Conversation AI Agent

The Conversation AI Agent is the text-channel responder. Once an agent is authored in AI Studio, this module is where you decide which channels it listens on and under what conditions it replies. It reads inbound messages from the website chat widget, SMS, email, Facebook and Instagram DMs, WhatsApp and other connected inboxes, composes a reply, and sends it back down the same thread the contact used.

What separates it from a generic chatbot is that it runs inside the conversation view with the contact record attached. It can see the contact's name, tags, custom field values, appointment history and prior messages across every channel. That means it does not restart the relationship on every message: a contact who filled a form on Tuesday and texts on Thursday is answered by an agent that already knows what they asked for.

Think of it as a front-desk person who never leaves the desk. It handles the first contact, answers the predictable questions, collects what you need, books where it can, and passes anything unusual to a human with the thread intact. It is not a replacement for your team and it is not an orchestration engine. It is the thing that keeps someone from waiting twenty minutes for a reply.

Channel enablement

• Which channels the agent answers on: web chat, SMS, email, social DMs, messaging apps
• Whether each channel is on separately, so you can pilot on chat before letting it text
• Per-channel reply formatting, since SMS needs short plain text while email tolerates structure

When the agent replies

• Operating hours and out-of-hours behavior
• Whether it answers all inbound messages or only those matching a condition such as a tag, source or pipeline stage
• Whether it replies to first contact only, or continues through the thread
• Suppression rules so it stays silent on conversations a human has taken over

Reply behavior

• Delay before the first reply, so the agent does not answer in under a second and give itself away
• Typing-style pacing and whether long answers are split across messages
• Maximum messages per conversation and what happens when that cap is hit

Data capture

• Which fields the agent should try to fill during conversation, such as service interest, timeline or location
• Tags to apply based on what the contact says
• Whether the agent may create or update an opportunity

Booking

• Which calendar or calendars the agent may offer
• Whether it can reschedule and cancel as well as book
• What it does when nothing suitable is available

Handover and escalation

• Keyword, intent and sentiment triggers that stop the agent
• Who is notified and how, and whether a task or pipeline move accompanies it
• Whether a human replying in the thread permanently pauses the agent for that contact

1. Finish the agent in AI Studio first. Do not enable channels on an agent you have not tested.
2. Confirm every channel you intend to use is actually connected and verified at the account level: a phone number able to send and receive, an authenticated sending domain, and the social accounts linked with messaging permissions.
3. Enable one channel to start. Web chat is the safest, because the audience is self-selected and the stakes are lowest.
4. Set the reply conditions. Decide deliberately whether the agent answers every inbound message or only those from a defined source, and write that decision down.
5. Configure the suppression rule that silences the agent once a human replies in the thread. Do this before go-live, not after the first awkward double-reply.
6. Set the first-reply delay to something human, and set per-channel length so SMS answers stay short.
7. Map the fields the agent should capture, and make sure those fields exist on the contact record with the types you expect.
8. Connect the booking calendar and run a real booking end to end, checking the confirmation, the calendar entry, the timezone and the reminder sequence.
9. Set the message cap and the stop condition so a confused contact cannot loop the agent indefinitely.
10. Configure escalation: triggers, destination, notification, and the sentence the agent says as it hands off.
11. Test from a real device on the real channel, not only in the Studio preview. SMS rendering and social DM behavior differ from the preview.
12. Go live, then read every conversation for the first few days and tune instructions or knowledge where the agent guessed, over-explained or missed a handover.
13. Add the next channel only once the first is behaving. Each new channel brings its own formatting and etiquette problems.
14. Set up a recurring review so someone reads a sample of transcripts weekly rather than assuming silence means success.

This module is the conversational front door of a build. It consumes agents and knowledge from AI Studio, triggers and is triggered by Workflows, writes to Contacts and Opportunities, and books through Calendars.

Web chat to booked appointment

A visitor opens the chat widget on a landing page. The agent greets them, asks the two or three qualifying questions you configured, writes the answers to custom fields, offers available slots and books. A workflow picks up the booking, creates the opportunity at a booked stage, and starts the reminder sequence.

Missed-call text-back

An inbound call goes unanswered. A workflow sends an SMS apologizing and asking what the caller needed. The agent takes the reply and runs the same qualification it would on chat, so a missed call becomes a tracked lead instead of a lost one.

Reactivation campaign with live replies

A campaign texts dormant contacts with an offer. Replies land in the conversation view, where the agent handles the straightforward ones and escalates anything that sounds like a complaint. Without the agent, a reactivation send to a large list buries the team; with it, only the real conversations reach a human.

• Double-replies are the most visible failure. If the suppression rule is wrong, a staff member and the agent will both answer the same message, often contradicting each other. Test this deliberately.
• Channel delivery is not the agent's problem but it looks like it. Unregistered messaging numbers, unauthenticated sending domains and expired social tokens all present as the agent going quiet. Check the channel before you blame the agent.
• SMS length and cost discipline. Long AI replies split into multiple segments, read badly and consume quota. Constrain length per channel rather than hoping the agent keeps it short.
• Social DM windows. Some platforms restrict how long after a contact's last message you may reply. An agent that is slow or paused can miss that window and simply fail to send.
• Context can mislead. Because the agent sees the contact record, stale field values become stale assumptions. A contact tagged from a campaign two years ago may be greeted about something they no longer want.
• Opt-out and consent are yours to handle. The agent does not grant you permission to text someone. Keep consent capture in forms and workflows, and make sure opt-out keywords are honoured ahead of the agent.
• Group and forwarded threads. Shared inboxes and forwarded email chains can present multiple speakers in one thread; the agent will usually treat it as one contact.
• Silence is not success. An agent can politely fail for weeks. Build a review habit and watch handover rate, booking rate and conversation length rather than only complaints.
• One agent per job. Running one agent across sales and support channels produces an agent that is mediocre at both. Separate agents with separate knowledge scopes do better.