According to an article presenting the findings of Salesforce's "State of Agentic AI in the Enterprise" report, based on a global study of more than 2,000 executives and AI decision-makers, the question occupying many boardrooms—whether the organization is moving fast enough—is the wrong question. According to the data presented in the report, what separates organizations achieving real return on investment (ROI) from those that remain stuck in expensive pilots is not related to speed, but to actions taken prior to launch: making data trustworthy for the specific task, defining the point where a human remains in the loop, and building guardrails in advance rather than retrospectively.
Task-Ready Data Instead of Perfect Data Unification
According to the report, perfect data is not a prerequisite for implementing AI agents. Only 31 percent of organizations that deployed AI agents fully unified their data prior to deployment, reaching ROI in 7.3 months. Meanwhile, 34 percent of organizations chose an iterative deployment—they started with the data they had and added further integrations over time. These organizations reached ROI in 8.2 months. The gap of less than a month in reaching ROI is described in the report as a modest gap that does not constitute a reason to freeze initiatives.
The article notes that less than a year before joining Salesforce, while working in the CIO's office at a large financial services company, skepticism prevailed regarding another platform's ability to fix what appeared to be a data problem rather than a tooling problem. This stance changed after observing what happens when data is scoped and governed for a specific task, instead of being treated as one giant data migration project that must be completed before anything can start.
Salesforce's Experience with Help Agent
In the article, this principle is exemplified through Help Agent, Salesforce's AI agent operating on help.salesforce.com. The agent was built on Data 360 and scoped to a single task: resolving support inquiries. The deployment did not wait for a fully unified view of all customer systems across the organization.
According to the provided data, the agent handled close to five million customer conversations across seven languages and three different portals. Help Agent resolved 68 percent of inquiries without human involvement, generating calculated annualized savings of $100 million. The key lesson noted from operating the agent is that there is no need to prepare all systems, but rather to prepare the data needed for the specific use case and get to work.
Predictive Success Factors: The Setup Around the Model
The study's findings show that out of ten measured success factors, the three most predictive factors are unrelated to the underlying technology:
- Clean, accessible data and a narrowly scoped use case share the top spot at 36 percent each.
- Escalation paths to a human defined before launch rank next at 35 percent.
- Model quality, platform choice, and a unified orchestration layer each received only about 30 percent.
The report notes that most organizations that deployed agents already work with capable models and platforms. The gap between strong and weak performers manifests in what is built around this infrastructure, rather than the infrastructure itself. According to the Agentic Maturity Model mentioned in the article, the most advanced organizations are not those using the most advanced model, but those that matched deployment to their organizational maturity level instead of assuming the organization was ready to run agents unsupervised from day one. This principle underpins the Agentforce platform, where agents are defined with a narrow focus for a specific role, with a built-in handoff to a human that is designed in advance rather than added only upon failure.
Integrating Agents into Complex and High-Stakes Tasks
Agents built on large language models (LLMs) operate probabilistically: they analyze uncertainty by weighing possibilities instead of following a predetermined script, making them suitable for complex and unstructured work. However, the article emphasizes that probabilistic logic cannot constitute the entire system; it must be anchored in deterministic components, such as structured logic and governed workflows that behave identically every time, to prevent situations where an agent's decision cannot be explained when an error occurs.
According to the report's data, agents currently operate in high-impact domains:
- Employee-facing decision support and customer-facing transactional work are the most common places where agents operate, each at 49 percent.
- Low-stakes internal workflows account for only 17 percent.
- 40 percent of organizations that deployed agents operate them in high-stakes or regulated tasks, such as financial transactions, compliance-sensitive processes, and decisions where an incorrect answer carries legal or financial weight.
In addition, the company's Agentic Enterprise Index indicates that agents are tackling more complex tasks: today, the average agent is capable of acting on six skills, compared to only two skills at the beginning of 2025.
The Organizational Advantage: Pre-Launch Decisions Instead of Speed
The report emphasizes that all organizations participating in the study began from a similar starting point two years ago, without playbooks or existing agents. Today's leading organizations did not reach their status through a faster pace, but because they made a specific set of decisions in advance: selecting the right use case, defining the relevant data, establishing where human involvement is maintained, and defining responses to failure scenarios.
Regarding autonomy levels, the report reveals that fully autonomous deployment of AI agents without any human involvement is extremely rare, reported by only 1 to 2 percent of organizations that deployed agents. The findings indicate that organizations that define the agent's responsibilities, data, and escalation mechanisms in advance achieve better results than those waiting for perfect starting conditions.