- When roadmap stability is critical
- Microsoft documentation indicates planned retirement activities for the Azure Custom Vision service; if a multi-year, stable platform without migration work is required, consider alternatives or plan to adopt Microsoft’s successor offerings.
- When full model control or best-possible accuracy is required for a hard domain
- Advanced teams may prefer an end-to-end framework (for example, Azure Machine Learning with custom models or open-source training stacks) that provides granular control over architectures, augmentation, and training recipes.
- When very large label taxonomies are required
- If thousands of categories are required within one model, community-reported limits on tags per project could be a blocker; alternative architectures or product choices may be required.
- For always-on, high-FPS video or continuous camera analytics in the cloud
- Cost and latency considerations make sending every frame to a cloud API impractical; edge inference, on-device models, or specialized video analytics pipelines are more appropriate.
- For tasks requiring segmentation, pose/hand tracking, or advanced video understanding
- Custom Vision is primarily focused on classification and object detection. For richer video analysis and tracking tasks, other stacks or models are likely needed.
- Product lifecycle and migration risk
- Microsoft documentation indicates planned retirement activities for the Azure Custom Vision service; buyers should obtain clear migration guidance and timelines.
- Successor capabilities
- Microsoft is positioning a newer model customization capability in Azure as an evolution of Custom Vision with different accuracy and few-shot learning characteristics; buyers should evaluate feature parity and migration effort.
- Developer experience friction
- Multiple keys and related calling conventions have been reported as a developer friction point.
- Limited low-level tuning and certain fine-grained tasks
- The service abstracts model internals, which can limit the ability to perform custom tuning required for some high-precision tasks.
- Service limits
- Community reports indicate tag limits per project in certain SKUs; buyers with large taxonomies should validate limits and support channels.
- Availability of third-party review data
- Some review platforms did not surface detailed review text consistently in the materials reviewed; sentiment analysis in this report relies more heavily on available review platforms and public community discussion.