Salesforce hopes its new AI tailored for CRM will be a boon for marketers and a moat against generic models from companies like OpenAI, Anthropic and Google.
At Dreamforce, the company introduced Koa, its first CRM reasoning model, developed with Nvidia to manage complex, multi-step sales, service and customer workflows. Instead of simply answering questions or generating content, Koa is designed to determine what tools and actions are needed to complete a job.
For marketers, this distinction is important since AI agents do more operational work. Writing an email is relatively simple. Deciding whether a lead is eligible, checking their account history, applying business rules, updating the CRM, and triggering the appropriate sales or nurturing workflow requires the model to understand how the business operates.
Koa is Salesforce’s attempt to put some of that operational knowledge directly into the model.
“The most valuable thing Salesforce has built is not our platform, but the accumulated knowledge of how enterprise business actually works,” Marc Benioff, president and CEO of Salesforce, said in a statement. “With Koa, knowledge is fed into the model itself. We’ve trained a reasoning engine that understands the structure of a deal, the lifecycle of a service case, and workflows that vary across industries. It’s a different kind of intelligence.”
Koa learns how CRM work is done
Salesforce built Koa on Nvidia’s Nemotron 3 Super and trained it using a proprietary synthetic dataset modeled after the company’s decades-long experience with CRM implementations. No customer data was used to train the model.
Its training scenarios cover more than 14 industries and recreate workflows such as lead generation, opportunity qualification, and service case resolution. Each scenario maps the actions and tool calls needed to complete a task, teaching Koa to take the steps necessary to achieve a result rather than simply producing a response.
Salesforce also checks model weights and runs inference within its own infrastructure, so customer data doesn’t cross the trust boundary when Koa is used. This is especially relevant when an agent goes beyond simply generating marketing assets and starts accessing customer information, editing records, or triggering workflows.
10x your SEO with Semrush for business.
The most powerful SEO platform in the world, created specifically for businesses.
Request a demonstration
The company says it is “moving into customer pilots” with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. General availability in US regions is expected this winter.
Marketers may need more than one AI model
Koa is not replacing the generic templates already used by Salesforce customers. Salesforce is expanding these choices while developing its own specialized model.
Agentforce customers can use Gemini templates through Salesforce’s Google Cloud partnership. An expanded AWS integration adds templates available through Amazon Bedrock, including templates from Anthropic, Nvidia, and OpenAI.
This points to a different way of thinking about AI in the martech stack. Instead of choosing one model to handle everything, companies could use different models for different tasks.
A generic template might be a good choice for campaign brainstorming, research analysis, or content writing. A specialized model like Koa could handle jobs that require knowledge of CRM processes, business rules, customer records, and the sequence of actions needed to complete a workflow.
For marketing operations teams, model selection could become another orchestration decision. Cost, accuracy, speed, access to customer data, governance requirements, and consequences of a failure could determine which model gets which job.
AIforce gives these models a place to run
Koa also completes another piece of the Salesforce strategy already underway.
Last month, the company’s partnership with Claudeforce made the Salesforce interface optional by putting data, business logic, and actions inside Claude. AIforce, formally introduced at Dreamforce, expands this approach across multiple AI interfaces.
AIforce makes Salesforce data, workflows, semantics, permissions, security, governance, and actions available via API so they can be used in other AI environments.
The company’s expanded partnerships with AWS and Google Cloud further extend that architecture. Salesforce capabilities can emerge within Amazon Quick and Gemini Enterprise, while models and agents from those ecosystems can work with Salesforce data and workflows.
For marketers, the interface where the work gets done and the technology that gets the work done are becoming separate choices. Salesforce can provide customer context and business rules, Claude or Gemini can provide general reasoning, and Koa can handle the work where specialized CRM knowledge is most useful.
The same change is reaching customers
Salesforce’s expanded partnership with Google shows what this separation can look like on the customer side.
Starting this fall, Commerce Cloud merchants will be able to surface products in Google Search, including AI and Gemini mode. Customers can complete purchases via Google’s Universal Commerce protocol while payments, compliance and order management remain on the merchant’s Commerce Cloud infrastructure.
The customer can interact with Google while Salesforce operates underneath the experience.
This is a significant shift for marketers accustomed to thinking about customer journeys in terms of websites, apps, e-commerce stores and other brand-controlled destinations. As AI interfaces become another place where discovery and transactions happen, the technology that determines what customers see and what happens next is largely invisible to them.
Likewise, marketers may do less directly with the applications in their stack since agents handle most of the navigation between them.
This gives more weight to what lies beneath the interface: accurate customer data, consistent definitions, permissions, business rules, APIs and governance. It also places more emphasis on choosing the right reasoning for the job.
In a world where artificial intelligence makes information cheap and access to expertise easier, Salesforce is betting that expertise is harder to copy.
