60-70% of the inquiries the support team receives every day are the same type of questions over and over again: ``password reset,'' ``plan confirmation,'' ``how to use questions,'' and so on.Breeze Customer Agent is an AI support agent that autonomously responds to these repeated inquiries 24 hours a day after learning from the company's knowledge base, website, and PDF.is. It is designed to support 9 channels, record behavior with a complete Audit Card, and appropriately escalate complex inquiries to humans, minimizing the risk of ``leaving it to AI and damaging your brand''.
Customer Agent is more than just a "chatbot." Unlike rule-based flow trees,"Inferential AI" that understands questions in natural language, searches your company's knowledge sources, and generates answers tailored to the context.is. It can also respond to ``phrases that are not written in the manual'' and ``complex questions.''
Since INBOUND in 2025, the supported channels have been significantly expanded to cover 9 channels including web chat, email, WhatsApp, SMS, and voice calls (beta).
The billing unit for Customer Agent is "conversation."1 conversation = 1 entire chat sessionRequired plan: Each Hub Professional or higher50,000 credits for 500 conversations per monthCheck the credit limit for each plan in advance.
The customer agent's answer accuracy isDetermined by “what you learned”. The quality and comprehensiveness of the knowledge source is directly linked to the quality of the agent. There are three types of knowledge sources that can be configured.
| Knowledge source type | Setting method | Best use | Points to note |
|---|---|---|---|
| HubSpot Knowledge Base | Automatically reference KB articles in Service Hub. New KB articles are automatically updated to the Agent. | FAQ/Usage guide/Troubleshooting. Information updated regularly | If the quality of the KB is low, it will cause incorrect answers. First, prepare the KB and then start the Agent. |
| Public website | Crawl and learn up to 1,000 pages by specifying a URL. Re-crawl regularly to reflect the latest information | Product information/company profile/price page. Complete information not written in KB | Pages that require login and dynamic content cannot be learned. Set the crawl range appropriately |
| PDF/Document | Upload manuals, specifications, and policy documents. Save to Breeze Intelligence Knowledge Vault | Product manuals, SLA, contract terms, and internal policies. Information not posted on the public web | PDF quality (scanned image vs. text) affects accuracy. Requires regular updates |
The most important design point for Customer Agent operation is “When should it be passed on to humans?"is. If the escalation rules are inappropriate, problems will occur, such as ``AI will not take over even if it cannot solve the problem'' or ``Inquiries that can be solved will be immediately passed on to humans.''
Audit Card is a logging feature that records all decisions, actions, and changes made by Customer Agents.“What was the basis for the AI to give this answer?” “Which knowledge base article did it refer to?” “Which field in CRM did it change?”Make everything transparent. This is especially important for compliance in regulated industries (financial, medical, legal).
| Information you can check with Audit Card | How to use it |
|---|---|
| Referenced knowledge sources | Check which KB articles, web pages, and PDFs the Agent based their answers on. Identify the cause of incorrect answers and improve the KB |
| CRM property change history | When the Agent automatically changes a contact's properties, the old and new values are recorded. Detect and rollback unintended changes |
| Basis for escalation decisions | Record which conditions were detected to determine escalation. Check if the rules are working properly |
| Resolved vs. unresolved classification | You can categorize and report conversations that were resolved by AI, unresolved and escalated conversations, and conversations where customers left midway through. |
| Correlation with CSAT scores | Contrast customer satisfaction scores with AI answer quality to identify which types of questions lead to lower satisfaction rates |
Check the following weekly:① Identify question categories with low resolution rates and add/improve KB articles.② Identify the source of the incorrect answer and remove/correct the incorrect information from the KB.③ Review categories with high escalation rates and verify that rules are working correctly.“Customer Agents who do not look at the Audit Card” are at risk of deteriorating quality over time.. It is strongly recommended that monthly quality reviews become an operational standard.
Simultaneously supports 9 channels including web chat, email, WhatsApp, SMS, and voice (beta). Aiming for a design in which ``AI solves 60-70% of FAQs.'' Humans can concentrate on complex, emotional, and legal inquiries.
If you start the Agent with an incomplete KB, customers will receive incorrect answers. First, prepare the KB article for "Frequently Asked Questions TOP 20" and then start it. It is safer to design the Agent to "escalate" rather than "don't answer what you don't know."
Emotional, Legal, VIP, Repeated Unresolved, Price Negotiation, and Non-Verbal categories are always set to be escalated to a human. This is where you can prevent the biggest risk of damaging your brand by relying too much on AI.
All actions of the AI are recorded on the Audit Card. Create a cycle to improve KB by checking incorrect answers, escalation rate, and unresolved categories on a weekly basis. It is important to recognize that AI is not something that ``starts and ends'' but ``is something that is continuously developed.''