Data access is where many AI agent risks become real. A model mistake is annoying; a model mistake with broad customer data, write permissions, and persistent memory can become an incident.
List every data source
Include application databases, uploaded files, CRM records, support tickets, logs, vector indexes, temporary files, memory stores, and third-party APIs. Teams often forget that retrieved context is also data access.
Separate read, write, and remember
The agent may be allowed to read a document but not write a record. It may use context for one session but not store it in memory. Treat read, write, retrieve, and remember as different permissions.
Check tenant and role boundaries
A safe agent should not rely on prompt instructions to respect tenant boundaries. User role, account scope, and object-level authorization should be enforced outside the model path.
Review exposure paths
Ask where sensitive data can leave the system: chat output, tool arguments, logs, analytics, email notifications, exported reports, and support screenshots. Logging can become a second data leak.
Define deletion and retention rules
Decide how long prompts, traces, retrieved snippets, uploaded files, embeddings, and generated reports are retained. A launch review should not leave retention as an implementation detail.
Quick checklist
- Inventory data sources before adding new tools.
- Separate read/write/memory permissions.
- Enforce tenant boundaries server-side.
- Mask sensitive data in logs where possible.
- Document retention and deletion behavior.
How to use this
Use this guide as a working review aid, not as a compliance certificate. A team preparing an AI agent for production should pair it with traces, test cases, policy decisions, and a named owner for each unresolved risk.
Related IBBS resources
AI Agent Data Access Review Template: A practical template for reviewing what data an AI agent can read, write, remember, retrieve, and expose before launch. Read it at https://ibbs.ai/2026/07/07/ai-agent-data-access-review-template/
If your team is preparing an AI agent for production and wants an outside review, start with the AI Agent Readiness Self-Assessment or review the sample audit report.