# AI Audit Trail: What It Is and What Auditors Ask For > An AI audit trail records what an AI system decided, why, and under which rules. See the six parts of a record that holds up with an auditor. - **URL:** https://aisquare.studio/blogs/ai-audit-trail - **Published:** 2026-11-13 - **Author:** Omkar Shendge - **Category:** Agentic Governance - **Tags:** ai-audit-trail, ai-agent-compliance An AI audit trail is a record of what an AI system decided, why it decided it, which rules applied, and proof that the record has not been changed. Application logs cover the first part. Auditors and regulators ask about all four. ## The question that exposes the gap A regulator asks why your AI agent made a specific decision last year. Is that a five-minute answer or a weeks-long scramble? For most teams it is a scramble. Engineers pull traces from one system, prompts from another, and the policy document from a shared drive. Then they write an account that connects them. That account is a reconstruction, and an auditor will treat it as one. ## Why logs are not an audit trail - **Logs record events.** They show a tool was called. They do not show the reasoning that led to the call. - **Logs can be edited.** A record the operating team can change is weak evidence. - **Logs miss the rules.** They rarely show which policy version was active and whether the agent obeyed it. - **An agent’s own summary is not evidence.** A system cannot grade its own work. Logs and traces are still the raw material. An audit trail is what you get when you add reasoning, rules and a seal on top. For the difference between traces and reasoning, see [agent observability](/blogs/agent-observability). ## The six parts of an AI audit trail 1. **Input and context.** What the agent was asked and what data it had at that moment. 2. **Reasoning.** The claims the agent made, the evidence behind each one, and the assumptions it relied on. 3. **Rules in force.** Which policies applied, which version, and whether each check passed. 4. **Action taken.** Every tool call and every response sent. 5. **Human involvement.** Who reviewed, approved or overrode the decision. See [human in the loop](/blogs/human-in-the-loop). 6. **Signature.** Cryptographic proof that the record was sealed at the time and has not been altered. ![Diagram of an AI audit trail record with six layers: input and context, reasoning, rules in force, action taken, human involvement and signature](/lovable-uploads/blog-ai-audit-trail-diagram.webp) If any one of these is missing, someone will have to fill the gap from memory during an audit. ## The liability stays with you Air Canada’s support chatbot gave a passenger wrong information about a bereavement fare. The airline argued the chatbot was responsible for its own words. The tribunal disagreed and held Air Canada liable. New York City’s MyCity chatbot told business owners they could do things that were illegal in New York. In each case the organization had to explain what its AI said and why. A complete record turns that into a lookup. ## Which frameworks expect an AI audit trail? | Framework | What it expects | | --- | --- | | NIST AI RMF | Documented, traceable AI risk management | | EU AI Act | Record keeping and transparency for high-risk systems | | SR 26-2 (replaced SR 11-7) | Model risk management and documentation for banks (generative and agentic AI are out of its scope) | | DORA | Operational resilience and incident records in EU finance | Check with your counsel on which apply to you. The common thread is a defensible record of AI decisions. ## What good looks like Use this as a checklist against your current setup. - **Captured at runtime.** The record is written as the agent runs. Nobody rebuilds it later. - **Independent.** It is produced outside the agent, so it does not depend on the agent’s account of itself. - **Readable.** A compliance lead can follow the reasoning without an engineer in the room. - **Versioned.** You can tell which model, prompt and policy version made the decision. - **Sealed.** Each record is signed, so tampering is detectable. - **Exportable.** It comes out in formats an auditor’s tools can read. - **Durable.** It verifies the same way next week and five years from now. ## How long to keep records Retention depends on your sector. Financial and healthcare records often need to be kept for years. Plan for records that stay verifiable for the full period, including after you change models or vendors. ## Where AISquare fits AISquare captures each agent run as it happens and stores the reasoning in RML, proposed as an open format for reasoning evidence. Records are signed at the time of the run and export as PDF, JSON, in-toto statements and DSSE envelopes. Each record sits alongside an [AI BOM](/blogs/ai-bom) that shows which components were involved. See how it fits into [AI agent governance](/blogs/agentic-governance). ## FAQ ### What is an AI audit trail? It is a tamper-evident record of an AI system’s inputs, reasoning, applied rules, actions and human approvals for each decision. ### Are logs enough for an AI audit? No. Logs show what happened. An audit also needs why it happened, which rules applied, and proof the record is intact. ### Who needs an AI audit trail? Any company using AI for decisions that affect customers, money or compliance, and especially teams in finance, healthcare, insurance and government. ## Next step See a signed decision record for one of your own agents. **[Book a demo](https://calendly.com/d/dws4-msf-8jc/aisquare-gtm-team)**. **Related reading:** [AI agent governance](/blogs/agentic-governance) · [AI agent guardrails](/blogs/ai-agent-guardrails) · [Agent observability](/blogs/agent-observability) · [AI governance tools](/blogs/ai-governance-tools) · [AI hallucination examples](/blogs/ai-hallucination-examples) · [AI governance framework](/blogs/ai-governance-framework) · [Human in the loop](/blogs/human-in-the-loop) · [AI BOM](/blogs/ai-bom) --- **Machine-readable AISquare** - Site index for LLMs: https://aisquare.studio/llms.txt - Agent instructions (when to use AISquare): https://aisquare.studio/agents.md - OpenAPI specification: https://aisquare.studio/openapi.json - MCP endpoint (Streamable HTTP, POST): https://aisquare.studio/mcp - Sitemap: https://aisquare.studio/sitemap.xml - Blog feed (RSS, full posts): https://aisquare.studio/blogs/rss.xml - Any page on this site also serves markdown: append `.md` to its path (the homepage is `/index.md`) or send `Accept: text/markdown`.