At Dreamforce 2026 (Sept. 15), Salesforce and NVIDIA announced Koa—Salesforce’s first CRM reasoning model for Agentforce, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic corpus modeled on nearly three decades of CRM deployments. Unlike a chat wrapper on a general LLM, Koa is trained to reason through multi-step enterprise actions—updating opportunities, routing cases, scheduling follow-ups—and Salesforce reports it matches or beats leading models on those CRM actions with about three times fewer errors.
That matters for industrial CRM because configure-to-order quotes, spare-part eligibility, and aftermarket case work are multi-step tool chains, not single prompts. Koa’s training scenarios spanned more than 14 industries—including manufacturing—pairing personas with task sequences and tool calls. Post-training mixed supervised fine-tuning with Group Relative Policy Optimization (GRPO) via NVIDIA NeMo RL so the model learns to take the right next action, not just sound confident. Weights stay under Salesforce control; training and inference run inside its trust boundary, with no customer data in the training set.
Availability is pilot-now, GA later. Select customers can already run Koa inside Agentforce; Salesforce expects general availability in U.S. regions in Winter 2026. Named pilots stretch across sectors—Formula 1, Engine, Xero, UChicago Medicine, Baxter Credit Union, 1-800Accountant—while manufacturing sits inside the synthetic scenario set that taught the model how industry workflows actually branch. A companion Missionforce track brings Nemotron-based models into private clouds and air-gapped networks for regulated buyers who cannot send product or customer truth onto public infrastructure.
The industrial contrast is the ownership boundary. A CRM reasoning model that lives only inside a SaaS seat can compress lead-to-case cycles—and also lease the judgment layer that used to sit with your sellers and service engineers. Koa is not the Siemens Teamcenter-into-Agentforce integration story; it is the reasoning engine Agentforce itself will run on. One announcement connects engineering data into front-office workflows; this one specializes the model that sequences those workflows. For aftermarket-heavy manufacturers—where service revenue grows faster and carries higher margin than new equipment—getting multi-step case and quote reasoning right is the difference between a chatbot demo and a seller who can close a valid upgrade without pinging engineering twice.
Manufacturing teams that already stitch PLM product truth into sales and service will want the agent’s decision trail—what tool it called, which opportunity it mutated, which policy it applied—vaulted where engineering and commercial history already live. Renting the reasoning pass without owning the audit trail is how shops wake up unable to reconstruct why a quote was valid last quarter.
For N23D, the lesson is local-first ownership of the commercial spine. Putting a CRM reasoning model on multi-step sales and service work is the right product instinct—and so is keeping deal, case, and quote history under your control. Shops still need opportunity state, service cases, and configure-to-order decisions that survive when a cloud model seat or Agentforce entitlement changes. N23D ships as one native Rust binary: local-first CRM with vaulted BUS workflow history you own, not a reasoning pass that only works while the SaaS clock is ticking.