Salesforce Chief Data Officer argues enterprises need “choice” across AI models, not one-size-fits-all

Salesforce is making the case that enterprise AI should not be tied to a single large language model. In an interview, Salesforce Chief Data Officer Michael Andrew argued that different business tasks call for different kinds of models, because some workflows require precise multistep reasoning while others need fast, cost-effective responses. The company’s message is that customers should be able to select the model that best matches the job and then keep adapting as AI models evolve.

Andrew framed the shift in terms of software reliability. He compared deterministic business systems to familiar repeatable behaviors, like a TV turning on when a power button is pressed, and contrasted that with AI’s nondeterministic nature, where outcomes can vary even when the user request is similar. He did not suggest that AI becomes fully deterministic, but he said enterprises need tools that balance reliability, speed of resolution, and cost.

At the center of Salesforce’s strategy is Koa, which Salesforce describes as its first CRM reasoning model built for the work enterprises do. Salesforce says Koa was developed in collaboration with NVIDIA, runs inside Salesforce’s “trust boundary,” and was trained using simulations drawn from 27 years of Salesforce experience running CRM workflows. Andrew characterized Koa as “purpose-built,” and said the model fits into a broader system rather than replacing other model options.

Salesforce’s model orchestration, according to Andrew, is designed to work alongside frontier models instead of swapping them out. The practical point, he said, is flexibility. “We believe in giving our customers choice,” Andrew said, emphasizing that enterprises should be able to pick the model that fits their particular use case and constraints and also keep changing what they use over time.

The company’s flexibility message extends beyond CRM workflows into team collaboration inside Slack. Andrew described a “multiplayer” agent approach, where instead of a single agent interacting with one person in a private tab, people and agents work with shared context in open collaboration. In his description, the agent is not operating in isolation, but alongside the team, with access to the shared discussion history so it can respond in line with what the group has been discussing.

Andrew also used Slack as an example of why customer-facing AI needs to stay inside everyday tools rather than pushing users into separate applications. He said Salesforce has about 85,000 employees using Slack to talk with colleagues about projects and accounts, and he argued that people do not want to leave their work environment to interact with an AI assistant. He added that the agent experience includes security and permissions, tied to what a user allows and what is available in relevant Slack channels.

In his comments about the wider model landscape, Andrew pointed to the competitive split between frontier foundation models from partners and other options such as open-source and custom models. He referenced models produced by companies including Anthropic and OpenAI, as well as open-source models, and he also described Salesforce’s custom model approach, including a model he said was recently released in partnership with NVIDIA and customized from Salesforce.

Andrew linked the “choice” approach to a broader partner strategy. He said AI is “fuel” for business processes, and Salesforce wants that fuel to come from multiple sources, including partners associated with Claude and ChatGPT, as well as additional partners it said it is announcing for AIforce. That framing suggests Salesforce wants to avoid locking customers into one model provider while still offering enterprise controls around data and workflows.

What Salesforce did not provide in the interview are details that enterprises typically ask for when evaluating model choice, such as benchmark performance numbers for specific workflows, end-to-end latency targets, or guidance on the circumstances where one model is likely to outperform another in production. The company also did not disclose pricing or contractual implications in the remarks, including how model selection may affect operating costs in different environments.

For customers and competitors, the immediate takeaway is that Salesforce is positioning AI model choice as a product principle, not just a technical backend feature. What to watch next is whether Salesforce expands the range of purpose-built or custom models available for CRM workflows, how it integrates additional partner models under AIforce, and whether the “shared context” multiplayer agent experience in Slack delivers measurable improvements in task resolution and user adoption across enterprise teams.