Overview
Instead of answering extraction questions tile by tile in Agent view, you can finish an order by prompting the agent in plain language — or dictating your request with voice-to-text. Describe what you want — set a field, swap a product, split an address — and the agent applies the change to the order.Good to know
We’re still learning how best to grow and shape this feature, so there are a few rough edges worth knowing about while we do.- Table edits don’t always look like they applied. When you edit a table, the column labels, line numbers, and splits may not visibly update — even though the agent has made the changes. Double-check the product cart if something looks out of step.
- The agent can take longer than you’re used to. Rather than waiting on a single request, open another mail in a separate tab and keep working there. You can move through several requests at once instead of one at a time.
Open the chat
With a request open in Flow and the Agent view panel visible:- Click the Finish with agent pill in the Agent view footer. The pill shows how many questions are still unresolved.
- Click Go back at the top of the chat to return to the regular Agent view at any time.
What the agent can do
Click the What’s supported button in the top-right corner to see the current capability list. Today the agent can:- Fill in and update — set field values, or batch-edit a value across all product lines at once.
- Product catalog — look up and select products by number or name, and add or remove product lines.
- Smart lookups — pick from catalogs and reference data (account numbers, cost centers, etc.).
- Delivery addresses — parse and split one-time delivery addresses into the right fields.
- Query the documents — ask about the attached documents and pull values from them into the order.
- Undo and refresh — revert changes, and re-run the simulation to refresh values that depend on others.
- Submit — confirm and place the order when everything is ready.
Prompting or dictating
Type your request in the composer and press Enter, or use the microphone button to dictate instead of typing. While the agent works, a thinking indicator shows its progress. If a message fails to send, an error banner appears above the composer with the reason. The composer placeholder adapts to your browser language when dictation is available. It reads Type or record in {Language} — for example, Type or record in English or Type or record in Deutsch — so you know which language voice-to-text will transcribe. When dictation isn’t supported in your browser, the placeholder falls back to Type a message…. Example prompts:Choosing from suggestion options
When the agent needs your input to continue — for example when several products or accounts match your request — it attaches an option list to its message:- The list header shows the agent’s question and the number of options, for example (8).
- Long lists scroll, so all options stay reachable without stretching the chat.
- Each option can include extra details and a Source line showing where the value comes from.
- Click an option to select it. Your selection is sent as your reply and the list collapses automatically.
- Text inside options is selectable. Highlight and copy a value, like a product number, without sending the option as your reply.
- Click the header to collapse or expand the list yourself, for example to re-check an earlier question. Options from earlier turns are shown but can no longer be clicked.
Refreshing guidance
When you re-open the agent chat on a request you’ve been editing elsewhere — for example after making changes in the product cart or agent view tiles — a translucent overlay covers the conversation while the agent catches up with the latest state. The overlay cycles through short status labels (Refreshing…, Catching up on your latest changes…, Reviewing recent updates…, Bringing guidance up to date…) so you know work is in progress. The conversation stays visible underneath and becomes interactive again as soon as the refresh finishes. You’ll also see a full-screen loader with playful labels (Waking up the agent…, Fetching AI data…, Warming up the neural engine…, Syncing your conversation…) the first time you open the chat on a request, while the initial history is fetched.Tracking agent changes
Every change the agent makes is recorded in History, so you can see exactly what it did and roll any of it back. Agent entries are easy to pick out from the ones you made yourself:- They appear on blue cards, while your own changes sit on plain cards marked Manual.
- They’re attributed as By Agent · prompted by [your name], so it’s clear both that the agent made the change and whose instruction prompted it.
Confirmed requests
If a request was confirmed without using the agent chat, reopening this page shows a short notice instead of a conversation.What’s next
- Agent view for the tile-based review experience.
- Review and confirm orders in Flow for the complete order processing workflow.
- Work with the product cart for managing extracted product data.