Control What AI Can See and Do

Meaning improves an answer. Permissions decide whether the agent was allowed to produce it. Object access, field access, and row access are separate checks, and all three affect the result.

Run as the asking user

A regional rep and a revenue leader should get different totals when their CRM visibility differs. That is correct behavior. The answer should state whose permissions were used and avoid presenting a scoped result as the whole company.

A hidden value looks like a blank value

Field-level security may return null instead of an error. The model cannot distinguish a suppressed renewal date from a genuinely empty one. Convert silent suppression into an explicit state before the value reaches the model.

Hard controls belong outside the prompt

Read-only means the tool handler has no write path or rejects every write operation. Object and field allowlists are checked against the final plan. A sentence asking the model to behave is useful guidance, but it is not enforcement.

Increase action scope one rung at a time

Start with supplied context, then live reads, recommendations, proposed writes for approval, and finally narrow automatic writes. Most useful CRM work stays in the middle. A polished demo is not a reason to skip the approval rung.

Try to force a write in the GTM Lab. Completion requires a refusal at the tool boundary.

Reference

• Permissions and safety for CRM AI agents →
• Production readiness checklist →