GoodData splits AI observability into aggregate metrics and on-demand interaction traces

GoodData.AI has launched an observability product for enterprise AI that combines organization-wide usage analytics with step-by-step traces of individual agent interactions. The company says the product tracks adoption, quality signals, token consumption and cost, then lets operators inspect the skills, retrieved knowledge, memory, model calls, failures and timing behind a particular response. It can run in GoodData’s cloud or on customer-controlled infrastructure, according to the launch announcement.

The more consequential design choice is not the dashboard. It is the boundary between aggregate monitoring and conversation inspection.

Broad metrics without broad transcript access

GoodData’s product page says its Usage Analytics layer works from aggregate usage and performance metrics rather than conversation content, which the vendor presents as suitable for broad rollout without first making a transcript-privacy decision. When an operator needs to investigate a particular answer, the underlying interaction is opened separately and on demand.

That split matters for natural-language analytics systems. A platform team may need to know that one workspace’s token use doubled, that an agent’s negative-feedback rate rose or that a skill stopped being adopted. None of those questions necessarily requires every dashboard viewer to read the business questions users asked or the data returned to them. GoodData’s published material describes this separation, though it does not specify the role matrix, retention defaults or redaction controls that determine how strong the boundary is in practice.

Tracing the path, not only grading the answer

For an individual interaction, GoodData says operators can inspect which skills were considered and activated, what knowledge and memory were retrieved, which model calls ran, where execution failed, and the timing, iterations and token use for each step. Its technical explainer argues that agent observability must capture tool calls, retrievals and orchestration steps rather than only the final input and output; it also recommends treating prompts, responses, tool arguments, retrieved content and memory as sensitive telemetry that may require redaction and access controls (GoodData engineering guide).

For text-to-SQL teams, that can separate several failures that look identical to an end user. A wrong answer may begin with the wrong skill, the wrong semantic definition, an irrelevant retrieval, a failed query or a model-generation error. A single quality score cannot identify which component needs repair; a trace can at least locate the suspect step.

GoodData also says the product can identify recurring issues across conversations and recommend changes to knowledge, semantic models or configuration. Those are vendor claims, not reported benchmark results. The launch materials reviewed by nl2sql.ai do not publish accuracy gains, customer outcomes, pricing or a reproducible evaluation of those recommendations.

The practical buying question is therefore narrower than whether the new dashboards look complete: can administrators prove that aggregate viewers cannot reach sensitive interaction content, while investigators retain enough governed detail to reproduce a bad answer? GoodData has articulated that architecture. Enterprise evaluations should now test the permissions, redaction, retention and audit behavior behind it.