Salesforce shipped Koa at Dreamforce on September 15. It is the company's first CRM reasoning model, built on NVIDIA Nemotron 3 Super and trained entirely on synthetic scenarios drawn from nearly three decades of CRM deployments. No customer data. Fourteen industries. Routing cases, updating opportunities, scheduling follow-ups, qualifying leads. In Salesforce's own benchmark, Koa matches or outperforms leading general-purpose models on CRM tasks with roughly three times fewer errors.
That is not a minor upgrade. That is Salesforce baking the AI layer into the product itself and running the weights on their own infrastructure. Three days earlier, on September 14, Zendesk announced specialized AI agents purpose-built for business workflows inside its platform. Same direction, different stack.
Two enterprise platforms. Five days. Both moving from "you wire up your AI" to "our AI ships with the product."
Let me explain what this actually means for the people who operate these platforms, because it is not what the press release says.
What Koa Actually Is (and Why It Matters Architecturally)
Before Koa, the standard playbook for AI in a CRM was: bring your own model, connect it via API or a no-code automation tool, point it at your records, and prompt your way to something useful. The model was external. Your CRM data was piped into it at query time. The integration was the work.
Koa flips this. Salesforce built their own model, trained it on synthetic CRM scenarios built from the patterns of how CRM work actually operates at scale, and deployed it as a native capability inside Agentforce. The model runs entirely within Salesforce's own infrastructure. No customer data crossed any trust boundary during training. The intelligence came from Salesforce constructing the training corpus themselves from thirty years of knowing what case routing, opportunity management, and lead qualification look like across fourteen industries.
The result is a model that understands case routing as a domain problem, not a generic text completion problem. It knows that a support case with certain attributes typically goes to a certain team because that pattern was baked into training. That is architecturally different from anything that came before in the enterprise CRM space.
Pilot customers include Formula 1, UChicago Medicine, and Xero. General availability is expected this winter in US regions.
The Zendesk Parallel: Both Happened the Same Week
Three days before Koa landed, Zendesk announced specialized AI agents purpose-built for business support workflows. The framing was agents that understand the structure of support conversations at a product level, not a generic language model grafted on top.
Two companies, five days apart, same direction: AI baked into the platform, no integration project required on your side.
This is not a coincidence. It is the leading edge of a wave where enterprise software vendors stop waiting for operators to wire up external AI and start shipping their own model layer as a platform-native default. The question for your team is no longer whether your CRM or support tool will have AI in it. It is what that AI will do with the configuration you give it.
What Vendor AI Is Good At
Koa is genuinely good at the generic version of CRM work. It has seen how routing, qualification, and case escalation look across thousands of simulated enterprise scenarios across fourteen industries. Zendesk's agents are good at the standard support handoff and clarification patterns that show up reliably in most support workflows.
For teams that still do this work manually, that is real value. Routing a case to the correct queue by hand, categorizing an inbound lead, looking up and updating an opportunity record: these are the repeating, low-judgment tasks that take time and produce errors when humans do them all day. Vendor AI takes that layer off the desk.
And that is worth celebrating. A well-configured Koa or Zendesk agent handling the generic routing and triage means the support team spends more time on the cases that actually require judgment, context, or a difficult conversation. The agent who would have spent thirty minutes per shift manually routing tickets can spend those thirty minutes resolving the cases that need a real person. That is the right use of the technology.
What Vendor AI Cannot Know
Here is where I want to slow down, because this is where the real configuration work lives and where most teams will underinvest.
Koa knows how a generic CRM workflow operates. It does not know that your company issues a credit automatically for any shipping delay past four days. It does not know that accounts flagged as Enterprise go directly to a named account manager regardless of case category. It does not know that any ticket mentioning "contract" or "legal" requires a supervisor to review before any response goes out. It does not know your escalation path. It does not know your retention SLA. It does not know the three product edge cases your team discovered in 2024 that changed how you handle billing disputes.
None of those rules are in Koa's training data. They are in your team's tribal knowledge, your documentation, your pinned Slack messages from two years ago that nobody has looked at since.
Zendesk's agents face exactly the same gap. The agent can escalate correctly when a customer says "I want to cancel" because that pattern appears reliably in support training data. It cannot know that your cancellation process requires a 48-hour cooling-off call with your retention specialist, who needs to review the last three billing notes before placing it. That is configuration. That is your work.
The gap between what vendor AI knows and what your business requires is not a failure of the vendor. It is the work. And it is increasingly the most important work an operator can do.
How the Operator's Job Changes
Before Koa, the question was: "Should we add AI to our CRM workflows?" That question required a real answer, a real project, and real integration engineering to act on.
After Koa, the AI is shipping whether you build it or not. It arrives with the platform upgrade. The question shifts.
What business rules does this model need to follow? These rules are what separate Koa behaving generically from Koa behaving correctly for your specific business. Get them out of people's heads and into documentation. The model will follow what you configure. If your configuration is vague, its behavior will be vague.
Where does a human stay in the loop? The cases where the rule is ambiguous, where the customer context is unusual, where a wrong AI action creates a bigger problem than it was solving: those checkpoints do not disappear when you turn on vendor AI. If anything, they become more important because the model is now handling the simple cases at scale, leaving the human team with the harder ones. The escalation path from AI to human needs to be explicit and tested.
What does success actually look like? Set your baseline before you turn it on. If Koa routes cases and you have no data on how often cases were misrouted by hand, you cannot tell whether Koa is better or worse in practice. You do not need precise numbers. You need something to compare against, even if it is directional.
The automation specialist who can answer those three questions and translate the answers into a vendor platform's configuration tools is now more valuable than the one who could build the integration layer from scratch. Because building the layer from scratch is no longer the hard part. The hard part is knowing what rules the model should run on, and making sure a person reviews the outcomes where it matters most.
Sources
Something in here worth pushing back on, or a configuration mistake you've already made with platform AI? Drop me a line.
Comments