Salesforce Introduces Koa: Training on CRM Workflows Leaves Exceptions to Be Tested
TL;DR
Built on NVIDIA Nemotron, Koa enters Agentforce customer pilots. Workflow specifications become training material, but internal benchmarks, real exceptions and general availability remain distinct.
Salesforce introduced Koa on 2026-09-15, adding a reasoning model specialized for multistep CRM work to Agentforce. The announcement leaves a concrete question: can it handle exceptions absent from its training scenarios? If unfamiliar workflows still require substantial human correction, familiarity with existing processes would not justify replacing models throughout a business.
The official announcement carries a September 15 date without a precise publication time. IT Pro’s report was published at 12:00 UTC that day, or 20:00 in Taipei. This article is dated September 16 in Taipei and uses the previous day’s product announcement within the 48-hour research window. The article date is not an activation date. The technical paper appeared earlier, on September 14.
Salesforce developed Koa with NVIDIA. The company describes turning 27 years of CRM experience into synthetic scenarios, without training on customer data. The paper specifies public and synthetic training data and a Nemotron-3-Super-120B foundation. Workflow specifications generate multiround tasks, with training rewards tied to successful tool-mediated task completion. Specifications that configure agents can therefore also train the model.
Access currently covers selected Agentforce pilot customers; general availability in U.S. regions is expected in winter 2026. It is not a broadly available release. IT Pro reports internal use in Salesforce’s Slack employee assistant and customer pilots including Formula 1 and Xero. It provides no customer task-success rates or savings figures. The official product page also provides no public price sufficient for comparing complete-task costs.
The paper reports improved multiround tool use while acknowledging that overall performance remains below the strongest frontier models. The announcement’s claim of fewer CRM-action errors comes from vendor evaluation; it cannot be converted directly into production customer success rates. The evidence supports the usefulness of specialization, not a conclusion that every reasoning task should move to Koa.
Reuse workflow specifications, but define completion correctly
I find the implications for software companies particularly useful. A company that has documented workflows, tool permissions and completion conditions could reuse those specifications for training, reducing dependence on explaining the work through prompts each time. That inference assumes the specifications actually describe the customer’s intended outcome. Years of documentation or the volume of data do not establish that the training objective is correct.
Suppose a service agent arranges a replacement. A successful tool response does not establish that the warehouse has stock, the address is correct or the customer accepts the new date. A reward that recognizes only an executed action could teach the model to finish a system operation without resolving the request. This is a hypothetical illustration of how completion criteria affect training, not a reported Koa customer incident. Workflow designers would also need to specify when an unresolved request must be handed over.
Let differences in work determine model choice
Salesforce’s product page describes model selection by agent or subagent. That flexibility has a practical use: a specialized model can handle established workflows while other options remain available for unfamiliar problems. This is a product judgment about allocating models by task, not an assumption that Koa is cheaper or that a pilot roster proves adoption at scale.
Useful next evidence would compare completion rates on the same work with identical tools and permissions, separating familiar scenarios from unseen exceptions. Cost comparisons should include retries and human handovers. If Koa improves only familiar tasks, its use should remain limited to those workflows. Stable improvements on exceptions would provide a stronger basis for extending its role.
The cover reuses an existing NVIDIA campus photograph to represent the collaborator. It does not show the Koa launch or product.
Sources:
Related Articles
NVIDIA Expands Agent Safety: Runtime Permissions Do Not Guarantee Correct Work
NVIDIA launches its Open Agent Safety Platform, with OpenShell 0.1.0 enforcing file, network and tool permissions outside the model. That can contain mistakes, but more than 100 participating organizations do not establish deployment results or correct work.
OpenAI Launches Dots: Ongoing Work Should Not Become a Growing Review Queue
OpenAI is rolling out dots to eligible paid users for work that continues between conversations. Read-only background research and limited memory controls make missed commitments and accumulated review work more useful tests than activity alone.