What the AI agent answers in DMs, and when it hands over

For agents

Guide: What the AI agent answers in DMs, and when it hands over

This page: https://commentwatchdog.com/help/keyword-automations/what-the-ai-agent-answers-in-dms-and-when-it-hands-over

Topic: Automation

Time: about 4 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: You know which messages the agent answers, how each answer is checked, and when a person gets the chat instead.

Context: Your AI agent is your business's knowledge and voice, set up once under AI agents and placed where it should answer. In DMs it answers in two places: on every message no keyword automation or greeting answered (the Let your AI agent answer messages automation), and inside a flow, where the AI agent answers step replies to what the person last wrote and the flow goes on. It answers in the customer's language, shorter than on comments, with the last few messages of the chat as context. Every automatic answer is checked against what your sources state before it is sent; anything it cannot back, and some subjects always, go to a person.

Before you start

1. An AI agent is on Starter and above (1 agent on Starter, 3 on Growth, 10 on Pro, unlimited on Elite).

2. Each answer or draft uses 5 AI credits; an answer it could not give uses none.

3. Only one automatic answer goes to one message.

Steps

Step 1 of 6: Start the agent automation.

Where: Automation tab, New automation. Under When someone messages you, pick Let your AI agent answer messages.

Link: https://app.commentwatchdog.com/en/inbox (Open the app)

Screenshot: https://commentwatchdog.com/media/shots/dm-gallery.en.webp (New automation: Let your AI agent answer messages, under When someone messages you.)

Screenshot: https://commentwatchdog.com/media/shots/dm-gallery-more-phone.en.webp (On a phone, scroll down to When someone messages you.)

You should see: The editor opens on its first step, Which agent.

Step 2 of 6: Pick the agent.

Where: Which agent: pick the agent that should answer (here Northwind barista, the default); New agent makes another one. Under it, What Northwind barista knows shows the start of its knowledge and its tone.

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent.en.webp (Which agent: the agent that answers this page's DMs.)

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-phone.en.webp (The same on a phone.)

You should see: Any message the keyword automations did not answer reaches this agent.

Step 3 of 6: Know who answers first.

Where: Before the agent: one automatic answer goes to each message: a keyword automation answers before a greeting, and a greeting before the agent, which catches the rest. After the agent: Who is asked, in order shows the path, the agent first (answers first), Add a fallback agent for a second one, then You, if nobody knows. In a flow, the agent answers only where its step sits.

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-path.en.webp (Who is asked, in order: the agent, a fallback agent if you add one, then you.)

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-path-phone.en.webp (The same on a phone.)

You should see: Keywords first, then a greeting, then the agent, then a person.

Step 4 of 6: Give it the knowledge it answers from.

Where: Edit the agent, under the agent you picked, opens it under AI agents. Its knowledge is one list: web links, text and FAQ pairs on every plan with an agent, files and images from Growth. Questions it could not answer are collected there too, so you can add the answer.

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent.en.webp (Edit the agent, to its knowledge.)

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-phone.en.webp (The same on a phone.)

You should see: The agent's own page opens. Back in this editor, What Northwind barista knows shows the start of what it answers from.

Step 5 of 6: Know when it hands the chat to a person.

Where: Hand to a person when, the editor's third step, has three boxes: They ask for a human, The agent is not confident, They say one of these words (your own list). Besides those, the chat always goes to a person when the answer would state a fact your sources do not (the second look), when the subject is one that always needs a person (an upset customer, refunds and wrong charges, health, legal or safety matters, abuse, private details), and when your AI credits ran out.

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-more.en.webp (Hand to a person when, and its three boxes.)

Screenshot: https://commentwatchdog.com/media/shots/dm-recipe-agent-more-phone.en.webp (The same on a phone.)

You should see: The message waits for your team. In Runs it is under Needs a person, with the reason (for example Left to you: refund).

Step 6 of 6: Read its answers in Messages.

Where: An answer the agent sent carries a robot and the automation's name under the bubble; your team sees it, the customer sees a normal reply from the page. In the chat list such a row starts with AI agent:. When the automation is set to draft, the answer waits in the thread under the agent's name until someone sends, edits or discards it.

Screenshot: https://commentwatchdog.com/media/shots/dm-inbox.en.webp (The AI agent's answer: a robot and the automation's name under it.)

Screenshot: https://commentwatchdog.com/media/shots/dm-thread-phone.en.webp (The same on a phone.)

You should see: Who answered, and any draft waiting for a person.

If something is off

Problem: It handed over a question I expected it to answer

Fix: The answer would have stated something your sources do not. Add it to the agent's knowledge or FAQ; Unanswered lists these questions.

Problem: It stopped answering

Fix: Your AI credits ran out: the agent stops answering, the message waits for a person, and the owner gets one email. Or the chat is paused for a person.

Problem: A keyword reply went instead of the agent

Fix: One answer per message: a keyword automation always goes before the agent.

Common questions

Question: Does the AI chatbot make things up?

Answer: It is built not to. Every automatic answer is checked by a second reader that lists the answer's business facts and whether your sources state each; one unstated fact and the chat goes to a person instead.

Question: Which language does the agent answer DMs in?

Answer: The customer's, whatever language its knowledge is written in.

Question: Which subjects always go to a person?

Answer: An upset customer, refunds and wrong charges, health, legal or safety matters, abuse, and private details posted in public.

Question: Can the agent and keyword automations work on the same page?

Answer: Yes. A keyword automation answers first, then the greeting, and the agent catches every message neither answered.

Related guides

Let your AI agent answer messages: https://commentwatchdog.com/help/keyword-automations/let-your-ai-agent-answer-messages

AI credits for DM answers: https://commentwatchdog.com/help/keyword-automations/ai-credits-for-dm-answers

Pause the automation in one chat: https://commentwatchdog.com/help/messages/pause-the-automation-in-one-chat

Plain text of this guide: https://commentwatchdog.com/help/keyword-automations/what-the-ai-agent-answers-in-dms-and-when-it-hands-over.md

All guides for assistants: https://commentwatchdog.com/help/llms.txt