Salesforce Turns Its CRM Into AI’s Backend Layer

Salesforce used its Dreamforce conference in San Francisco on September 15, 2026, to make a consequential claim about the future of enterprise software: the CRM no longer needs to be the place where users work. Instead, Salesforce wants its data, workflows, permissions and business actions to sit behind whatever artificial-intelligence interface employees or customers prefer.

The company calls that layer AIforce. It is designed to expose Salesforce’s core capabilities to external models, agents and interfaces, allowing those systems to retrieve business context and take governed actions without forcing users to open Salesforce itself. Independent industry reporting described AIforce as a connective layer for Salesforce’s Data 360, Customer 360 and Agentforce foundations, available across “any AI interface.” (salesforce.com)

That is more than another product launch. It is a strategic attempt to defend the economic position of enterprise software as AI assistants become the place where work begins. If users increasingly ask Claude, Gemini, Slack, voice assistants or internal agents to summarize accounts, update opportunities, resolve cases or launch workflows, the application that owns the underlying data may become less visible. Salesforce is betting that visibility is optional if control of the system of record is not.

What changed

Traditional enterprise applications are built around a user interface. A salesperson logs into Salesforce, navigates through records, follows prescribed screens and triggers actions from within the platform. Salesforce’s new direction separates the interface from the application’s underlying capabilities.

The company says AIforce can bring Salesforce data, workflows, business logic, actions and governance into other agentic interfaces. That includes the permissions needed to determine what an agent may see and do. In theory, a worker could ask an assistant in another application to prepare a customer briefing, identify a stalled deal or initiate a service response while Salesforce handles the authoritative data and execution layer in the background. (investor.salesforce.com)

The announcement arrives alongside a wider Salesforce push to make Agentforce operate across professional workflows rather than as a chatbot attached to a CRM screen. The company is also expanding partnerships with Google Cloud, AWS, Anthropic, NVIDIA and Siemens. AWS said its new collaboration would bring Salesforce context and actions into Amazon’s AI tools, while enabling model choice that includes Anthropic and NVIDIA systems, with OpenAI models planned for later access through Amazon Bedrock. (press.aboutamazon.com)

That partner strategy matters because Salesforce is not trying to win the model race. It is trying to become indispensable to the companies deploying models from several vendors. The more fragmented the model market becomes, the more valuable a neutral, governed business context layer could be—provided Salesforce can make that layer easy to access without weakening security or data quality.

Why it matters

AIforce is a direct response to the emerging “headless enterprise” model: software capabilities are consumed through agents and APIs rather than through a single application interface. The shift could alter how enterprise software is bought, measured and priced.

For customers, the promise is flexibility. Companies would not need to force every employee into one interface or rebuild integrations whenever they adopt a new model. A service agent could work in a contact-center system, a sales representative could use an assistant in Slack, and an executive could ask a voice interface for a pipeline analysis—all while drawing from the same governed records.

For Salesforce, the opportunity is defensive as much as offensive. A powerful external assistant could otherwise become the primary relationship with the customer, reducing the importance of Salesforce’s screens and potentially pressuring seat-based pricing. By exposing its capabilities everywhere, Salesforce hopes to preserve its role even if its interface disappears from daily view.

The danger is that the company may be giving away the layer customers actually interact with while retaining a more difficult-to-explain backend subscription. If users no longer associate value with Salesforce’s application, the company will need to prove that governance, data quality, workflow reliability and action permissions justify continued spending. That could push the market toward usage-based pricing tied to automated actions, transactions or “agentic work units,” rather than conventional seats.

The strategy also places Salesforce closer to infrastructure than application software. Its competitive advantage would depend on maintaining clean customer data, dependable APIs, fine-grained authorization and audit trails. A model can generate a plausible answer with imperfect context; an enterprise agent that changes pricing, contacts a customer or updates a forecast cannot be allowed the same margin for error.

The unresolved governance problem

Salesforce presents governance as a core feature of AIforce, not an afterthought. But centralizing business actions behind many outside interfaces creates a difficult accountability question: when an agent makes a wrong or unauthorized decision, which layer is responsible—the model provider, the interface vendor, the customer or Salesforce?

The answer will depend on implementation details that remain unclear. Salesforce has not yet demonstrated how consistently permissions will carry across different models and agent frameworks, how customers will audit multi-step decisions, or how conflicts between an external agent’s instructions and Salesforce’s business rules will be resolved.

There is also a practical distinction between exposing information and exposing action. Reading an account summary through an outside assistant is relatively low risk. Issuing refunds, changing contract terms, sending regulated communications or modifying customer records is different. Enterprises will likely demand approval gates, policy enforcement and reversible actions before allowing agents to operate broadly.

Salesforce’s partnerships may accelerate adoption, but they could also complicate governance. Each model provider has different safety systems, context windows, tool-calling behavior and data-handling policies. A single Salesforce permission model may not be enough to make those environments operationally equivalent.

What to watch next

The next test is not whether AIforce can produce compelling conference demonstrations. It is whether customers deploy it for consequential work without recreating the integration complexity Salesforce says it can eliminate.

Three indicators will matter. First, Salesforce must show whether external agents can use its capabilities with genuinely granular permissions, not merely broad API credentials. Second, customers will need evidence that actions are auditable across model providers and interfaces. Third, the company must prove that its economics improve when the CRM becomes a backend rather than a destination.

If it succeeds, Salesforce could become less visible but more deeply embedded in enterprise AI. The company would no longer compete primarily to own the screen; it would compete to own the trusted business context and the permissioned actions behind every screen. That would make AIforce one of the clearest early signals that enterprise software is being reorganized around agents—not by replacing systems of record, but by turning them into programmable infrastructure.