note · 2026-10-12

46% of large companies have AI agents in production, and only 52% have defined when a person steps in

Schellman surveyed 525 US professionals who govern AI at companies with 500 or more employees. Agents reach production far more often where governance is mature, 78% against 22%. The escalation path is what lets an agent go live.

For IT directors and CTOs putting an AI agent into operationyaab1 min read
46% of large companies have AI agents in production, and only 52% have defined when a person steps in

The first question we write for any agent an IT department puts into operation is who can stop it and when. A survey says half of large companies skipped it.

Schellman, an audit and compliance firm, surveyed 525 US professionals who govern AI at companies with 500 or more employees between April 13 and May 11, 2026, and published the results on July 29, 2026; the results are not weighted. 86% say they have tested or piloted AI agents and 46% have them in production: 78% among companies with mature governance, against 22% among those still developing it. The survey shows the two go together; it does not say which came first. Schellman's note of September 9, 2026 adds that only 52% of organizations with agents in testing or production have defined human oversight and escalation requirements.

BCG, the consulting firm, quoting its chief AI ethics officer Steve Mills, on X: many governance programs were designed for two or three centrally managed pilots, not for the scale of agentic AI deployment today.

In one IT department the list took an afternoon: what the agent does alone (reads, drafts, routes), what it proposes (any change to a record), who gets paged when it is unsure, and how to undo an action. The agent went live the next week. Before the list it had waited four months.

The escalation path is not the brake on the agent. It is what lets you turn it on.

Sources