AI Agent Audit Intake Checklist

Paste what you have. Missing items are fine; the audit will mark confidence and unknowns.

Pick The Surface

• Code repo: local files, GitHub repo, or uploaded source.
• Workflow builder: n8n, Zapier, Make, Retool, Voiceflow, Botpress, or similar.
• SDK/API: OpenAI Agents SDK, Assistants, LangGraph, LangChain, CrewAI, AutoGen, Semantic Kernel, or custom stack.
• Black-box: transcripts, screenshots, demos, or observed behavior only.
• Client report: audit output needs to become a formal deliverable.

If you are unsure, start with the general sections below and include whatever screenshots, exports, or traces you have.

Agent Overview

• What does the agent do?
• Who uses it?
• What actions can it take?
• What would a bad failure look like?
• Where does it run?
• What is the highest-stakes thing it can do without a human?

Prompts

• System prompt:
• Developer prompt:
• User prompt template:
• Tool-selection instructions:
• Refusal/escalation instructions:

Tools

For each tool:

• Name:
• Description shown to the model:
• Inputs:
• Outputs:
• Permissions:
• Side effects:
• Can it read private data?
• Can it send data outside the system?
• Is there a human approval step?

Retrieval / Memory

• What sources can the agent retrieve from?
• How are documents chunked?
• How are chunks ranked or reranked?
• How is freshness handled?
• Are citations required?
• Can user-supplied or web content enter the same context as privileged instructions?

Evals

• What test cases exist?
• What metrics are tracked?
• Are failures sliced by task type, customer, language, or risk?
• Is there a regression suite?
• Are real production traces reviewed?

Logs / Traces

Include 1-3 successful runs and 1-3 failed runs:

• User input:
• Retrieved context:
• Model output:
• Tool calls:
• Final result:
• What went wrong or right:

Constraints

• Latency target:
• Cost target:
• Autonomy level:
• Human review requirements:
• Sensitive data involved:
• Compliance or legal constraints:

n8n / Workflow Builder Exports

• Exported workflow JSON or blueprint:
• Screenshots of AI/prompt nodes:
• Screenshots of action/tool nodes:
• Credential scopes, described without secrets:
• Error branches and retry settings:
• Sample execution log:

Code Repo / SDK Artifacts

• Agent entrypoint files:
• Tool/function schema files:
• Retrieval or memory files:
• Eval/test files:
• Trace/log examples:
• Deployment config:

Black-Box Evidence

• Public docs or screenshots:
• 5-10 transcripts:
• Known complaints or incidents:
• Probe questions you are allowed to run:
• What you cannot inspect: