IBM Maximo Visual Inspection: implementation guide for safety-critical workflows

IBM Maximo Visual Inspection is one of the few AI-led MAS-suite components where the payback is straightforward: faster inspections, more consistent classification, and an evidence trail that survives a change of inspector and a regulatory review. The catch, as with most AI capabilities, is that the model only earns trust when it is built inside a real workflow with the inspectors who actually do the work. This guide is for the operations or safety leader who is sponsoring a first programme.

What Visual Inspection actually does

Visual Inspection applies image and video models to inspection tasks: defect detection on equipment, condition classification, anomaly spotting on assets that are routinely photographed or filmed. The output feeds into the inspection workflow on the inspector’s device, and into the work record in Maximo Manage. The intent is to make the inspector faster and more consistent, not to remove them.

It is not a generic computer-vision platform. The integration with Manage and the inspection workflow is what makes it earn its licence.

When Visual Inspection pays off

Three conditions usually have to be true.

The inspection is frequent enough that consistency between inspectors varies materially. A weekly safety walkdown on a large estate. A pre-flight check on a fleet. A daily condition check on rolling stock. The volume is what produces the consistency problem and the data that solves it.

The cost of a missed defect is high. Safety-critical, regulatory, or operationally expensive. A rolling-stock crack missed in inspection costs more than the model that catches it. A runway pavement defect missed in an FOD walk costs more than the model that flags it.

The inspection target is well-defined enough that a model has a fair chance of learning it. Cracks on a defined asset surface. Vegetation encroachment in a defined corridor. A defect class with a reference image. A model cannot learn “anything that looks wrong”; it can learn “this defect class on this surface, scored against this reference”.

Where those three hold, Visual Inspection is one of the safer AI bets in MAS. Where they do not, it produces a model that the inspectors quietly stop trusting.

Building training data the inspectors will trust

The single biggest determinant of success is the training data, and the single biggest determinant of training data quality is the inspector who normally does the inspection. They have to be in the room, labelling the images, calibrating the model and challenging the false positives.

Practical training-data discipline:

Start with historical defect data, if it exists in a usable form. Many operators have years of inspection photographs sitting in attachments or shared drives.
Label against an explicit defect taxonomy, agreed with the inspectors. Vague labels produce vague models.
Calibrate inter-inspector variability before you train the model. If two inspectors classify the same defect differently, the model will learn whichever one labelled the most images.
Keep a hold-out set the inspectors agree represents the field. Validate against that, not against a clean lab set.
Plan for retraining from day one. Defect populations change; assets change; weather changes; the model has to be operated, not deployed.

Inspector workflow design

The model lives inside the inspector’s day. If the inspector becomes slower with the model in the loop, the model is wrong, no matter what the accuracy number says.

Patterns we see work:

The inspector takes the image. The model classifies. The inspector confirms or overrides on-device. Confirmed and overridden classifications are equally valuable as feedback.
The override flow has to be easier than the confirm flow on the day-1 model. As trust builds, the balance shifts.
The audit trail records who saw what, when, and what they decided — for every image, every classification, every override. This is what makes the model defensible to a regulator and to internal safety review.
The inspector is part of the model retraining cycle. They are not a passive consumer.

This typically pairs with a Maximo Mobile rollout, because the practical workflow lives on the inspector’s device.

Audit and evidence trail

Safety-critical inspections eventually meet a regulator. Visual Inspection helps if the audit trail is designed in from the start.

What good looks like:

Every classification recorded against the work order, the asset, the inspector, the timestamp, the model version
Every override recorded with reason
Every model version retained, with the training data that produced it
A reproducible answer to the question “why did we classify this defect as low-severity in March”, three years later

This is not glamour work. It is the work that turns “we have a model” into “we have an inspection regime”.

Scoping a credible first programme

A credible first Visual Inspection programme has six elements.

Workflow scoping with the inspectors who actually do the inspection
Image and video labelling against historical defect data, with inspector calibration
Model build and validation using the MAS Visual Inspection lifecycle
Manage integration so classification becomes part of the work record, not a parallel system
Mobile and field workflow design so inspectors are faster, not slower
Audit and evidence trail so the basis of every decision is reproducible

Questions to ask a supplier

“Which inspection workflow would you start on, and why that one?”
“How would you build the training set, and which of our inspectors would be in the room?”
“How do you handle inter-inspector variability before training the model?”
“How does a classification get into the work record in Manage?”
“Who retrains the model after go-live, on what cadence, and what triggers a retrain?”
“What does the audit trail look like in three years, when a regulator asks why we classified a defect a certain way?”

Closing position

IBM Maximo Visual Inspection is one of the few AI-led MAS components where the payback can be measured in inspection cycles, not in dashboards. It rewards discipline in training data, in inspector engagement, and in audit trail design. It punishes the supplier pattern of training a generic model in a lab and dropping it on a workflow it has never seen.

For where Visual Inspection sits inside the wider suite, see the MAS suite overview. For the implementation pattern in detail, see IBM Maximo Visual Inspection: implementation and managed services.