_edited.png){kind=small}
What Is Koa? Salesforce’s First-Party CRM Reasoning Model, Explained
In 50 words: Koa is Salesforce’s first CRM-specific reasoning model, built by post-training NVIDIA’s Nemotron 3 Super on synthetic data modeled on 27 years of CRM deployments. It runs inside the Agentforce Reasoning Engine alongside Claude and Gemini, and is currently pilot-only, with general availability expected Winter 2026 for U.S. regions.
At Dreamforce 2026, alongside the AIforce interface layer, Salesforce and NVIDIA announced Koa - Salesforce’s first reasoning model built specifically for CRM work. It’s a notable move: Salesforce already has Claude powering reasoning inside Agentforce (via Claudeforce) and Gemini now generally available in the Agentforce Reasoning Engine too. Koa is Salesforce building its own model on top of that, not replacing the others.
If you’re running Agentforce, here’s what Koa actually is, how it was built, what it’s being used for today, and - just as important - what’s still pilot-only and not something to plan around yet.
Table of Contents
What is Koa, exactly?
Koa is a reasoning model purpose-built for CRM tasks - the kind of multi-step work Agentforce agents do every day: qualifying a lead, routing a case, updating an opportunity, scheduling a follow-up. Rather than being a general-purpose model adapted for business use, Koa was built specifically around how CRM work actually happens.
It sits inside the Agentforce Reasoning Engine as a model option, alongside Claude and Gemini - not as a separate product you install, and not as a replacement for either.
How Koa was actually built
Two details matter more than the marketing language:
The foundation is NVIDIA Nemotron 3 Super, an open model. Salesforce didn’t build a model from scratch - it post-trained an existing open foundation model for CRM-specific reasoning.
The training data is synthetic, not customer data. Salesforce says no customer information was used to train Koa. Instead, training scenarios were modeled on what it describes as 27 years of Salesforce CRM deployment knowledge, spanning more than 14 industries (manufacturing, financial services, healthcare, travel, and others), with each scenario built around a persona, a sequence of actions, and the tool calls needed to complete a task.
Salesforce also states that it controls the model weights and runs Koa inside its own infrastructure, so customer data doesn’t cross the trust boundary during inference. If your org has strict data-residency or governance requirements, that architecture detail is worth confirming directly with your Salesforce contact rather than taking the summary version at face value - “runs inside our infrastructure” is a meaningfully different claim from “runs inside your infrastructure,” and the distinction matters for compliance conversations.
What Koa is being used for right now
Koa isn’t purely theoretical - it’s already live in a few places:
Inside Salesforce itself, powering a Slack-based agent that helps Salesforce employees find information and complete everyday tasks.
In customer pilots with a named group of early adopters, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero.
Extending into Missionforce, Salesforce’s unit for government and regulated organizations, where Nemotron-based models are being positioned for private-cloud and air-gapped deployments - relevant if your org operates in a regulated or public-sector environment, less relevant otherwise.
Koa vs. Claude vs. Gemini: why build a third model at all?
Worth asking directly, since Salesforce already had two frontier models inside Agentforce before Koa existed. Salesforce’s own framing is that general-purpose models are strong at broad reasoning and language, while Koa is narrower by design - trained specifically on CRM process knowledge, workflows, and tool sequences, rather than general-purpose tasks like content generation.
Whether that specialization is worth it in practice - and when you’d actually choose Koa over Claude or Gemini for a given Agentforce use case - is a big enough question that it deserves its own post. We’ll cover that comparison directly once more pilot detail is public.
What’s live now vs. what’s still coming
Capability | Status as of this blog |
Koa inside the Agentforce Reasoning Engine | Pilot customers only |
Koa general availability | Expected Winter 2026, U.S. regions only at launch |
Koa in EU or APAC regions | Not confirmed for 2026 |
Missionforce / Nemotron for regulated industries | Generally available now (Missionforce Operations); post-trained NVIDIA models to select customers starting October 2026 |
Public pricing for Koa specifically | Not announced |
As with Claudeforce, don’t build a 2026 budget or roadmap around Koa yet. Pilot access and a named group of early customers is a meaningful signal, but it’s not the same as general availability.
A word on Salesforce’s own benchmark claim
Salesforce states that Koa “matches or exceeds leading model performance on CRM actions with three times fewer errors,” based on its own CRM Benchmark. Worth being precise about what that claim is and isn’t: it’s a vendor-run result, measured against a vendor-defined task set, without a published methodology you can independently replicate against your own org’s workflows. That doesn’t make it untrue - but it’s not independently verified either, and it’s reasonable to wait for third-party benchmarking or your own pilot results before treating “3x fewer errors” as a settled fact.
What this means if you’re running Agentforce
Nothing changes in your org today unless you’re actively part of the pilot. Koa is additive to the Reasoning Engine, not a required upgrade.
If you’re US-based and already deep into Agentforce, it’s reasonable to ask your Salesforce account team about pilot eligibility - particularly if your use cases are CRM-process-heavy (lead qualification, case routing) rather than open-ended reasoning or content generation.
If you’re outside the US, or have EU/APAC data residency requirements, Koa isn’t a 2026 planning item - treat this as background context for now, not a roadmap input.
Model choice inside Agentforce is becoming a real decision, not a fixed default. Claude, Gemini, and soon Koa are all options in the same Reasoning Engine - which one fits best will depend on the specific task, not a one-size-fits-all answer
Don't expect ROI numbers for Koa specifically yet. It's too early - no independent benchmarking, no pricing, pilot-only. If you're building a business case for Agentforce investment more broadly, lean on what's already measurable rather than waiting on Koa: see our breakdown of how to measure Agentforce ROI for a framework that works regardless of which model is doing the reasoning.
FAQ's
Is Koa replacing Claude or Gemini inside Agentforce?
No. Koa is a third reasoning-model option inside the Agentforce Reasoning Engine, not a replacement for either. Salesforce’s stated direction is a multi-model approach, letting different models handle different parts of the work.
Was my Salesforce data used to train Koa?
Salesforce states that no customer data was used in training - the training set was synthetic, built from modeled CRM scenarios rather than real customer records.
When will Koa be generally available?
Salesforce has stated general availability is expected Winter 2026, limited to U.S. regions at launch. No date has been confirmed for other regions.
How is Koa different from a general-purpose model like Claude or Gemini?
Koa was trained specifically on CRM process knowledge, workflows, and tool-use sequences, rather than broad general-purpose reasoning and language tasks. Whether that specialization outperforms a general-purpose model for a specific use case is still an open, use-case-dependent question - one we’ll dig into in a follow-up comparison post.
Where to start
Koa is early - pilot-only, US-first, and still leaning on a vendor-run benchmark rather than independent results. That’s exactly the stage where getting clear on your own Agentforce reasoning-model strategy matters more than rushing to add another pilot.
Comments