What happened
Salesforce and Nvidia announced Koa, described by Salesforce as its first CRM reasoning model for its Agentforce platform. Salesforce built Koa by post-training Nvidia's open-weight Nemotron 3 Super model on a large synthetic dataset modeled on roughly 27 to 30 years of Salesforce's own CRM deployment patterns, covering more than 14 industries. Salesforce says no actual customer data was used to train Koa. Instead, the company built simulated scenarios, a fictional customer service call, a fictional sales negotiation, that mirror the structure of real CRM work without using anyone's real records. Salesforce controls the model's weights and runs both training and day-to-day use (inference) inside its own infrastructure, what the company calls its trust boundary, so customer data does not have to leave Salesforce's systems to get a reasoning answer. On Salesforce's own internal benchmark, the company reports that Koa matches or beats the performance of leading general-purpose models on CRM tasks like updating a sales opportunity or routing a support case, while making about a third as many errors. Salesforce AI executive Jayesh Govindarajan told TechCrunch that before Koa, any task that needed real reasoning, not just following a script, got routed to a frontier model like Claude or ChatGPT through Agentforce's internal AI "gateway," the system that decides which model handles which request. Koa gives Salesforce a cheaper, narrower option for a large share of that traffic. Salesforce did not walk away from frontier labs to do this. The company separately confirmed at Dreamforce that Claudeforce, its expanded partnership with Anthropic announced in late August 2026, continues, with Claude remaining the default model inside Slack and for other reasoning tasks in Agentforce. Salesforce is now running at least three tiers of models side by side: its own narrow Koa model for high-volume routine CRM tasks, Claude and other frontier models for broader reasoning, and older, smaller task-specific models it already had.
Why it matters for business owners
Almost no small or medium-size business will ever train its own AI model the way Salesforce just did. That is not the point. The point is what this reveals about how the AI-powered software you already buy is likely to be built going forward: a single product, sold under one name, quietly running on more than one model depending on the task. That matters for two practical reasons. First, cost: a narrow model trained for one repeated job is usually far cheaper to run than a general frontier model, because it uses fewer tokens and less compute to do the same task. Vendors have a direct financial incentive to route as much routine work as possible to the cheaper model, whatever they call it in their marketing. Second, data: where your information goes depends on which model actually handles a given request, not on the product name on the invoice. A tool that says it is "powered by AI" or even names a specific frontier model in its marketing may still route different tasks to different models behind the scenes, some kept inside the vendor's own systems, some sent to a separate model provider. This is already common in enterprise software. Salesforce's own Agentforce gateway has been routing tasks across multiple models for some time. Koa is simply the first case big enough, and public enough, to make that architecture visible to people outside the industry.
What owners should not misunderstand
Koa is not evidence that specialized, narrow AI models are always better than frontier models. Salesforce's own reporting is clear that Koa is meant for a specific, repeated category of task, CRM actions like updating records or routing cases, not for open-ended reasoning or unfamiliar problems. Trade publications covering the announcement noted directly that Koa does not match the performance of top frontier models on general tasks. A narrow model is a good fit for a narrow, repeated job, not a replacement for a general one. The three-times-fewer-errors figure is Salesforce's own number, measured on Salesforce's own internal benchmark, on Salesforce's own defined tasks. No independent, outside organization has reproduced that figure. That does not make it false. It does mean a business should treat it the same way it should treat any vendor's self-reported benchmark: a claim to note, not a fact to repeat as settled. This also is not a case of a vendor cutting off frontier labs to save money. Salesforce is expanding its use of Claude in other parts of the same platform at the same time it introduces Koa. The lesson is not "specialized models replace frontier models." It is that a serious AI vendor now runs a deliberate mix, and picks the model per task, not per product.
The operational lesson
The product name on an AI tool does not tell you which model is doing the work, or where your data goes when it does. A single AI feature inside a piece of software you already pay for can be quietly backed by more than one model: one kept fully inside the vendor's own systems for routine, high-volume tasks, and another, from an entirely separate company, handling the harder or less frequent requests. That split usually happens for the vendor's benefit, cost control and data-handling convenience, not necessarily as a feature the vendor advertises to you. It is not something to be alarmed about by default. It is something to ask about, the same way you would ask which subcontractor is doing the actual work on a project you hired a general contractor for. This pattern will keep spreading. As specialized, cheaper models become easier to build and license, more of the software you already use will quietly mix models behind one interface. The businesses that get surprised later are the ones that never asked the question in the first place.
What a serious business should do next
Ask any AI vendor you pay for a direct question: for the specific feature you use, which model or models actually handle the task, and does that ever change based on cost or load. A vague answer, or a name that only mentions the product, not the underlying model, is itself useful information. Ask where your data goes when a task is handled. Does it stay inside the vendor's own infrastructure, as Salesforce describes for Koa, or does it get sent to a separate model provider for some tasks and not others. Get this in writing where it matters, in your data processing agreement, not just in a sales conversation. Do not assume a vendor's marketing name for a feature, or even the name of a well-known frontier model mentioned in their materials, describes everything that feature actually does. Large platforms increasingly mix models by design. If you are evaluating a new AI-powered tool for a repeated, high-volume task, ask whether the vendor offers, or plans to offer, a narrower and cheaper option for that specific job, the same tradeoff Salesforce is making internally. You do not need to build a model to benefit from this shift. You need to know it is happening inside the tools you already rely on.
The Atlacis view
Atlacis helps business owners see past a vendor's product name to what is actually running underneath it: which model handles which task, where the data goes, and what that means for cost and exposure. Koa is a useful, visible example of a shift that is already happening more quietly across enterprise software, vendors mixing specialized and frontier models inside one product to control cost and keep sensitive work inside a defined boundary. Atlacis helps owners map which models power the AI features they already pay for, check whether that matches what their contract and their risk tolerance actually require, and decide where a narrower, cheaper, more controlled option would serve the business better than routing everything through the biggest available model.
The short version
- On September 15, 2026, Salesforce and Nvidia announced Koa, Salesforce's first in-house reasoning model, built by training Nvidia's open-weight Nemotron 3 Super on synthetic (non-customer) data modeled on Salesforce's own CRM deployment history.
- Salesforce controls Koa's weights and runs it entirely inside its own infrastructure, so routine CRM tasks no longer have to be routed to a frontier model like Claude or ChatGPT to get a reasoning answer, per Salesforce and TechCrunch.
- Salesforce reports Koa matches or beats leading models on its own internal CRM benchmark with three times fewer errors. That figure is self-reported and has not been independently reproduced by any source covering the announcement.
- Salesforce did not stop using frontier models. Claude remains the default model elsewhere in the same platform through the separate Claudeforce partnership with Anthropic, confirmed as ongoing at the same event.
- The pattern behind Koa, a single product quietly running on more than one model depending on the task, is spreading across enterprise software. The product name and marketing rarely disclose which model handles a given request or where the data goes.
- Ask any AI vendor which model handles the specific feature you use and where your data goes when it does. Get the answer in your contract or data processing agreement, not just a sales conversation.
Where ATLACIS can help
- Read: what private AI actually means before you evaluate a vendor's claims
- Read: how to reduce AI token costs without cutting capability
- Read: private AI versus public AI tools, and how to tell which one you are actually using
- Review Private AI Implementation
- Book a call to map which models power the AI tools your business already runs
Sources
- Salesforce: Announcing Koa, Salesforce's First CRM Reasoning Model, Built on NVIDIA Nemotron (September 15, 2026)
- Salesforce: Koa, built on Nvidia Nemotron (product page)
- TechCrunch: Salesforce and Nvidia's new reasoning model is everything the AI labs should fear (Julie Bort, September 15, 2026)
- CIO: Salesforce, NVIDIA unveil CRM domain-specific reasoning model (Thor Olavsrud, September 15, 2026)
- SiliconANGLE: Salesforce debuts Koa, a specialized model built to reason about CRM data (Mike Wheatley, September 15, 2026)