TL;DR:

  • Agentforce is Salesforce’s native AI agent platform — deploy autonomous agents for sales development, service case resolution, commerce, and marketing directly inside Salesforce without building your own infrastructure
  • Agents are configured via Agent Builder (a declarative low-code tool) with access to Salesforce data, flows, Apex actions, and external APIs via MCP connections
  • The main advantage over custom frameworks is that Agentforce agents already have access to your CRM data and permissions model; the tradeoff is that complex reasoning tasks still require prompt engineering discipline and the platform has real limits outside the Salesforce ecosystem

Salesforce has been putting AI into CRM features since Einstein launched in 2016. Agentforce, introduced properly at Dreamforce 2024 and significantly expanded through 2025, is a different proposition — not AI-assisted features bolted onto existing workflows, but an autonomous agent layer that can take multi-step actions across your Salesforce environment with minimal human intervention.

By mid-2026, Agentforce has shipped in enough production deployments that the hype cycle is over and the practical picture is clearer. Here is what it actually is, how it works, and when to use it.

What Agentforce Is

Agentforce is a platform for deploying AI agents within Salesforce. An agent in Agentforce terms is a configured system with:

  • A role and persona — the agent has a defined job and a name (e.g., “Service Agent for EMEA Customers”)
  • A set of topics — domains the agent can handle, each with instructions and available actions
  • Actions — the tools the agent can invoke: Salesforce Flows, Apex methods, API calls, Data Cloud queries, or MCP server connections
  • A model — by default Salesforce’s own Einstein models, with options for GPT-4o, Claude, or custom models via Model Builder
  • A channel — where the agent operates: Salesforce messaging, email, Slack, a website chat widget, or a custom API integration

The key distinction from earlier Salesforce AI features is autonomy. An Agentforce agent decides which action to take based on the conversation and the customer’s record — it does not just suggest the next best action to a human; it executes it.

What You Can Build

Salesforce ships four pre-configured agent templates that cover the most common use cases:

Sales Development Representative (SDR) Agent — handles inbound lead qualification, responds to prospect enquiries, books meetings, and creates Salesforce activities. The agent has access to Account, Contact, and Lead objects and can create Tasks and Events.

Service Agent — resolves customer service cases autonomously by querying knowledge articles, looking up order or case history, processing returns, and escalating to a human when the case requires it. Connects to case management flows already built in Service Cloud.

Commerce Agent — helps customers navigate product catalogues, check order status, apply promo codes, and complete checkout. Integrates with Salesforce B2C Commerce or third-party commerce via API.

Marketing Agent — answers questions about campaigns, helps segment audiences, and drafts marketing copy using brand guidelines stored in the org.

Each template is a starting point. Agent Builder lets you modify the topics, instructions, and available actions without code.

How Agent Builder Works

Agent Builder is the low-code configuration environment for Agentforce. The experience is closer to configuring a service workflow than writing code — you define topics in plain language (“this agent handles return requests”), write instructions (“when a customer requests a return, look up their order and check if it’s within the 30-day return window using the CheckReturnEligibility flow”), and attach available actions.

Actions are the interesting part. Out of the box, agents can call any Salesforce Flow, invoke an Apex class as a tool, query or write to any object the running user’s profile has access to, or call Data Cloud for real-time customer profile data. External tool connections are handled via MCP — Agentforce supports MCP server connections so agents can reach external APIs, databases, or services without building a custom integration layer.

Guardrails are set per-topic: you can restrict which objects an agent can write to, require human confirmation for high-value actions (e.g., any refund over £500), and set escalation triggers that hand off to a human agent queue.

The Einstein Trust Layer

Salesforce is explicit that Agentforce agents do not use your CRM data to train third-party models. All LLM calls are routed through the Einstein Trust Layer, which strips PII from prompts before sending to external models and does not retain data for model improvement unless explicitly opted in.

For organisations in regulated industries — financial services, healthcare, public sector — this matters. The trust layer means you can point an agent at customer records without a separate data processing review for each LLM call. Whether your legal team accepts this architecture depends on your compliance environment, but Salesforce publishes the data flow documentation and it has passed review at a significant number of enterprise deployments.

Where Agentforce Works Well

Salesforce-native workflows are the obvious sweet spot. If your sales or service process already lives in Salesforce — leads in CRM, cases in Service Cloud, orders in B2C Commerce — an Agentforce agent can participate in those workflows immediately without integration work. The data access is already there.

High-volume repetitive tasks with defined decision trees suit Agentforce well. Lead qualification with clear criteria, return eligibility checking, order status responses — these are tasks where the agent’s reasoning burden is low and the action library is well-defined.

Teams already invested in the Salesforce platform benefit from Agentforce because the skills and tooling carry over. If your org has Salesforce admins who know Flow and a developer who knows Apex, they can build and maintain Agentforce agents without learning a new framework.

Where the Limits Show

Complex reasoning tasks — multi-hop research, open-ended analysis, tasks where the right answer requires synthesising information from many disparate sources — are not where Agentforce shines. The platform is built for transaction execution, not deep thought.

Data outside Salesforce requires integration. Agentforce agents can reach external systems via MCP connections, but you need to build those connections. For organisations where the key data lives in a data warehouse, a homegrown system, or a competitor’s CRM, the integration overhead reduces the advantage over a custom agent framework.

Non-Salesforce channels need more work. Deploying an Agentforce agent inside Salesforce messaging or a connected Slack workspace is straightforward. Deploying one inside a custom web application or an internal tool built on another stack requires the Agentforce API, which adds complexity.

Pricing

Agentforce is priced on a per-conversation model — roughly $2 per conversation in the standard tier as of 2026, with volume discounts at scale. A “conversation” is defined as a full agent session from initiation to resolution or escalation. For high-volume service use cases, this pricing model rewards efficient resolution; agents that resolve cases without escalation cost the same as agents that pass cases on to humans.

If you are replacing high-volume human agent work, the economics can be compelling. A tier-1 support agent handling 30 conversations per day at fully loaded cost is significantly more expensive than $2 per conversation at scale. If you are augmenting a small team handling complex cases, the per-conversation model is less compelling.

Agentforce vs. Building Your Own

The decision comes down to where your data lives and what your engineering capacity is.

If 80%+ of the data your agents need is in Salesforce, Agentforce will get you to production faster than building on LangGraph or CrewAI. The platform handles memory, session management, tool calling, guardrails, and audit logging. You configure rather than build.

If your data is distributed across multiple systems and your use cases extend well beyond CRM workflows, a framework-based agent gives you more flexibility at the cost of more engineering work. The integration overhead of MCP connections can be significant when you have six systems to wire up rather than one.

For Salesforce shops with clear CRM-centric use cases, Agentforce is the lowest-friction path to deployed agents in 2026. For organisations building general-purpose AI automation, it is one platform among several rather than the default choice.