Salesforce pushes an “agentic” customer service workforce at Dreamforce, with new AI interfaces, voice automation, and governance

Salesforce used its Dreamforce keynote to make a blunt claim: the next phase of customer experience will not be powered by better chatbots, but by governed AI agents that behave like a digital workforce. The company’s messaging to CX leaders focused on a harder question than automation. It emphasized whether AI can resolve the right problems, carry out approved actions across enterprise systems, and know when to stop and route work to a human.

At the center of the pitch was a vision of enterprise AI that is “agentic,” not just conversational. Salesforce described an approach built on four layers: connected data, applications and business semantics, agents, and a new AI interface. In this framing, an agent should not simply fetch an answer from a knowledge base. It should understand a customer’s situation, determine a permitted next step based on policy and permissions, and then execute that work or escalate when it cannot do so safely.

Salesforce also introduced AIforce, described as an umbrella for AI interfaces designed to bring Salesforce context and actions into Claude, Slack, Salesforce Lightning, and other environments via an SDK. For CX teams, the company argued this shifts the value of CRM away from its user screens and toward the governed customer data, business rules, workflows, permissions, and audit trail underneath. Salesforce co-founder Parker Harris highlighted the governance tradeoff implied by that shift, saying CX leaders will need to treat AI interfaces as new operational surfaces with clear boundaries on what information each user or agent type can access and what actions it can take.

The keynote extended the agentic argument to voice and contact centers, including Salesforce’s plans for its Contact Center as a Service (CCaaS) offering. Salesforce said CCaaS is scheduled to go live in October, pitching contact center, telephony, and CRM as a unified system designed to reduce “tool stitching,” the operational friction created when agents bounce between separate telephony, CRM, knowledge, workflow, and customer history systems. The company presented Agentforce Voice through a live example tied to Adecco’s recruiting workflow, with a caller receiving a real-time answer about why an application is paused, providing a new certification number, having it verified, and then seeing the application updated during the call.

Salesforce’s speakers also emphasized that the quality target is resolution, not containment. Patrick Stokes, described in the keynote remarks as connecting trust and security to Salesforce’s role, framed the platform as the place where customer data, workflows, permissions, and security rules are held. The company’s message to buyers was to measure whether AI can resolve a customer’s underlying issue safely, while preserving a complete record and a clear recovery path if escalation is required.

A recurring theme was that scaling useful agents is difficult, especially because decision-making systems can be probabilistic. Salesforce said it is responding with out-of-the-box, role-specific agents that business users can deploy alongside IT teams, rather than requiring every customer to build everything from scratch. In an example intended to illustrate the target experience, Salesforce said Adecco’s Ada recruiting agent can manage a 24/7 voice conversation, recognize candidate identity and application status, identify an expired certification, validate renewal information, and update the application. Adecco told the keynote it has handled 2.6 million agent interactions this year, and its CEO Denis Machuel said recruiters saved 35–40% of their time through agentic AI and that the business placed 20,000 more people year to date, attributing that to improved conversations.

Salesforce also pointed to scale inside its own support experience. The company said Casey, its AI agent, has resolved five million service conversations since launch. Still, Salesforce warned through its own framing that scale alone is not the metric that matters for CX. The keynote stressed that CX leaders should examine outcomes such as repeat-contact behavior, satisfaction and complaint levels, escalation quality, and which categories of issues remain unsuitable for automation.

Beyond interfaces and agent demos, Salesforce introduced Koa, described as its first CRM reasoning model for Customer 360 built using NVIDIA Nemotron technology and trained on synthetic data. Salesforce said Koa is designed to support long-running agents that execute multiple tasks and reach complex outcomes, citing “next-best action” decisions in complicated support cases. The company also stressed model choice, saying customers should select models based on price-performance needs for a specific result rather than relying on a single universal model.

Finally, Salesforce tied its agentic push to governance and security, presenting Salesforce Guardian as a new security suite positioned as an evolution of Salesforce Shield and Trust capabilities. The keynote described the suite as focusing on additional guardrails around AI and actions, though the extracted announcement text in this release cuts off before the full list of functions is shown. Still, the throughline was consistent: trusted customer experience automation depends on policy enforcement, action auditability, and clear human handoffs when the system cannot act safely.

For CX leaders evaluating what to do next, the company’s advice was operational. It urged teams to stop judging AI by deflection and instead judge it by whether agents resolve the right problems, improve customer journeys, and escalate appropriately. Salesforce also urged a context-mapping exercise for complex service cases, identifying which systems are the authoritative sources for customer, asset, product, entitlement, policy, operational status, and knowledge, then deciding what the agent is allowed to recommend, complete, or escalate. What remains less concrete is how broad the promised capabilities will be across regions, channels, and integrations at launch, and which governance constraints will be configurable versus fixed. The keynote also did not provide independent validation of Adecco’s productivity and placement outcomes, instead citing company-provided figures.

Looking ahead, the most immediate milestones to watch are whether CCaaS functionality and telephony integrations deliver on the “single system” promise in the October timeframe, how effectively role-specific agents perform on real, representative customer cases, and how Salesforce Guardian’s AI governance capabilities are defined and measured in practice. CX teams will likely be looking for clearer, testable evidence on resolution quality, escalation handling, and the reduction of repeat contacts once these agentic workflows are deployed across customer-facing channels.