Pricing for Arize AI can become a discussion once prediction volume grows, especially for companies with very high inference traffic. Also, some advanced configuration still felt documentation-heavy. Junior engineers sometimes struggled understanding how to structure data sets correctly for meaningful monitoring. And honestly, alert tuning took more effort than expected. At first, we had way too many noisy alerts. The documentation for Arize AI explains APIs reasonably well, but operational scenarios were missing sometimes, such as how to monitor LLM hallucination drift or how to handle delayed ground truth labels. Those practical examples help a lot more than API reference pages. I think integration could still be smoother in some areas with Arize AI. We spent more time than expected normalizing schemas and mapping metadata between different ML platforms. If your organization has multiple teams with inconsistent naming conventions, our onboarding got messy pretty fast. On the user experience side, the dashboards are good overall, but some advanced workflows felt a little overwhelming for newer engineers. Our data scientists adapted quickly, but back-end developers sometimes struggled understanding which metrics actually mattered. I would also like tighter integration between infrastructure observability and ML observability. During an incident, we still jump between Arize AI, DataDog, Kubernetes logs instead of having one clear investigation flow.
