What do you like best about Salesforce Agentforce?
The most compelling aspect of Salesforce Agentforce is its shift from passive retrieval to autonomous execution grounded directly in native platform metadata.
Instead of functioning as an isolated conversational wrapper that merely summarizes text or requires complex external API middleware, Agentforce operates inside the core CRM layer:
Autonomous Orchestration via the Atlas Engine: Rather than following rigid, hardcoded decision trees or basic prompt-and-response patterns, the Atlas Reasoning Engine evaluates user intent, formulates a multi-step plan, refines it against real-time context, and decides which actions to trigger dynamically.
Direct Reusability of Existing Business Logic: It leverages platform tools you already have in place—Flows, Apex classes, and MuleSoft APIs—treating them as modular skills. You do not need to rebuild your operational logic; the agent executes the exact validation rules, triggers, and automations your org already enforces.
Native Security and Metadata Inheritance: Because it runs within the Salesforce trust boundary, it natively inherits Object-Level Security, Field-Level Security, sharing rules, and user-context permissions. The agent cannot query, manipulate, or expose records that fall outside the configured security profile.
Unified Grounding with Data Cloud: It natively pairs structured CRM records with unstructured external data—such as PDFs, documentation, and communication histories—giving the agent a unified context without brittle custom ETL pipelines.
Deterministic Handoffs: When an issue falls outside guardrails or hits predefined business thresholds, it packages the conversation history, attempted actions, and active context for a clean handoff to human representatives. Review collected by and hosted on G2.com.
What do you dislike about Salesforce Agentforce?
Salesforce Agentforce has real enterprise utility, but deploying it exposes several sharp friction points:
High Total Cost of Ownership and Metering Uncertainty: The conversation-based billing model (paying per conversational turn or session) makes budgeting unpredictable compared to traditional user-license seats. If an agent enters an unexpected loop or handles low-value interactions, costs accumulate rapidly.
Heavy Dependency on Data Cloud Hygiene: Agentforce is only as reliable as the data backing it. If an org lacks a clean, consolidated Data Cloud setup with standardized schemas and unified identity resolution, the Atlas engine hallucinates context or fails to retrieve relevant data. It turns into an expensive test of an organization's existing data debt.
Debugging and Observability Gaps: Tracing why the reasoning engine chose one action over another during an edge-case failure can be opaque. While prompt logs and execution traces exist, they lack the granular, step-by-step determinism of traditional Flow or Apex debug logs. Troubleshooting non-deterministic runtime paths remains frustrating.
Action Granularity and Guardrail Fragility: Exposing standard Flows or Apex directly to an LLM requires strict input validation. If a Flow was not originally designed with explicit semantic constraints, an agent might supply plausible but invalid parameter combinations, triggering unhandled faults or partial transaction rollbacks.
Platform Lock-In: Because the intelligence layer is tightly bound to Salesforce’s proprietary metadata, Atlas engine, and Data Cloud constructs, custom agent logic cannot be easily ported to multi-cloud or open-source agentic frameworks (like LangGraph or AutoGen). You are fully married to Salesforce's ecosystem evolution and pricing tiers. Review collected by and hosted on G2.com.
