Salesforce is betting that decades of CRM knowledge can give its artificial intelligence an edge that general-purpose models cannot easily copy. At Dreamforce, the company introduced Koa, its first CRM reasoning model, developed with Nvidia to work through complex, multistep sales, service and customer workflows.
Unlike models built primarily to answer questions or generate content, Koa is designed to determine which tools and actions are needed to complete a job. For marketers, that is a meaningful shift as AI agents take on more operational responsibilities.
Why operational reasoning is different
Writing an email is relatively straightforward. Deciding whether a lead qualifies, checking its account history, applying company rules, updating the CRM and triggering the right nurture or sales workflow is a different kind of task. Koa is Salesforce’s attempt to place that operational knowledge inside the model itself.
“The most valuable thing Salesforce has built isn’t our platform — it’s the accumulated knowledge of how enterprise business actually works,” Salesforce chair and CEO Marc Benioff said in a statement.
Salesforce says Koa was built on Nvidia’s Nemotron 3 Super and post-trained using a proprietary synthetic dataset modelled on the company’s experience with CRM deployments. No customer data was used. Training scenarios cover more than 14 industries and recreate workflows such as generating leads, qualifying opportunities and resolving service cases, mapping the actions and tool calls required to finish each task.
A multi-model approach to martech AI
Koa is not replacing the general-purpose models already in use. Salesforce is expanding model choices at the same time. Agentforce customers can use Gemini models through the Google Cloud partnership, while an expanded AWS integration adds models available through Amazon Bedrock, including Anthropic, Nvidia and OpenAI.
The practical implication for marketing operations teams is that model selection could become another orchestration decision:
- Use a general-purpose model for brainstorming, research and content drafting.
- Use a specialised model such as Koa for work that needs CRM process knowledge, customer records, company rules and action sequences.
- Weigh cost, accuracy, speed, data access, governance and error consequences when assigning each AI task.
What changes under the interface
Koa also fits into Salesforce’s wider AIforce strategy, which exposes Salesforce data, workflows, semantics, permissions, security, governance and actions through APIs for use in other AI environments. The result is that the interface where work happens and the technology doing the work become separate choices.
The customer side shows a similar shift. Starting this fall, Commerce Cloud merchants will be able to surface products in Google Search, including AI Mode and Gemini, with payments, compliance and order management remaining on Commerce Cloud infrastructure. Customers may interact with Google while Salesforce operates underneath.
For marketers, that puts more weight on accurate customer data, consistent definitions, permissions, business rules, APIs and governance—and on choosing the right reasoning model for each job.
Customer pilots are underway with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. General availability in U.S. regions is expected this winter.
Source: MarTech
