Salesforce launches Koa, a CRM reasoning model built on NVIDIA Nemotron

Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Salesforce says it post-trained NVIDIA Nemotron 3 Super on proprietary synthetic enterprise scenarios — not customer data — so agents can reason through multistep CRM work inside Salesforce’s trust boundary. Pilot customers named include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero; general availability is expected in winter 2026 in U.S. regions.

Same-day company primary that a major enterprise platform is shipping its own CRM reasoning model on NVIDIA open weights — specialized for sales and service workflows, hosted inside Salesforce’s trust boundary, with named pilots — instead of only routing hard tasks out to frontier labs.

Salesforce and NVIDIA announced Koa on 15 Sep 2026 in a Salesforce news story titled “Announcing Koa: Salesforce’s First CRM Reasoning Model, Built on NVIDIA Nemotron.” Koa is Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. CRM means customer relationship management software — the sales and support systems companies use to track deals and cases. A reasoning model here is AI that can work through a multistep job, not just one-shot answers. Agentforce is Salesforce’s platform for building AI agents. That company page, dated September 15, 2026 and datelined San Francisco, is the filing event. These are company claims. This desk did not run Koa.

How it was built, as Salesforce tells it: the company post-trained NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on enterprise knowledge from nearly three decades of CRM deployments. Post-trained means they started from NVIDIA’s published model and taught it extra CRM skills. Synthetic here means made-up training scenes, not live customer records. File the base model and the synthetic-corpus line as Salesforce’s. How the training stack is wired stays in Sources.

Training data, still the company’s — a critical line: no customer data was used. The training corpus was built entirely from synthetic scenarios that simulate agent work across the customer lifecycle — generating leads, qualifying opportunities, and resolving service cases. Salesforce says those scenes cover more than 14 industries, including manufacturing, financial services, healthcare, and travel, and that they draw on “27 years of Salesforce CRM intelligence.” File “no customer data,” the 14-plus industries, and the 27-year line as Salesforce’s. This desk did not inspect the training set.

Control, company: Salesforce controls the model weights and runs post-training and inference inside its own trust boundary — its controlled systems — so customer data does not cross that line during inference. Open-weight / Nemotron, as used here, means NVIDIA published the starting model weights so a partner can customize them. File the weights-and-hosting line as Salesforce’s. This desk did not sit on the inference path.

Benchmark claim, company — do not independently verify: on Salesforce’s CRM benchmark, including tasks such as updating an opportunity, routing a case, or scheduling a follow-up, Koa “matches or exceeds leading model performance on CRM actions with three times fewer errors.” File that as Salesforce’s scoreboard, not an independent lab result and not a claim this desk reran.

Pilots named on the same page: 1-800Accountant, Baxter Credit Union (BCU), Engine, Formula 1, UChicago Medicine, and Xero. Salesforce also says Koa is already in use inside the company, including a Slack agent that helps employees find information and complete everyday tasks. File those names and the internal Slack agent as Salesforce’s. This desk did not interview the pilots.

Named voices on the company page: Marc Benioff, Salesforce chair and CEO; and Jensen Huang, NVIDIA founder and CEO. Benioff, quoted: “The most valuable thing Salesforce has built isn’t our platform — it’s the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself.” And: “That’s a different kind of intelligence, and it runs entirely inside your trust boundary.” Huang, quoted: “NVIDIA Nemotron open models give Salesforce the foundation to turn decades of enterprise expertise into specialized AI with Koa, creating a CRM model that can reason and securely take action.” File the quotes as theirs, via Salesforce. They are not an independent product review. Customer quotes stay in Sources.

Missionforce, still company: Salesforce and NVIDIA are bringing Nemotron-based models and accelerated computing into Missionforce for government and regulated organizations — private clouds, air-gapped networks, and other secure environments that do not touch public infrastructure. An air gap means the system is kept off the public internet. Missionforce Operations, a product that digitizes government back-office work such as procurement, supplier management, and logistics, is generally available now in U.S. regions. Post-trained NVIDIA models for that product are available to select customers in October 2026. File those windows as Salesforce’s. This filing is not a federal procurement award.

Availability, company forecast: Koa is available to select pilot customers now in Agentforce; general availability is expected in winter 2026 in U.S. regions. File that as Salesforce’s. Koa is not a global GA product today, and this desk is not inventing a price, a token cost, or a parameter count.

Plain English for the rest of the card: CRM = sales and support software companies use to track customers. Reasoning model = AI that can work through several steps, not just one answer. Agentforce = Salesforce’s AI-agent platform. Open-weight / Nemotron = NVIDIA’s published model weights that partners can customize. Trust boundary = customer data and inference stay inside Salesforce’s controlled systems.

PRIMARY here: Salesforce’s 15 Sep 2026 news story — Tier A company source, the original record. Discovery / corroboration: an arXiv paper, “Salesforce Koa: An Enterprise Language Model for Agentic Tool Use” (2609.15066), submitted 14 Sep 2026 — not the sole source and not a second originating company text. The Koa announce, the Nemotron 3 Super post-train, the no-customer-data / synthetic-scenario corpus, the 27-year CRM-intelligence line, the trust-boundary hosting, the company CRM-benchmark claim, the named pilots, the Benioff and Huang quotes, the Missionforce windows, and the winter 2026 U.S. GA aim are company-attributed. NOT claimed: a price, token costs, a parameter count, that Koa replaces Claude or GPT, independently verified error rates, that this desk tested Agentforce, a federal award, a stock tip, or investment advice. Distinct from the already-filed salesforce-trusted-enterprise-ai-harness, salesforce-tsa-ace, and nvidia-crowdstrike-safemind.