Unveiled at Dreamforce 2026, AIforce exposes Salesforce data, business logic, semantic context, actions and governance through open interfaces while retaining the access controls already configured by customers. Salesforce said requests continue to run through existing permissions and business rules, so users and agents can see only information they are authorised to access and actions are routed back through Salesforce.
The company is positioning AIforce as the connection layer between its core platform and a widening range of AI applications. It launches with Claudeforce, Slackforce and Agentforce Coworker, while Salesforce’s Headless Toolkit provides Model Context Protocol servers, APIs, plug-ins, skills and development tools for other interfaces.
Its AgentExchange ecosystem lets partners build interfaces and agents on that architecture, with Salesforce listing Anthropic, AWS, Google and OpenAI among companies participating in integrations around the toolkit globally today.
Salesforce in Claude, part of the expanded partnership with Anthropic, packages a prebuilt MCP server with 37 sales skills covering tasks such as prospecting and pipeline management. The integration is available to Salesforce customers in beta. The companies also plan to add Tableau analytics and capabilities spanning service, marketing, commerce and industry-specific workflows.
Slackforce brings Salesforce context into Slack conversations and workflows. Its Surfaces feature can assemble interactive views from Salesforce, Slack and connected tools, while Slack CRM allows users to create accounts, record call notes and update records through natural-language prompts. Slackbot can draw on both Slack conversations and governed Salesforce context to identify account changes, examine support cases and initiate follow-up actions.
Salesforce has also broadened its collaboration with Amazon Web Services, making Salesforce data, context and skills natively accessible inside Amazon Quick. The integration, which uses MCP, can retrieve pipeline information, account summaries, service cases and operational analytics without a custom integration or data migration. Salesforce said the Amazon Quick connection to its headless architecture is available now.
The AWS agreement also brings the AWS DevOps Agent into Slack, with Security, FinOps and Partner Central agents due to follow. Agentforce customers can access models offered through Amazon Bedrock, while Salesforce and AWS have expanded zero-copy data connections across services including Amazon Aurora, Amazon RDS, AWS Glue-managed Apache Iceberg tables and SageMaker Lakehouse.
Alongside AIforce, Salesforce and Nvidia introduced Koa, Salesforce’s first CRM reasoning model for Agentforce. Koa is based on Nvidia’s Nemotron 3 Super and was post-trained by Salesforce using a proprietary synthetic dataset modelled on enterprise tasks and workflows developed from nearly three decades of CRM deployments.
Salesforce said Koa is designed for multistep work across sales, service and other customer operations rather than general-purpose reasoning. Its training scenarios cover more than 14 industries and simulate tasks such as generating leads, qualifying opportunities, routing service cases and selecting the correct sequence of business actions.
The company says it controls Koa’s model weights and runs the model within Salesforce infrastructure, keeping customer data inside its trust boundary during inference. Koa can be selected as a managed model in Salesforce’s Data Cloud generative models catalogue and configured for agents and sub-agents in Agentforce.
Salesforce said internal testing on its CRM Bench showed Koa was 11 per cent more precise at choosing the correct action, delivered 2.1 times greater reliability and improved long-context retention by 15 per cent compared with the general models used as its baseline. Those performance figures are Salesforce’s own benchmark results rather than independently verified measurements.
The AIforce architecture also includes a zero-data-retention policy under which business data sent to model providers is used to answer the immediate request and is not retained by the provider. Salesforce said organisations do not need to create a separate permissions model when connecting supported interfaces.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
