This AI Customer Agent answers repeat questions in your client portal and chat, grounded in your knowledge base. A classifier triages tickets in Breeze and handles ticket deflection on the top questions. Sentiment detection escalates to a human as soon as things fall outside scope. Deflection rate up, CSAT at the same level or better.
The Customer Agent is not a standalone chatbot and not a sentiment tool nobody knows how to use. It is a working unit that runs continuously on your knowledge base, catches repeat questions, and knows when to hand off to a human.
Service teams are swamped by AI demos: a classifier that is 60 per cent accurate, a chatbot that has confidence but no knowledge. We do it the other way around. First get your knowledge base in order, then build an agent with tight grounding, then set up an escalation path, and only then go live. CSAT and deflection as KPIs, not tokens.
A Customer Agent is only as good as its knowledge base. Four things you want to know before you go live: what knowledge it uses, which tools, how escalation works, and what we measure.
Grounding
The agent runs on your HubSpot Knowledge Base, supplemented with product documentation, FAQ pages and internal procedures where relevant. Before we start we do a knowledge base audit: which articles are current, which need to be merged, where are the gaps. A good agent on a half-built knowledge base produces frustration, so this is not an optional step.
Stack
Default stack: HubSpot Service Hub as the system of record for tickets and conversations. Breeze Ticket Triage for classification and routing. Breeze Customer Agent or a custom Claude or OpenAI flow for chat responses, depending on complexity and data volume. For the client portal we can use native HubSpot or a custom front end. For sentiment detection, a lightweight classifier pipeline.
Escalation
Three default triggers: out-of-scope (the question falls outside the knowledge base), negative sentiment (frustration or urgency detected), and explicit request for a human. On each trigger the ticket is created or routed to the right service rep, with the conversation context added as a note. The customer receives a notification that a human is taking over, no silent hand-off. Your service lead decides per category when escalation should apply.
KPIs
Four standard KPIs: deflection rate (what percentage of questions the agent catches), CSAT comparison of AI conversation versus human, escalation rate (how often the agent needs to hand off), and a gap report (which questions the agent could not answer well). Monthly review with service lead to resolve the top 10 gaps through knowledge base updates or prompt tuning.
A SaaS organisation had a service team of four that started every morning with 80 tickets. A large portion were repeat questions that were in the knowledge base, but customers preferred to ask quickly via chat. We first reviewed and restructured the knowledge base. Then a Customer Agent on Breeze with an escalation path to a human.
After six weeks the agent catches 40 to 50 tickets every morning. The service team can focus on the tickets that really matter: integration issues, custom questions, customers at risk of churning. CSAT on AI conversations is within 5 per cent of CSAT on human conversations.
On a Customer Agent engagement you work with Carel as architect and Kim for service process design. Carsten joins for data integration and grounding maintenance.
Service process design
Works on knowledge base structure, escalation paths and the collaboration between the agent and the service team. Ensures adoption lands on your side.
Questions every service team asks before bringing AI close to their customers.
No. The agent catches repeat questions so your service team has time for the tickets that matter: integration issues, custom questions, customers at risk of churning. A service person with us does only work that requires a human. The rest goes to the agent, with an escalation path back to a human if it falls outside scope.
Two safeguards. First: the agent is grounded in your knowledge base and instructed to say "I don't know" when the question falls outside its scope. Better to hand off to a human than to guess. Second: every conversation is logged as an activity on the Contact record, your service lead can spot-check and resolve gaps in the monthly review. Not foolproof, but with an audit trail and an improvement loop.
A working HubSpot Service Hub (or willingness to set it up properly), a knowledge base with at least 30 to 50 articles (we then do an audit and restructure), a service lead who takes ownership and is willing to invest two hours per week in reviews during the first month. A Breeze licence is also needed, we advise on that in the scope.
By default the agent states this at the start: "I am the Customer Agent of [company]. I can answer a large part of questions directly. For more complex matters I will transfer you to a colleague." Transparency over the illusion of a human. In our experience customers appreciate that, they want to know where they stand. EU AI Act compliance has also become stricter on this point.
No. The use case fits any B2B organisation with a knowledge base and a service team that receives repeat questions. We have done it for SaaS, for industrial B2B and for professional services. What differs is the complexity of the knowledge base and the escalation path. The architecture is the same.
A Customer Agent falls under limited risk under the AI Act in most configurations, provided there is transparency towards the customer (see above) and an escalation path to a human is available. We classify per use case and document the data flow, so your compliance team can assess it. For heavier decisions (financial or legal) we maintain human-in-the-loop as default.
Book a strategy call. We look at your current service work, knowledge base and HubSpot setup and give an honest scope. Not a sales call.