88% of enterprises use or test AI agents. Most still keep a human in the loop.

    Zapier’s 2026 survey of enterprise leaders, published in November, has a clear headline: agents have moved from experiments to production. The detail underneath is more interesting, because it shows how much people are still involved.

    How the survey was run

    Centiment ran it for Zapier in September 2026 with 513 completed responses. Respondents were US managers and above at organizations with at least 1,000 employees that have used AI professionally. The data are unweighted, with a margin of error of about four percentage points. Zapier sells automation and agent products, which is worth keeping in mind.

    Production is now normal

    Figure 1 · Adoption year over year

    Use or test agents, 202572%
    Use or test agents, 202688%
    In production, 202540%
    In production, 202671%
    Share of leaders whose organization uses or tests agents, and whose agents are in production. Source: Zapier 2026 AI agents survey.

    Autonomy is another matter

    Only 30% of leaders at agent-using organizations said their agents complete more than 60% of tasks without human intervention. The most cited barriers were data privacy, security, and compliance (27%), then accuracy and reliability (24%).

    One finding stands out. When more than half of an organization’s agents were part of broader automated workflows, 77% of leaders reported that higher level of autonomy. When half or fewer were, the figure was 8%. The article does not show that workflows cause autonomy, and the figures are self-reported, so treat it as a lead, not a proof.

    Figure 2 · Autonomy and workflows

    Agents mostly inside workflows77%
    Agents mostly standalone8%
    Share of leaders reporting agents that finish most tasks without intervention, by how connected their agents are. Self-reported, correlation only.

    Where to put the human

    If most teams still review agent work, the useful question is where review adds the most value. A few patterns work well:

    • Before anything irreversible. Sending, paying, deleting, or publishing should wait for a person.
    • On low confidence. Let the agent flag what it is unsure about, so people look at the hard cases.
    • By sampling. Review a share of routine work to catch drift, and raise the share when something changes.
    • At the plan stage. For long tasks, approving a plan is cheaper than fixing a finished mistake.

    Review points are a design choice, not a failure of automation. They are a core part of how we think about Elsor Harness.

    Sources

    Want help putting this into practice?

    Book a free consultation and tell us about your team or school.

    Book a free consultation