Analysis
Most of the news is infrastructure for agents. For marketers the value is in what it implies about data access, governance and cost.
In 60 seconds
- Salesforce introduced AIforce, which lets AI agents use Salesforce through APIs and metadata without going through the user interface.
- It also introduced Koa, a CRM-optimized model built with NVIDIA Nemotron 3 Super, and an "Enterprise AI Harness" covering trusted models, context, actions, governance and security.
- The performance numbers are Salesforce's own internal tests; customer results are vendor-reported.
- Marketers should focus on data quality, permissions and what agent activity will cost.
Impact Layer
- What changed
- Salesforce used Dreamforce 2026 to introduce AIforce, the Koa model and the Enterprise AI Harness, positioning Slack as the hub for agent work.
- Why it matters
- If agents reach Salesforce data directly through APIs rather than screens, the quality of marketing data, the permission model and the audit trail become the main controls on what agents can do.
- Who is affected
- CMOs and marketing operations teams on Salesforce; any team deciding how far to let agents act on customer data.
- Martech stack area
- CRM, AI, Data
- Stack impact
- Expect agent access to flow through APIs and metadata, with governance applied in a central layer. Review which of your data and workflows are exposed that way.
- Business impact
- Moderate The cost model for agent activity is the unanswered question for most marketing budgets.
- Related vendors
- What to watch
- Independent evaluation of Koa, how the Enterprise AI Harness is priced and packaged, and how many of the announced customer results hold up in marketing use cases.
Salesforce’s Dreamforce 2026 announcements were dominated by infrastructure for AI agents. Here is what was introduced, in terms a marketing team can use. All figures below are Salesforce’s own.
What was announced
- AIforce. Salesforce says agents can use its platform through APIs and metadata, without going through the user interface a person would use.
- Koa. A model optimized for CRM work, built with NVIDIA’s Nemotron 3 Super. In Salesforce’s internal testing it produced 15% more relevant answers and 11% more precise tool calls; the company has not made that testing public in the material we reviewed.
- Slack as the hub. Slack is positioned as the place where people and agents work together.
- Enterprise AI Harness. A framework covering trusted models, context, agency, actions, governance and security.
Marc Benioff and Patrick Stokes, president of applications and marketing, were among the Salesforce executives presenting.
Customer results Salesforce cited
Salesforce says Fulton Bank built its first agent in one hour and reached 3,000 users in six weeks, and that Southwest saw “seven times the ROI”. These are vendor-reported results, and neither is a marketing example, so they say little about marketing outcomes.
Our read
Three practical points for marketers.
- Data quality becomes a safety issue. An agent that acts through APIs does not see the cautionary notes a person would. Duplicate records, stale consent flags and unclear field definitions turn into wrong actions.
- Permissions need a deliberate design. Decide what an agent may read, write and send, and who approves it, before the first deployment.
- Ask about cost per action. Agent activity is the new unit of spend. The announcements we reviewed did not give a marketing-specific cost model.
The Koa figures come from internal testing, so wait for independent comparison before treating the numbers as a benchmark.
Sources
- Salesforce: Five Dreamforce 2026 takeaways (opens in a new tab)
Summary of the announcements and customer examples.
- Salesforce: AIforce announcement (opens in a new tab)
Details on AIforce and the Koa model.