Safety & good practices

The assistant is built to be a safe, focused investigator that can also take a few specific actions when you ask. This page explains the guardrails that protect you, and how to get the best out of it.

What the assistant will and won't do

• Explain mode is genuinely read-only. It has no ability to change anything — not a limit it's asked to respect, but a capability it doesn't have. Ask it to change something and it tells you to switch modes.
• Edit mode does a short, fixed list. Notes, interactions, metafields, and resolving or reopening a claim. For anything else it stays read-only and says so, rather than improvising a workaround.
• It never acts without your confirmation. Notes and interactions are shown to you before they're saved; metafield, resolve and reopen changes are described and confirmed in the chat first. Naming a claim isn't confirming a change.
• It can't exceed your permissions. Every change goes through the same Role Manager permission the manual action needs.
• It acts on one claim at a time. Bulk changes through the assistant aren't possible by design.
• It stays on the record in front of it. It won't act on claims outside the account you're on, and it steers off-topic questions back to what it can answer.
• It doesn't invent evidence. When the data won't support an answer it says so. See Evidence & grounding.

Getting the best results

• Ask before you act. Check the current state — "what is riskLevel set to right now?" — before asking for a change. The assistant does this itself, but it's how you catch a mistaken assumption early.
• Name the claim. Use an @mention for the exact claim rather than describing it, especially on accounts with several similar claims.
• Be specific about the field and value. "Set riskLevel to HIGH" beats "mark this as risky".
• Ask follow-ups. The conversation keeps its context, so drill in rather than restating the whole question.
• Ask for the evidence. If an answer is surprising, ask which event it came from.
• Read the confirmation. It names the exact change. It's the last checkpoint before something real happens.

Where judgement is still yours

The assistant is strong on what the record says — dates, values, who changed what, which claims are overdue. It's interpreting, not reporting, when it explains why a claim is late, since no reason is stored anywhere.

It's also honest about the boundary of its evidence. An answer of "I can't find that" means the record doesn't show it — not that the assistant gave up. Treat that as a signal to look elsewhere.

tip

Think of the assistant as a fast, thorough colleague who has read the entire timeline. You bring the judgement about what should happen; it brings the evidence about what did.