Salesforce says agentic AI workforces are more than doubling each year

AI agent workforces are growing fast, while deployment time is falling

Salesforce customers increased their average number of production AI agents from five in February 2025 to 13 in April 2026. That is a 7% compound monthly growth rate, according to Salesforce’s second annual Agentic Enterprise Index. Over the same period, the average time required to deploy an agent into production fell 53% to 1.9 days.

These two numbers matter together. Companies are not simply adding more agents. They are also becoming faster at putting them into operational use. Salesforce based its analysis on customers that had agents active in production during every month of the study period, from February 2025 through April 2026. It also incorporated data from Salesforce research conducted in May 2026.

The trend points to a shift from experimentation toward repeatable deployment. Once a company has the data connections, permissions, workflows and governance needed for one agent, it can reuse parts of that foundation. This can reduce the work required to launch the next agent. A 1.9-day average deployment time suggests that, among the customers studied, agent creation is becoming a routine software and operations process.

For executives, however, deployment speed is not the main measure of success. An agent in production creates value only when it improves an outcome: lower service cost, shorter processing time, higher revenue, fewer errors or better employee productivity. Rapid deployment without controls can instead increase operational and security risk.

The key management task is therefore moving from “Can we deploy an agent?” to “Which processes should an agent control, and how do we measure the result?” Companies that solve this problem can scale with discipline. Agent count alone is not a useful business target.

AI agents are doing more work and moving across business functions

The larger change is happening after deployment. Salesforce found that the average number of agent actions per customer account grew at a 31% compound monthly rate during the 15-month analysis period. This growth was much faster than the 7% monthly increase in the number of agents.

That difference is important. It indicates that existing agents are being used more intensively rather than growth coming only from companies creating additional agents. Their role is also expanding. Caila Schwartz, Head of Agentic Commerce Insights at Salesforce, said the agents are “expanding beyond their initial scope to really become cross-functional.”

A cross-functional agent can participate in processes that span several applications or departments. A service interaction, for example, may require retrieving customer data, updating a record, initiating another workflow and communicating the result. This is materially different from an AI system that only summarizes a document or generates a response.

For management teams, the constraint shifts as agents gain permission to act. Model capability becomes only part of the problem. Access control, reliable business data and clear authorization rules become critical because an agent that can change records or trigger workflows can also make consequential errors at greater speed. Companies therefore need to define which actions an agent may take independently, which require human approval and how every action is recorded for review.

The 31% monthly growth in actions is evidence of expanding usage, not evidence by itself of higher productivity or return on investment. Executives should connect agent activity to business metrics such as resolution time, cost per transaction, error rates and completed workflows. The strongest agent programs will not be those that generate the most actions. They will be those that convert those actions into measurable business results.

Salesforce’s AWU metric measures agent activity

Salesforce says its Agentforce agents had completed 734 million Agentic Work Units (AWUs) as of April 2026. The total was growing by about 15% each month. Salesforce created the metric to measure work performed by agents instead of relying on token consumption, which mainly reflects the amount of AI computation or text processing involved.

The distinction is useful. Token counts can help track technical usage and cost, but they say little about what an agent actually accomplished. Salesforce is trying to move measurement closer to completed work. AWUs provide one way to quantify the volume of agent activity across deployments.

But the metric has a clear limitation. Analysts have criticized AWUs because they are not directly tied to business outcomes. Completing more work units does not establish that a company reduced costs, increased revenue, improved customer service or raised employee productivity. Activity and value are different measures.

This distinction should shape executive reporting. AWUs can help teams understand adoption, workload growth and system utilization. They should not become the primary measure of an agent program’s success. Management needs to connect agent activity with outcomes such as cost per completed process, resolution time, conversion rates, error rates and employee hours saved.

The 734 million AWUs demonstrate substantial activity within Salesforce’s Agentforce ecosystem. The 15% monthly growth rate shows that this activity is increasing quickly. Neither figure, on its own, establishes return on investment. For C-suite leaders, the relevant question is how much economically useful work each agent produces relative to its total cost and risk.

Cross-system agents require architecture built for actions

Salesforce’s research found that agents are operating across multiple cloud domains. This matters because useful business processes rarely remain inside one application. An agent may need to retrieve information from one system, process a request, update another system and trigger a separate workflow.

Salesforce argues that this pattern strengthens the case for a “headless” architecture. In this design, the agent’s underlying logic is separated from a traditional graphical user interface. The agent can interact with services and workflows directly rather than depending on the screens designed for human users. Salesforce says this allows agents to “process tasks, execute actions, and trigger workflows anywhere.”

For executives, the critical issue is not whether an architecture carries the “headless” label. The real requirement is controlled access to business systems. Agents need reliable application programming interfaces, consistent data, clear identities and tightly defined permissions if they are expected to take actions across applications.

This becomes more important as agent autonomy increases. An AI system that only answers a question creates one class of risk. An agent authorized to change customer records, initiate transactions or trigger operational workflows creates a larger control requirement. Companies need authentication, authorization, audit logs and limits on what each agent can change. High-impact actions may also require human approval.

A decoupled architecture can make agents easier to deploy across channels and applications, but flexibility must not remove accountability. Architecture should make every action attributable, observable and reversible where practical. For business leaders, that foundation is what allows cross-functional agents to scale without creating uncontrolled operational dependencies.

Salesforce reports broad internal agent adoption and significant employee time savings

Salesforce says 83% of its employees have adopted Slackbot, its AI agent in Slack. Joe Inzerillo, President of Enterprise & AI Technology at Salesforce, said Slackbot saves the average employee five hours per week. He also reported that internal AI-agent sessions tripled between February 2025 and April 2026.

These figures show that Salesforce is using agents at significant scale inside its own organization. The increase in sessions indicates growing usage, while the 83% adoption rate suggests that the technology has reached a large share of employees. More important, Salesforce is connecting usage to a claimed productivity result: time saved.

For executives, time savings are more useful than measures such as sessions, prompts or tokens. They are closer to an economic outcome. But saved time does not automatically become lower operating costs or higher output. The business benefit depends on what employees do with those hours and whether the agent reduces total effort rather than moving work elsewhere in the process.

The five-hours-per-week figure should also be interpreted as a Salesforce-reported internal result. Leaders evaluating similar tools should establish a baseline before deployment and then measure completed work, processing time, quality and employee effort after adoption.

The broader lesson is that adoption and productivity need to be measured separately. A company can have high adoption without meaningful returns, while a narrowly deployed agent can produce substantial value in a high-cost process. Management should focus on where employee time is being saved and how that capacity translates into measurable business output.

Manufacturing, finance and healthcare are leading on agent sophistication

The most advanced agent deployments are not concentrated in the industries commonly associated with AI leadership. Salesforce found that manufacturing, financial services, and healthcare and life sciences have developed more sophisticated agent networks than technology and retail.

Salesforce measures this through its five-point Sophistication Index, which scores the cognitive complexity of agent actions. Levels 1 through 3 cover tasks such as looking up records, drafting emails and summarizing documents. Levels 4 and 5 cover more complex functions, including updating database fields. Caila Schwartz, Head of Agentic Commerce Insights at Salesforce, said industries are approaching agentic AI differently, with some operating at higher levels of sophistication than others.

The distinction matters because deployment volume does not reveal how much operational responsibility agents have. An organization running many agents that retrieve information may have a less advanced implementation than one using fewer agents to execute controlled changes in business systems. The ability to act changes both the potential value and the risk profile.

The results also suggest that AI maturity should not be judged by industry reputation. Manufacturing, finance and healthcare operate complex processes where structured workflows, records and transactions create opportunities for agent-based automation. At the same time, these sectors often face strict requirements around accuracy, security, privacy and accountability. More sophisticated agents therefore require stronger controls, not less oversight.

Salesforce’s index should still be treated as a measure of functional complexity rather than business performance. A level 5 action is not necessarily more valuable than a level 2 action. Executives should select the lowest level of agent autonomy needed to produce the required business result, then increase autonomy where the expected return justifies the added control requirements.

The more useful benchmark is therefore not sophistication for its own sake. It is whether an agent can perform the required process accurately, securely and at a lower total cost or higher service level. Salesforce’s findings show that advanced agent deployments are already appearing across heavily operational and regulated industries, but their business value must ultimately be demonstrated through outcomes.

AI agents are moving from answering questions to completing business tasks

Customer service is the most common starting point for agentic AI, according to Salesforce. Joe Inzerillo, President of Enterprise & AI Technology at Salesforce, described service as providing “far and away the best ROI to start with.” The underlying reason is practical. Service operations contain large volumes of repeated requests, defined processes and measurable outcomes, making it easier to identify where an agent creates value.

The more important shift is from information to action. Early AI assistants largely answered questions, summarized information or explained procedures. Inzerillo said users are increasingly asking agents to perform the underlying task. Instead of asking an employee agent how to submit a vacation request, for example, an employee can provide the required details and instruct the agent to submit it.

This changes the role of enterprise AI. Answering a question primarily requires access to reliable information. Completing a task also requires access to business systems and permission to change data or trigger workflows. The agent must understand the request, select the correct action, execute it and confirm the result. This creates greater potential for automation, but it also raises the consequences of errors.

For executives, that makes authorization the central control issue. An agent should have only the permissions required for its assigned work. Companies also need clear rules for when an agent can act independently, when human approval is required and how completed actions are logged and reviewed. These controls become more important when agents handle customer records, financial processes, employee data or other sensitive operations.

Service is a strong place to begin because outcomes can be measured directly. Leaders can track resolution time, cost per case, first-contact resolution, escalation rates and customer satisfaction. Similar measures should follow agents as they expand into employee services and other functions. This makes it possible to distinguish automation that completes useful work from AI activity that simply increases usage.

Salesforce’s evidence supports a clear direction: enterprise agents are becoming execution systems, not only information tools. Inzerillo described the growing “bias towards action” as an evolution in agentic use. The opportunity for businesses increases as agents gain the ability to complete work, but so does the need for reliable data, controlled system access and measurable accountability.

Concluding thoughts

Salesforce’s data points to a clear change in enterprise AI. Companies are deploying more agents, doing it faster and giving those agents more work. The next phase will be defined less by agent count and more by how much useful work those systems can complete.

For executives, that changes the management priority. Deployment speed and adoption rates are useful operating measures, but they do not prove business value. Agent activity needs to connect to outcomes such as lower cost per transaction, shorter resolution times, fewer errors, higher revenue or measurable employee capacity.

The shift from answering questions to executing actions also raises the control requirement. Agents that can update records and trigger workflows need clear permissions, reliable data, audit trails and defined points for human approval. Greater autonomy should follow proven reliability.

The strongest approach is to start with processes where value is measurable and actions are well defined. Service is one such area, as Salesforce’s findings suggest. Scale from there when the economics and controls are clear. The competitive advantage will not come from having the largest agent workforce. It will come from turning agent autonomy into controlled, measurable business performance.