OpenTelemetry GenAI semantic conventions

A shared set of attributes for representing model calls, token usage, tool invocations, and related metadata in a portable trace format. They let different observability backends read the same AI execution data without custom schemas or one-off instrumentation.

Expanded Definition

OpenTelemetry GenAI semantic conventions are a standardised attribute set for describing generative AI activity inside traces, spans, and related telemetry. They provide a common vocabulary for model requests, responses, token counts, prompt and completion metadata, tool calls, and error conditions so observability data can move between backends without being remapped each time. In practice, these conventions sit at the boundary between application observability and AI governance, because they make GenAI behaviour visible in a way that can be measured, searched, and correlated.

Definitions vary across vendors and implementation guides, because OpenTelemetry evolves through community specification rather than a single regulatory standard. That makes the term especially useful, but also easy to overstate: semantic conventions do not secure a model, prove correctness, or guarantee safety. They simply make GenAI telemetry more interoperable and more structured. For governance-oriented context, the NIST AI 600-1 GenAI Profile is a useful reference point for aligning observability with risk management expectations. The most common misapplication is treating semantic conventions as a complete GenAI monitoring strategy, which occurs when teams instrument traces but fail to define the risks, alerts, and response actions those traces should support.

Examples and Use Cases

Implementing GenAI semantic conventions rigorously often introduces instrumentation overhead and governance work, requiring organisations to weigh richer visibility against added engineering effort and telemetry volume.

• A chatbot service records prompt length, response length, model name, and latency in a consistent span structure so SRE teams can compare behaviour across environments.
• A retrieval-augmented generation workflow logs tool invocations, document retrieval events, and token usage so investigators can reconstruct how an answer was produced.
• An internal agent platform tags each model call with tenant, workflow, and tool identifiers so usage can be segregated for auditing and chargeback.
• A security team correlates GenAI span attributes with IAM and API gateway logs to detect abnormal tool access or prompt injection attempts.
• An engineering group adopts the OpenTelemetry patterns described in the OpenTelemetry GenAI semantic conventions to avoid custom fields that break dashboard portability later.

These uses are strongest when the organisation needs consistent tracing across multiple services, model providers, or orchestration layers. They are less effective when telemetry is collected without a clear purpose, because the resulting data can be detailed yet still operationally ambiguous. The conventions help teams answer what happened, where it happened, and with which model-related components involved, but not whether the output was safe, compliant, or correct.

Why It Matters for Security Teams

For security teams, OpenTelemetry GenAI semantic conventions turn generative AI from an opaque application layer into something that can be investigated, governed, and trended. That matters when models access secrets, call tools, or participate in workflows that affect privileged actions. Standardised traces make it easier to detect abnormal invocation patterns, identify which agent or service used a tool, and preserve evidence when a prompt injection, data leakage, or misrouted request needs review.

The identity connection is especially important in environments using non-human identities and agentic AI, because the telemetry may be the only reliable way to show which service account, API key, or delegated workflow initiated a model action. This is where observability and NHI governance overlap: trace data can support accountability, but only if the surrounding controls already define ownership, least privilege, and escalation paths. The OpenAI function calling guidance is not a control framework, but it illustrates why tool-use visibility matters when model outputs trigger downstream execution. Teams should also align telemetry design with broader AI risk practices described in the NIST AI 600-1 GenAI Profile. Organisations typically encounter the urgency of this term only after an incident review cannot explain which model call initiated a risky action, at which point semantic conventions become operationally unavoidable.

Related resources from NHI Mgmt Group

• What makes GenAI usage part of the same secrets problem?
• When does a GenAI integration become a non-human identity risk?
• When does GenAI create more identity risk than business value?
• How can organisations reduce the blast radius of compromised GenAI credentials?