What Koa Actually Does (and Doesn't Do)
Most AI announcements in the CRM space are about generating text: draft this email, summarize this call, write this follow-up. Koa is different. Built on NVIDIA's Nemotron 3 Super at 120 billion parameters, post-trained on synthetic scenarios drawn from 27 years of Salesforce CRM deployments across 14+ industries, it is a reasoning model — meaning its job is to determine which tools and actions are required to complete a multistep workflow, not to produce content.
That is a meaningful architectural shift. Lead qualification, case routing, opportunity updates — these are sequences of decisions, not documents. A model optimized to reason through those sequences, rather than one adapted from a general-purpose language model, has a structural advantage for CRM operations specifically. On Salesforce's own CRM Bench, Koa matches or exceeds leading general models with 3x fewer errors.
Take the benchmark with the usual caution — proprietary benchmarks measure what the vendor trained for. But the error-rate claim is the number operators should pressure-test in pilots.
The Deployment Reality Right Now
Koa is live internally at Salesforce and in active pilots with Formula 1, Xero, Baxter Credit Union, and UChicago Medicine. General availability is targeted for U.S. regions in winter 2026. That timeline is tight enough to matter for Q1 planning.
For non-tech companies already on Agentforce or evaluating it: the relevant move this quarter is requesting pilot access now. Pilot cohorts for purpose-built reasoning models typically shape the feature roadmap. Being in the room early has outsized returns compared to adopting at GA.
The zero-customer-data-in-training detail also matters for regulated industries — healthcare, financial services, credit unions. Koa was trained entirely on synthetic scenarios. That removes a significant compliance objection that has stalled AI adoption in those verticals.
What to Actually Do With This Information
If you run revenue operations on Salesforce, three concrete actions:
Audit your highest-error workflows first. Koa's design targets multistep sequences where mistakes compound — lead routing logic, case escalation rules, pipeline stage updates. Map where your team corrects CRM errors most frequently. Those are your pilot use cases.
Don't conflate the reasoning model with Agentforce broadly. Koa is the decision-making layer; Agentforce is the agent orchestration layer. Conflating them leads to vague pilots with no measurable outcome. Scope them separately.
Request access before GA. Salesforce's pilot program for Agentforce customers is the fastest path to Koa access. If your implementation partner hasn't raised this, raise it yourself.
The underlying bet Salesforce is making — that a domain-specific reasoning model outperforms a general one on CRM-specific multistep tasks — is directionally correct. The execution risk is in deployment complexity, not the model architecture. That is where your due diligence should focus.