Salesforce today published an interview in which Chief Data Officer Michael Andrew argued that enterprise customers should be able to switch between frontier, open and custom AI models, including Koa, the company's first CRM reasoning model, which it says was built with NVIDIA and trained in simulations drawn from 27 years of running CRM workflows.

In Short

Salesforce posted an interview with its data chief explaining why its software lets companies pick which AI model runs each task, from Anthropic's Claude to OpenAI's ChatGPT to Salesforce's own new model, Koa. That matters to the marketing and sales teams that live inside Salesforce and Slack, because the model behind an agent shapes how fast it answers, how much it costs and how reliable it is. The piece describes a philosophy rather than a product change, so nothing new switches on for customers today.

What Salesforce published

The article, titled "Why Salesforce Lets You Pick the Right AI Model for the Job," appeared on the Salesforce Newsroom on October 13, 2026, under the byline of Cody Snell. It is filed under the "Agentic Enterprise" topic and is labelled a four-minute read. The page carries a video interview with Andrew and a condensed transcript of three questions and answers. It is company-published material, written by Salesforce staff about Salesforce products, and every claim about Koa and Slack in it is a vendor claim that the page does not support with benchmarks, prices or customer data.

The framing is set out in the opening paragraphs. "Not every job calls for the same AI model," according to Salesforce. Some tasks, "like routing a case or updating an opportunity, need precise multistep reasoning," while others "might need to be done particularly quickly or cost-effectively." Rather than "forcing customers to stick with a single tool," the company says it gives them "flexibility to use whichever model works best for any given use case, whether it's a frontier, open, or custom model like Koa."

Koa itself gets a single paragraph. Salesforce describes it as its "first CRM reasoning model," "purpose-built for the work enterprises do," "built in collaboration with NVIDIA and running inside the Salesforce trust boundary." The training claim is the one hard number in the piece: Koa is, according to Salesforce, "trained in simulations drawn from 27 years of Salesforce experience running CRM workflows." The page also states that Koa is "orchestrated alongside frontier models rather than in place of them."

What the page does not say is just as telling. There is no parameter count, no benchmark score, no price per token, no availability date and no list of which Agentforce surfaces Koa runs in. The Newsroom piece links to a separate post titled "Why We Post-Trained Our Own Reasoning Model" and to another titled "How we cut inference spend by right-sizing our models," but those were not part of the material reviewed for this article, and their contents are not reported here.

Andrew on deterministic and probabilistic systems

Most of the transcript is Andrew explaining, in plain terms, why AI agents behave differently from the software businesses already run on. His analogy uses a television. "When you press the power button on your TV, it's always going to come on. That's deterministic," he said. "You tell your child, 'Hey, will you turn off the TV?' Maybe they do. Maybe they don't. Agents are the same way."

He extended the comparison to CRM records. "If you open a case about your customer, if a customer files an issue, every time you press that button, you're going to open a case," Andrew said. "Businesses have always run on deterministic systems, which guarantee you're going to get the same thing every time, everywhere in the world." AI, by contrast, behaves like a new hire. "You can train them, you can guide them, but they're not always going to be reliable," he said, describing the models as "basically big artificial brains."

That admission is notable coming from a vendor selling agents. It concedes the core risk that buyers have raised throughout 2026: an agent sitting on top of a system of record can produce a different answer to the same request. PPC Land examined the same tension in January, when analysts and executives debated why deterministic systems still win in AI's probabilistic world, concluding that data quality matters more, not less, once agents start acting on records that humans used to reconcile by hand. Andrew's argument lands in a similar place. The deterministic layer stays, and the probabilistic layer is plugged into it.

Why any model, according to Salesforce

The second question asked why a company's deterministic systems should connect to any model. Andrew's answer was that the market moves too quickly to bet on one supplier. "Different models have different strengths and weaknesses, and they're changing all the time," he said. "All the companies are competing to produce different models, and of course, you have big frontier foundation models produced from partners like Anthropic and OpenAI, but you also have open-source models. You have custom models, like we just released a model partnering with NVIDIA that's customized from Salesforce."

The stated selection criteria are use case, cost and value. "So we believe in giving our customers choice, that you can pick the model that's right for your particular use case and cost and value, but knowing that your data is secure, that you can run your process, and you have the flexibility to keep evolving and changing what model you use for the use case," Andrew said. "And it's based on that speed of resolution and cost."

He closed the answer with a metaphor. "AI is a fuel; it's more fuel into your engine, and we want to let that fuel come from anywhere, whether it's Anthropic and Claude, OpenAI and ChatGPT, or any of the many partners that we're announcing for AIforce." The page does not define AIforce, list its partners or say when those announcements happen.

How that sits with Claudeforce

The "choice" message lands six weeks after Salesforce tied itself more closely to one supplier. On August 26, 2026, Salesforce and Anthropic announced an expanded partnership branded Claudeforce, under which Salesforce put 37 prebuilt sales skills inside Claude for pilot customers. The same arrangement made Claude the default model for Agentforce Vibes and Agentforce Coworker, a reasoning model for the Atlas Reasoning Engine, and the default model for Slack, powering Slackbot. Neither company disclosed contract value, exclusivity or minimum commitments.

Defaults and choice are not contradictory, but they are not the same thing either. A default model is what most users get unless an administrator changes it, and in practice defaults tend to stick. The interview does not say how customers switch models in Slack, whether Koa is offered there, or whether switching changes price. Those are the operational details that would show how much choice exists in practice.

Koa, and the other Koa

Marketers who buy programmatic media will notice a naming overlap. The Trade Desk already uses Koa as the name for the AI layer inside its Kokai platform, and in August the company said its upgraded Koa Optimizations cut CPAs 32% across 62 campaigns in a comparison between model versions. The two products are unrelated. Salesforce's Koa is a reasoning model for CRM work; The Trade Desk's Koa is an assistant and optimization layer for media buying. The Salesforce page does not mention the overlap.

Custom models built on NVIDIA

Salesforce is not the only marketing technology company to train its own model with NVIDIA. On October 9, PPC Land reported that Zeta Global plans to add NVIDIA Nemotron-based models to its AthenaOS beta, describing a model still in development whose accuracy Zeta said was on par with leading models without naming a benchmark. Zeta's Athena had gone generally available on March 24, 2026, built on OpenAI models. The pattern is similar in both cases: a platform company that began on a general-purpose frontier model adds a domain-tuned model of its own, while keeping the general-purpose ones available.

The economics behind that pattern are about inference, the cost of running a model each time an agent answers a question or takes an action. Frontier models are capable but expensive per call. A smaller model trained for a narrow domain can, in principle, handle routine tasks such as case routing at lower cost, leaving the frontier model for harder reasoning. Andrew's reference to "speed of resolution and cost" points at this trade-off, and the Newsroom's own link to a post about right-sizing models suggests the company has examined it internally. The interview itself publishes no cost comparison.

Open models are part of the same calculation. On October 8, PPC Land reported that Ecosia dropped Mistral AI for open models hosted by a third party, with chief executive Christian Kroll saying costs had roughly halved, a figure the company did not substantiate with a baseline. Supplier switching of that kind is exactly what a multi-model platform is meant to make easier, and exactly what a deep partnership with one supplier can make harder.

Multiplayer AI in Slack

The third question turned to Slack. Salesforce's framing, in the introduction, is that "multiplayer AI" replaces "one person, one screen, and one agent in a private tab" with "people and agents work out in the open on shared context so the whole team learns from every exchange." The page adds that this works "no matter what model it's based on."

Andrew described the internal picture at Salesforce. "We use Slack every day. I think 85,000 employees are talking with their colleagues, talking about their projects, talking about their accounts," he said. The figure is offered with a hedge ("I think") and is not a disclosed headcount. "You don't want to have to go somewhere else. Now you can just talk to your agent and say, 'Hey, agent, help me out.' Suddenly, boom! The agent's in the chat, and it's your digital colleague right there in the flow of work with you."

On permissions, he said the agent's reach is bounded. "And importantly, in Slack, it also comes with security and permissions. It's to your channels, to what you want to allow it to do," Andrew said. He described the selling point as teamwork: "Now you have an agent that's working with your team together, and it has all the shared context. It can know what you were talking about with that deal or that customer issue."

Shared context is the point of the design, and it is also the part that raises questions for regulated teams. An agent reading a channel sees whatever is said there, including customer details that appear in passing. The interview says the agent is limited to the channels and actions a team allows, but it does not describe retention, whether channel content is used to improve models, or how the "trust boundary" applies to third-party frontier models as opposed to Koa. Slack has been central to Salesforce's agent plans for some time; when Salesforce agreed to acquire Doti in November 2025, it named Slack as the conversational interface for an agentic search layer across its products.

Why this matters for marketers

For marketing and sales teams, Salesforce is the CRM where leads, opportunities and service cases live, and the agents now attached to it decide how those records get updated. Which model sits behind an agent affects three things marketers notice: response time, the reliability of the action taken, and the per-interaction cost that ends up in a consumption bill. Salesforce moved Agentforce from per-seat licensing toward consumption-based pricing, a shift PPC Land covered when Salesforce Headless 360 opened its platform to AI agents in April. Under consumption pricing, model selection becomes a budget line.

Reliability remains the open issue. Salesforce's own CRMArena-Pro research, cited in PPC Land's coverage of the company's September agent releases, found enterprise agents succeeded in about 58% of single-turn tasks and 35% of multi-turn tasks. That reporting on how Salesforce agents gained a runtime that pursues goals over weeks noted that the company had not published an updated long-horizon benchmark. Koa is pitched as the model for "precise multistep reasoning" on tasks such as updating an opportunity, which is the multi-turn category where those success rates were lowest. Without a published score, there is no way to tell from today's material whether Koa moves that number.

Wider adoption data counsels caution too. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing cost, unclear value and weak risk controls. A pitch that leads with model choice and cost flexibility speaks to the first of those concerns. It does not, on its own, address the others.

There is also a market structure question. Andrew named Anthropic and OpenAI as the frontier partners, alongside open-source models and Salesforce's own. For agentic AI buyers, the practical test of "choice" is whether a team can move a live workflow from one model to another without rebuilding it, and whether the price of doing so is visible in advance. The interview asserts that flexibility exists. It does not show how it works.

Timeline

Summary

Who: Salesforce, through Chief Data Officer Michael Andrew, interviewed by Cody Snell of the Salesforce Newsroom. Anthropic, OpenAI and NVIDIA are named as model partners.

What: A vendor-published interview arguing that Salesforce customers should pick the AI model for each task, whether a frontier model such as Claude or ChatGPT, an open-source model or Koa, Salesforce's first CRM reasoning model, which the company says was built with NVIDIA and trained in simulations drawn from 27 years of CRM workflows. Andrew also described "multiplayer AI" in Slack, where agents work in shared channels with team context.

When: The article was published on the Salesforce Newsroom on October 13, 2026.

Where: The Salesforce Newsroom, in its Agentic Enterprise section, with a video and a transcript. The products discussed are Salesforce's CRM platform, Agentforce and Slack.

Why: Salesforce says models differ in strengths, speed and cost and change constantly, so customers need flexibility. For marketers, the model behind a CRM agent affects reliability and consumption costs, yet the interview publishes no benchmarks, prices or switching details, and it comes six weeks after Salesforce made Claude the default model in Slack.