Hexagon applies AI to measurement, inspection and equipment health

Hexagon’s AI work is closely tied to measuring the physical world. Two concrete examples are training models to segment industrial CT scans and monitoring the condition of metrology equipment. These are narrower, more operational tasks than general-purpose chat. They also make the company’s recent portfolio changes important: some software that used to belong to Hexagon now belongs to other companies.

01 / ProductThe current portfolio starts with measurement and inspection

Hexagon’s Octave separation page records completion on 28 May 2026. Asset Lifecycle Intelligence, Safety, Infrastructure & Geospatial, ETQ and Bricsys moved into the separate company, with no continuing Hexagon ownership. Separately, Hexagon completed the sale of Design & Engineering to Cadence on 23 February 2026. Older product lists can therefore misidentify the current supplier.

This blueprint concentrates on retained measurement-related products. VGTRAINER trains segmentation models from curated CT data without requiring users to write model-training code. The product page describes an on-premises application without cloud dependencies. Its output can be used in the company’s established inspection environment, connecting a learned model to the work of a quality specialist.

Deep Segmentation applies trained ONNX models in VGSTUDIO MAX and VGinLINE, producing regions of interest for subsequent analysis. APOLLO addresses a different problem: monitoring coordinate measuring machines and machine tools for anomalies and condition changes. Inspection of a part and monitoring of the equipment that measures it should be evaluated as separate workflows.

02 / AudienceQuality teams need stable tasks and a way to establish ground truth

A CT inspection team that repeatedly separates materials, internal features or defects from scan data has a concrete model-training problem. The best first task is one for which specialists can create reliable reference labels and explain the mistakes that matter. A visually convincing segmentation is insufficient if a small missed region changes the downstream quality decision.

The C3 AI blueprint offers a broader industrial-application comparison for teams investigating operational analytics. Hexagon’s specific advantage here is its connection to measurement and inspection tooling. The right starting point depends on whether the buyer needs to interpret scan volumes, monitor equipment or combine broader business and asset data.

The Cadence blueprint is relevant for readers following the transferred engineering-software portfolio. That corporate distinction is practical, not merely historical: it affects the supplier, product roadmap and support route a buyer should investigate. A proposal for simulation software should not be evaluated under Hexagon’s retained CT inspection offer simply because an older document uses the Hexagon name.

Organizations without labeled examples should budget expert preparation before expecting useful training. The scarce resource may be consistent interpretation of scan boundaries rather than computing capacity. Use disagreements between specialists to improve the inspection definition; feeding inconsistent labels into a convenient training interface does not resolve the underlying ambiguity.

03 / WorkflowA proposed CT segmentation pilot follows errors into the final measurement

Consider a manufacturer inspecting a recurring family of cast components. The following is a proposed evaluation, not hands-on testing. Choose a bounded segmentation task and define how the resulting region will be used. For example, the quality team might need to isolate a material region before applying an existing measurement or defect-analysis procedure.

Prepare representative scans from the real inspection process, including variation in geometry, material and scan quality. Keep a separate evaluation set that is not used for training. Specialists should agree on the labeling convention and document ambiguous boundaries. Avoid treating multiple near-identical scans of the same part as strong evidence of performance across the production population.

Use VGSTUDIO MAX to prepare and inspect the relevant CT data and labels, then train the intended segmentation task in VGTRAINER. Keep the data preparation, training configuration and resulting model associated with a version identifier. A later correction should be traceable to the examples or labeling rule that changed.

Apply the trained model through Deep Segmentation and inspect both the region and the downstream result. A segmentation can have a large amount of pixel overlap with a reference and still miss a small consequential feature. Have the quality team define acceptable errors in terms of the actual inspection decision, rather than adopting a generic accuracy percentage from a demonstration.

Include scans near the boundaries of the expected operating conditions. Examine cases with different noise, reconstruction artifacts or geometry variations, and decide when the system should defer to manual review. The pilot should identify the region in which the model is useful as well as the cases where its output cannot support the normal inspection process.

If the result is promising, repeat on another production batch before integrating it into a routine workflow. Track correction effort and false acceptance risk separately from processing time. A faster first pass can still be valuable, but only if the reviewer understands where it makes mistakes and the downstream process preserves that review.

APOLLO deserves its own evaluation if equipment reliability is the bottleneck. Compare its alerts with maintenance history and the team’s actual response process. A warning has value only when someone can investigate and act on it at the right time. Do not combine a successful segmentation trial with an unrelated assumption that predictive maintenance has also been validated.

04 / PricingLicense form, service coverage and evaluation restrictions affect the budget

The VG licensing page offers floating, node-locked and dongle arrangements, with availability varying by product. It lists VGTRAINER under floating and node-locked licensing and states that new licenses include a mandatory one-year update/service agreement. Public pages direct buyers to a quote rather than publishing a universal software tariff.

The VGTRAINER page says models created with evaluation licenses expire and that using them watermarks the project irreversibly. That is a consequential limitation for a pilot. Use disposable copies and establish the production conversion path before investing heavily in training an evaluation model that cannot become part of the final inspection process.

Hardware belongs in the estimate as well. The published VGTRAINER requirements include Windows 11 and at least 128 GB of RAM, with actual memory needs depending on the task. Validate the full workstation specification against the intended scan volumes and training workload. A license quote alone is not the cost of setting up a suitable training station.

For a distributed quality organization, decide whether training and inspection occur at the same site. Shared licensing may change how teams schedule access, while a production inspection machine needs a predictable operating arrangement. Ask for a quote that separates model preparation, training, inference, service coverage and any equipment-monitoring product so the responsibilities remain clear.

05 / DistinctionsLearned segmentation is attached to a measurement workflow

The meaningful distinction is the path from a trained model to an inspection region and then to an established analysis. That makes Hexagon relevant to teams that need more than a standalone computer-vision model. They need the output to remain interpretable inside the tools and procedures used to decide whether a part is acceptable.

Local training is another concrete distinction for organizations managing sensitive part data. It can reduce dependence on cloud services for this specific product, although the team still has to operate the workstation, protect the data and maintain the model. The absence of a cloud dependency is not an automatic guarantee that every configuration is suitable for every customer’s requirements.

APOLLO broadens the company’s AI relevance to the measurement equipment itself. Hexagon describes support for its own and third-party devices, with cloud and on-premises deployment options. Its claimed prediction horizons are not treated here as expected results. The relevant evidence for a buyer is useful warning time on their equipment and a maintenance process that can use it.

06 / QuestionsThe hardest questions concern labels, drift and deployment

Ask how performance changes when the scanner, reconstruction settings or part family changes. A model trained under one set of conditions may need additional evidence before it is trusted under another. Establish a review trigger when the incoming data moves beyond the pilot’s scope, and keep old model versions available for understanding previous results.

Examine false positives and false negatives separately. A highlighted region that wastes a reviewer’s time and a missed defect that affects acceptance are not equivalent outcomes. The evaluation should reflect the quality team’s actual decision and the consequences of each error, rather than reducing everything to a single overall score.

Finally, confirm the current supplier and support route for every quoted component. The completed corporate transactions make this especially relevant for mixed portfolios. A retained Hexagon product, an Octave product and a Cadence engineering tool may still appear together in historical material, but that does not make them one current commercial package.

07 / DecisionChoose a repeatable inspection task with an expert owner

Hexagon is a strong AI-related candidate for manufacturers because its learned models can operate close to physical measurement. Begin with a task whose inputs, labels and downstream acceptance criteria are understood. Expand only after the team can explain the model’s useful range, preserve the evidence behind each version and run the intended deployment under the appropriate licence.