At Dreamforce on Monday, Salesforce and NVIDIA introduced Koa, a reasoning model built to power Agentforce agents on multistep customer-relationship tasks without routing every decision to the largest frontier APIs.

Why a CRM-specific model now

Koa post-trains NVIDIA’s open-weight Nemotron-3-Super-120B using synthetic scenarios derived from decades of Salesforce implementation patterns—not live customer data, executives stressed. A simulation-to-reward pipeline expands workflow specs written in Agent Script into multi-turn tasks graded on successful tool use.

On Salesforce’s internal CRM benchmark, Koa matched or beat larger general models on actions like updating opportunities and routing cases while logging about three times fewer errors, according to company materials. The model runs inside Salesforce’s trust boundary with weights controlled by the vendor.

Economics and availability

Enterprises chafing at per-token bills from frontier providers see specialized models as a way to keep agents on rails. Koa is available to select pilot customers now, with broader U.S. availability slated for winter 2026.

SiliconANGLE reported that chief platform officer Rohan Kumar emphasized inference stays within Salesforce infrastructure, addressing data residency concerns that slow agent adoption in regulated industries.

Technical trade-offs

The arXiv paper notes Koa surpasses a strong proprietary baseline on tool use but remains below the strongest frontier systems on open-ended reasoning. That gap may be acceptable if agents mostly orchestrate CRM APIs rather than draft strategy memos.

Because the base Nemotron weights are openly licensed, sophisticated customers could attempt their own post-training. Salesforce is betting its synthetic data moat and hosted guardrails matter more than raw parameter access.

Competitive context

Microsoft’s Monday AI constitution and weekend pacing debate highlight safety costs that specialized models might amortize across narrower tasks. Oracle and ServiceNow are racing similar offerings tied to their data models.

For CIOs, the question is whether Koa reduces total cost of ownership or simply shifts spend from OpenAI invoices to Salesforce SKUs. Early pilots will be judged on mean time to resolve tickets, not benchmark leaderboard spots.

Dreamforce reception

Attendees at Dreamforce demo sessions said Koa handled nested CRM updates—closing a case, logging a call, and scheduling follow-ups—with fewer hallucinated record IDs than general models. Failures still appeared on ambiguous natural-language requests, reminding buyers that human review remains necessary.

NVIDIA highlighted the partnership as evidence that Nemotron fine-tunes can anchor vertical clouds, a pitch to other software giants weighing whether to train proprietary reasoning layers.

Customer advisory boards asked whether Koa will support on-prem GPU clusters for regulated banks. Salesforce executives said the first releases remain multi-tenant cloud only, with Missionforce options for defense clients later in 2026.

Analysts estimated that if Koa displaces even a fraction of external API calls inside Agentforce, gross margins could improve, though training and inference on Nemotron still carry substantial NVIDIA fees.

Consultants helping Fortune 500 firms negotiate enterprise software renewals said Koa gives Salesforce leverage to bundle reasoning into existing seats, potentially raising average revenue per account even if list prices stay flat.