Salesforce launches 7 AI agents built for specific jobs

For two years the usual way to sell AI agents was to hand businesses a platform and let them build. Last week Salesforce, the customer relationship management (CRM) vendor familiar to large enterprises, changed the pitch: no more bot-building toolkits, but agents already "trained" for a specific job. Vietnamese businesses that do not use Salesforce should still read closely, because the way it divides and measures the work applies directly on Zalo OA.

What happened

According to Salesforce's announcement of 11 September 2026, the company introduced a portfolio of agents it calls job-ready. Each agent ships with the skills, actions and data models for one kind of work, and businesses tailor it to their own processes, including giving it their own name. Salesforce says it has delivered 7 billion Agentic Work Units across Agentforce and Slack, 3.2 billion of them in Q2 alone.

Agent	Job	Status
Casey	Customer service across voice, SMS, WhatsApp and web chat: FAQs, returns, account management, human escalation	Generally available
Carter	Helps shoppers discover and compare products and check out inside the chat	Generally available
Piper	Engages and qualifies inbound leads from websites and inboxes	Generally available
Hunter	Outbound sales: research, outreach and pursuing opportunities over weeks	Pilot, GA planned November 2026
Paige	Employee IT and HR requests	Generally available
Marshall	Orchestrates supply chain processes with an audit record of every action	Generally available
Fin	Complex customer experience workflows across every channel	Generally available

The news came with a technical shift: Hunter is the first agent on a new long-horizon runtime that lets it pursue a goal across days and weeks, with memory between sessions, instead of finishing a single conversation. Sellers keep approval at the steps where the rules require it.

Dreamforce is Salesforce's annual conference for product announcements; photo taken in 2022. Photo: InvadingInvader / Wikimedia Commons, CC BY-SA 4.0.

What the customer numbers say, and what they do not

The most useful part of the announcement is not the list of agents but how Salesforce proves value. There is no "enhanced experience" or "optimised operations". Each example is a specific share of work the agent completed on its own:

Two examples are not rates: Asana's website agent Piper drives 4x the conversation volume, and Hibbett went live with its shopping agent in six weeks; customers deploy Piper in 45 days on average.

Read these at the right weight: they are figures the vendor chose to publish, each customer measures a different kind of work so they cannot be compared with each other, and these companies already had their customer data in Salesforce. The numbers say nothing about projects that fell short.

What really changes: from building bots to assigning jobs

The old question when adopting AI was "what can our bot answer". The way Salesforce packages its agents shows the question moving to "which jobs can we hand to an agent, and how much of each does it finish". The two questions lead to very different approaches:

• Narrow, clear scope. Casey only does customer service, Carter only does shopping. An agent that juggles advice, closing, complaints and reminders is hard to measure and hard to fix when it goes wrong.
• The metric is the autonomous resolution rate, not how many questions the bot can answer. A bot that answers 500 questions but still hands 70% of chats to a human is not carrying the work.
• Outcome-based pricing is appearing. Since June 2026, Salesforce's customer service agent charges only when it resolves an issue on its own; if the customer asks for a human or leaves unhappy, there is no charge.

A customer panel at a Salesforce event. Stock photo (CC0), Raysonho / Wikimedia Commons.

Chạm AI's view: how small businesses can apply it

Most of our clients sell through Zalo OA, Facebook Pages and websites; they do not use Salesforce and do not need to. The way the work is divided, though, can be brought home right away in three steps:

1. Pick exactly one highly repetitive job. Open last month's Zalo OA inbox and count what customers ask most: price, stock, booking, delivery fee. That is the first job, not "answer everything".
2. Set the target as an autonomous resolution rate before signing, for example "the agent fully resolves 50% of price and stock questions after 8 weeks", with a definition of done and of when to hand off to a human. A vendor unwilling to commit to a measurable metric deserves a second question.
3. Expand once the first job is stable, one job at a time. Hibbett's six weeks is a reasonable reference for one job when the data is clean; if the data still lives in screenshots and old messages, add time to organise it.

Costs for each level of agent, including the monthly AI fee based on chat volume, are in our chatbot development pricing guide for 2026. If Zalo is your main channel, the Zalo OA chatbot guide explains which OA plan can host an agent. The steps from picking a job to running it are in our AI agent implementation process, and last month we looked at why many companies run AI agents but few manage to scale them; narrow jobs are part of the answer.

Sources: Salesforce press release "Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work", 11 September 2026 (agent names, availability, customer figures); Salesforce press release on Agentforce Help Agent, 25 June 2026 (pay-per-resolution pricing); cross-checked with Enterprise DNA and AI Agent Store. The rates are figures published by Salesforce and have not been independently verified.

Frequently asked questions