Agentforce and AI Capabilities for Service
AI with guardrails
Sort capabilities into two buckets, because questions exploit the confusion.
Predictive / classification: Einstein Case Classification (predicts field values from case text), Einstein Article Recommendations (ranks existing Knowledge), and routing or priority suggestions. These learn from historical records and need volume plus clean data.
Generative: case summaries, work summaries, reply recommendations, service replies, and Agentforce agents that hold a conversation, reason over a topic, and invoke actions (Flows, Apex, prompt templates) to do real work. Agentforce is configured with topics (what the agent handles), instructions, and actions grounded in your data.
Platform plumbing: the Einstein Trust Layer (masking, zero data retention with the model provider, toxicity detection, audit trail) must be enabled; prompt templates ground output in records; agents need an escalation path to a human through Omni-Channel and a queue.
Traps: assuming generative features are on by default or available in every edition; pointing an agent at topics with no actions, so it can talk but not do; skipping human handoff; expecting a bot to compensate for poor Knowledge - retrieval is only as good as the articles.
Rule of thumb: deploy AI where volume is high, data is clean, and the cost of a wrong answer is low; escalate everything else.
Worked example: A password-reset agent deflects 40% of contacts but users get stuck. Adding a verification action and a queue-based escalation to a human agent fixes the real failure - not a model change.