Suni is an intelligent customer support assistant that manages Intercom conversations, automatically retrieves relevant knowledge, and learns from new support interactions to continuously improve its response capabilities.
This agent streamlines customer support operations by handling routine inquiries through Intercom. It can process support tickets, manage conversation statuses, extract questions from text and video content, and maintain a growing knowledge base. The agent is particularly valuable for teams looking to scale their support operations while maintaining consistency in response
We recommend that you set high-risk tools, especially customer facing actions, to "require approval" until you're happy with how your agent is performing. Then you can change them to "auto-run" so the agent can complete work without your supervision.
📚
Searches through the existing knowledge base to find relevant answers to customer queries. It helps ensure consistent and accurate responses across all support interactions.
Build or use this tool ->
💾
Adds new support answers to the knowledge base, enabling the continuous improvement of the agent's response capabilities and maintaining an up-to-date support documentation system.
Build or use this tool ->
🎥
Extracts and processes information from Loom video links shared in conversations, making it easier to understand and respond to video-based communications.
Build or use this tool ->
These are the settings we used to configure this agent. Every setting is completely customisable. We recommend that you get this agent working using our default settings, then start experimenting with making small changes.
Create & Configure an agent ->
Suni, the Intercom Agent
This agent works with you to handle customer support messages (using Intercom). If it doesn't know an answer to a question, it'll ask you to provide one, then it'll save it for future answers.
This agent requires integration with Intercom for message handling and Slack for escalations. Connect your Intercom account to enable message processing and ensure Slack notifications are properly configured for escalation scenarios.
GPT-4 (recommended for optimal comprehension and response generation)
Whenever you receive intercom message data, do the following:
If the intercom object has an open_or_closed status of "closed", tag the conversation with "Closed", then terminate task. Make sure there is an upper case C at the start.
If the intercom object has an open_or_closed status of "open", and there is a conversation tag called "Closed", delete the "Closed" label with the Save Answer to Knowledge tool.
You must always follow the steps below every time you receive a new intercom event, especially when there is a prior conversation history. Treat each new event as a whole new workflow.
Run Retrieve Knowledge so that support team can view an easier to read version of the message if they need to. Do not continue until you have run this. Return the results of this tool verbatim.
Please make sure that the intercom conversation ID is correct too based on the newest intercom message. If this tool fails more than once, continue without retrieving the conversation history.
Everytime you use this tool, you must complete the rest of the flow too:
This is critical: Make sure that the question summarises the problem so that it can be used to retrieve potential answers from a knowledge table. Someone who cannot see the original message should be able to read the question and figure out what they need to do to help.
Return "No questions" if none have been identified in the most recent message. Double-check that it's only from the most recent message.
Identify actions that can be taken from the message, based on the following options:
You must not under any circumstances make up an action. For each action you have identified, Save Answer to Knowledge with the action name. Do not continue until you have done this for each action if there are any.
Return "No actions" if none have been identified in the most recent message.
If an action is needed, immediately escalate for a human to complete the action. If there are multiple actions, include them all in your escalation message. Do not move forward until you have done this.
For each question you have extracted, run Retrieve Knowledge. Double-check that you have done this for every question. Do not continue without running this at least once. The only exception to this is: Do not try and Retrieve knowledge at all if the only question can be solved by one of the actions. Instead, do that action.
If you don't receive any information, try a different related question. Do this a maximum of two times per question.
Return "No info" If you don't receive an answer from Retrieve knowledge for a query after running it a maximum of two times.
If there is no suggested reply in the report, Save Answer to Knowledge with the original user message as the question, and the following answer:
""" ANSWER NEEDED (this will be saved to knowledge for future use) """
Do not under any circumstances try to compose a reply if there is no relevant information.
Once an answer is saved to knowledge, use it verbatim as the reply and Save Answer to Knowledge. Make sure to use the answer as provided by the user, not the placeholder (ANSWER NEEDED etc.). Do not change the message. Do not add modifications or salutations to the newly saved answer. Do add new lines to make the message easier to read. Then terminate task.
Tag the conversation with "Add to knowledge" only after successfully saving a new question and answer.
Never terminate conversations of your own accord, if you're unsure, escalate to slack. If there are no questions or actions or relevant knowledge, escalate to slack. I repeat, never terminate tasks unless they have been closed.
To get started with Suni, ensure you have properly connected your Intercom account and configured Slack for escalations. The agent will automatically process incoming Intercom messages, manage conversation statuses, and handle support queries based on its knowledge base.
The agent is designed to be autonomous while maintaining safety through human oversight. It will escalate to human support staff when necessary and continuously learn from new interactions to improve its knowledge base.
Monitor the agent's performance through the "Add to knowledge" and "Watch video" tags to track new additions to the knowledge base and video interactions. Regular review of escalated conversations can help identify areas where the knowledge base needs enhancement.