Salesforce survey finds “starting first” is not what delivers ROI for enterprise AI agents

Salesforce says the enterprise race to launch AI agents is underway, but its latest survey suggests the advantage is not going live first. In findings from a global study of 2,025 agentic AI decision-makers, Salesforce reports that agent deployments more than doubled over the prior year and that retailers using AI agents are seeing stronger commercial results than those that are not. Yet among companies already realizing benefits, Salesforce finds the most predictive differentiators are operational preparation and governance rather than deployment pace.

Salesforce’s “State of Agentic AI in the Enterprise” is built around a double-blind survey administered May 14–28, 2026 across 20 countries and five continents. Respondents include leaders who influence AI agent purchasing decisions, and Salesforce groups them as deployed (30% of respondents), piloting (47%), or evaluating (23%). Unless otherwise noted, Salesforce presents rounded figures that reflect the deployed cohort when discussing outcome measures such as time to ROI and performance lifts.

For organizations that have agents in production, Salesforce reports that deployments reach ROI in about eight months. It also reports an employee adoption rate of 53% and a 29% average lift in customer satisfaction among those running agents. Those outcomes are presented as self-reported, and Salesforce does not offer a separate, audited measurement of ROI or satisfaction gains in the released findings.

One of the survey’s core comparisons is between retailers using AI agents and those that are not. Salesforce says retailers running AI agents grew online sales at four times the rate of retailers that did not deploy agents. Salesforce frames the broader question for early deployers as a sharper one: if the pace of launch does not explain the best ROI timelines, what is different about the organizations that get meaningful returns?

Salesforce’s answer centers on preparation. The company says early movers were not necessarily the first to reach meaningful ROI, and that operational factors such as having clean, well-governed data available to agents at the time of use, and defining each agent’s scope clearly, were most predictive of success. In other words, Salesforce argues that agents can only be effective when the information they need is trustworthy and when the tasks they are allowed to perform are explicit.

In the company’s messaging, data readiness is not portrayed as an all-at-once exercise. Joe Inzerillo, President of Enterprise & AI Technology at Salesforce, is quoted saying that organizations do not need to “boil the ocean” and perfect all data in one place before starting. Instead, Salesforce says teams can go use case by use case, making data accurate, mechanized, and semantically described so agents understand what the data is and how it should be used. Salesforce also links semantic description to unlocking value, suggesting that how information is labeled and connected matters as much as whether it exists.

Salesforce also highlights the design choices around where and how an agent operates. It says most deployers started with some level of oversight and then built out additional controls over time. Shibani Ahuja, Salesforce SVP for Data & AI Strategy, is quoted emphasizing that the advantage was not in starting first, but in starting deliberately, including decisions about when a person stays “in the loop,” and the guardrails the organization builds before they are needed. Salesforce further cautions against overbuilding guardrails, with Inzerillo quoted saying that “good enough to learn” should be the bar.

Salesforce illustrates its view with an internal example from Asymbl, a workforce orchestration company. Asymbl is described as having experienced rapid growth in 2025, with candidate volume outpacing its recruiting team and applicant records stored across disconnected systems. To address this, the release describes Asymbl building Rosa, a digital recruiter agent that drafts job descriptions, screens candidates, schedules interviews, and sends offers. Salesforce says Rosa works inside Slack, with Data 360 used as a centralized knowledge foundation so Rosa runs on consistent, trusted records, while MuleSoft pulls in data from systems including GitHub, Hira, and Google Drive so recruiters do not have to search across systems mid-pipeline.

Salesforce’s release provides additional methodological context but also leaves key details unquantified beyond the headline outcomes. The company says outcome metrics, including time to ROI and the 29% customer satisfaction lift, are self-reported, and that figures may not sum cleanly due to rounding. It does not disclose the specific definitions used by respondents to classify ROI, nor does it provide comparative timelines for different kinds of industries beyond the retailer example. The study’s scope covers agentic AI decision-makers, but Salesforce does not include a breakdown of which agent types or governance models produced the largest differences, beyond general themes like data governance, scope definition, and guardrails.

Even so, the findings offer a practical steer for organizations moving from pilots into production. If Salesforce’s pattern holds, the near-term focus for teams may shift from how quickly they can deploy to whether they can make their data usable at the moment agents need it, define what agents should and should not do, and determine when humans must remain responsible. The next thing to watch is whether later deployments increasingly reflect these “starting deliberately” themes, and whether follow-up research quantifies how different governance and data practices affect ROI timelines across industries and agent categories.