Standard event schema for AI product analytics (www.rudderstack.com)

A new AI Product Spec proposes a standard event schema and privacy-aware workflow to help teams measure usage and impact of chatbots, assistants and copilots. The guide defines three core events—ai_user_prompt_created (when a user submits a prompt), ai_llm_response_received (LLM output plus performance and cost metrics), and ai_user_action (user feedback or downstream actions)—and shows how to tie them together with a conversation_id. That standardization makes cross-product metrics, cost analysis, and failure/latency tracking practical, closing the “measurement gap” where AI features are operational but not optimizable or tied to business outcomes. Technically, the spec prescribes required properties (conversation_id, latency_ms, token_count, response_status, model_used, etc.), distinguishes sessions from conversation threads, and recommends intent classification to avoid storing raw prompts. The recommended privacy pattern: classify prompts into 5–10 business intents (plus “other”) using LLMs or lightweight classifiers, enrich events via RudderStack Transformations (example uses OpenRouter) and only deliver classified intents to warehouses/CDPs. The guide also covers model selection (cheaper small LLMs or DistillBERT for simple classification), prompt engineering, and retaining raw text only in dev. For AI/ML teams, this lowers analytics complexity, reduces data sensitivity risk, enables accurate cost/usage tracking, and provides actionable signals for product iteration.