Agentforce is Salesforce’s platform for building AI agents that work inside your Salesforce data and workflows: answer a customer, update a record, look something up, or hand a case to a person. Salesforce’s own product page calls it “the AI agent platform that delivers 24/7 autonomous support at enterprise scale.” Customer service is where the platform has the most published results, and it’s the use case Agentforce service consulting covers.
Under the hood, an agent reads a request, picks the job it fits, pulls company data from Data 360, runs actions through Flows or APIs, checks what happened and repeats until the goal is met or a person takes over. The Einstein Trust Layer wraps that loop with zero data retention and toxicity detection.
The reported results are real but vendor-sourced: Nexo cites 62% case resolution, Engine says 50% of inquiries resolve automatically, and Salesforce puts Agentforce ARR above $1.5 billion. What decides your own result is groundwork: clean data, tight permissions, a narrow first job and a test set of real cases.
TL;DR
- Agentforce is Salesforce’s platform for AI agents, which choose each next action from the request in front of them.
- The Atlas Reasoning Engine routes each request to a subagent, formerly called a topic, and keeps looping on it until a person steps in or the job is done.
- Prebuilt agents cover service, sales, commerce, marketing and employee support, and they work on the web, by voice and in messaging apps.
- Salesforce reports Agentforce ARR above $1.5 billion, though the figure now counts more products than earlier quarters did.
- Launch with one high-volume queue on clean data, and test it against real past cases first.
What an Agentforce Agent Does
A chatbot follows a decision tree somebody drew in advance. An agent chooses its next step from the request in front of it, using the instructions, data and actions you’ve given it.
In practice, an agent can look up an order, check a return policy in your knowledge base, send a status update and close the case without a person touching it. Customer service is where Salesforce and its customers publish results, which makes a service agent the natural first build and an easy one to measure.
Agents earn their keep on high-volume, repeatable requests (order status, billing questions, policy lookups and account changes) where the right answer lives in your data. Rare, high-stakes requests belong with a person, and a good agent design sends them there on purpose.
How Does Agentforce Work, Step by Step?
1. A request arrives. The Atlas Reasoning Engine sorts it into a topic, which Salesforce’s engineering blog now calls a subagent: one job to be done, with its own instructions, policies and actions.
2. The agent grounds itself. It runs semantic search over structured and unstructured data in Data 360, so answers draw on your records and documents.
3. It picks and runs actions. Actions are Flows, Apex, prompt templates and APIs, with MuleSoft and Slack as native tools alongside Data 360.
4. It loops. Salesforce describes a reason, act, observe cycle that repeats until the goal is met, with clarifying questions or confirmation requests along the way.
5. Guardrails apply. The Trust Layer adds zero data retention, toxicity detection, secure data retrieval and an audit trail around the model.
Step 1 deserves the most care. If two topics overlap, the engine has to pick between them, so write each topic’s instructions in a way a new hire could follow and keep the list short at launch. Name topics the way your team names queues, such as billing or order status, so the routing is easy to audit later.
The Atlas Reasoning Engine post publishes no accuracy or latency figures, so run Agentforce testing on real past cases before the agent talks to a customer. Agent Builder lets you generate test cases and inspect potential responses before deployment, and Salesforce’s how-it-works page covers each layer in more detail.
Agentforce Agent Types and Channels
Salesforce ships prebuilt agents for the teams that talk to customers or employees, and your admins and developers can build their own from the same pieces in Agent Builder.
- Service Agent: replaces the traditional chatbot and handles a wide range of service issues without preprogrammed scenarios.
- Sales Development Representative: engages prospects around the clock, answers product questions, handles objections and books meetings.
- Sales Coach: runs role-play sessions so sellers can practice pitches and objections for specific deals.
- Commerce agents: a Merchandiser for the ecommerce team, a Personal Shopper for recommendations and a Buyer Agent that helps B2B buyers find products and track orders.
- Marketing and employee agents: a campaign agent that analyzes and personalizes campaigns, and an employee agent (formerly Einstein Copilot) that searches data and executes action plans.
Agents reach people on the web, on the phone through Agentforce Voice and in messaging apps, with Slack as one of the native tools. Grupo Falabella, for one, scaled its WhatsApp support 3x with Agentforce.
Pick the channel where your volume already sits and build the first agent there. A second channel is easier once the first agent’s topics and test set have settled.
What Salesforce and Its Customers Report
Salesforce’s second-quarter fiscal 2027 results, released August 26, 2026, put annual recurring revenue from Agentforce above $1.5 billion, up over 240% year over year. Agentforce plus Data 360 reached nearly $3.9 billion, up over 210%, and Salesforce reports 7.0 billion Agentic Work Units delivered across Agentforce and Slack. From that quarter, Agentforce ARR includes Salesforce’s AI offerings, Slackbot and Headless 360, so older figures aren’t like-for-like.
On the customer side, Nexo reports 62% case resolution, and Grupo Falabella says it resolved 60% of inquiries autonomously. Engine cut handle time 15%, resolved 50% of inquiries automatically and deployed its customer-facing agent in 12 days. Safari365 passed 30% efficiency against a 15% goal and went live in six weeks.
Xero reports 65%+ of its online queries resolved autonomously across about 500,000 Service Agent conversations since launch. 1-800Accountant said in March 2025 that Agentforce resolved over 1,000 client engagements in its first 24 hours, with up to 50% of incoming requests resolved.
All of these figures come from Salesforce’s own material, and most carry no timeframe. Treat “up to” as a ceiling and the rest as claims to test against your own cases, and budget time for writing the knowledge articles your agent will read if yours are thin.
What to Have in Place Before You Build One
- Clean data. Grounding retrieves what your records say, so duplicates and stale fields come back in the answers.
- Tight permissions. An agent can touch whatever its actions can touch, so give each action only the access it needs.
- A narrow first job. Pick one high-volume queue where a wrong answer does little damage, and keep the list of topics short.
- A test set. Collect real past cases the agent must handle correctly before launch, and rerun them after every change.
Plan the handoff as carefully as the answers. Salesforce treats “transfer to human agent” as an action like any other, so decide which topics trigger it, how many failed attempts count, and what context passes along so the customer doesn’t repeat themselves.
After launch, Agentforce Observability monitors agent performance in near real time, and the Trust Layer’s audit trail gives your team a record to review each week. That review is a good place to find early fixes. An agent acts on whatever your data says, so any cleanup you’ve been postponing is the first task in the project.
Also READ-Non-Contact Sensing Technologies in Industrial Automation
