The cardinal rule of customer-facing bot rollouts is "fail to a human cleanly." Customers tolerate a bot that punts. They do not tolerate a bot that loops, gaslights, or hallucinates a policy. The sequence below puts that rule first.
- Audit Knowledge before turning on the agent
Pull every Knowledge article in the public-facing channel. Archive articles older than 18 months unless they are timeless. Confirm article titles match how customers actually phrase the question. Without this step the agent confidently cites wrong answers from stale content.
- Enable Agentforce for Service and provision both agents
Setup, Einstein, Agentforce, Enable. Accept terms. Provision Service Agent and Service Coach from the gallery. Both land in Agent Builder as editable copies of shipped templates.
- Customize Returns and Refunds for your actual policy
The shipped topic uses generic e-commerce policy. Replace the instruction and the Knowledge Data Library scope with your company's actual return windows, restock fees, and exceptions. This topic is the single biggest source of customer complaints when shipped without customization.
- Wire escalation to your existing Omni-Channel queues
Open the Escalation topic. Point its Transfer to Agent action at the queue you currently use for chat. Do not invent a new queue; the rep team must see escalated conversations in the same place they see all other work.
- Soft-launch the Service Agent on one channel with a beta opt-in
Pick Web Chat or Messaging. Put a small "Try our AI assistant (beta)" link rather than replacing the default chat entry point. Watch deflection and customer satisfaction for two weeks before expanding.
- Roll out the Service Coach to one rep team in parallel
The Coach is lower risk because it never speaks to a customer directly. A pilot team of five reps with weekly retrospectives identifies prompt and topic refinements before broad rollout.
- Watch the Plan Trace and a weekly conversation review
Pull a random sample of 50 conversations a week, read them with the service ops lead, and adjust topics where the agent picks wrong. This work never ends. Budget for it permanently.
The list of channels (Messaging, Web Chat, WhatsApp, SMS, Slack, etc.) the Service Agent is exposed on. Most rollouts start with one.
Which Knowledge articles, data categories, and channels the agent can read. Drives the FAQ Lookup topic accuracy.
The queue or routing path that catches conversations the agent cannot handle. Must integrate with the team's existing Omni-Channel setup.
The Coach's reply draft prompt template. Highest-leverage customization on the rep-facing agent.
When the Service Agent is active vs falls straight through to a queue. Useful when the agent's quality is uneven and you want human coverage during peak.
- The Service Agent without a Knowledge audit confidently cites stale articles. Treat the audit as a hard prerequisite, not a nice to have.
- Conversation pricing on the customer-facing agent surprises orgs that did not estimate volume. A B2C site with 100k monthly chat sessions and 60 percent agent coverage is 60k conversations a month.
- Escalation that drops the transcript is a customer satisfaction killer. Verify the handoff carries history before any external rollout.
- The Service Coach is mostly inert without Knowledge and prior case data. Empty Knowledge means generic suggestions that reps stop using.
- Voice is not yet a standalone bot path. Service Cloud Voice integrates the Coach for human reps, but a fully autonomous voice agent is not a shipped configuration as of mid-2026.