Inventory the systems
List source systems, available APIs, record ownership and access permissions. Check whether information is accurate and current before connecting an assistant.
- Authorised data sources
- Read-only vs write access
- Retention and audit requirements
- Fallback when a connection fails
Design the trust boundary
Separate retrieved information from instructions. Treat incoming documents as untrusted content; validate structured outputs before passing them to operational tools.
Use human approval for irreversible actions. Log enough to investigate errors without retaining more personal data than necessary.
Test the end-to-end workflow
Test ambiguous requests, missing records, access-denied states and provider outages. A successful demo on one clean example is not evidence of production readiness.
Questions worth asking
Can AI connect to our CRM?
Often, if access is authorised and a suitable integration interface exists. Scope it around specific reads or actions rather than broad access.
Advice here is general; the appropriate implementation depends on the people, systems and constraints involved.