Dreamforce returned to San Francisco’s Moscone Center from September 15 to 17, organized around the theme of the “Agentic Enterprise.” Across more than 1,600 breakout sessions, the event focused on moving AI agents into production while keeping data access, permissions, and business actions under control.
Across the agent announcements, Salesforce made an architectural argument, drawing its platform as four connected layers that run from Data 360 at the base through Customer 360 and Agentforce to the new AIforce interface on top.
Key Announcements
AIforce Pushes Salesforce Into More Interfaces
Salesforce introduced AIforce as a “live interface layer” that carries governed Salesforce data and actions into the tools where employees already work. It launched with Claudeforce, Slackforce, and Agentforce Coworker, and the Claude integration alone ships with 37 prebuilt sales skills.
Organizations can now expose Salesforce functions through conversational interfaces while retaining the platform’s established identity, authorization, and workflow controls, which reduces the temptation to build disconnected AI experiences with separate governance models.
Koa Targets CRM-Specific Reasoning
Koa is Salesforce’s first CRM reasoning model, post-trained from NVIDIA Nemotron 3 Super on synthetic scenarios modeled on nearly three decades of CRM deployments, with no customer data used in training. Salesforce says it matches or beats leading models on CRM actions with three times fewer errors, a figure drawn from Salesforce’s own benchmark. Koa is in pilot now, with general availability expected in U.S. regions this winter.
Because Salesforce controls Koa’s weights and runs inference inside its own infrastructure, customer data stays within Salesforce’s trust boundary during a transaction. That gives regulated buyers a reasoning option they can judge on data control as well as accuracy, although the error figure comes from a vendor-run benchmark that pilot teams should reproduce against their own workflows.
Prebuilt Agents Shift the Starting Line
Salesforce expanded Agentforce with seven job-ready agents, each built for a specific role. Casey covers customer service, Paige takes IT and HR requests, and Marshall handles supply chain workflows, while Carter, Piper, Fin, and Hunter handle commerce, inbound pipeline, customer experience, and outbound sales. Six are generally available now, and Hunter remains in pilot until its November release.
These agents give implementation teams a defined starting point, although buyers still need to adapt permissions, business rules, escalation paths, and data access. Prebuilt roles can shorten the distance between a controlled test and a useful deployment.
The Enterprise AI Harness Targets Control
The Enterprise AI Harness, announced ahead of Dreamforce, groups six capabilities across context, agency, action, governance, security, and models, including routing each task to the right model by cost and accuracy. It is aimed at the case every enterprise worries about, an agent that holds permission to act and then makes the wrong decision.
Its controls travel with the data and workflow, including policy enforcement and visibility into agent activity, which makes cybersecurity a product requirement from the first stage of an agent project. Many of the underlying components ship today, and the rest begin rolling out early in Salesforce’s fiscal 2028, which starts in February 2027.
Cloud and Model Partnerships Expand
Salesforce expanded its partnerships with AWS and Google Cloud while continuing its Anthropic relationship through Claudeforce. Hyperforce will run on Google Cloud, with general availability in North America set for November, and Salesforce data can surface directly inside Gemini Enterprise. The AWS agreement brings Salesforce context into Amazon Quick, AWS agents into Slack, and a wider set of models to Agentforce through Amazon Bedrock.
The partner mix suggests that Salesforce wants to own trusted business context without requiring every customer to standardize on one model provider or cloud. That stance can protect customer choice, although it also increases the integration and oversight burden for security and architecture teams.
Strategic Insights
The Interface Is Becoming Fluid
Marc Benioff called AIforce an “interface revolution,” comparing it to the jump from DOS to graphical interfaces. Employees may increasingly reach Salesforce records and actions through Claude, Slack, Lightning, or another approved surface, which makes consistent identity and authorization controls essential.
A permission granted inside CRM can now influence actions initiated through several interfaces, so security teams reviewing application access must follow the transaction from request through model reasoning, data retrieval, tool invocation, and final action.
Governance Moves Into the Architecture
Salesforce introduced a new AI Control Plane alongside the Enterprise AI Harness, giving companies one place to register agents, set identity and policy, and track performance and cost across Salesforce and third-party AI. Safety took center stage in the main keynote, where Anthropic CEO Dario Amodei and NVIDIA CEO Jensen Huang staked out different positions on how fast model development should move.
Each deployment has to settle which decisions an agent can make independently, which require approval, and what evidence must be retained. Those rules need named owners in security, legal, and the business unit receiving the automation.
Data Quality Sets the Ceiling
Data 360 remained central to Salesforce’s story because agents require current customer context, governed access, and dependable relationships between records. A polished reasoning layer cannot repair conflicting account histories, missing consent records, stale product data, or inconsistent case classifications during an active transaction.
Technology leaders should treat agent programs as data accountability programs from the first planning session. Each use case needs documented data sources, freshness expectations, access boundaries, quality thresholds, and a response plan when the supporting information is incomplete.
Production Economics Enters the Conversation
Model routing in the Enterprise AI Harness and the event’s emphasis on job-ready agents showed Salesforce pushing buyers toward repeatable deployment patterns. Live Nation’s Venue Agent now fields fan questions across 120 venue websites and answers 95% of them without a handoff to staff. Adecco Group announced a rollout of Agentforce Coworker to 27,000 employees in more than 40 countries following pilots in the UK and France, and CEO Denis Machuel joined Marc Benioff on the main keynote stage.
A credible internal proposal should now identify transaction volume, inference cost, and exception rates, along with the business metric the agent is expected to change.
The Undercurrent
Dreamforce showed a productive tension between expanding access and tightening control. AIforce, Claudeforce, Slackforce, and cloud partnerships push Salesforce capabilities into more places, while the Enterprise AI Harness and Agent Fabric seek to keep those distributed interactions observable and governed.
Each new interface creates another path to sensitive customer data and business actions, which turns inherited permissions, prompt manipulation, excessive agency, and third-party model handling into board-level deployment questions.
What Changes for Technology Buyers
The path from isolated demonstrations to production starts with a tightly bounded process and a short list of permitted actions. From there, teams connect trusted data, set approval thresholds, and record every consequential step so the operational result can be measured against the cost of model usage and human review.
Procurement decisions also need to account for architecture across vendors. Salesforce may supply business context and workflow controls while AWS, Google Cloud, Anthropic, NVIDIA, or OpenAI supplies infrastructure or intelligence, leaving the customer responsible for understanding how data, identity, logs, and contractual obligations cross those boundaries.
What’s Next
Roadmaps should prioritize one or two workflows where data quality, process ownership, and success metrics are already clear, so security teams can test those workflows end to end, from interface access and tool permissions through escalation and recovery.
Agent deployment is becoming an architecture discipline, and Dreamforce made that case in product form. Deciding where agents belong now starts with settling who holds authority, what evidence gets kept, and what each automated decision costs.
