With Koa, Salesforce introduces an AI model trained specifically for multi-step CRM processes. Instead of writing emails, Koa works through structured chains of actions, such as updating opportunities or assigning cases. In Agentforce it replaces standard LLMs for these routine tasks. That lowers costs and makes autonomous agents more reliable, because generic language models often fail at the exact order of system updates.
Training on synthetic data
Koa is based on NVIDIA’s Nemotron 3 Super. Salesforce post-trained this base model, but entirely without real customer data. A synthetic data set simulated realistic business scenarios from more than 14 industries, including manufacturing, financial services and healthcare.
The training via supervised fine-tuning and Group Relative Policy Optimization (GRPO) followed a fixed pattern: a persona received a task, and the model learned the exact sequence of the required actions and tool calls. Because Salesforce controls the model weights and Koa runs in its own infrastructure, your data never leaves the Salesforce trust boundary during inference.
The AI gateway saves token costs
Most CRM tasks are recurring and clearly defined. That is why Salesforce uses an AI gateway in Agentforce to distribute tasks. When an agent starts a routine task, the gateway routes it to the cheaper Koa model. Only when a request requires complex text analysis or genuine judgement does the system hand over to large frontier models such as Claude or ChatGPT.
How does Salesforce’s business model work in the AI era?
Traditionally, Salesforce sells cloud-based software subscriptions for sales, customer service and marketing. With Agentforce and models like Koa, this shifts towards usage-based AI capacity. Revenue comes from the successful execution of business processes, no longer just from user licences. Already in the second quarter of fiscal year 2027, annual recurring revenue (ARR) for Agentforce exceeded 1.5 billion US dollars. Value moves away from pure token prices towards intelligent routing between inexpensive models like Koa and expensive models like Claude, plus secure execution in Salesforce’s own infrastructure.
Three times fewer errors, with caveats
According to Salesforce’s in-house CRM Benchmark, Koa makes three times fewer errors on CRM actions than leading standard models. Since Salesforce defined the benchmark itself, that is a strong indicator but no guarantee for every individual setup.
In addition, the synthetic training data only covers standard patterns. The edge cases that have grown historically in your company are missing. Koa’s cost advantages also only apply when the model completes tasks. If it has to abort and escalate to Claude too often, the savings melt away. How high that escalation rate turns out will only become clear in broad real-world use.
Your Agentforce roadmap
Large enterprises such as Siemens already use Agentforce intensively and qualify more than 2,500 inbound leads from 132 countries every month. The DACH mid-market still has time to prepare: general availability of Koa in the US regions is planned for winter 2026. It will take longer until European data centres follow.
Use this time to make your CRM processes agent-ready. Koa can only derive logical steps if the underlying tools and interfaces in Salesforce are cleanly documented. Check your workflows for clear input and output parameters.
If you want to know which processes are suitable for Agentforce and how to prepare your data architecture, talk to our consultants at Aviando.
