Skip to content
Aymar Agents

Use cases

On-screen assistant

View as Markdown

Who it’s for: the internal team operating the Platform (or your own website’s “Assistant” panel, for your employees) who wants help getting around and filling in forms without memorizing where every screen lives.

What the end customer sees

Section titled “What the end customer sees”

The copilot is a tool for your team, not your end customers, unless you explicitly choose to open one of its actions to “Both” — something that only makes sense if you genuinely want a visitor to be able to create or edit something themselves. Every proposal is prepared by the AI, but you’re the one who makes the final click: the confirmation dialog itself makes clear the assistant prepared the step, never hiding that an AI is involved.

Steps in the product

Section titled “Steps in the product”
  1. Turn on the copilot for a channel. In Channels, find your Web widget instance and open the “Copilot” section of its card; turn on the “Enable the copilot for this channel” switch. Without this step, the widget keeps working as a normal chat.
  2. Your developer installs the manifest code on that website (the on-screen assistant SDK guide) — or, if the channel is the Platform’s own “Assistant” panel, the manifest already comes prepared by your provider.
  3. Review and approve the manifest version. The manifest is the closed list of screens and actions that website declares; the diff (what changes) is shown before you approve it. Any later change on the website leaves the approved version marked “outdated” until you review it again.
  4. Turn on the actions you want, one at a time, and choose their audience (Internal or Both). All of them start off — nothing turns on just because you approved the manifest.
  5. Open the “Assistant” panel and ask for something specific, for example “create a lead for Ana García, phone 600 123 456”. The assistant asks whether you’d rather “Guide me” (just navigate and highlight) or “Let’s do it together” (also fill in fields and offer to click).
  6. The copilot fills in the form in front of you and tells you it’s about to — it never says it’s already saved. If a value fails your own form’s validation, you see the exact same error you’d get typing it by hand; it never forces the submission through.
  7. You confirm (or correct and confirm) and the action actually runs. The assistant confirms it back to you with the real result, never from memory.
  8. Check the plan history at the bottom of the settings screen: every plan the copilot ran on that channel, with its status (Proposed, Delivered, Completed, Cancelled, Failed, Expired or Rejected) and a link to the conversation.

What to configure

Section titled “What to configure”
  • Channel: the copilot turns on per Web widget instance.
  • Manifest: installed by your developer (or the one the internal panel already ships with) and approved by an Admin.
  • Actions: each one, turned on individually, with its own audience.

Limits and honesty

Section titled “Limits and honesty”
  • The copilot never goes outside the manifest: if a screen, field or action isn’t declared, it doesn’t exist for it — it never improvises buttons or screens.
  • It never saves, sends or deletes without your explicit confirmation, even if you chose “let’s do it together” — that’s a product floor that can’t be turned off from any setting.
  • Your provider’s support staff, from their own internal panel, only guide: they never fill in or run anything for you.
  • It doesn’t act against a third-party system outside your own website (your WhatsApp provider, say) — for those stretches, it guides and waits for you to do it.
  • Revoking the manifest turns off the copilot instantly across every open session of that website.

See also

Section titled “See also”