Salesforce Hires Seven AI Agents: What It Signals for SMEs

TL;DR: On September 11, 2026, Salesforce launched seven Agentforce AI agents named like employees (Casey, Paige, Carter, Hunter, Marshall, Piper, Fin), each dedicated to one business function; six are generally available now, and Hunter, the outbound sales agent, debuts a runtime that can pursue a goal over several weeks.

Key takeaways

• Salesforce announced seven "job-ready" Agentforce agents on September 11, 2026, one per function: customer service, IT and HR, e-commerce, outbound sales, back office, inbound leads, and customer experience.
• Six agents are generally available today; Hunter (outbound sales) remains in pilot, with general availability planned for November 2026.
• The Fin agent comes from Salesforce's $3.6bn acquisition of Fin (formerly Intercom), announced on June 15, 2026.
• Salesforce claims 7 billion "Agentic Work Units" delivered since Agentforce launched, including 3.2 billion in Q2 2026 alone.

Four days before Dreamforce, held September 15-17, 2026 in San Francisco under the "agentic enterprise" banner, Salesforce moved the line: you no longer buy a module, you "hire" an agent. That vocabulary choice is not a detail; it tells you where enterprise software money is heading.

Seven first names, six agents shipping: the facts

Salesforce's official announcement of September 11, 2026 presents seven agents described as "job-ready": they arrive with the skills, actions, and data models of their function, then get customized to each company's workflows. The table below sums up the lineup.

Fin deserves its own sentence: this agent is built on technology from Fin, the company formerly known as Intercom, which Salesforce agreed to acquire for $3.6bn on June 15, 2026. Three months between signing and catalog integration: the pace of assimilation says a lot about competitive pressure. According to SiliconANGLE, most features shipped immediately, with the rest rolling out by year-end. No pricing was disclosed in the announcement.

Why give a first name to a piece of software?

Because the first name changes the budget line. A customer service module gets bought from the tools budget; "Casey, your customer service agent" gets compared to a headcount. Salesforce now measures its activity in "Agentic Work Units": 7 billion delivered since Agentforce launched, including 3.2 billion in Q2 2026, by its own count. The vendor also cites customers such as Engine, which credits Agentforce with resolving 50% of its requests automatically, and Anthropic, with 79% autonomous resolution on support. Those figures come from the vendor and deserve the usual caution, but the commercial logic is crystal clear: sell completed work, not the right to access a screen.

This shift moves enterprise software closer to a labor market. The same questions come up as in any hiring decision: what job description, what probation period, what control over what the person, here the agent, actually does. Salesforce answers that last question with Agent Script, a language that combines model reasoning with deterministic rules: certain decisions follow fixed logic rather than inference, precisely to limit unpredictable behavior.

An agent that works for weeks, not one chat session

The most structural piece of the announcement is not a first name; it is the "long-horizon" runtime that Hunter debuts. According to the analysis by PPC Land, the runtime rests on three building blocks: memory that carries context between sessions, durable execution that keeps a plan alive over time and corrects course when circumstances change, and dynamic steering that adjusts behavior based on user feedback.

Concretely, an agent of this kind no longer answers a question and vanishes: it holds a goal, say building a sales pipeline, over days or weeks, collaborating with human sellers. That is a change in kind, not degree. Most "AI agents" deployed so far were improved chatbots whose life ended with the conversation. An agent that persists raises new questions: who supervises its plan, who approves its follow-ups, what happens when it goes wrong on day 12 of a 30-day cycle. Marshall's audit trail and Agent Script's deterministic rules exist precisely because those questions have no standard answer yet.

What does this change for a small or mid-sized company?

Directly, very little: most companies with 10 to 100 employees are not Salesforce customers, and nothing in this announcement pushes them to become one. Indirectly, three concrete things.

First, normalization. When the world's largest CRM vendor sells agents by job function, the whole ecosystem follows: HubSpot is already pushing Breeze into the same slot, and the tools SMEs actually use will ship equivalent agents within a few quarters. The question "do we need an AI agent?" will fade in favor of "which one, on which process, with what controls".

Second, the real prerequisite sits elsewhere. A "job-ready" agent is only ready if the process you hand it is defined and the data it consumes is clean. A company whose customer requests land in three untriaged inboxes, or whose customer file lives in four contradictory spreadsheets, has nothing to gain from an agent, whatever its first name. The preparatory work, mapping the process, centralizing the data, deciding what the agent may do on its own, is where most of the value sits; the agent is only the last layer.

Third, budget discipline. For a standard process carried by a tool already in place, the off-the-shelf agent will almost always be the right call: there is no point building a custom customer service agent if the tool you already use ships one tomorrow. A custom agent makes sense in the other cases: a process specific to your business, a chain that spans several tools, or a hosting and data-control requirement. And sometimes the right answer is to automate nothing at all: a process that handles ten requests a week will never pay back its agent.

What I take away from the field

I have worked inside the Salesforce ecosystem as an administrator for the French nonprofit e-Enfance (Apex, LWC, 3CX telephony integration): the platform is powerful, and its complexity is exactly what reserves it for organizations with the means to operate it. For everyone else, the lesson of this announcement is not "adopt Salesforce" but "prepare your processes": the same scoping work is what makes an off-the-shelf agent or a custom tool genuinely useful. My own website runs a public demo AI agent; what took time was not the model, it was defining what it is allowed to say and do.

Salesforce just gave its software first names so that buyers compare it to employees. The useful comparison lies elsewhere: like a hire, an AI agent succeeds or fails on the quality of the job description. Write the job description first; the first name can wait.

Frequently asked questions

Are Salesforce's Agentforce agents within reach for SMEs?

Technically yes, but they require being a Salesforce customer, and the total cost of that ecosystem exceeds the tooling budget of most companies with 10 to 100 employees. Salesforce disclosed no pricing in the September 11, 2026 announcement. For an SME, the real question is which process to hand to an agent, not which platform to adopt.

What is a job-ready AI agent?

According to Salesforce, a job-ready agent ships with the skills, actions, and data models of its function, then gets customized to each company's workflows. In practice, the agent only works if the process and the data it consumes are clean: configuration and data quality remain the customer's job.

Off-the-shelf AI agent or custom AI agent: how do you choose?

An off-the-shelf agent fits when the process is standard and the tool carrying it is already in place, such as customer service inside an existing CRM. A custom agent makes sense when the process is specific to your business, spans several tools, or requires controlled hosting. In both cases, a poorly defined process will produce a useless agent.