Dreamforce 2026 Positions Salesforce for the Post-CRM-Interface Era
Last year, Info-Tech advised that the traditional CRM interface was losing ground to conversational AI. Microsoft positioned Copilot as the new user interface, with Dynamics 365 reframed as a contextual data environment rather than the central application where work happens. ServiceNow took a similar path with Now Assist across IT, HR, supply chain, and customer service. Salesforce positioned Slack as the conversational interface for CRM. Consider also SAP Joule, Amazon Quick, and Oracle’s Digital Assistant for the broader picture. Dreamforce 2026 showed this trend entering a new stage: the race to own the interface may now be too fragmented to win. Salesforce’s strategy is now shifting away from interface ownership and toward keeping its data, semantics, permissions, actions, and workflows central, even when customer work begins in another conversational interface.
Major announcements at Dreamforce reflected this reimagined architecture, in which employees and customers may no longer interact with Salesforce primarily through a conventional CRM interface – or even via the Slack interface. Of note was “AIforce,” which exposes Salesforce data, semantics, workflows, permissions, and actions through conversational interfaces such as Claude (in addition to Slack), as well as through external models, agents, and applications. Underneath this interface layer, Salesforce’s “Enterprise AI Harness” is intended to supply reusable context, reasoning, actions, governance, security, and model choice. These capabilities are then applied to roles such as customer service, sales, employee support, commerce, and supply chain.
AIforce Separates the Salesforce Platform From the Salesforce Interface
AIforce is Salesforce’s new interface layer for bringing its data, workflows, business logic, semantics, permissions, and governance into third-party AI tool environments where workers (and their AI agents) have shifted.
AIforce launches with three principal experiences. Claudeforce brings Salesforce into Claude through a prebuilt Model Context Protocol (MCP) server and 37 sales skills. Slackforce brings Salesforce data and actions into Slack, including Slackforce Surfaces that can generate interactive dashboards, reports, and working artifacts from conversational requests. Agentforce Coworker provides a similar AI teammate inside the Salesforce Lightning interface. A Headless Toolkit exposes Salesforce capabilities through MCP, APIs, plug-ins, skills, and developer tools. Salesforce describes this architecture as bringing its platform to “any AI interface.”
AIforce is the result of Salesforce realizing its users will no longer treat the CRM interface as the center of their work. Microsoft Copilot, Claude, Gemini Enterprise, Slack, and other conversational environments increasingly compete to become the primary point of interaction.
Salesforce’s response is therefore to separate the value of its platform from its ownership of the interface. If a seller prefers to work in Claude, Salesforce still wants to provide the customer context, permissions, business rules, and executable actions. If an organization standardizes on Gemini Enterprise or an AWS environment, Salesforce wants its capabilities available there without surrendering its role as the governed customer platform.
The expanded AWS and Google Cloud partnerships reinforce this strategy. Salesforce capabilities can now be surfaced within Amazon Quick and Gemini Enterprise, while models from Anthropic, Nvidia, Google, and eventually OpenAI can operate within Salesforce workflows. AWS agents can also work inside Slack, while Salesforce and Google are connecting their respective data and agent environments through MCP and zero-copy access. AWS integrations are at different stages of availability, while Salesforce’s Google Cloud partnership includes a mixture of generally available, beta, preview, and planned capabilities.
Of course, Salesforce’s premium per-user pricing has historically reflected ownership of both the customer system of record and the interface through which employees perform their work. Once Claude, Copilot, or Gemini owns that interface, customers may reduce Salesforce seats or question why a largely invisible data and execution layer commands the same price. Salesforce is attempting to offset that pressure by shifting monetization toward Agentforce actions, outcomes, Data 360 consumption, governance, and Headless 360 capacity; its new editions bundle these capabilities with Slack, Tableau, security, and included Flex Credits to defend overall contract value. The upside is that Salesforce can earn revenue whenever an external agent invokes its context or actions, potentially reaching far beyond its licensed users. The downside is greater scrutiny of the cost of every query, action, and transaction. Salesforce can preserve or expand revenue if its context and workflows remain indispensable; if customers begin to view it as a replaceable back-end data service, its pricing power will diminish.
The Enterprise AI Harness Is About Governance
The Enterprise AI Harness was a significant Dreamforce announcement for large organizations. It organizes Salesforce’s AI platform around six capabilities:
- Trusted Context combines customer data, metadata, semantics, knowledge, memory, and real-time signals.
- Trusted Agency provides reasoning, planning, state, memory, and orchestration.
- Trusted Action connects agents with applications, APIs, workflows, and business processes.
- Trusted Governance manages data quality, lineage, policies, and guardrails.
- Trusted Security applies identity, permissions, privacy, and runtime controls.
- Trusted Models routes work according to model accuracy, performance, cost, and business requirements.
Salesforce is also developing an AI Control Plane through which organizations will be able to register agents, assign identity and policy, manage lifecycles, evaluate performance, observe behavior and outcomes, and control costs across both Salesforce and third-party AI. The architecture draws upon Data 360, Informatica, MuleSoft, Agent Fabric, Tableau, Agentforce, Salesforce Guardian, and the broader Salesforce Platform. Salesforce says the unified experience and new capabilities will begin rolling out in early fiscal 2028.
The Enterprise Harness addresses an important enterprise problem. Individual AI agents are relatively easy to prototype; governing hundreds of agents that use different models, access overlapping data, and take actions across multiple systems is considerably harder. Context, identity, evaluation, observability, cost management, and lifecycle controls need to be reusable rather than rebuilt for each deployment.
However, buyers should think of the Enterprise Harness more as Salesforce architectural direction, rather than a fully available control plane today. The unified experience, precise packaging, cross-platform governance depth, and upgrade paths remain under development.
Organizations should consequently ask Salesforce to demonstrate which controls work across third-party agents. Important questions include whether external agents can be discovered automatically, whether policies are enforced at runtime or merely reported, whether decision and tool-use traces can be exported, and how Salesforce manages agents that act across non-Salesforce systems.
Job-Ready Agents
Salesforce introduced a portfolio of job-ready agents for customer service, employee support, shopping, outbound sales, supply chain, and inbound pipeline generation. These agents are pre-packaged with relevant skills, actions, and data models, giving customers a faster starting point than building each agent from scratch. Most are generally available, although the Hunter outbound sales agent remains in pilot until its planned November 2026 release.
The product names are less important than the change in how Salesforce is encouraging adoption of this technology. Salesforce initially emphasized Agentforce as a platform on which customers could construct agents. Salesforce is now packaging agents around out-of-the-box, recognizable jobs and business outcomes. This should reduce design effort and accelerate deployment, particularly for organizations with conventional Salesforce processes.
However, users should remain mindful about the limits of out-of-the-box agents. Salesforce can prebuild the task structure, standard actions, and data model, but the agent’s effectiveness will still depend heavily on the customer’s data and knowledge maturity. A service agent needs current, non-duplicative knowledge articles and clearly documented policies; a sales agent needs reliable account, contact, activity, and opportunity data; a commerce agent needs accurate catalog, inventory, pricing, entitlement, and returns information. Salesforce’s own implementation guidance requires customers to connect and maintain data libraries, configure retrieval, define permissions, and test responses against their knowledge sources. Pre-packaging therefore shortens agent design, but it does not remove the work of curating knowledge, standardizing processes, resolving conflicting sources, and assigning content ownership. For less mature organizations, faster deployment may simply automate inconsistent answers and poorly defined workflows.
Another operational risk will be these agents’ new “long-horizon runtime.” The benefit of long-horizon runtime is persistent memory, durable execution, and dynamic steering so an agent can pursue an objective over days or weeks, adjust its plan as conditions change, and resume work across sessions. However, an agent pursuing a goal for several weeks may encounter changing data, revoked permissions, conflicting instructions, unavailable tools, and decisions whose consequences only become visible later. Customers should require explicit goal boundaries, time limits, approval checkpoints, escalation rules, revocation controls, compensating actions, and records of how plans changed. Multi-agent orchestration makes these controls more important because accountability can become unclear when work passes between specialized agents. Hunter will be the first agent to use this runtime, with additional Salesforce and customer-built agents expected to follow.
Fin and Contentful Fill Product Gaps
Salesforce’s acquisitions of Fin and Contentful fill two areas that Agentforce previously lacked: a proven, rapidly deployable customer agent and a composable content layer.
Fin, formerly Intercom, brings to Salesforce a customer-agent platform, proprietary customer-experience models, a specialized technical team, and more than 30,000 customers. Salesforce reports an average autonomous resolution rate of 76% across Fin deployments. Fin will continue serving customers through Salesforce AI Labs while complementing the more customizable Agentforce platform. Salesforce completed the acquisition on September 10, 2026.
The acquisition gives Salesforce a faster route into organizations that want an out-of-the-box service agent without first undertaking a broader Agentforce transformation. However, buyers will need clarity on the eventual relationship among Fin, Casey, Agentforce Help Agent, Agentforce Voice, and existing Service Cloud automation. Separate products may support different adoption paths today, but overlapping administration, analytics, knowledge, pricing, and roadmap ownership would create complexity if Salesforce does not rationalize the portfolio.
Contentful supplies the content infrastructure required for agent-generated customer experiences. Its structured, API-first content can be combined with Salesforce customer context and Agentforce decisioning to assemble experiences dynamically across websites, mobile applications, commerce, marketing, and service channels. Salesforce completed the acquisition on September 1, 2026, after initially describing Contentful as the native content layer for Headless 360.
Notably, Contentful is one of Salesforce’s more consequential acquisitions. Salesforce already offered content management capabilities, but its modern composable-content strategy relied substantially on partners. Contentful entered the Salesforce ecosystem through a Commerce Cloud partnership, prebuilt Composable Storefront integrations, and an AppExchange connector. Ownership of Contentful now moves Salesforce directly into the content curation, management, and multichannel-delivery layer, strengthening a proprietary digital experience platform play spanning Data 360, Agentforce, Marketing Cloud, Commerce Cloud, and Experience Cloud. Contentful also supplied Salesforce as a significant market asset: It supports more than 4,800 brands and was named a 2026 SoftwareReviews Data Quadrant Champion, with an 8.6 Customer Experience score, a 97% plan-to-renew score, and a +92 Net Emotional Footprint.
Koa Gives Salesforce Its Own CRM Reasoning Layer
Salesforce and Nvidia introduced Koa, Salesforce’s first reasoning model designed specifically for CRM and enterprise workflows. Koa is based on Nvidia’s 120-billion-parameter Nemotron 3 Super model and was post-trained using synthetic enterprise workflows developed from Salesforce’s experience across more than 14 industries.
Salesforce argues that general purpose frontier models reason from first principles each time they encounter tasks such as lead qualification, case routing, or refund approval. Koa is instead trained to perform multistep, tool-using enterprise work consistently, including recognizing when the required tool is unavailable and stopping rather than falsely claiming that an action occurred.
Koa runs within the Salesforce trust boundary, and Salesforce says customer data and reasoning traces are not used to train the model. It is currently being tested with a limited group of customers, with broader availability expected in winter 2026. Salesforce reports that early CRM-specific benchmarks are encouraging, but it has not yet published sufficient methodology for buyers to assess those comparisons independently. Koa is already operating within some internal Salesforce workflows and is moving into customer pilots.
Koa does not invalidate Salesforce’s model choice strategy. Frontier models can remain appropriate for broad reasoning and creative work, while Koa and smaller specialized models handle narrower enterprise functions. Customers should evaluate Koa through task-level testing rather than accept a general claim of CRM superiority. Tests should compare completion quality, consistency, latency, cost, tool selection accuracy, refusal behavior, and human-escalation rates against the other models available through Agentforce.
Our Take
Dreamforce 2026 should not be read as Salesforce pulling away from its competitors. It showed the company adapting to an agentic platform race already underway, in which major enterprise vendors are competing to provide the context, governance, and execution layer beneath a growing population of AI agents.
Each major competitor enters this contest from a different position. Microsoft has the strongest productivity distribution and can extend identity, security, and compliance controls to registered agents through Agent 365, Entra, Defender, and Purview. ServiceNow starts from cross-system operational workflows and is expanding AI Control Tower across first- and third-party agents. SAP and Oracle approach agents from the transactional core, grounding them in financial, HR, procurement, manufacturing, and supply-chain processes. AWS and Google provide the models, infrastructure, and runtime services on which many of these agents operate. Microsoft, ServiceNow, SAP, and Oracle are consequently building overlapping claims to enterprise agent governance.
Salesforce’s defensible position is customer context. It holds customer data, CRM metadata, permissions, workflows, and actions across sales, service, marketing, and commerce. AIforce is intended to make these assets available even when work begins in Claude, Slack, Gemini, Amazon Quick, or another environment. The Enterprise AI Harness would then provide common context, security, governance, and model selection underneath those experiences.
However, buyers should distinguish this architectural direction from a fully delivered control plane. Salesforce says many underlying technologies are available, but the unified Enterprise AI Harness experience and additional capabilities will begin rolling out in early fiscal 2028. Salesforce in Claude is currently in beta, while the availability of other AIforce components varies. Salesforce’s Enterprise AI Harness announcement and AIforce announcement should therefore be read partly as roadmap statements.
Three wider market shifts can be inferred from this context:
- Conversational environments increasingly provide an alternative starting point for retrieving information and initiating work. Traditional application interfaces will remain important, but vendors can no longer require users to enter their application before realizing value from its data and functions.
- As enterprises gain access to multiple capable models, differentiation moves toward proprietary data, semantics, permissions, workflow knowledge, agent identity, evaluation, and policy enforcement.
- MCP, Agent2Agent, APIs, and zero-copy access make platforms easier to connect, but they do not guarantee semantic, policy, or configuration portability. Dependency can shift into agent registries, business definitions, workflow logic, permission structures, evaluation records, and operational memory.
Salesforce’s strategic adjustment is therefore understandable. If it cannot ensure that customer work begins inside Salesforce, it must keep Salesforce important to completing that work. The challenge is remaining the system that agents consult before customer work can be understood, authorized, and completed.