Zowie

Zowie — Complete AI-agent customer-service traces with ground-truth outcomes: a normalised Users → Tickets → Interactions → Events hierarchy in which the Events table exposes every message, every reasoning step, every knowledge retrieval and every tool invocation the AI agent made — joined to intent and topic labels, channel (chat/email/voice), AI-handled vs transferred-to-human status, SLA/first-response/resolution timings, and final_contact_reason / final_topic outcome fields.. Top-tier for this pool: large row counts AND the most sought-after record class in agent training — full reasoning…

Zowie offers (Alternative, Sentiment, Fundamental) — Very large and unusually deep: every conversation multiplied by its event stream — each interaction contains many events (messages, reasoning steps, retrievals, tool calls), so at hundreds of thousands of interactions per month the Events table plausibly holds 10^8+ rows per year. Zowie absorbs the BigQuery storage cost for the shared dataset, implying material volume..

Agentic trajectory training data: (thought, retrieval, tool call, observation) sequences with a terminal success/failure label — the scarcest dataset in agent development; escalation prediction (when the agent should hand off to a human); tool-selection and function-calling supervision; RAG retrieval-relevance labels (retrieved passage vs whether it resolved the ticket); intent/topic classification across ecommerce verticals; conversational preference/RLHF data from human-vs-AI resolution comparisons; voice-support ASR + intent corpora; autonomous-resolution-rate benchmarking

The data is with 6 years of history.

Coverage spans US; Application Software; alternative, sentiment, fundamental; equities.