Salesforce Introduces AIforce: Breaking Down Fixed Interfaces and Operational Barriers in Traditional CRM Software to Let Enterprise Data and Business Logic Flow Freely Across Diverse AIs

Using Salesforce historically relied on a fixed set of entry points: employees opened the CRM web interface or app, navigated through pages, fields, forms, and reports pre-configured by administrators to locate information, and then manually wrote results back to the system. Although data, business logic, and approval permissions resided in Salesforce, they were locked within Salesforce's own interface; when employees transitioned to work environments like Claude or Slack, these capabilities could hardly follow in a seamless manner.

AIforce, introduced by Salesforce at Dreamforce, is designed to dismantle this "interface wall." It is neither a new foundation model nor another agent performing standalone tasks; rather, it is a dynamic interaction and capability delivery layer situated between Salesforce data and diverse AI interfaces. It delivers customer data and historical context from Data 360, business semantics and workflows from Customer 360, and task-executing agents from Agentforce into Claude, Slack, Lightning, and future supported AI work environments, strictly in accordance with original permissions and business rules.

Therefore, "enabling enterprise data and business logic to flow freely across diverse AIs" referenced in the title does not mean replicating databases to different models, much less letting external AIs bypass Salesforce to manipulate enterprise data directly. Instead, it refers to a governed, bidirectional closed loop: the AI interface passes the employee's natural-language intent to AIforce; AIforce determines which data, workflows, and agents should be invoked; Salesforce executes the read or action according to original permissions; and the results return to the current conversation, dynamically generating on-demand interfaces for viewing, filtering, modifying, and confirming tasks. The underlying records remain in Salesforce; what changes is the access point through which employees invoke these capabilities.

This is also what makes AIforce fundamentally different from simply "adding a chat box to CRM." It seeks to transform Salesforce from an application that must be opened into an enterprise data and execution engine powering various AIs: wherever an employee asks a question, Salesforce's context, rules, and action capabilities can surface right there.

What AIforce Truly Changes Is the Interaction Boundary of Enterprise Software

For decades, the format of most enterprise software has been static. Administrators configured page layouts, mandatory fields, and fixed views in the backend, while sales reps or service agents filled in information entry by entry, much like working on an assembly line. When frontline employees needed a complete view of a customer, they typically had to separately open contact pages, historical opportunity lists, service ticket records, and various analytical dashboards.

The problem with fixed interfaces is not merely "too many clicks." It simultaneously limits who can use the system and how much scattered information can be mobilized for a given task:

1. High barrier to entry: Ordinary employees without specialized training find it difficult to adeptly extract multi-dimensional, cross-cutting insights from backend systems, leaving large amounts of business data effectively dormant.
2. Constrained information throughput: In graphical interfaces, humans can only browse a few cards and tables at a time, unable to instantly digest and correlate the logic behind hundreds of interaction history entries.

AIforce reshapes this interaction boundary: the user first describes the task they wish to accomplish, and the system then determines which context should be retrieved, which business actions should be invoked, and what interface format is best suited for the moment. Employees no longer need to understand CRM menu hierarchies before finding the correct data and operational entry points. For instance, sales reps can request client meeting prep within Claude, and the system automatically aggregates opportunities, historical communications, and team discussions; alternatively, they can generate shared dynamic surfaces in Slack to filter data, comment, and write confirmed changes back to Salesforce directly. Interfaces are generated dynamically on demand by the task, while underlying data, permissions, and workflows continue to be governed by Salesforce.

---

Where Does AIforce Stand in Enterprise Architecture?

To enable external AI interfaces to safely interpret and operate enterprise systems, Salesforce established four collaborative tiers across its overall architecture, with AIforce positioned at the topmost interaction output tier:

1. Unified Data Layer (Data 360): Responsible for cross-system integration of internal and external customer data, metadata, and historical memory, providing AI agents with complete contextual perspective.
2. Business Semantics and Logic Layer (Customer 360): Houses core business logic, approval workflows, action execution permissions, and security compliance rules across sales, service, marketing, and commerce.
3. Digital Workforce Layer (Agentforce): Provides out-of-the-box as well as custom-built autonomous agents responsible for executing role-specific business processes.
4. Dynamic Interaction Layer (AIforce): The newly added top-level architecture. Its core function is to project all the data, semantic, and execution capabilities residing in the underlying three layers into diverse external AI interfaces without alteration.

Where AIforce Sits in the Architecture

Enterprise semantics and permissions remain at the foundation; capabilities reach every work interface at the top

1. 4AIforceDynamic interface and external connection layer
2. 3AgentforceSpecialized agents that execute job-specific tasks
3. 2Customer 360Business semantics, processes, permissions, and actions
4. 1Data 360Unified data, metadata, and historical context

Read with the current user's permissions → Actions return to Salesforce for execution

• ClaudeClaudeforce
• SlackSlackforce
• LightningAgentforce Coworker

The point is not to copy data into three new places. These interfaces read and act through Salesforce's existing semantics, permissions, and workflows.

Under this architecture, external AIs do not simply read data via public web pages; instead, they strictly adhere to the pre-existing security and permission inheritance model. Every request is bound to the initiating user's original access permissions, ensuring AI agents can only view fields the employee is authorized to see; all business modifications initiated by AI must likewise write back along business rules defined by Salesforce. Furthermore, Salesforce commits to a Zero Data Retention policy: enterprise data is used solely to answer the active request and is not retained by model providers. This represents a vendor-stated security commitment at launch, rather than an independently audited finding.

---

Model and Runtime Foundation: NVIDIA Full-Stack Infrastructure and Salesforce Koa Reasoning Model

When evaluating AIforce's four-layer architecture, focusing solely on the top-level interaction can lead to a misconception: assuming that simply connecting general foundation models to enterprise systems will automatically execute complex business processes, or conflating partner technology stacks with components of AIforce itself. In reality, AIforce addresses the interaction layer, securely projecting underlying accumulated data, permissions, and action capabilities onto external AI interfaces. Beneath this layer, the models enterprise agents rely on for business reasoning, along with the underlying compute power and runtime environments, constitute a parallel technical foundation. The Koa specialized reasoning model introduced by Salesforce and the infrastructure provided by NVIDIA complement this very foundation.

On stage at Dreamforce, Salesforce co-founder Marc Benioff and NVIDIA CEO Jensen Huang held a public conversation. Rather than a specific product demo of AIforce, this exchange showcased the strategic and computational underpinnings behind enterprise-grade agents:

• Full-stack AI infrastructure: Huang described NVIDIA as a full-stack AI infrastructure company, proposing that "AI factories" transform electricity and proprietary enterprise data into intelligence.
• Open model and runtime portfolio: He enumerated NVIDIA's open model portfolio, including Nemotron, Cosmos, Alpamayo, BioNeMo / Proteina-Complexa, and the humanoid robot model GR00T; he also highlighted an open agentic harness and a secure sandbox and runtime environment named OpenShell. Crucially, these represent NVIDIA's own technological foundation and must not be conflated with components of AIforce.
• Coexistence of closed-source and custom models: Huang stated clearly that while general closed-source frontier models will continue to exist across the industry, enterprises will simultaneously build proprietary custom models—the precise direction of the NVIDIA and Salesforce collaboration.
• Safety perspective and on-stage remarks: Regarding AI safety, Huang referred to it on stage as an "engineering problem," emphasizing the creation of rigorous testing environments, comprehensive product evaluations, and withholding release when uncertain; he also expressed a personal stance that "no new laws or regulations are needed." To be clear, this represents Huang's personal opinion expressed on stage and cannot be framed as an established product guarantee or broad industry consensus.

It is precisely along this path of custom models and vertical reasoning that Salesforce introduced its first specialized CRM reasoning model for Agentforce—Koa:

• Model positioning and complementary relationship: Post-trained on NVIDIA Nemotron 3 Super, Koa is an optional specialized reasoning model for Agentforce. While AIforce delivers data and actions to external front-ends, Koa focuses on handling complex CRM business logic reasoning at the foundational layer; the two complement each other but are not the same product.
• Training data and trust boundaries: Salesforce stated explicitly that no customer data was used to train Koa, with training corpora constructed from proprietary synthetic data simulating enterprise knowledge across nearly thirty years of CRM deployments. Salesforce controls the model weights, and both post-training and inference take place within Salesforce's own trust boundary.
• Training technology stack: Post-training of Koa utilized Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO), leveraging toolsets such as NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel.
• Performance claims and availability: Salesforce claimed Koa achieved "3x fewer errors" on benchmarks, noting immediately alongside that this figure originates from Salesforce's proprietary CRM benchmark and linked paper, representing vendor benchmark results rather than independent third-party evaluations. Koa is currently deployed internally at Salesforce and in pilot with select customers, with general availability in the US region expected in winter 2026.

---

Claudeforce: Official MCP Server Packaging CRM into Claude

Connecting external foundation models with internal enterprise business systems historically required weeks of manual configuration across APIs and gateways. Viewing the Salesforce and Anthropic partnership merely as "calling APIs in a chat box" overlooks how sales reps execute specific workflows in real-world scenarios, how multiple connectors collaborate, and where enterprise permissions and human intervention boundaries lie. Claudeforce, co-developed by Salesforce and Anthropic, debuts in the form of the Salesforce in Claude Beta add-on co-built with Salesforce. It directly brings customer accounts, opportunities, and sales pipelines into the Claude interface, complete with 37 out-of-the-box sales skills. Upon initial use, the add-on's setup skill automatically identifies available tools and connectors, generating a Claude Artifact tailored to the sales rep's specific role and assigned territory to rapidly establish a dedicated sales workspace.

The specific sales workflows disclosed cover core stages of daily operations:

• Daily personalized briefing: Automatically organizes daily priorities and action items based on real-time customer activity and follow-up cadences.
• Pre-call preparation: Automatically pulls open opportunities, customer team discussion records, unanswered threads, and unresolved questions from prior calls, enabling sales reps to grasp full context rapidly before customer interactions.
• Deal evaluation and mutual close plans: Assesses deal progression against sales team methodologies and automatically generates mutual close plans.
• Post-call follow-up and write-back: Drafts follow-up emails, Slack channel summaries, and extracts record updates to be written back to Salesforce following customer meetings.
• Interactive pipeline and forecast surfaces: Generates sales pipeline and forecasting surfaces that support drill-downs, interactive filtering, and direct write-backs to Salesforce.

In terms of connection architecture, the add-on includes two core connectors: the Salesforce connector and the Slack connector. The Salesforce connector handles reading and updating records, logging call details, and creating follow-up tasks; the Slack connector is used to read and write across deal-specific channels and account team discussion threads. Although the initial setup interface in the official demonstration displays connection statuses for Salesforce, Slack, and Gmail simultaneously, Anthropic's official blog explicitly clarifies that the add-on itself contains only the two aforementioned connectors; Gmail must not be mistakenly characterized as an included connector.

Regarding data governance and security controls, Salesforce remains the system of record. Sales reps log in using their own Salesforce credentials, and Claude can access only data permitted by the user's native permissions. To mitigate risks from automated actions, the system enforces item-by-item approval by default for all data modifications intended for write-back to Salesforce, executing only after manual confirmation by the sales rep; notably, however, this functions as an interaction-layer human confirmation mechanism and does not guarantee the elimination of foundation model hallucinations or user errors.

Addressing enterprise concerns regarding data privacy, Anthropic officially states that Team and Enterprise plans do not use customer data to train models by default. However, it must be emphasized that this is not equivalent to "zero data retention under all circumstances," nor does it replace the Zero Data Retention product guarantee committed by Salesforce for AIforce. When assessing deployment, enterprises need to verify terms against their specific purchased plans, commercial agreements, and administrative settings. Regarding deployment progress, Anthropic disclosed that GitLab, Siemens, and Legora have deployed the add-on within their organizations, alongside 7,000 Salesforce sales reps using it in daily work; these figures reflect vendor-reported adoption.

The Beta version is currently available across all paid Claude plans. Sales reps can install the Salesforce MCP from Claude's marketplace, though organizational administrators must first request access on AgentExchange and complete the Salesforce connection setup for the organization. Additionally, capabilities mentioned in Salesforce's initial announcement—such as integrating Tableau data analytics and expanding into service, marketing, commerce, and vertical industries—remain on future roadmaps and are not included in the current sales Beta release.

Alongside Salesforce in Claude designed for office and sales environments, the two companies concurrently launched a dedicated developer plugin for the command-line tool Claude Code. This is an independent developer product that should not be confused with the sales add-on described above: it includes over 40 built-in development skills, connects directly to skill repositories on GitHub, and supports dynamically loading dedicated sub-plugins to assist developers in executing system integration and code development tasks directly within terminal environments.

---

Slackforce: Turning Chat Windows into Multi-User Real-Time Workbenches

Slack itself is the primary operational hub for internal collaboration across many enterprises. By introducing Slackforce, Salesforce aims to thoroughly integrate operations that previously required dedicated CRM windows into everyday conversations:

Slackforce Surfaces dynamic surfaces: Allows teams to pull real-time data directly from Salesforce, Slack, and other integrated tools within discussion channels, assembling them on the spot into interactive multi-user interfaces. Team members can collaboratively filter metrics, post comments, and trigger business modifications within cards without requesting specialized reports from IT. For example, service company Engine, which handles over 800,000 customer inquiries annually, used this capability to let frontline staff generate operational interfaces using natural language without toggling between disparate systems.

Slackbot cross-system personal assistant: Breaks down data silos between chat histories and CRM records. The built-in Slackbot can read conversation context in Slack channels and Salesforce service tickets simultaneously. For instance, it can identify accounts with declining engagement, correlate recent support cases and internal discussion threads to determine root causes, reassign owners, create follow-up tasks, and draft retention emails.

Slack CRM streamlined entry: Simplifies tedious logging into brief natural-language interactions. After concluding a customer call, employees only need to send a single prompt in the chat window to create account profiles, log call highlights, or advance opportunity stages in Salesforce, all without ever leaving Slack.

Slack Code collaborative multi-user coding: Expands AI coding from a standalone personal assistant into a multi-user team paradigm. Engineering teams can collaborate with coding agents on development projects in dedicated Slack channels, where all operations and debugging context are shared in-channel by default. This ensures project context transfers seamlessly among team members and agents, ultimately producing runnable code, prototypes, and documentation directly inside Slack.

---

Agentforce Coworker: Out-of-the-Box Co-pilot in the Native Backend

For users accustomed to remaining within Salesforce's native web interface (Lightning), the corresponding solution in this announcement is Agentforce Coworker.

Unlike Claudeforce, which functions as an external connector, Coworker operates natively inside the Salesforce platform. It requires neither data migration nor a redesign of permission models, allowing system administrators to activate it with one click in the backend. Its core mechanism lies in directly invoking specialized agents that enterprises have previously constructed and deployed on the Agentforce platform, functioning as a unified orchestration front-end for existing enterprise agent assets.

According to data published by Salesforce, Coworker reached 100,000 active users within the first 35 days of launch. Fulton Bank stated that it launched more than 20 production use cases covering approximately 3,000 employees within weeks, observing that Coworker helped staff locate information and prepare for business tasks more quickly. These figures derive from vendor announcements and customer quotes, rather than independent evaluations of business returns.

---

Architectural Decoupling: Headless Toolkit and External Open Ecosystem

The foundational pillar enabling CRM capabilities to decouple from native UIs is the Headless Toolkit introduced by Salesforce. This open architecture encapsulates underlying elements across the entire Salesforce platform—including object data, business flows, field semantics, and action interfaces—into standardized MCP interfaces, APIs, add-ons, and skills for external developers to invoke freely.

Leveraging this toolkit, Salesforce established the AgentExchange partner ecosystem and distribution marketplace, bringing together diverse partners:

• Agent interaction interfaces: Anthropic, Amazon Web Services (AWS), Google, and Microsoft;
• AI build and design platforms: Lovable and Vercel;
• Vertical business tools and agents: DocuSign, Gamma, Jasper, Rippling, and others.

Alongside the rollout of this open architecture, Salesforce specifically highlighted its 20-year-old Trailblazer developer and administrator community. With the platform open, the 12 million applications accumulated by community developers and experience from 9.6 billion daily API calls can be further translated into skills and plugins adapted for various AI front-ends.

---

Boundaries and Observations: Shifting from 'Walled Gardens' to 'Protocol Delivery'

A core commercial moat of traditional enterprise SaaS software has historically relied on high front-end interface stickiness to retain users within its proprietary application ecosystem. As this announcement illustrates, the rise of foundation models and agents is shifting the competitive epicenter of enterprise software: front-end interfaces are rapidly being replaced by dialogue boxes, collaboration tools, or autonomous agents, while the true moat shifts to irreplaceable underlying business data, approval permissions, and compliance logic.

By introducing AIforce, Salesforce actively steps behind the scenes, transforming itself into an "enterprise data and execution engine" governed by strict permissions, willingly yielding the interaction front-end to Claude and Slack.

However, enterprise procurement and implementation teams must remain mindful of the operational boundaries noted in official materials:

• Billing and terms of service: Pricing models and product packaging for related features may change in the future, with actual availability governed by regional rollouts and specific customer contracts;
• Data governance prerequisites: Although the system claims zero migration and automatic permission inheritance, the quality of dynamic interfaces and external agent outputs remains entirely contingent on the quality of an enterprise's existing metadata standards, field cleanliness, and workflow completeness within Salesforce. If underlying data suffers from gaps or inconsistencies, the intelligent experience projected onto external interfaces will be similarly constrained.