At Dreamforce 2026 in San Francisco, Salesforce introduced Koa, its first purpose-built CRM reasoning model, developed by post-training NVIDIA's Nemotron 3 Super base model on a proprietary synthetic dataset derived from approximately 27 years of CRM domain knowledge. The announcement represents a strategic bet that enterprise AI needs domain-specific models rather than general-purpose large language models. But the launch arrives alongside a Bloomberg investigation documenting that enterprise customers are reporting Agentforce outcomes that materially miss what Salesforce's marketing promised.
What Koa Is and How It Works
Koa is a reasoning model built specifically for CRM tasks. The training data was synthetic, generated from Salesforce's domain expertise across more than 14 industries, and notably used no real customer data. The underlying NVIDIA Nemotron 3 Super architecture uses a hybrid Mixture of Experts (MoE) design, which activates only a subset of specialized parameters per request rather than the full model capacity. This keeps inference costs low, which matters because Agentforce's base pricing is $2 per conversation.
Koa also features a one-million-token context window, enabling an agent to reason across an entire customer history or a complex multi-step workflow without losing context. The model is optimized for the specific CRM failure modes that damage enterprise trust: updating opportunity stages incorrectly, misrouting support cases, and scheduling follow-ups for the wrong contacts.
The Benchmark Problem
Koa's headline claim is that it "matches or exceeds leading model performance" on Salesforce's internal CRM benchmark while producing three times fewer errors on specific CRM tasks. That benchmark is Salesforce's own creation, administered by the same organization that benefits from strong results. No independent third party has verified the figures.
This matters because the AI industry has a documented benchmark reliability problem. A UC Berkeley research team published findings in April 2026 showing that researchers had manipulated eight industry-standard AI agent benchmarks to near-perfect scores without solving a single real task. Apple's GSM-Symbolic research found performance drops of up to 65% when models that scored highly on one math benchmark were tested on minor variations of the same problems. The international AI safety community's 2026 report concluded that "performance on pre-deployment tests does not reliably predict real-world utility or risk."
None of this proves Koa's benchmark is wrong. It proves that a single vendor-administered benchmark on a vendor-developed evaluation set is insufficient evidence to justify a purchasing decision, especially one with the cost profile described below.
The Bloomberg Investigation and ROI Reality
A Bloomberg investigation published before Dreamforce found that enterprise customers are reporting Agentforce outcomes that materially miss Salesforce's marketed expectations. A Goldman Sachs analysis cited alongside those findings noted that agent token economics are shifting rapidly, creating unstable ROI calculations across different deployment contexts.
Independent reviews are consistent with this picture. A 2026 technical review of Agentforce found implementation timelines of four to six weeks for single-use-case deployments and eight to sixteen weeks for complex rollouts, with Year 1 total costs reaching $270,000 to $540,000 when Data Cloud licensing, implementation services, and training are included. These figures place Agentforce firmly in enterprise-tier spending, well beyond what most SMBs would consider for a single AI tool.
Koa is currently in pilots with 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, and Xero, with U.S. general availability not expected until winter 2026.
AIforce and the MCP Integration
The more architecturally significant announcement may be AIforce, a new layer that makes Salesforce's CRM data, permissions, and business logic accessible from inside third-party AI environments including Anthropic's Claude, Google Gemini Enterprise, and Amazon Q, without employees opening a Salesforce tab.
Patrick Stokes, Salesforce's president of applications, articulated the strategy candidly: "The value of Salesforce is in the data and the metadata. What we are doing now is exposing it to a new UI." That statement is an unusually direct acknowledgment that Salesforce's competitive future no longer rests on the interface it spent 25 years building.
AIforce uses the Model Context Protocol (MCP), the open standard Anthropic introduced in November 2024. Salesforce's Headless Toolkit now exposes more than 60 MCP tools, more than 30 coding skills, more than 4,000 APIs, and more than 220 CLI commands to external builders. Three initial AIforce products launched: Claudeforce (inside Claude with 37 ready-made sales skills), Slackforce (Slack integration with Salesforce context), and Agentforce Coworker (within Lightning).
What This Means for SMBs
For small and midsize businesses evaluating whether Salesforce's AI roadmap affects them, the Koa launch and the Agentforce ROI gap reveal three practical lessons.
1. Self-Administered Benchmarks Are Marketing, Not Evidence
When a vendor claims its model outperforms competitors on a benchmark it designed, the claim should be treated as marketing copy, not independent evidence. Demand third-party validation, independent benchmarks, or pilot results with metrics you can verify before committing budget.
2. Enterprise AI Costs Are Not SMB Costs
Agentforce's Year 1 costs of $270,000 to $540,000 place it firmly in enterprise territory. SMBs should understand that the same AI capabilities, often powered by the same foundational models, can be deployed through smaller platforms at a fraction of the cost. The price premium buys the Salesforce ecosystem, not the AI itself.
3. MCP Is the Real Signal
The most important trend in Salesforce's announcement is not Koa. It is the adoption of MCP as the standard interface for connecting AI agents to enterprise data. This is the same protocol that WSO2, Anthropic, Google, and dozens of other vendors are building around. For SMBs, MCP means that the AI tools you choose, including lower-cost alternatives, can increasingly connect to the same data sources that Salesforce exposes through its own tools.
The Buy-vs-Build Question
Salesforce's Dreamforce announcements reinforce a pattern we see across the enterprise AI market: large vendors are building domain-specific models and open-protocol data access, but their pricing and implementation timelines remain tuned for enterprise buyers. For SMBs, the practical question is not whether to adopt Salesforce's AI but whether the MCP-based integration layer makes it possible to achieve similar outcomes with a leaner stack.
At Vivid Visuals, we help businesses evaluate exactly this trade-off: enterprise-grade capabilities delivered through open standards at a cost that fits the business.
Read next from the blog : The AI Visibility Gap: Why Your Website Is Not Enough Anymore · WSO2 Agent Manager: Open-Source Governance for the AI Agent Sprawl Era · Meta's Stilla.ai Acquisition and WhatsApp Business MCP: The Automation Play for SMBs · Google Ads AI Video Dubbing: Why Multilingual Video Just Got Fast
