Choose work that can be checked
Start with classification, retrieval suggestions and draft generation. Outputs should lead back to original material rather than just provide a confident answer. Keep explicit human approval for payments, legal commitments and official filings. Fluent output is not a reason to expand permissions.
Design sources and permissions together
Identify which material may be used, who may access it and when it expires. Retrieval must respect existing access permissions. Before sending data to an external service, confirm its processing boundaries and minimise personal information. Use redacted samples in testing.
Test failure cases
Test missing material, conflicting sources, stale information and misleading instructions. The system should express uncertainty, identify conflicts and request verification. Record error types and human corrections, and compare work saved with review effort rather than counting answers alone.
Expand automation in stages
Move from read-only suggestions to reversible actions and then approval-gated execution, with a separate threshold for each stage. Keep logs, permission boundaries and a stop mechanism. The business owner should know what the system can actually do and who takes over when correctness cannot be established.