Multi-turn Dialogue

A single question and answer is easy. The hard part of a chat or voice agent is the exchange that takes six messages: the customer wants to book, the agent needs a date, a time, a name and a phone number, the customer changes the date halfway through, and then asks a side question about parking. Holding all of that together, remembering what has been collected and what is still missing, is what this term describes. It is the difference between a FAQ bot and an agent that completes tasks.

How it works. The agent keeps a conversation state: the identified goal, the slots required for it (date, time, service, contact), which of them are filled, and the history of what was said. Each new message is interpreted in that context, so "make it Thursday instead" updates the date slot rather than starting over. The language model handles the interpretation; the state is held in the session and, for long or resumed conversations, in storage keyed to the customer. When all slots are filled the agent confirms and calls the tool (a calendar write, a CRM update). Side questions are answered and the agent returns to the open task. Context windows have limits, so long conversations are summarized rather than replayed in full.

Why it matters for task completion. Most useful customer tasks need more than one piece of information. Without state, the customer repeats themselves, the agent re-asks answered questions, and completion rate collapses. With state, the agent can be interrupted, corrected and resumed, which is how real conversations go. It also matters across channels: a customer who started on web chat and continues on WhatsApp the next morning expects the agent to remember.

Example. A Shopify store's chat agent gets: "I want to return the blue jacket." The agent asks for the order number. The customer: "it's 48213, actually can I exchange for a medium instead?" The goal shifts from return to exchange; the order slot fills; the size slot fills. The agent checks stock through the store integration, finds a medium, and asks for confirmation. The customer asks "how long does shipping take?", gets the answer, and says "yes, do it." The exchange is created and a confirmation is sent. Five turns, one task, two mid-stream changes, no repetition.

Multi-turn dialogue vs contextual awareness. Contextual awareness is the broader capacity to use everything known about the customer and the situation: past orders, the page they are on, the time of day. Multi-turn dialogue is specifically about carrying state within one conversation toward one goal. An agent can be contextually aware and still fail at multi-turn tasks if it does not track slots.

Common pitfalls.

• Re-asking for information already given, usually because state was not persisted.
• Losing the open task after a side question.
• Treating a correction as a brand-new request.
• No limit on turns, so a stuck conversation loops instead of handing off.
• Resetting state on channel switch or page reload.

Botcadence chat agents answer website chat, WhatsApp, Instagram, Telegram, Slack, SMS and email from one platform and hand off to a person with the full transcript; see chat agents and the appointment scheduling chatbot use case.