Follow-up suggestions

Most conversations don't end because the visitor got what they wanted. They end because the visitor finished reading a good answer, had nothing in particular to type next, and closed the tab.

Turn on follow-up suggestions and the agent writes up to three short questions after every reply. They appear as buttons above the message box, and clicking one sends it as if the visitor had typed it.

What makes a good suggestion

The agent writes them from the exchange that just happened, in the visitor's own voice — "Can I use this for my whole class?", not "Enquire about classroom licensing". They have to read as something a person would say, because clicking one sends it verbatim.

Three rules are enforced rather than suggested:

• They move forward. A question that re-asks what was just answered wastes a turn, so the agent is told to go deeper or sideways, never sideways-into-the-same-place.
• They stay inside what the agent can answer. The agent's own task description is part of the instruction. A suggestion that leads to a refusal is worse than no suggestion — the agent walked the visitor over its own boundary.
• None is a valid answer. When the exchange has genuinely finished, no buttons appear. Padding the screen with "Anything else?" is not help.

They also step aside for a form. If the agent's reply already asks something — a set of options to pick from — no suggestions are offered that turn. Two prompts on one screen compete, and the one that loses is usually the form you actually wanted answered.

Turning it on

In the console, open the agent → Basic → Follow-up suggestions. Off by default.

Underneath the switch is an optional line of steering: where you'd like the questions to lead.

Move the visitor toward starting the free trial, or toward talking to the
schools team. Prefer questions about pricing and setup.

Leave it empty and the questions are written from the conversation alone, which is usually fine. Fill it in when the agent has a job beyond answering — pre-sales, qualification, booking.

Where they show up, and what they cost

They render in the embedded widget and on share links — the same strip that page playbooks use for starter questions. The two never collide: starter questions break the ice before the visitor has spoken, follow-ups take over once they have.

Each turn costs one extra short model call, made in parallel with the rest of the turn's wrap-up. It does not delay the answer or hold the message box locked — the visitor can start typing before the buttons arrive, and typing dismisses them. While they're being written, the same progress indicator the agent uses for thinking appears where they will land, so the visitor can see the turn isn't quite finished.

Chat channels without buttons — Telegram, WhatsApp — are unaffected.

They stay with the conversation

The suggestions are stored with the reply that produced them, not thrown away after they're shown. Three consequences:

• Refreshing keeps them. The buttons are part of that turn, not a one-off screen effect.
• Shared conversations carry them. A replay shows the buttons appearing and — when the visitor took one — that button being pressed, rather than the question being typed out. Which is what actually happened, and it makes the shared link show how the conversation was steered, not just what was said.
• Whether a question was clicked is recorded, not guessed. The visitor might well have typed the same sentence themselves; the platform stores what happened rather than inferring it from the text.

Related

• Page playbooks — the questions offered before the visitor says anything, chosen by which page they opened the chat from.
• Keeping an agent honest — why the boundary in the agent's own instructions is what stops a suggestion leading somewhere it shouldn't.