At Dreamforce 2026 in San Francisco, Salesforce shifted the test for AI agents from conversation quality to completed work and measurable return. Standalone toolkits for quick proofs of concept are giving way to integrated platforms tied to enterprise data, workflows and managers. Interviews led by George Gilbert and Alison Kosik framed the agenda around pricing, digital labor, Slack and governance. The signal for business is direct: pricing for Agentforce service agents now depends on resolved customer issues, which turns agent performance into a budget item rather than an experiment.
What Salesforce showed on outcomes and scale
Salesforce now charges for Agentforce service agents only when a customer issue is resolved, not when the agent hands the case to a human. Built-in analytics let companies keep tuning agent behavior as business processes change, explained Kishan Chetan, executive vice president and general manager of Agentforce and Service Cloud. The model removes payment for deflection without resolution and links cost to a countable service result. For support operations, vendor comparison moves from answer quality to resolution rate, cost per closed case and the discipline of continuous tuning.
Slackbot has become the entry point for agents inside Slack, acting as a Model Context Protocol client with access to every agent a user is permitted to use. It pulls in trusted enterprise data and builds shareable interfaces on demand, without forcing a switch of tools. The product has reached 1.5 million monthly active users since January, the fastest adoption in Salesforce history, noted Ryan Gavin, EVP and chief marketing officer of Slack. The design treats chat as the control layer where permissions, data and agent actions meet in one working surface.
Governance is handled by the Enterprise AI Harness, which combines Data 360, MuleSoft and Agentforce into an open, composable platform. Agents inherit only the task-level policies defined by their owners, rather than broad system access. That granularity reflects chief information officers putting controls ahead of demonstrations, as agent activity heads toward creating 150 to 200 zettabytes of data, said Savinay Berry, EVP and general manager of data management. Principal analyst George Gilbert described the approach as putting data, applications and agents into one harness that moves beyond request and response to outcomes.
What this means for companies adopting agents
Asymbl Inc. shows how digital labor is managed when it counts as capacity. Digital workers make up half of its workforce across 13 functions, each assigned to a human manager with performance metrics. A recruiter agent helped the company hire about 100 people in 100 days, according to chief executive Brandon Metcalf. Productivity impact rose from about $5 million last year to about $13 million this year. For companies, the pattern is clear: define the role, assign ownership and track output in hires or money, not in chat volume.
The move to outcomes does not remove selection risk. Buyers still need to verify how Salesforce defines a resolved issue, which events trigger billing and how disputes are logged. Permission mapping in Slackbot and task-level policies in the Harness require clean data access rules, or agents will stall or overreach. Built-in analytics help only if someone owns retuning when processes change. The announcements alone do not prove return in a different data landscape, so pilots should fix a baseline for resolution rate, handoff rate and handling cost before scaling.
Confirmation will come from usage and billing data rather than roadmaps. If Slackbot holds above 1.5 million monthly active users and more customers report resolution-linked spend inside service budgets, the outcome model will have moved from the Dreamforce stage to procurement standard. The next check is whether task-level governance appears as a requirement in enterprise tenders alongside accuracy and price.
