Guide: Let the AI agent answer in a flow
This page: https://commentwatchdog.com/help/flow-steps/let-the-ai-agent-answer-in-a-flow
Topic: Flow steps
Time: about 5 minutes
Who can do it: Owner, Administrator or Manager
Author: https://anantgupta.xyz (Anant Gupta, founder of CommentWatchDog)
Start in the app: https://app.commentwatchdog.com/en/inbox
Result when done: Your agent answers once, or holds a whole conversation towards a goal, and every way it can end has a next step.
Context: AI agent answers hands the conversation to your agent. With One answer it replies to what the person last wrote, from what it knows and in its voice, and the flow goes on at once. With Keep the conversation going it replies message after message towards a goal you write, saves what they tell it, and the steps after it wait until a way out fires. When it cannot answer, nothing is sent and the flow takes the way that says why: Wants a person, Couldn't answer or Didn't run.

Before you start:
1. AI agents come with Starter and every plan above it.
2. Each answer it sends uses 5 AI credits; a turn that sends nothing uses none. A conversation costs at most its answer limit times 5.
3. It needs an open chat: in a comment flow it goes after the private reply's tap.

Steps:
Step 1 of 6: Add AI agent answers to your flow.
Where: Open the flow (Automation, then the flow's row). In the Steps panel on the left, under Send, drag AI agent answers onto the canvas, or select the step it should follow and click AI agent answers. On a phone: Add step at the top left of the canvas, then AI agent answers under Send.
Link: https://app.commentwatchdog.com/en/inbox (Open the app)
Screenshot: https://commentwatchdog.com/media/shots/dm-flow-palette.en.webp (The Steps panel: AI agent answers, under Send.)
Screenshot: https://commentwatchdog.com/media/shots/dm-flow-palette-phone.en.webp (On a phone, Add step lists the same steps: AI agent answers, under Send.)
You should see: The AI agent answers card is on the canvas and its settings panel is open (a sheet on a phone).

Step 2 of 6: Choose the agent.
Where: Panel, Which agent: pick one of your agents, or Create an agent if you have none. Edit what it knows opens the agent's knowledge.
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent.en.webp (Which agent: the agent that answers here.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-phone.en.webp (The same on a phone.)
You should see: The card names the agent.

Step 3 of 6: Choose how it answers.
Where: Panel, How it answers, under Which agent. One answer (the default) replies once, then the flow goes on. Keep the conversation going keeps answering until its goal is met, they go quiet, or a person is needed. Rest the pointer on either option to see a sample chat of it.
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent.en.webp (How it answers: One answer, the default.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-phone.en.webp (The same on a phone.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk.en.webp (How it answers set to Keep the conversation going: the goal and Save what they tell it open under it.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-phone.en.webp (The same on a phone.)
You should see: With One answer the card says it answers what they last wrote. With Keep the conversation going it says Keeps the conversation going and the most answers it may send, and the panel opens the settings of the next two steps.

Step 4 of 6: For a conversation, write its goal and what to save.
Where: Panel, What should it get done?: one or two sentences in your own words, an instruction your customers never see. The agent answers what they ask first, then works towards it. Then Save what they tell it, Add something to save (up to 6). Each row starts with where it is kept: Contact detail (their Name, Email or Phone), Custom field (one of your fields, or New field…), or Conversation only (a name you type; it is never stored on the contact, and later steps can check it in a Condition or write it as {{name}}). Add a few words on what to ask for, and switch Required to finish on for what the conversation cannot end without. An email or a phone number is saved only if they typed a valid one.
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk.en.webp (The goal, then one thing to save: Contact detail, Email, Required to finish.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-phone.en.webp (The same on a phone.)
You should see: Under each row, one line says where the value is kept and how later steps use it.

Step 5 of 6: For a conversation, set when it stops.
Where: Panel, Stop after: the most answers one conversation may send (6 unless you change it, 20 at most); after the last one it leaves by Couldn't finish. If they go quiet for: how long it waits for their next message (1 hour unless you change it, from 5 minutes to 23 hours, never past the chat's 24-hour window); then it leaves by No reply. The line under them says the most AI credits one conversation can use.
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-limits.en.webp (Stop after 6 answers.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-limits-phone.en.webp (The same on a phone.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-limits.en.webp (If they go quiet for 1 hour.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-limits-phone.en.webp (The same on a phone.)
You should see: Up to 30 AI credits per conversation (5 per answer), for 6 answers.

Step 6 of 6: Wire its ways out.
Where: Panel, Its four ways out (five for a conversation) explains each one; rest the pointer on a way to see a sample chat. One answer: Answered (it replied), Wants a person (they asked for one, or the subject needs one: a refund, an upset customer), Couldn't answer (it didn't know), Didn't run (AI credits ran out, the agent was paused or removed, or the reply failed). A conversation: Done (everything required is saved, or with nothing required the goal is met), Wants a person, Couldn't finish (it didn't know an answer, they wouldn't give what was required, or its answers ran out), No reply, Didn't run. Under Next steps, each way has a row: Add a step gives it its next step (or drag from the way's dot on the canvas). A way you leave unconnected takes the Couldn't answer way; if that one is unconnected too, the flow stops there and your team is asked.
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent.en.webp (One answer: Its four ways out.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-phone.en.webp (The same on a phone.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-ways.en.webp (A conversation: Its five ways out.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-ways-phone.en.webp (The same on a phone.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-next.en.webp (Next steps, one row for each way: Done leads to Notify the team, Wants a person and Couldn't finish to Hand to a person; No reply and Didn't run still say Add a step.)
Screenshot: https://commentwatchdog.com/media/shots/dm-node-agent-talk-next-phone.en.webp (The same on a phone.)
You should see: Under Next steps each way you wired names its next step instead of Add a step.

If something is off:
Problem: It took Couldn't answer for a question it should know
Fix: Every automatic answer is checked by a second reader against the agent's knowledge; a fact the knowledge does not state makes it Couldn't answer instead of a guess. Add the fact to the agent's knowledge.
Problem: A conversation ended as Couldn't finish
Fix: It met a question its knowledge does not cover, the person would not give something marked Required to finish, or it used all the answers in Stop after. The step's run history says which. Raise Stop after, make the item optional, or add the fact to the agent's knowledge.
Problem: It did not save the email or phone they gave
Fix: An email or a phone number is saved only when it reads as a valid one and the person typed it in the chat. The agent asks for it again on its next answer.
Problem: This agent is paused by your plan
Fix: The step takes Didn't run until the plan includes the agent again.
Problem: AI agents come with a higher plan
Fix: AI agents start on Starter.

Common questions:
Question: What does an answer cost?
Answer: 5 AI credits for each answer it sends. A turn that sends nothing uses none, so a conversation costs at most its Stop after number times 5.
Question: What happens when AI credits run out?
Answer: The step takes Didn't run and nothing is sent. With nothing wired to Didn't run it takes the Couldn't answer way, and with nothing wired there either the flow stops and your team is asked.
Question: Can the agent invent a price or a policy?
Answer: Every automatic answer is checked by a second model that lists its business facts; one the knowledge does not state turns the answer into Couldn't answer (Couldn't finish in a conversation).
Question: Do the steps after it wait during a conversation?
Answer: Yes. With Keep the conversation going, nothing after the step runs until one of its five ways out fires, and no other automation answers the person meanwhile.
Question: Can a conversation go on for ever?
Answer: No. It leaves by Couldn't finish after the answers in Stop after, and by No reply when they stay quiet for the time you set.
Question: What happens if someone on my team replies during a conversation?
Answer: The agent says nothing more in that conversation and the run goes to your team as Handed to a person.
Question: Where can I use what it saved?
Answer: A Contact detail or a Custom field stays on the contact: it shows in People and works in segments and other flows. A Conversation only value lasts for that conversation: a Condition after the step can check it and a message can write it as {{name}}.

Related guides:
- Hand the chat to a person: https://commentwatchdog.com/help/flow-steps/hand-the-chat-to-a-person
- Create an agent: https://commentwatchdog.com/help/ai-agents/create-an-agent
- Let your AI agent answer messages: https://commentwatchdog.com/help/keyword-automations/let-your-ai-agent-answer-messages

Plain text of this guide: https://commentwatchdog.com/help/flow-steps/let-the-ai-agent-answer-in-a-flow.md
All guides for assistants: https://commentwatchdog.com/help/llms.txt