Right now, Agentforce agents make decisions with general-purpose AI models. Those models are clever, but they were never trained for the specific work of a CRM — routing a support case, qualifying a lead, deciding which tool to call next.
At Dreamforce 2026, Salesforce announced Koa, its first AI reasoning model built specifically for CRM work. Here is what it is, how it was trained, what the test numbers actually show, and what it means for your org.
Inside an AI agent, the reasoning model is the part that works out what to do: which steps to take, in what order, and which tools to reach for. Everything else in the agent executes; the reasoning model plans.
Koa is that planner, trained on CRM work itself. You will never "open Koa" — it runs behind the scenes, and admins pick it the same way they pick any other model for an agent.
Salesforce likes to compare it to onboarding a new employee. A bright new hire still has to learn your company's rules before they can be trusted with real work. A general-purpose model faces the same problem on every task: it reasons everything out from scratch, and it may not do it the same way twice. CRM work, though, is mostly repetitive and rule-based — the same kinds of decisions, over and over. Salesforce's bet is that a model trained on exactly that work will be steadier.
There is a safety angle here too. An agent that says "I can't do that" is far safer than one that pretends it did the job. A reasoning model trained on CRM rules should know its limits better.
Salesforce says Koa makes three times fewer errors on CRM tasks in its own benchmark. What it doesn't say is which models it was compared against.
Its own research paper tells a more mixed story:
| Test (Salesforce internal) | Koa | GPT-4.1 | GPT-5.5 |
|---|---|---|---|
| Multi-step CRM tasks completed | 87% | 82% | 91% | | Single-step CRM score | 0.87 | 0.81 | 0.90 |
In plain terms: Koa beats GPT-4.1 on CRM tasks, but trails GPT-5.5 slightly. And the "three times fewer errors" figure doesn't appear anywhere in the paper itself.
All of these results come from Salesforce testing its own model. There are no independent tests yet, so treat the numbers as a direction, not a verdict.
Was Koa trained on customer data? No. Salesforce says it used simulated and public data only.
Is Koa better than GPT or Claude? It depends on the task. In Salesforce's own paper it beats GPT-4.1 but sits just behind GPT-5.5.
Does Koa replace the other models? No. It's one more option in the list — other providers stay available, and admins choose per agent.
Koa is still in pilot, so most orgs can't switch to it yet. That waiting time is useful, though: the smart move is to build a small test set from one of your own agents now — a handful of real tasks with known-good outcomes — so you can compare Koa fairly the day it becomes available.
If you want the wider picture on where Salesforce's AI is heading, start with [how the Anthropic partnership changes things for Salesforce teams](https://way2force.com/claudeforce-in-salesforce/) and [the Agentforce Coworker announcement](https://way2force.com/agentforce-coworker/). And before any AI rollout, remember that [clean data comes first](https://way2force.com/ai-strategy-needs-data-cleaning/) — a reasoning model is only as good as the records it reasons about.
*Originally published at [Way2Force](https://way2force.com/salesforce-koa-explained-the-crm-reasoning-model-behind-agentforce/).*