MANILA, Philippines — Salesforce is escalating the enterprise AI race with a new generation of “job-ready” agents designed to do something far more ambitious than answer questions: take responsibility for business objectives that can unfold over days or even weeks.
The software giant has introduced specialized Agentforce agents for customer service, IT and human resources, shopping, sales, supply chains and lead generation, arguing that the next stage of corporate AI will be defined not by better chatbots but by systems capable of actually performing work across multiple applications and business processes.
That distinction is becoming increasingly important.
Companies have spent the past several years experimenting with generative AI that can draft emails, summarize documents or answer employee questions.
Salesforce now wants businesses to trust AI with something much bigger:
an objective.
A sales agent, for example, could be told to rescue at-risk deals before the end of a quarter, determine what needs to happen, research accounts, prepare outreach, monitor new information and keep revising its plan as circumstances change.
The work may continue long after the original conversation ends.
And that could represent one of the most consequential shifts yet in the battle over how AI changes white-collar work.
Salesforce Introduces Seven Specialized AI Agents
Instead of asking customers to construct every AI agent from scratch, Salesforce is packaging agents around specific business roles.
The portfolio currently includes:
Casey — Customer service: Handles requests through channels including voice, SMS, WhatsApp and web chat, including FAQs, returns, account management and escalation to human agents.
Paige — IT and HR: Handles employee requests through Slack, portals and other workplace systems.
Carter — Commerce: Helps customers search for and compare products, answer shopping questions and potentially complete purchases inside a conversation.
Hunter — Outbound sales: Researches prospects and works through sales opportunities over extended periods.
Marshall — Supply chain: Coordinates back-office workflows while maintaining records of actions performed.
Piper — Inbound sales: Engages website visitors and inbound prospects, qualifies potential customers and attempts to convert them into pipeline.
Fin — Customer experience: Handles more complex customer-service workflows across different channels.
Salesforce says each agent arrives with skills, data models and actions relevant to its designated job but can still be customized for an individual company’s processes and policies.
The pitch is straightforward: rather than giving companies an empty AI toolkit and telling them to figure out what to build, Salesforce wants to sell something closer to a digital employee template that already understands a category of work.
But Not Every Agent Is Fully Available Yet
That distinction is important because “introduced” does not necessarily mean every feature is immediately available to every customer.
Casey, Paige, Carter, Marshall, Piper and Fin are generally available, according to Salesforce.
Hunter—the agent that best demonstrates Salesforce’s new long-duration work model—is currently in pilot, with general availability scheduled for November 2026.
The company is also rolling out other pieces of the Agentforce system on different schedules.
Multi-Agent Orchestration, which allows specialized agents to hand work between one another, is already generally available.
AI Skills, which lets employees teach Agentforce Coworker how to perform tasks, is in pilot and is expected to become generally available in October.
Agent Optimizer, designed to test agents and identify ways to improve their performance, is also scheduled for general availability in October 2026.
Salesforce itself cautions that some announced services and features remain under development and advises customers to make purchasing decisions based on functionality currently available.
That caveat matters in an AI industry where announcements can sometimes run months ahead of real-world deployment.
The Biggest Technical Change: AI That Remembers What It Was Doing
Salesforce’s most significant announcement may be its new long-horizon Agentforce runtime.
Traditional chatbots are largely session-based.
Ask a question.
Receive an answer.
Close the conversation.
The new runtime is designed to keep an objective alive much longer.
Salesforce says it does this through three main capabilities:
Memory retains context and progress across sessions.
Durable execution allows a plan to continue over time and resume when circumstances change.
Dynamic steering allows users to provide new instructions and alter the agent’s direction while work is underway.
Hunter is the first Salesforce agent built on this long-horizon runtime.
Imagine a sales manager telling Hunter:
Find which major deals are likely to slip this quarter and improve the odds of closing them.
Instead of returning a one-time list, the agent could theoretically analyze opportunities, identify accounts requiring attention, research the companies, recommend steps, prepare communications and continue monitoring progress over several weeks.
As new information appears, it could modify the plan.
That is a fundamentally different model from simply asking ChatGPT or another AI assistant, “Which deals look risky?”
Why Salesforce Calls Them ‘Job-Ready’
The phrase “job-ready” is marketing language, but it describes an important shift in enterprise AI architecture.
Salesforce is trying to move the unit of automation from a task to a role or workflow.
A conventional automation might perform one predetermined action:
When a customer submits a form, create a CRM record.
An AI agent could theoretically interpret the customer’s situation, determine which actions are appropriate, interact with several systems and decide when human approval is necessary.
That is why Salesforce is emphasizing that its agents operate using information stored in Customer 360, where companies already keep customer records, sales opportunities, service histories and business processes.
Salesforce’s argument is that general-purpose AI intelligence alone is becoming increasingly commoditized.
The more defensible advantage is giving AI access to a company’s proprietary business context, workflow rules, customer data, permissions and governance systems.
That is also why Salesforce has been building what it calls an Enterprise AI Harness: infrastructure designed to connect AI models and agents to corporate information while maintaining security and control.
Companies Are Already Using Earlier Versions
Salesforce is backing its pitch with several customer case studies.
Travel platform Engine says an Agentforce help agent now completely resolves about 50% of chat inquiries.
Salesforce’s detailed Engine case study also says customer-support handling time declined 15% while customer satisfaction increased 16%.
Corporate travel platform Perk, formerly TravelPerk, says its prospecting agent is responsible for approximately 60% of its North American sales pipeline and helped increase call volumes threefold.
Salesforce also says Autism Queensland’s employee-service agent resolves roughly 70% of administrative requests, while Anthropic’s use of Fin autonomously resolves about 79% of the conversations the system sees.
Those figures are individual customer results rather than independent benchmarks for what every company should expect.
The outcomes can depend heavily on the workflow being automated, the quality of underlying data, how narrowly the agent’s role is defined and how much human intervention remains necessary.
Still, they provide evidence that agentic AI is moving beyond small laboratory demonstrations.
Even U.S. Airport Security Is Using an AI Agent
One of Salesforce’s most prominent recent deployments is at the U.S. Transportation Security Administration.
The TSA has deployed an AI agent called Ace to answer traveler questions about airport-security procedures.
Salesforce says Ace handles around 100,000 traveler conversations each month and resolves 96% of routine inquiries without human escalation, including questions involving liquids and traveling with medical equipment.
That example helps illustrate where AI agents are currently most mature.
Highly repetitive queries with well-defined information and clear escalation rules are much easier to automate than decisions involving significant ambiguity, financial consequences or judgment.
Salesforce’s latest launch attempts to push agents further into that second category.
Salesforce Says Agent Activity Is Exploding
The company now says Agentforce and Slack have delivered 7 billion Agentic Work Units, Salesforce’s measurement of discrete production tasks completed by AI.
About 3.2 billion of those units occurred during its latest fiscal second quarter alone.
Salesforce defines an Agentic Work Unit as a discrete AI-executed task such as resolving a support case, updating a record or triggering a workflow.
The number therefore should not be interpreted as seven billion full jobs or seven billion customer interactions.
It is Salesforce’s internal measure of agent activity.
Even with that qualification, growth is rapid: Q2 work units increased 97% quarter over quarter.
And Salesforce is beginning to generate meaningful revenue from the technology.
Agentforce ARR Has Surpassed $1.5 Billion
In its fiscal second quarter ended July 31, Salesforce reported that Agentforce annual recurring revenue exceeded $1.5 billion, up more than 240% year over year under its current definition.
Agentforce and Data 360 combined reached nearly $3.9 billion in ARR, up more than 210%.
There is an accounting nuance worth noting.
Beginning in Q2, Salesforce broadened its Agentforce ARR measurement to include AI offerings such as Slackbot and Headless 360, so the growth figure is not necessarily a pure apples-to-apples measure of the original Agentforce product alone.
Still, Salesforce’s overall financial performance suggests enterprise customers are increasing spending.
Second-quarter revenue reached $11.3 billion, up 11% year over year, while subscription and support revenue rose 12% to $10.8 billion.
Salesforce also raised its full-year fiscal 2027 revenue guidance to between $46.1 billion and $46.4 billion.
Independent Reuters coverage noted that Salesforce’s stronger results and Agentforce momentum have helped calm some investor anxiety over whether AI will destroy the economics of traditional business software.
Because Salesforce Is Fighting a Bigger AI Threat
That context is crucial.
Salesforce is not merely trying to capitalize on AI.
It is also defending itself from it.
Investors have increasingly questioned whether powerful general-purpose AI models could eventually bypass traditional software applications altogether.
If an AI agent can understand what a user wants and interact directly with databases, APIs and business systems, companies may wonder why employees need to navigate dozens of software interfaces.
Those concerns resurfaced sharply this month after OpenAI released GPT-6 Astra, sending shares of major software companies—including Salesforce, Intuit and ServiceNow—down roughly 4% to 5% in a single trading session amid fears of increased AI competition.
Salesforce’s response is effectively:
AI does not make CRM obsolete if CRM becomes the place where AI does the work.
That explains the urgency behind Agentforce.
Salesforce Is Also Deepening Its Anthropic Partnership
Salesforce is not attempting to build every underlying AI model itself.
Instead, it is positioning its platform as the business layer through which different models interact with corporate data and workflows.
In August, Salesforce and Anthropic expanded their partnership through Claudeforce, integrating Anthropic’s Claude more deeply across Salesforce and Slack.
Salesforce says Claude is now its preferred model across several internal and customer-facing applications, including Slackbot and Agentforce Coworker.
The company’s strategy increasingly resembles an AI operating layer rather than a single-model bet:
Salesforce owns customer data, workflows, permissions and applications while models from companies such as Anthropic provide portions of the underlying intelligence.
That approach could allow Salesforce to benefit even as the most powerful foundation model changes from year to year.
But More Autonomous Agents Create New Risks
The same capabilities that make AI agents more useful also make mistakes potentially more consequential.
A chatbot that gives an incorrect answer creates one kind of problem.
An AI agent that takes an incorrect action creates another.
An agent could potentially contact the wrong customer, modify a record incorrectly, apply an inappropriate discount, misunderstand an employee request or trigger a workflow that should have required approval.
The risk increases as agents gain access to more systems and remain active over longer periods.
Salesforce says its agents operate within corporate rules, permissions and security controls and that its Agent Script system allows companies to combine AI reasoning with deterministic rules.
Its new Agent Optimizer is also intended to let companies analyze agent behavior and session traces to identify failures and improve performance.
But the broader AI industry is simultaneously confronting growing concerns about autonomous-agent behavior.
Recent incidents involving AI agents accessing external systems without authorization have intensified debates over how much independence these systems should receive and what safeguards should be mandatory.
That makes governance particularly important as enterprise agents move from simply producing text to taking real-world actions.
Human Approval Isn’t Disappearing
Despite the “digital labor” narrative, Salesforce’s own architecture still assumes people will remain involved.
The Hunter example explicitly includes guardrails defining when the agent can act autonomously and when it must obtain seller approval.
Even earlier Salesforce data showed that increased AI adoption did not eliminate escalation to human employees.
In Salesforce’s Agentic Enterprise Index, customer-service escalations to humans actually increased from 22% in Q1 2025 to 32% in Q2 2025 as AI agents handled a broader variety of interactions.
That suggests the more realistic near-term model may not be humans versus AI.
It is AI performing the repetitive and procedural portions of work while humans remain responsible for exceptions, judgment, relationships and accountability.
The exact division will differ by industry.
Data Quality May Be the Real Bottleneck
There is another problem no sophisticated agent can completely solve:
bad corporate data.
An agent deciding what to do based on incomplete customer histories, duplicate records or disconnected databases can produce poor decisions regardless of how intelligent its underlying model is.
Salesforce itself has acknowledged this challenge in its research.
Its recent sales survey found that disconnected systems were slowing AI initiatives at many organizations and emphasized the need for companies to clean and unify their data before expecting agents to deliver reliable results.
That is one reason Salesforce has tied Agentforce so closely to Data 360 and Customer 360.
The company wants the agent not simply to have a powerful model but to know:
Who is this customer?
What did they buy?
What happened previously?
What is the company allowed to offer them?
Who needs to approve the next action?
Without those answers, an autonomous agent is essentially guessing with confidence.
The Bigger Question: When Does an AI Assistant Become an AI Worker?
For several years, corporate AI has mostly been sold as a copilot.
It sits beside an employee.
The human asks.
The AI responds.
Salesforce’s new strategy moves further toward delegation.
An employee states a goal.
The agent figures out the work.
It remembers what happened yesterday.
It returns tomorrow.
It coordinates with other agents.
It executes approved actions.
And it continues until the objective is completed—or until a person intervenes.
That is why the long-horizon runtime may ultimately matter more than the playful names Casey, Paige, Hunter or Marshall.
If AI can reliably retain context, operate across systems and pursue a business objective for weeks, the economics of many office workflows begin to change.
Salesforce is betting that businesses are ready to make that transition.
Its growing Agentforce revenue and billions of reported work units suggest companies are at least willing to experiment.
But the next phase will require Salesforce to prove something much harder than whether an AI can answer a customer’s question.
It has to prove that companies can safely give AI responsibility for getting the job done.
And when that job takes weeks, touches multiple business systems and affects real customers, the difference between an impressive demonstration and a trustworthy digital worker becomes much bigger.
Leave a Reply