An image can show the condition of a cabinet, part or workstation. By itself, however, it does not explain what the inspector was expected to see, which criterion was applied or who authorized the final decision. That distinction separates a computer-vision demonstration from an inspection that can participate in a quality process.

Industrial AI visual inspection becomes more useful when it turns each image into reviewable evidence. The system must control capture, compare against approved criteria, communicate uncertainty and preserve the connection between a finding and the original image.

ReadyMark is Avira Agent’s first implementation for exploring that approach with industrial cabinets. Its flow illustrates the components a traceable visual-inspection solution needs without assuming that AI replaces physical tests or the quality team’s release authority.

Capture is part of the inspection

A model cannot evaluate what the camera does not show. Glare, perspective, shadows, occlusion or insufficient resolution can hide the component under review. Capture should therefore not remain an informal step before analysis.

A capture contract defines which views are required, how they are identified and which conditions make an image unsuitable. A cabinet, for example, may require several complementary views: an overview plus additional angles that cover areas a single position cannot reveal.

The interface should guide the operator through those views and request a new image when evidence is insufficient. Marking a part as incorrect from an ambiguous photo is more dangerous than acknowledging that the result cannot yet be verified.

Criteria must exist before the model

AI should not invent a definition of conformance. The inspection needs a checklist, product reference, technical documentation or a combination of sources approved by engineering and quality.

Each check should state what must be visible. It may verify component presence, visible identification, an apparent configuration or the absence of a foreign object. The criterion must also declare what stays out of scope: torque, conductor gauge, electrical measurements or properties that require instruments and physical access.

Versioning the checklist and its references lets the team identify which rules were used for every inspection. This matters when a product family changes, a drawing is revised or a result is reviewed weeks later.

Passed, failed or needs verification

Forcing every image into a binary result hides an important part of the risk. A responsible inspection needs at least three states:

  • Passed: available evidence is sufficient to verify the criterion.
  • Failed: a visible, localizable deviation is present.
  • Needs verification: capture or criteria do not support a reliable conclusion.

The third state is not a system failure. It is the correct output when glare, occlusion, missing detail or a property that only a person can verify prevents a conclusion. It also keeps uncertainty from becoming either a false alarm or an approval without sufficient evidence.

Localize the finding without changing the original

A sentence such as “component missing” forces the next person to repeat the search. The result becomes more operational when it identifies the evidence view, exact location and technical reason behind the finding.

Annotations should be drawn over a presentation copy, not over the file that serves as evidence. The original remains intact; the annotated version adds a marker, outline or callout that makes the deviation easy to find.

This separation also prevents a common mistake: a generated or reinterpreted image should never be presented as proof of what the camera captured. Industrial evidence requires the original pixels to remain available for review.

From finding to a reviewable decision

An inspection does not end when a model responds. The result must reach a workflow where production and quality can act. A useful report brings together:

  • Original views and their annotated counterparts.
  • The applied criterion and reference.
  • The state of every check.
  • The finding’s explanation and location.
  • The corrective action or pending verification.
  • Fields for review, ownership and closure.

That report allows another person to replay the decision without relying on an informal conversation or shift memory. It also creates a basis for comparing the solution with the current manual process during a pilot.

Taking the solution into a cell

Computer vision does not live in isolation. An operating project may involve a fixed camera, a robot that positions the sensor, product identification, a PLC or a quality system. Every connection changes scope and must be designed with automation, operations and security.

The AI assistance layer should remain separate from critical control. A pilot can begin with uploaded images or capture in a controlled environment. Once criteria, failure modes and uncertainty handling have been validated, the team can decide which information needs to move between the inspection and the cell, and which permissions are truly necessary.

The goal is not to connect everything on day one. It is to take only a proven, reviewable and safe flow into operation for one concrete use case.

What to validate in your plant

Before starting an AI visual-inspection pilot, confirm:

  • The included asset, product family and inspection zone.
  • Required views and minimum capture conditions.
  • The approved checklist and versioned references.
  • Which properties are visually observable and which need manual measurement.
  • Ambiguous cases that must return “needs verification.”
  • How originals are preserved and annotations presented.
  • Who reviews, corrects and releases the work.
  • Necessary camera, robot, PLC or quality-system integrations.
  • Criteria to proceed, adjust or stop the pilot.

The best AI visual-inspection system is not the one that always produces an answer. It is the one that shows what it saw, which criterion it used, where it found a deviation and when a person must complete the decision.