AI WORLD HQ
Imagine asking a tool to sort your inbox, look up a customer record, and draft a reply, all without you clicking through each step. That kind of tool is what people mean by an AI agent: software that can take a series of actions toward a goal, not just answer one question. Zapier, a company that builds workflow automation tools, has surveyed business leaders about these agents for a second year. Its findings show fast growth, but also a big gap between using agents and trusting them to run alone.
What the survey found
Zapier says the share of organizations running AI agents in production (meaning in real, everyday work rather than a trial) rose from 40% in 2025 to 71% in 2026. The share still testing them fell from 32% to 17%.
Use grew in every department the survey tracked. According to Zapier, these were the biggest jumps in the share of organizations using agents:
- Finance: 24% to 41%
- HR: 30% to 43%
- Operations: 47% to 59%
- Sales: 26% to 37%
The most common jobs stayed the same: summarizing documents and pulling out or entering data.
Why people still stay in the loop
Zapier's central finding is about autonomy, meaning how much an agent can finish without help. Only 30% of leaders say their agents complete more than 60% of assigned tasks without human intervention.
The survey points to a clear reason. When more than half of an organization's agents are connected to broader automated workflows (a fixed series of steps that software runs in order), 77% of leaders say their agents handle more than 60% of tasks alone. When half or fewer are connected that way, the figure drops to 8%.
Here is how that could look. An agent reads a new support ticket and labels the problem. Then fixed rules take over: billing disputes go to a person, and routine questions go to the right team. The agent makes the judgment call, while the workflow decides when a human needs to step in. Zapier describes this as putting human oversight at chosen checkpoints, not whenever someone happens to notice a problem.
Where the concerns are
Some barriers remain. Leaders most often named data privacy, security, and regulatory compliance (27%), followed by accuracy, reliability, and errors (24%).
Zapier suggests pairing AI with deterministic automation, meaning rule-based steps that behave the same way every time. In its example, the AI interprets an unclear request, while fixed rules control where information goes and which actions need approval. Zapier is clear that this does not make the AI error-free. It just limits how far one mistake can spread.
Wrap-up
Zapier's survey suggests that AI agents are moving from testing into everyday work, and that most organizations still keep people involved at key moments. The practical lesson is to map one task you repeat each week, mark the points where a person should review the result, and note which steps could run automatically. Then you can decide how much freedom an agent should have.