The News
Koa Arrives At Dreamforce, Trained On Decades Of CRM Work
On September 15, 2026, Salesforce and NVIDIA announced Koa, Salesforce's first CRM reasoning model for Agentforce, built on NVIDIA Nemotron and designed to help agents reason through complex, multistep workflows.
It was not trained from scratch. Koa was built by post-training NVIDIA Nemotron 3 Super, an open-weight model, using a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments, with scenarios spanning more than 14 industries including manufacturing, financial services, healthcare, and travel. Salesforce says no customer data went into training it.
The performance claim is careful. Salesforce says Koa matches or exceeds leading model performance on CRM actions with 3x fewer errors, and on its internal CRM Bench evaluation Koa averaged 0.86 against 0.87 and 0.90 for two leading general-purpose models.
Read that again: on the company's own test, its new model came in slightly behind the frontier models it is meant to displace. Koa runs internally at Salesforce today and is in customer pilots including 1-800Accountant, Formula 1, and UChicago Medicine, with U.S. general availability expected in winter 2026.
A new type of AI called "Accelerated AI" is about to take the world by storm…
And stocks connected to it are already breaking out: 133%… 210%… and even 320% or more just in the last few months.
But it's just getting started…
If you want to learn more about "Accelerated AI" – and get name and ticker of the #1 pick to play this opportunity…
The Company Behind It
A Vendor That Is Now Customer, Partner, And Competitor To The Labs
Salesforce sells the systems where sales and support work happens, and its durable advantage is not its interface but the accumulated data, workflow logic, and permissions built up over 27 years of enterprise deployments.
Koa is one piece of a wider strategy. The same week, Salesforce announced AIforce, which opens its data, business logic, permissions, and actions to outside AI tools through more than 60 MCP tools and 4,000-plus APIs, and it continues to route customers to frontier models for harder reasoning work.
That leaves Salesforce occupying three roles at once. It buys frontier models, distributes them to its customers, and now competes with them on the specific tasks its platform handles most often.
Why This Matters Financially
Good Enough, Narrower, And Cheaper
The economics favor specialization. A model tuned to CRM work should need fewer tokens to complete a task than a general-purpose model, which lets Salesforce price a CRM-specific option aggressively without disturbing its frontier arrangements for harder work.
Margin is the real prize. Every repetitive sales or support query answered by a model Salesforce controls is a query it does not pay a frontier lab to handle, and at enterprise volume that difference compounds.
Control matters as much as cost. Running its own model keeps pricing, availability, and behavior in Salesforce's hands rather than subject to another company's roadmap.
Limits and Uncertainty
The Catch: The Benchmark Belongs To The Company Selling The Model
CRM Bench was designed by the organization that benefits from strong results, and no independent auditor has verified them. The AI industry has a documented problem with benchmark reliability, which makes self-administered scores a weak basis for a purchasing decision.
Delivery is unproven. Koa is in pilots with general availability not due until winter, Salesforce disclosed no pricing for the AIforce layer it sits within, and a Bloomberg investigation reported that Agentforce, the platform Koa powers, has been falling short of its marketing promises in real deployments. Every enterprise software vendor is attempting the same specialization, so the approach is not proprietary.
The launch matters because it shows enterprise software companies concluding that frontier capability is not what most business tasks require. The real impact depends on whether independent buyers confirm the cost advantage in production, and whether a slightly weaker model at a lower price is a trade enterprises actually want.
Disclosure: This content is for educational and informational purposes only and does not constitute investment advice or recommendations. You should always conduct your own research or consult a qualified financial advisor before making investment decisions.
