What Agentforce Actually Is
Einstein Copilot, launched in early 2024, was Salesforce's AI assistant layer: a chat interface within Salesforce that could answer questions about your CRM data, draft email responses, summarize accounts, and execute basic actions through natural language. Useful, but fundamentally reactive — you asked, it answered or acted.
Agentforce represents a genuine architectural evolution beyond this. The core addition is the ability to define autonomous agents that can run multi-step workflows without user prompting at each step. An Agentforce service agent can monitor a case queue, classify incoming support requests by severity and product area, draft initial responses, escalate cases that match certain criteria, and update case fields — all without a human approving each action.
Salesforce built Agentforce on the Atlas reasoning engine, a proprietary model layer that sits above the underlying LLM (which can be Salesforce's own xGen models or third-party models via the Einstein Trust Layer). Atlas handles the planning, tool selection, and step orchestration that makes multi-action workflows possible. The Trust Layer enforces data residency, PII masking, and audit logging — essential for enterprise customers operating in regulated environments.
Agent Builder: The Developer Story
The primary interface for building Agentforce agents is Agent Builder in Setup. You define the agent's role, its persona (how it presents itself to users), its actions (which Salesforce objects and flows it can interact with), and its topic scope (the types of requests it's authorized to handle). The configuration is declarative for most common patterns, with Apex and Flow extensions available for custom logic.
The action library is the most important part of the developer story. Agentforce ships with pre-built actions for common Salesforce objects: creating and updating records, querying with SOQL, sending emails, creating tasks, updating opportunities. Custom actions can be defined through Flow Actions (for declarative logic) or Apex Actions (for imperative code), making the platform extensible to essentially any Salesforce capability.
The testing experience improved significantly at Dreamforce — Agentforce Testing Center provides a sandbox environment where you can run agents against synthetic data and evaluate their action sequences before deploying to production. This was a gap in the Einstein Copilot offering that the community had complained about loudly, and the platform team addressed it directly. The test suite includes a simulation mode that shows you each step an agent would take without actually executing the actions, which is useful for reviewing complex multi-step workflows.
The Use Cases That Actually Work
After evaluating Agentforce through its September 2025 launch period, the use cases that most clearly delivered value were high-volume, well-structured workflows where the action space is bounded and the business rules are well-defined. Service case routing and initial response drafting was the canonical example — the inputs (incoming cases with description, category, and product) are structured, the actions are clearly scoped, and the quality bar for initial routing is achievable with current model capabilities.
Sales development workflows — qualifying inbound leads, enriching contact data, scheduling discovery calls — also emerged as strong fits. The workflow steps are clear enough that an agent can execute them reliably, the cost of occasional errors is low (a slightly wrong qualification score gets corrected by the sales rep), and the volume of work makes automation economically compelling.
The use cases that struggled were open-ended or required judgment about edge cases outside the training distribution. Agents asked to handle complex billing disputes or negotiate contract modifications ran into the limits of what the model could reliably reason about, and errors in those contexts had material consequences. The discipline of scoping Agentforce agents tightly — clear topic boundaries, explicit fallback to human queues for out-of-scope requests — separated successful deployments from problematic ones.
Pricing and the Salesforce Reality
Agentforce launched with a consumption-based pricing model at $2 per conversation, with volume discounts for enterprise agreements. This pricing model is fundamentally different from seat-based Copilot pricing and reflects Salesforce's bet that autonomous agents handling thousands of customer interactions per day will generate enough business value to justify the cost.
For high-volume service operations — a telecom company handling 50,000 support interactions per day — $2 per conversation adds up quickly. The business case depends on what it displaces: if each Agentforce interaction saves 10 minutes of human agent time at $30/hour, the ROI at $2 per conversation is straightforward. If Agentforce handles the first 5 minutes and then escalates 80% of cases to human agents anyway, the math is different.
The pricing model also reflects Salesforce's confidence in the product's ability to handle meaningful volumes of conversations end-to-end. Organizations evaluating Agentforce should build a careful model of containment rate — the percentage of conversations the agent handles fully without human escalation — before committing to the pricing model. The difference between 60% and 80% containment rate has large implications for total cost.