Dreamforce 2026: Moving Agentic AI from Capability Building to Enterprise-Scale Execution

Two years into the agentic AI cycle, the industry’s focus is shifting from demonstrating what agents can do to determining where they can operate reliably at scale. The challenge is no longer simply whether an agent can complete a task, but whether it can act on live business data, coordinate across systems, operate within enterprise controls, and deliver outcomes at a predictable cost.

Avasant’s Salesforce Services 2026 Market Insights finds that 53% of Agentforce deployments remain concentrated in foundational use cases, primarily fixed-rule automation and information retrieval. More advanced deployments require agents to reason across systems, orchestrate multistep workflows, and execute business actions with appropriate oversight.

The challenge is increasingly less about model capability alone than about enterprise readiness. Business data, workflows, permissions, and governance must be aligned to enable agents to operate reliably within established controls. Scaling agentic AI will, therefore, depend on adapting existing enterprise systems to support autonomous execution without compromising operational control.

Dreamforce 2026 showed how Salesforce is preparing for this evolution across four areas:

Table 1: Key barriers to scaling agentic AI adoption and Salesforce’s initiatives

AI Is Extending Beyond Application Boundaries

As enterprises adopt multiple AI models and agents, they will expect AI capabilities to be available in the environments where employees already work. Salesforce’s AIforce initiative reflects this shift. Claudeforce brings selected Salesforce data and workflows into Claude, while Slackforce connects Salesforce intelligence and workflows to Slack. Headless 360 extends the approach by allowing authorized external agents to discover and invoke Salesforce capabilities through APIs and the Model Context Protocol (MCP).

The strategic implication is not that enterprise application interfaces will disappear, but that they will no longer be the only way to access business capabilities. By making its data, business logic, and workflows accessible to authorized agents, Salesforce is positioning its platform to support agent interactions across multiple interfaces. For enterprises, the priority will be to enable this cross-system access without duplicating business logic or weakening governance.

Purpose-Trained Models Are Bringing Domain Context to AI Reasoning

AI models are becoming increasingly capable of reasoning through complex tasks, but enterprise execution requires more than general-purpose intelligence. A model working within CRM, for example, needs to understand the business processes, terminology, objectives, and outcomes associated with sales, service, marketing, and commerce workflows.

Salesforce’s introduction of Koa, its purpose-trained CRM reasoning model, reflects this shift. Built on NVIDIA Nemotron 3 Super and trained on synthetic enterprise datasets modeled from nearly three decades of Salesforce CRM intelligence, Koa is designed to reason across CRM workflows spanning the customer life cycle. The model covers synthetic scenarios across more than 14 industries and is designed to perform tasks such as updating opportunities, routing cases, and scheduling follow-ups.

Jensen Huang, founder and CEO of NVIDIA, highlighted the rationale for this shift, noting that enterprises are increasingly building their own custom AI as every software company and enterprise becomes an AI company. Salesforce’s advantage lies in combining a purpose-trained reasoning model with its proprietary domain and process knowledge, enabling AI to reason within the specific context of CRM workflows rather than relying solely on general-purpose intelligence. As agents take on increasingly complex business processes, this ability to embed domain, workflow, and operating context into AI could become a key differentiator.

The AI Harness Is Connecting AI Reasoning to Enterprise Execution

As AI agents move from providing recommendations to executing actions, they need to understand not only enterprise data but also the relationships, permissions, business rules, and context surrounding that data. Salesforce’s Enterprise AI Harness reflects this requirement by bringing together trusted models, context, agency, actions, governance, and security. The architecture connects data, models, agents, identity, security, monitoring, and governance to establish controls around how AI operates. Salesforce Guardian and Agent Fabric further support agent governance through capabilities such as security, agent discovery, identity, and control over agent interactions.

At Dreamforce 2026, Salesforce expanded Data 360 through zero-copy connectivity, enabling access to data across enterprise data platforms without duplicating information. Headless Data 360 makes data and insights available directly to external agents, while Salesforce has also expanded its ability to bring unstructured information, including documents, emails, transcripts, and images, into the data environment.

These developments align with enterprise demand for Data 360 identified in Avasant’s Salesforce Services 2026 Market Insights, where enterprises are using Data 360 to establish the foundation for AI through data connectivity across fragmented systems, harmonization of customer and business information, governance and access controls, and activation of data for AI-driven decisions and actions.

The significance of the AI Harness is that these components are no longer operating as separate layers. Models provide reasoning, Data 360 provides context, applications provide business logic, and governance and security establish the boundaries within which agents can act. Together, they create the foundation for moving from AI that can reason to AI that can execute.

Agent ROI is Becoming a Core Enterprise Requirement

The economics of agentic AI are evolving as AI moves from human-paced interactions to machine-paced workflows. Avasant’s research, When AI Gets Cheaper but Costs More, highlighted this shift: falling model prices do not necessarily translate into lower enterprise bills because a single agentic workflow can trigger multiple model calls, retrieval operations, tool use, validation, and retries. The economic boundary is, therefore, shifting from the individual user to the workflow, making the cost of producing an accepted business outcome more relevant than the price of a single AI interaction.

Salesforce’s pricing evolution illustrates how this shift is beginning to appear in commercial models, across packaging, licensing, consumption, and unit economics. The new Editions, particularly Max, bundle the capabilities a customer needs into a single package, with included Flex Credits as the consumption unit for agentic workloads. The Agentic Enterprise License Agreement (AELA) offers fixed-price, uncapped usage over a multiyear term, providing customers with cost predictability as agentic workloads scale. At the unit level, Salesforce’s earlier model charged $2 per conversation regardless of outcome. Under the Help Agent model, the $2 charge applies only when the issue is resolved end-to-end. Together, these moves shift economic risk, at least partly, from the customer to Salesforce. This points to a broader change in how agentic AI may be commercialized. As agents take on more autonomous work, providers are beginning to move beyond charging for access or activity toward sharing greater responsibility for the outcome delivered.

What This Means for Enterprises

Context is becoming the foundation of AI execution

The value of an agent will depend on the quality of context it can access, not just the underlying model. Enterprises will need to make business data, processes, policies, and institutional knowledge accessible to AI in a structured and trusted way.

Accessibility will determine how widely AI can be embedded

AI capabilities will need to be available wherever employees and workflows operate, rather than remain confined to individual applications. Enterprises should prioritize architectures that allow agents to securely access and orchestrate capabilities across applications, data sources, and workflows.

Control will become critical as agents move from assistance to action

Greater autonomy increases the need for controls around identity, permissions, data access, business rules, monitoring, and human intervention. Enterprises will need to determine not only what agents can do, but what they should be allowed to do.

ROI will determine whether agentic AI scales

As agents perform increasingly autonomous, multistep work, enterprises will need to measure the business outcomes generated relative to the cost and risk of producing them. The focus will shift from AI usage metrics toward measurable improvements in productivity, revenue, service quality, and operational efficiency.

By Kanak Singh, Research Analyst, Premal Shah, Principal Analyst, and Chandrika Dutt, Research Director, Avasant