Wonderful Raised $550M in 20 Months to Build the Enterprise AI Operating System. Here's What That Actually Means — and Why Salesforce Backed It.

On September 22, 2026, Salesforce opened Dreamforce with a thesis that inverted two decades of enterprise software product strategy: AI agents don't need a user interface. They need the capabilities beneath it.

The product built from that thesis is AIforce — Salesforce's new interface layer that makes the company's full enterprise knowledge stack (data, workflows, business logic, permissions, security, governance) accessible from Claude, Slack, or any AI interface, without requiring anyone to open the Salesforce application itself. The headline framing at Dreamforce was "AI replaces the UI," and Salesforce's CEO Marc Benioff positioned AIforce as the company's most significant architectural move since the original launch of the cloud CRM.

The skeptical read — and a credible one, based on what Salesforce actually shipped versus what it announced — is that "AI replaces the UI" is a 2026 thesis with a 2027 product schedule. Two of AIforce's three surfaces are in beta, Koa is in pilot, and several announced capabilities are on the October and November roadmap. But the strategic direction is clear, and for enterprise technology leaders, the implications of where Salesforce is heading are worth examining before the full stack is generally available.

What Dreamforce 2026 Actually Shipped

The headline announcements from Dreamforce 2026 sort into four categories: generally available products, products in beta, products in pilot, and products on the roadmap. Enterprise buyers should treat them differently.

Generally available as of Dreamforce 2026: - Agentforce Coworker — the AI teammate embedded in the Lightning interface, allowing employees to interact with Salesforce data through conversation without navigating the standard UI - Expanded integrations with Google Cloud, AWS, and Siemens for enterprise AI workflows - AgentExchange, the marketplace for pre-built Agentforce agents and integrations

In beta or pilot: - Claudeforce (beta) — the AIforce surface connecting Claude to Salesforce enterprise data and workflows - Slackforce (beta) — the AIforce surface making Salesforce callable from Slack - Koa (pilot, select customers) — Salesforce's first purpose-built CRM reasoning model built on NVIDIA Nemotron 3 Super, expected to reach general availability in US regions in winter 2026

Pricing with unresolved variables: - Headless Platform Interactions (HPIs) — the metering unit for agent calls to Salesforce via MCP or API — are tracked through Flex Credits with a multiplier listed as TBA; Salesforce has committed to 30 days' notice before billing begins

The "AI replaces the UI" framing was a thesis at Dreamforce, not a complete product delivery. That is important context for enterprise buyers evaluating purchase decisions based on Dreamforce announcements. The direction is clear; the fully deployed, production-ready, priced version of the AIforce stack is still being built.

What AIforce Is — and Isn't

AIforce is not a product customers buy as a standalone. It is an architectural layer that makes Salesforce's existing enterprise platform callable from non-Salesforce AI interfaces.

The business problem AIforce is designed to solve is context fragmentation: the friction cost of employees switching between AI interfaces and enterprise applications to complete work. An enterprise that has deployed Claude for internal knowledge work, Slack for team communication, and Salesforce for CRM is operating three separate contexts. Every time a sales rep needs to look up a customer record, update an opportunity, or trigger a workflow, they leave the AI interface they were using and navigate into Salesforce — breaking the agent-assisted workflow and adding latency to every task.

AIforce eliminates that friction by making Salesforce's data, workflows, and actions directly callable from the AI interface. The employee never opens Salesforce. The agent queries the enterprise data, executes the workflow, and returns results inside the interface the employee is already using.

The architectural mechanism is Salesforce's implementation of the Model Context Protocol (MCP), the emerging standard for connecting AI agents to external data sources and tools. Each agent that calls Salesforce over MCP or direct API is registered in Salesforce's Digital Wallet system and consumes Flex Credits as Headless Platform Interactions. The rate card for HPI consumption is not yet published, which creates the central budget uncertainty for enterprise procurement teams.

What AIforce is not: - It is not a replacement for Salesforce's core CRM platform — it extends that platform's reach into AI interfaces rather than replacing its data model - It is not an agent builder — Agentforce remains the layer where agents are built and deployed inside the Salesforce ecosystem - It is not currently available in full across all use cases — the Claudeforce and Slackforce surfaces are in beta, and Koa is in pilot

Koa: The CRM Reasoning Model

Koa is Salesforce's first purpose-built reasoning model for enterprise CRM tasks. It is built on the NVIDIA Nemotron 3 Super architecture and fine-tuned specifically for the Salesforce data schema: customer records, opportunity pipelines, case management, service workflows, and multi-step CRM reasoning.

The significance of Koa is what it represents architecturally, not just what it can do in individual tasks. Enterprise software vendors have historically relied on general-purpose frontier models adapted for their domains through prompt engineering, retrieval, and fine-tuning. Koa is a different bet: a model built from the ground up for the specific task distribution of enterprise CRM, trained on the data schema, business logic, and usage patterns that Salesforce customers generate at scale.

A CRM reasoning model that natively understands the Salesforce data schema — that knows what an "Account" object means in context, how "Opportunity" relates to "Contact," what "Close Date" and "Stage" fields represent in a pipeline forecast — reduces the error rate and prompt engineering overhead associated with adapting general-purpose models to these tasks. For enterprise customers, this translates into more reliable agent outputs, fewer hallucinated CRM actions, and lower latency on reasoning-intensive tasks like pipeline analysis and account summarization.

Koa is currently in pilot for select Agentforce customers. GA for US regions is expected in winter 2026. International availability is on the 2027 roadmap.

Agentforce Coworker: The Thesis in Practice

The clearest documented production deployment of AIforce's agent-first thesis is Fulton Bank, a $30 billion regional financial services company that presented at Dreamforce with specific operational metrics.

Fulton Bank went from zero to more than 20 Agentforce Coworker use cases in production in approximately six weeks: setting up its first agent in one hour, deploying to power users within one week, and reaching 3,000 active users in six weeks. The bank attributes 80,000 hours of saved work and $389 million in loans and deposits to its Agentforce deployment.

The $389 million revenue attribution figure deserves scrutiny. Attribution methodology for AI agent revenue impact is not standardized, and the figure likely reflects a combination of agent-assisted deal acceleration, reduced friction in lending workflows, and some portion of correlation rather than direct causation. Salesforce has not published the attribution methodology, and the number should be read as an internal customer estimate rather than an independently audited finding.

The 80,000 hours saved figure is more directly verifiable: it represents a measurable reduction in time-per-task across 3,000 users over a defined deployment period. If accurate, it translates to roughly 27 hours per user over the deployment period — not transformative on a per-employee basis, but meaningful at scale when multiplied across a 3,000-person user base and compounded over a full year.

The Fulton Bank case illustrates the deployment model Salesforce is selling: rapid time-to-first-agent (one hour), fast rollout to power users (one week), and expansion across the organization driven by demonstrated value rather than top-down mandate. That deployment playbook is important context for enterprise buyers evaluating AIforce, because it reflects the operational reality of agent deployment at scale — the value is real, but it accumulates through use cases and user adoption rather than arriving as a step-change transformation on day one.

The Pricing Model Is the Real Story

The most consequential unresolved question from Dreamforce 2026 is not which surface is in beta or when Koa will reach GA. It is what enterprises will pay for agent-scale Salesforce interactions once HPI metering begins.

The current state of Agentforce pricing: Salesforce's legacy seat-based licensing model charges per user per month. The Agentforce conversational layer charged $2 per conversation at launch, which was later revised to $0.10 per action for simpler interactions. AIforce introduces a new consumption unit — the Headless Platform Interaction — for agent API and MCP calls. The HPI Flex Credit multiplier is listed as TBA; Salesforce has committed to 30 days' notice before metering begins.

For a 30-seat enterprise team, the realistic first-year total cost of ownership for an AIforce deployment — factoring in Service Cloud Enterprise Edition licensing, implementation services, and estimated Flex Credit consumption — ranges from $200,000 to over $450,000. The variance in that range is almost entirely explained by the unresolved HPI multiplier.

This pricing structure creates a material forecasting problem for enterprise buyers. An enterprise customer that deploys Claude as its primary internal AI interface, configured via Claudeforce to access Salesforce data, and runs that deployment across a sales team of 500 people doing 50 Salesforce calls per day, would generate 25,000 HPIs per day — 9 million per year. At a Flex Credit multiplier of 1x (comparable to a basic automation action), that volume is manageable. At a multiplier of 5x (comparable to complex reasoning actions in Salesforce's existing rate card), it becomes a significant line item that wasn't in the original enterprise contract.

Salesforce pricing history suggests the company will set the initial HPI multiplier at a level that encourages adoption, then adjust as the category matures — the pattern it has followed with Agentforce conversation pricing, Data Cloud credits, and Einstein analytics consumption. But enterprise procurement teams making multi-year commitments to AIforce architectures cannot rely on that pattern as a budget assumption.

The four questions enterprise buyers should answer before committing to an AIforce deployment:

1. Instrument consumption before signing. Deploy AIforce in a pilot environment and measure HPI call volumes per user per workflow before the multiplier is published. 30 days' notice gives you a window to model costs on real data rather than assumptions.

2. Model the multiplier range. Price your deployment at 1x, 3x, and 5x multipliers and assess which scenarios remain within budget. If any scenario exceeds tolerance, negotiate contract protections before the rate card is published.

3. Identify which surfaces you actually need. Claudeforce and Slackforce are in beta; Koa is in pilot. If your use case depends on those surfaces, build a timeline that assumes GA delays and evaluate alternative architectures for the gap period.

4. Benchmark the Fulton Bank ROI model. Fulton Bank's 80,000 hours saved across 3,000 users is a credible production benchmark. Scale it to your user count and task mix to build a bottom-up ROI case that doesn't depend on the $389M revenue attribution figure.

What "AI Replaces the UI" Means for Enterprise GTM

The "AI replaces the UI" thesis is a claim about where enterprise work happens in the long run, not just a product announcement. It is a bet that the center of gravity for employee interaction with enterprise software will shift from browser-rendered application interfaces to AI-native interfaces: conversational agents, orchestrated workflows, and headless API calls that happen without any human navigating a screen.

Microsoft made the same bet with its Copilot super app in September 2026: a unified AI interface for Home, Code, and Autopilot that pulls enterprise context from Microsoft's data stack (Teams, SharePoint, Dynamics 365, Exchange) into a single AI-native surface. The two approaches represent different views of how the enterprise UI transition happens: Microsoft's bet is "one AI interface to rule them all" — a Copilot surface that pulls everything into a single Microsoft-controlled experience. Salesforce's AIforce bet is "let enterprise data be callable from any AI interface" — a headless platform layer that is neutral about which AI interface the employee uses.

The neutrality of AIforce is both its competitive advantage and its strategic risk. An enterprise CIO who wants employees to use Claude, Salesforce, and Slack as separate best-of-breed tools, all connected through a shared data layer, finds AIforce's architecture compelling. A CIO who wants a single, governed AI interface for the whole organization finds Microsoft Copilot's "one app for work" model simpler to deploy and audit.

The enterprise agent coordination problem — the challenge of governing AI workflows across multiple vendors and platforms — is being attacked from multiple directions simultaneously. Wonderful's $550M raise at $5B built an explicit cross-vendor orchestration layer above every agent platform including Agentforce. AIforce is Salesforce's answer to the same problem from the inside: make Salesforce so deeply integrated into the AI interface layer that it becomes the system of record for enterprise context, regardless of which AI model is running.

The 81% of CIOs who reported lacking full oversight of their deployed AI agents in the Dataiku Harris Poll are the natural AIforce customer: organizations where agent proliferation has outrun governance, where the cost of context-switching between AI interfaces and enterprise applications is visible in productivity data, and where a trusted vendor with established compliance credentials can solve an architectural problem without requiring a complete infrastructure rebuild.

The Cross-Vendor Neutrality Question

AIforce's architecture is nominally AI-interface neutral: it exposes Salesforce capabilities via MCP to any compliant AI interface. Claude, Microsoft Copilot, Google Gemini, and custom enterprise agents can all call Salesforce over MCP if they are configured to do so.

In practice, Salesforce has prioritized three integrations for the AIforce launch: Claudeforce (Anthropic Claude), Slackforce (Slack, which Salesforce owns), and Agentforce Coworker (Salesforce's own Lightning interface). Claudeforce launched in beta in September 2026 with 37 prebuilt Salesforce skills and a governed MCP architecture that routes Claude actions through Salesforce's permission and audit layer.

The competitive implication is that AIforce is neutral in principle and Salesforce-first in practice. The best-supported, most-tested, and most feature-complete AIforce experience is via Claudeforce or Agentforce Coworker — both developed and maintained by Salesforce or a priority Salesforce partner. Third-party AI interfaces connecting via generic MCP will work but will lack the semantic pre-configuration and prebuilt skill libraries that the priority integrations provide.

For enterprise buyers, this distinction matters when evaluating which AI interface to standardize on. If the organization has already committed to Claude as its primary AI platform, Claudeforce provides a tighter integration path. If the organization is evaluating AI interfaces and Salesforce is a critical system of record, AIforce's integration architecture is a legitimate tiebreaker in favor of Claude.

What Salesforce's $27 Million Education Bet Signals

Alongside the AIforce announcement, Salesforce committed $27 million to AI education at Dreamforce 2026: funding for Trailhead AI curriculum, workforce development partnerships, and an expansion of the Salesforce career pathway program.

The education commitment is a distribution play, not a philanthropy signal. Salesforce's long-term competitive position depends on an enterprise workforce that knows how to build, deploy, and manage Agentforce agents and AIforce integrations. Every enterprise employee who learns Salesforce's AI tooling through Trailhead or a workforce development program becomes a potential internal advocate for expanding the organization's Salesforce footprint.

It is the same playbook that made Salesforce Trailhead one of the most effective enterprise distribution strategies in SaaS: educate the practitioner, create a talent pool that is credentialed in your platform, and watch the demand for your platform follow the talent.

At $27 million, the education commitment is small relative to Salesforce's revenue base — but directionally correct as a long-term distribution investment in the agentic enterprise era.

The Honest Assessment: What Dreamforce Delivered vs. What It Promised

Dreamforce 2026 delivered a coherent architectural vision and one production-validated deployment (Fulton Bank). It delivered a generally available Agentforce Coworker, expanded enterprise partnerships, and a clear product roadmap for the AIforce stack. What it did not deliver is a fully available, fully priced AIforce platform.

The two AIforce surfaces with the broadest enterprise interest — Claudeforce and Slackforce — are in beta. Koa, the CRM reasoning model that could fundamentally improve agent accuracy on Salesforce data tasks, is in pilot. HPI pricing, the variable that will determine the real TCO of agent-scale Salesforce deployments, is TBA.

That gap between vision and delivery is not unusual for a Dreamforce release cycle. Salesforce has historically used Dreamforce to announce directional bets and then shipped production-ready versions over the following six to nine months. The pattern for AIforce suggests general availability of the full stack in H1 2027.

For enterprise technology leaders, the operational planning implication is this: AIforce is worth evaluating now, worth piloting in a production environment to instrument HPI consumption volumes, and worth building into the 2027 enterprise AI architecture roadmap. It is not worth committing to a full-scale deployment today based on Dreamforce's stage demos alone.

Takeaway: Salesforce's "AI replaces the UI" thesis is directionally correct and architecturally coherent. The bet is that Salesforce's enterprise knowledge layer — not its application interface — is its durable competitive asset in the agentic era. AIforce exposes that asset to any AI interface through MCP, making Salesforce callable wherever work happens. The pricing question (HPI multipliers TBA) is the central unresolved variable for enterprise procurement teams. Pilot now. Price carefully. Plan for Q1 2027.