Customer Agent in practice.

Every head of service in Southeast Asia is being asked the same question: when does the AI agent go live? It is the wrong first question. The right one is which conversations the agent should own, which it should never touch, and how you will know it is working.

HubSpot's Customer Agent is now mature enough to answer real customers on real channels. That makes the design decisions more important, not less.

What can Customer Agent actually do today?

HubSpot describes Customer Agent as the agent that resolves inquiries automatically, and reports that teams using it close “77% more customer tickets” per month on average, with “39% faster ticket resolution” than teams not using it (HubSpot). Treat those as vendor averages: they show direction, not a forecast for your operation.

The practical facts matter more. According to HubSpot's documentation, the agent can be deployed to WhatsApp, Facebook, live chat, forms, email and custom channels, with calling in beta. It needs a Professional or Enterprise subscription of a HubSpot hub, and it runs on HubSpot Credits, which HubSpot says are consumed only when a conversation is resolved (HubSpot Knowledge Base).

Two controls in that same documentation shape every rollout we run. You can set the agent's working hours, and you can set the percentage of conversations on a channel it should handle. That means you do not have to choose between “off” and “everything”.

Which conversations should go to the agent first?

We score candidate intents on three things: volume, simplicity and risk. The first intents to automate are high on the first two and low on the third.

• Volume. Pull three months of conversations and group them by intent. In most service operations, a small number of intents carry a large share of the load: order status, opening hours, booking changes, product questions, document requests.
• Simplicity. Can the answer be given from content you already own and keep current? If your best agent would need to check three systems and ask a supervisor, it is not a first-wave intent.
• Risk. What happens if the answer is wrong? A wrong opening time is an annoyance. A wrong refund promise, medical instruction or contract term is a liability.

The first wave is usually informational: questions with a stable, written answer. The second wave adds simple transactions once the agent can read the right records. Complex cases stay with people for longer than most business cases assume, and that is fine.

What should stay human?

Some conversations should go to a person by design, not by failure. In our experience these include refunds and cancellations, complaints, anything involving health, money or legal commitments, VIP and key accounts, and any customer who is already upset.

HubSpot's handoff settings support this. By default, the agent hands over when it cannot answer, when the visitor asks for a human, or when the agent is paused. You can add your own triggers; HubSpot's own example uses terms such as “Cancellation, Refund, or Can't log in”. You can also choose a live handoff to the next available person or an asynchronous handoff for later follow-up, routed to an inbox, help desk, user or team (HubSpot Knowledge Base).

On WhatsApp, this is not optional. Meta's WhatsApp Business Messaging Policy requires businesses running automated experiences to have “prompt, clear, and direct escalation paths”, such as an in-chat transfer to a human agent, phone, email or web support (WhatsApp Business Messaging Policy). Design the handoff before the first message is sent.

What does the agent need to know before it answers?

An agent is only as good as the content it draws on. Before go-live, we check four things.

1. A current knowledge base for the first-wave intents, written in plain language, with a named owner for each article.
2. The policies the agent must respect: what it may promise, what it must never say, and when it must stop.
3. Language coverage. If customers write in English, Bahasa, Tagalog, Thai or Chinese, the content and the testing must reflect that.
4. Customer context. Which records the agent may read, and what it may write back to the CRM.

This is the same foundation argument we make in context before agents. Most poor agent answers trace back to missing or outdated content, not to the model.

How do you roll it out without risking customers?

We use a staged rollout on one channel and one market.

1. Shadow. Test the agent internally against real historical questions. Review every answer.
2. Partial. Assign a small percentage of live conversations during working hours, with live handoff on.
3. Expand. Raise the share as resolution holds and feedback stays healthy. Add intents one at a time.
4. Next channel. Only then move to the next channel or market.

Each step has an owner and an exit rule. If quality drops, the share goes down, not the standard.

How should you measure Customer Agent?

HubSpot's performance reporting for the agent includes resolutions, deflections (conversations handled without any human involvement), human handoffs and handoff rate, conversations by channel, and visitor feedback on whether the agent was helpful, with CX scores in beta (HubSpot Knowledge Base).

We report four numbers to the leadership team each week:

• Resolution rate on the intents the agent owns.
• Handoff rate, and the top reasons for handoff.
• Customer feedback on agent-handled conversations against human-handled ones.
• Cost per resolved conversation, including HubSpot Credits.

Avoid vanity measures such as “conversations touched”. An agent that touches everything and resolves little is adding a step, not removing one.

What changes for the service team?

The team's work shifts from answering the same question many times to handling the hard cases, and to improving the agent. Someone has to own the knowledge base, review handoff reasons and decide which intent comes next. That is a role, not a side task. Plan for it, and train the team on the new process rather than the new screen.

Start by pulling three months of conversations and grouping them by intent, or begin with an AI readiness assessment.