# Ask Ralti & review every change

Describe an outcome, inspect the evidence, then decide what to apply.

Ask Ralti can find information, explain records, propose updates, and design connected workbooks. Its context identifies the record, selection, sheet, workbook, or workspace in scope. A useful request names the destination, the desired result, and constraints: “Find overdue projects, then propose follow-up tasks without changing project status.”

## From request to approved change

- Start a conversation and answer any clarification questions. A question is not a failed request.
- For answers, open the cited records or conversations and inspect coverage notes.
- For changes, review the proposed operations, field mappings, record values, rules, and helpers. Adjust the draft or ask for a refinement.
- Apply the reviewed web proposal explicitly. Generating or previewing a draft does not change records.
- If a revision conflict appears, refresh or regenerate against current data, then review again.

## Recover a saved request

AI requests are saved on the server and can continue when a browser tab closes. Reopen the workspace to inspect progress or recover a completed proposal. Stop generating requests cancellation; inspect the resulting status. New chat begins another conversation. Conversations are private to their creator, including inside a shared workspace.

> **Provider and data boundary** Relevant workspace context is sent to the configured AI provider. Provider availability, indexing, and worker capacity affect response time. The built-in helper is an explicit, limited rules-and-templates option; it is not a local language model.

## Know the control boundaries

Viewers can ask read-only questions; server permissions reject mutations. AI retrieval is scoped to accessible records and permitted mailbox content. Search matches are not exhaustive evidence. External AI connections can save proposals but cannot directly apply them. Saved automatic agents are separately authorized helpers, with their own review rules.

If a request remains queued, the operator should check the background worker. If the provider fails, keep the current draft, retry later, or continue manually. Repeatedly submitting the same request can create additional jobs without resolving the underlying problem.