Definition
An AI audit trail is proof: a time-ordered, attributable sequence of AI events across the systems a team actually uses. Each event should answer: who invoked it, which product or model, what was asked or decided, which policy applied, and what happened downstream. Identifiers, hashes, and parent events make the sequence exportable. Customer-owned environments keep that proof yours.
A trail is not a chat history
Vendor chat histories are siloed and aimed at the person who typed. They drop shadow use and line-of-business assistants. A trail is operator-shaped: one table, many sources, policy on the row.
A trail is not LLM observability
Observability instruments applications you built: spans, tokens, evals, gateway metrics. Copilot in Outlook, JAX on a bank rec, and a public ChatGPT paste never pass that tracer. That is an audit problem.
A trail is not DLP
Data loss prevention can block a paste. It does not attribute an allowed Copilot draft to a policy and a hash chain. Different controls.
Why the category exists
AI embedded in the tools of record. Work happened. Proof did not. Taigris is the AI audit trail: customer-owned and hermetic.
Minimum viable record
- Source and AI system.
- User and timestamp.
- Policy, including shadow.
- Input / output or a faithful summary plus downstream action.
- Stable IDs and a hash chain for export.
Questions
Related