Salesforce Built a Model to Run Your Service Agents. It Has Never Met a Customer.
· CX Pulse
Salesforce's Koa will reason inside Agentforce from Winter 2026. Its training set was entirely synthetic, and its benchmark is Salesforce's own.
Salesforce and NVIDIA announced Koa on September 15. It's Salesforce's first CRM reasoning model, built by post-training NVIDIA's Nemotron 3 Super, and it's meant to be the brain behind Agentforce agents: the help agents, case routing and follow-up scheduling that sit between a company and the people contacting it. Pilot customers include UChicago Medicine, Baxter Credit Union, Xero and Formula 1. General availability is Winter 2026 in U.S. regions, with an open beta after that.
The coverage is running the vendor war, with Salesforce as customer, distributor and now competitor to the frontier labs. Anyone who owns a queue should read two sentences further down the release.
Every customer in the training set was written
The training corpus, in Salesforce's words, "consisted entirely of synthetic scenarios that reflect the reasoning, tool use, and decision-making skills Agentforce agents perform across CRM workflows." Each scenario "paired a persona with specific tasks and mapped the exact sequence of actions needed to complete them," across 14 or more industries.
Read that as a privacy officer and it's the best news of the week. Salesforce controls the weights, runs the model inside its own infrastructure, and says "no customer data ever crosses the trust boundary during training or inference." That's a real design choice and a good one. Nobody's support transcripts became somebody else's training run.
Read it as a support leader and it's the same sentence with a different weight. The model that will decide where a case goes has learned from personas that were written for it, on tasks with a mapped correct path. Neither the release nor the product page says how many of those scenarios go wrong on purpose: the customer who has already been through the bot twice, the message carrying two problems, the policy exception that the process has no branch for. Those are most of what an escalation queue is made of.
The benchmark is the vendor's
Koa's numbers come from CRM Bench, which Salesforce describes as "a benchmark built from real-world tasks like updating an opportunity, routing a case, and scheduling a follow-up." Salesforce calls itself customer zero for it. Against "today's default general intelligence models" the release claims three times fewer errors, 11% more precision "at calling the right action so agents take fewer wrong turns," 2.1 times more reliable recall of customer context, and 15% better memory in long back-and-forth conversations "so customers don't have to repeat themselves."
Those last two are customer experience claims, and they're exactly the ones a service organisation would want before switching. They were produced in the same building as the model, on tasks the same building chose, with synthetic conversations. None of that makes them wrong. It makes them the vendor's evidence rather than yours.
Where this lands at scale
If you run Agentforce, a model selection decision is arriving that used to belong to your platform team. Today the reasoning behind your agents is a frontier model you evaluated, with a contract and a rate card you negotiated. From Winter 2026 there's a Salesforce-hosted option that the vendor has already benchmarked on your behalf. The product page says you choose it per deployment, which means the choice is real and so is the default it will become once the pricing favours it.
Three things follow. Legal gets an easier conversation, because the trust boundary claim answers the data-residency question that has stalled more than one agent rollout. Your QA and evaluation function gets a harder one, because CRM Bench is the wrong test for your cases and the only published one. And your escalation design needs to be checked against a model that learned routing from scenarios with a known right answer. The cases that hurt are the ones without one.
The practical move is replay. Take the last fifty cases that escalated to a person, run them through the pilot next to the model you use now, and compare where each one went. That's an afternoon, it needs no vendor, and it produces the one number the release can't: how it behaves on your customers.
The smaller operation gets the same shift, quieter
A smaller business is unlikely to be evaluating reasoning models. It rents a help desk or a chat tool, and that tool will do what Salesforce just did, sooner or later: swap the model underneath for one it owns. There won't be a CRM Bench for it. There'll be a line in a changelog, if that.
The question is the same one and it fits in a sentence: which model is answering our customers, and what did it learn from? A vendor that can answer that in one sentence is telling you something. So is one that can't.
The line to keep
Koa's scores come from customers Salesforce invented. Before it meets yours, hand it fifty of the real ones and see where it sends them.