteristics only appear visual. A gap judged by eye is a dimension, and a fixture or gauge will decide it faster and identically every time. Presence and orientation of a component can be confirmed by a physical poka-yoke that makes the wrong condition impossible. Where the feature is genuinely cosmetic and subjective, human judgement remains the right tool and deserves proper conditions. Where automated inspection equipment is being considered, that is an engineering choice with its own capability and validation questions, and it does not remove the need for an agreed acceptance criterion.

Frequently asked questions

Is a full visual check of every part enough to protect a customer?
No, and relying on it is one of the more expensive assumptions in manufacturing. Human detection of a rare defect in a repetitive stream is far from complete even with attentive inspectors, so a proportion escapes regardless of effort. Full inspection is a reasonable containment while a process problem is being fixed, and a poor permanent control. If the defect must never reach the customer, the answer is to stop producing it or to detect it by a means that does not depend on attention.
How should limit samples be controlled?
Treat each as a controlled item: uniquely identified, agreed and signed by whoever owns the acceptance criterion including the customer where they hold approval, dated, stored so it does not degrade, and reviewed on a schedule. Parts that discolour, corrode or distort need replacing before they shift the boundary. Where a physical sample cannot be retained, calibrated photographs taken under the specified lighting are a workable substitute if the reproduction and the display are also controlled.
Our customer rejects parts we passed. How do we resolve the disagreement?
Compare methods before arguing about parts. Establish the lighting, distance, angle and time each side uses, then have both parties assess the same physical items under both conditions. Most disputes resolve into a difference in inspection conditions or an unstated criterion rather than a difference in diligence. The output should be a jointly agreed limit sample and a written condition statement, which prevents the same argument recurring with a different batch.

Data limitations

  • Standards are referenced, never reproduced. Pages describe what a standard governs and point to the issuing body; they do not restate its requirements, and conformity is determined by the standard itself and by an accredited assessment, not by anything here.
  • Manufacturing figures are operator-supplied inputs, not market data. GeoBusinessIQ holds no factory costs, production volumes, yields, cycle times, tooling prices or capacity data and does not estimate them — every result reflects only the figures you enter.

Explore the graph

Sources

  • International Organization for Standardization ISO (accessed )
    Covers: International standards for quality management, environmental management, occupational health and safety, and industrial processes.
    Does not cover: The content of any standard, conformity decisions, or certification status of any organisation.
    Why it matters: Cited so a reader can reach the issuing body's own public description of a standard. Standard text is never reproduced here.
    Review cadence: annual
  • NIST Manufacturing Extension Partnership NIST MEP (accessed )
    Covers: A public programme supporting small and medium manufacturers with operational, quality and technology adoption practice.
    Does not cover: Results attributable to any specific manufacturer, or improvement figures transferable to another plant.
    Why it matters: Cited for the operational practice it publishes for smaller manufacturers, not for benchmarks or outcome claims.
    Review cadence: annual

Educational and operational information only — not legal, engineering, safety, customs, tax, or financial advice. Requirements vary by jurisdiction, product, process, and contract; confirm with the relevant authority or a qualified professional before acting.

Last updated: