Salesforce and Nvidia have used knowledge accumulated from 27 years of customer relationship management deployments to build Koa, a specialist AI model designed to reason through the complex work of selling to and serving customers.
Unveiled at the Dreamforce 2026 conference in San Francisco, Koa is Salesforce’s first AI reasoning model built specifically for CRM. It can qualify a sales opportunity or route a service case, selecting software tools and acting within the company’s permissions and operating rules. Salesforce says Koa matches or beats leading models on CRM actions, with three times fewer errors.
“The most valuable thing Salesforce has built isn’t our platform,” said Salesforce chair and CEO Marc Benioff. “It’s the accumulated knowledge of how enterprise business actually works. With Koa, the knowledge is put inside the model itself.”
Koa uses Nvidia’s Nemotron 3 Super as its foundation. Salesforce trained it further on a proprietary synthetic dataset that recreates business processes across more than 14 industries, including manufacturing, financial services, healthcare and travel. Salesforce says no customer data was used.
Salesforce controls Koa and runs it on its own infrastructure. Customer information remains inside the Salesforce “trust boundary” when the model processes a request. Koa is currently undergoing pilot tests with six customers, including Formula 1 and Xero.
The technical work is more substantial than the cheery Koa name may suggest. Salesforce used reinforcement learning to teach the model how to select tools and complete multistep workflows. The same Agent Script specifications used to configure an Agentforce agent can be converted into simulated tasks for training and testing the model.
A Salesforce’s technical paper published this week describes Koa as a 120-billion-parameter model and provides the mathematical qualifications to the launch claims. For those who can read benchmarks, Koa scored 0.86 on Salesforce’s CRM Bench, against 0.87 for Claude Opus 4.8, 0.90 for GPT-5.5 and 0.81 for GPT-4.1. On the Berkeley Function Calling Leaderboard it scored 66.63%, ahead of GPT-4.1 and its Nemotron base. Salesforce admitted, however, that it was behind GPT-5.5 and Claude Opus 4.8 on this benchmark.
The figures support Salesforce’s case for a model trained around enterprise tasks. GPT-5.5 still led the CRM test, and the paper states that Koa remains below the strongest frontier models. However, specialisation gives Salesforce greater control over cost, deployment and data handling without needing it to win every general-purpose benchmark.
For now, Koa is available only to selected Agentforce pilot customers, with general US availability expected towards the end of the year. Salesforce has not announced a South African release date.
Salesforce leaves the Salesforce screen
Other Dreamforce announcements show how Salesforce plans to make its data and business processes available beyond its own applications. AIforce is a new interface layer that allows employees and AI agents to use Salesforce functions from services like Claude and Slack.
An employee working in Claude could ask for an account summary, review the health of a sales pipeline or update a customer record without opening Salesforce. The request still uses the business rules and security controls configured in Salesforce.
Salesforce in Claude is entering beta with 37 prebuilt sales skills, covering work from preparing for a meeting to reviewing a pipeline. The companies are also making more than 40 Salesforce development skills available through Claude Code, with additional functions planned for service and commerce.
Slackforce brings the same approach into Slack. Its Surfaces feature can generate an interactive dashboard, report or presentation from Salesforce and Slack data in response to a written request. Team members can filter the information, and act on it from the shared Slack workspace.
Slack CRM allows users to create an account, record notes from a call or update Salesforce records from a Slack prompt. Slackbot can draw on information in Slack and Salesforceto identify an account requiring follow-up, create a task and prepare an email.
Agentforce Coworker provides similar access inside Salesforce’s Lightning interface. Salesforce says 100,000 users activated Coworker during its first 35 days. It can call specialist Agentforce agents already deployed by the customer and works within existing Salesforce permissions.
The foundation for these services is Headless 360, which makes Salesforce functions available to authorised AI systems through the Model Context Protocol. This is an open technical standard that allows an AI service to find and use external tools. In Salesforce, those tools retain the customer’s existing permissions and workflows.
Salesforce says Headless 360 can be used from Agentforce, Claude, ChatGPT and other AI platforms. It also offers more than 100 reusable skills that package Salesforce business functions for use by authorised agents.
The partnerships announced at Dreamforce extend this approach across the major cloud providers. Salesforce data and functions are being added to Google’s Gemini Enterprise, while Gemini models can be selected inside Agentforce. Google Cloud is also running live Salesforce workloads through Hyperforce, with selected US customer migrations due to begin in the fourth quarter of 2026.
An expanded agreement with AWS brings Salesforce information into Amazon Quick and allows AWS specialist agents to work inside Slack. Agentforce customers will gain access to models offered through Amazon Bedrock, while Agentforce Voice will work with Amazon Connect Customer for real-time voice service.
The agreements allow a company to use Salesforce data and processes from the AI service or cloud platform it has chosen, while actions continue to follow the rules configured in Salesforce.
Salesforce leaves the Salesforce screen
Other Dreamforce announcements offer persepective of where Koa is positioned. AIforce is a new interface layer that exposes Salesforce data, workflows, permissions and business logic to other AI systems. An employee could ask Claude or Slack to review a sales pipeline, update a record or trigger a workflow without opening the conventional Salesforce interface.
This is a large concession from a software company whose interfaces have kept customers and employees inside its applications. It is also a defence against AI assistants becoming the primary route into corporate systems. Salesforce is offering up the screen so it can retain control of the data, permissions and business processes underneath it.
The first versions include Claudeforce, Slackforce and Agentforce Coworker. Salesforce in Claude enters beta with 37 prebuilt sales skills, covering work such as meeting preparation and pipeline reviews. In Slack, users can create live interfaces from Salesforce and Slack data, then filter the information and take action from the shared workspace.
Headless 360 provides the plumbing. Its Model Context Protocol server lets authorised AI systems discover Salesforce capabilities while retaining the identity rules, metadata, validations and workflows already configured by a customer. Salesforce says the same approach works with Agentforce, Claude, ChatGPT, Cursor and other AI platforms.
The partnerships announced at Dreamforce reinforce that approach. Salesforce data and capabilities will be available in Google’s Gemini Enterprise, while Gemini models remain available in Agentforce. Salesforce will also run selected customer workloads on Google Cloud through Hyperforce, beginning with US migrations in the fourth quarter of 2026. The expanded AWS agreement puts Salesforce context into Amazon Quick, brings AWS specialist agents into Slack and adds Amazon Bedrock model choice to Agentforce.
Salesforce is positioning itself as the governed business layer between corporate data and whichever model or interface a customer chooses. Koa gives it a model of its own, while the agreements with Anthropic, Amazon, Google and Nvidia reduce the risk of binding the strategy to a single AI provider.
Agents get longer jobs
Dreamforce also brought a portfolio of preconfigured Agentforce workers for service, employee support, shopping, sales and supply-chain work. The names, including Casey, Paige, Carter and Hunter, make the agents sound like new recruits. The commercial idea is simpler: customers can begin with an agent designed for a recognisable job instead of assembling every workflow from scratch.
Hunter, the outbound sales agent, introduces a more consequential capability. A new long-horizon runtime allows it to pursue a goal over days or weeks, preserve progress between sessions and modify its plan when new information appears. Earlier agents tended to complete a request in a single interaction. A sales agent can now research accounts, monitor an opportunity and continue follow-up work across a quarter, with defined points where a human must approve action.
Salesforce says Agentforce has processed seven billion Agentic Work Units, including 3.2-billion in its second quarter. Its customer examples include Engine, where half of chat enquiries are resolved by its help agent, and Autism Queensland, where the Paige employee-service agent handles 70% of administrative requests.
Greater autonomy increases the cost of errors. Salesforce’s new Enterprise AI Harness combines the company’s data, integration, analytics, security and agent tools with an AI Control Plane for managing agents. Agent Script can combine model reasoning with fixed business rules, which allows a company to decide where an agent may improvise and where it must follow an exact procedure.
Koa connects these announcements. AIforce gives agents access toSalesforce capabilities from different interfaces. The long-horizon runtime allows them to continue working. The harness imposes permissions and oversight, while Koa is trained to select the tools and actions used in CRM processes.
Salesforce’s claims will now meet production systems, imperfect corporate data and customers who expect a wrong answer to be corrected before it updates an account or authorises a refund. Koa’s pilots include business travel platform Engine, whose founder and CEO, Elia Wallen, supplied the most useful test for the model: “Business travel has countless moving parts, and what we need from AI isn’t a model that sounds confident — it’s one that can reason precisely through complex, multi-step problems.”
* Arthur Goldstuck is CEO of World Wide Worx, editor-in-chief of Gadget.co.za, and author of The Hitchhiker’s Guide to AI – The African Edge.