Ask AI

When a customer needs a reply or a client handoff needs work, tell Ask AI the result you want. It can read the relevant records, prepare changes and show their results in the same conversation. Open it from Today or the AI Workspace, then name the customer or record.

Ask something specific

A clear target and a clear result you need beats a vague prompt every time. Some real shapes that work well:

• "What needs my attention today?"
• "Show the recent timeline for this contact."
• "What is the next action on this opportunity?"
• "Draft a follow-up for this conversation, using the latest reply."

Choose the relevant record when the interface offers you that context, and resolve an ambiguous person or company before you ask for a change, so the assistant isn't guessing which one you meant either.

For recurring billing, Ask AI can explain only the current workspace records and registered capabilities. Stripe remains the payment authority. Use the customer account for checkout, invoices, payment details, cancellation, and plan changes; the assistant cannot charge or change a subscription unless a registered tool explicitly offers that action.

Reading and replying

Answers follow the latest response while you stay at the bottom. Scroll up to review an earlier message without being pulled away by new text. Choose Jump to latest to resume following the answer. Sending another question or switching conversations takes you to the latest message. Opening a change review pauses following and keeps the decision controls clear.

The input grows as you write, up to several lines. Enter sends your question; Shift+Enter adds a line. Confirming characters with an input method does not send unfinished text. Use Stop AI run to interrupt a response, then inspect any staged proposals before retrying.

Checking the evidence

Read the tool results or record links behind an answer before you act on it. If the answer is partial, look at what was unavailable rather than treating a gap as a fact the assistant confirmed; a disconnected provider or an unavailable tool can limit what it's able to answer, and the honest response to that is telling you it doesn't know, not guessing.

Reviewing anything it proposes

A request can produce exact proposals. Choose Review beneath the answer to inspect the recipient, wording or record changes inside the conversation. Approve or reject each change there. After opening one exact review, you can also type “approve” or “reject”; those words are handled as your decision, never as an agent tool.

Approval records your decision. The result records what happened, including a failed or partial action. Returning to the conversation restores pending proposals and their recorded statuses. Work → Approvals remains available for the full action receipt. If a run stops early, inspect its proposals before repeating it.

Delegate an ordered job

Ask for a plan such as “Read this client's records, prepare the handoff tasks, then draft a follow-up.” The agent previews up to eight ordered steps before starting saved work. It can continue after you close the browser when the workspace scheduler and a finite AI call budget are configured.

Use Read progress in chat to retrieve saved steps and receipts. Pause stops future work, Resume continues eligible steps, and Cancel stops future steps while retaining completed changes. A step awaiting an action decision cannot advance until the required actions have finished. Interrupted steps need reconciliation rather than being repeated automatically.

Allow bounded internal work

Workspace administrators can ask the agent to prepare permission for a specific internal operation on named records. The exact review includes allowed fields, an expiry within 30 days and a daily action limit. The requesting administrator must approve the permission themselves.

A new grant replaces the previous workspace permission for that operation, including its record and member scope. Include every record you want covered. The audit records the previous and new constraints; granting again does not reset actions already counted toward that day's limit.

Eligible operations cover tasks, notes, contact tags, opportunity next actions and open pipeline stages. Current membership, policy, record scope, permitted fields, expiry and remaining daily allowance are checked before execution. Messages, publishing, billing, deletion and permission changes still require human approval. If permission does not cover a proposed change, it stays pending for review.

Try the fictional workspace

The public demo offers a real model when its dedicated inference service is configured and within its shared $5 daily budget. It works with fictional records. Every task, reply and reminder effect is simulated in your browser session. If inference is unavailable, the workspace remains explorable and shows the failure plainly.

What to check

• AI tools lists every registered tool, what it needs to be connected, and its input schema, if you want to see the full capability behind an answer.
• Developers extending the registry with a new tool should read Add an AI tool.