The problem: the queue answers itself, slowly
Your agents answer the same ticket all day. Where is my order. How do I reset access. Can I move my appointment. The answer is in the knowledge base, but finding it, pasting it and closing the case takes longer than the question deserves. Meanwhile the ticket that needs a person, the angry customer or the edge case, waits behind the easy ones.
An AI customer service agent should take the easy ones and make the hard ones faster. That is the whole outcome. Not a chatbot that deflects, a colleague that triages, answers, hands over and writes it all down.
Triage: every ticket read, labelled and routed
The agent reads each email, chat or form as it arrives. It sets the category, the product, the urgency and the language, and routes the case to the right queue. This is the step most teams underrate. Good routing alone moves the hard cases to the right person sooner. We describe the pattern on support ticket routing.
Answers from your knowledge base, with citations
When the question has an approved answer, the agent replies with it and cites the article it used. Your team can click through and check. When the article is out of date, the agent flags it rather than improvising. Grounding comes from Salesforce Knowledge and, where it helps, Data Cloud for the customer's orders, entitlements and past cases.
Handed to a person when unsure
The agent hands over when confidence is low, when the customer asks for a person, when the topic is on your never-automate list, or when the tone says the customer has had enough. The person gets a summary, the sources checked and the suggested next step. The customer does not repeat themselves. We set the handoff rules before we set the answers.
Written back to the case
Every action lands on the case in Service Cloud: the classification, the reply, the articles cited, the handoff reason. Reporting then works on real data. You can see which ticket types the agent closes, which it hands over, and which articles keep failing.
Agentforce Service Agent or a custom agent
Agentforce Service Agent fits when the cases, the knowledge and the entitlements already live in Service Cloud. It runs inside Salesforce permissions, uses your topics and actions, and deploys to web chat and messaging without another platform to govern.
A custom agent on Claude or OpenAI fits when the conversation lives outside Salesforce: WhatsApp for Gulf customers, Slack for an internal help desk, voice on a telephony platform, or a product with its own support widget. You own the model choice and the prompts; Salesforce stays the system of record through the API.
The deciding questions are where the customer writes, where the record lives and who owns the control plane. Our Agentforce vs custom agents comparison walks through them.
Channels: Service Cloud, WhatsApp, Slack, web chat and voice
In the UAE, customers write on WhatsApp. In the US, many internal teams live in Slack. The agent should sit where the conversation already is. The setup for both is on the WhatsApp or Slack agent page. Voice adds a telephony layer such as Retell AI that transcribes the call and passes it to the same agent and the same handoff rules.
How we start
We read the org first, then pull a sample of your real past tickets. We pick one ticket type, write the handoff rules, ground the answers, and test against those tickets before a customer sees anything. The customer service automation use case shows the operator view, and the customer service ROI calculator lets you put your own volumes in. We will not quote a deflection rate before we have seen your queue.