OpenAI's Operator Agents Are Now Running Live Enterprise Workflows

OpenAI's Operator — its autonomous web agent — has quietly crossed from impressive demo into live enterprise use. Companies are deploying it to handle multi-step browser workflows that previously required a human to sit at a screen: filling procurement forms, navigating supplier portals, extracting data from legacy web interfaces, and scheduling across systems that don't have APIs.

What Operator Actually Does in Production

Operator is a browser-native agent. It sees a webpage the way a human does, identifies interactive elements, and takes actions — clicking, typing, submitting, navigating — without needing the target site to have an integration or API endpoint.

In enterprise deployments this matters enormously. Most large organizations run a patchwork of legacy web tools, vendor portals, and internal dashboards that were never designed for programmatic access. Operator doesn't need them to be. It works on the UI layer directly.

Reported production use cases include:

• Procurement automation — navigating supplier portals to pull quotes and submit orders
• HR workflow execution — completing multi-step onboarding forms across disconnected systems
• Data extraction — pulling structured data from web dashboards into internal reporting pipelines
• Travel and expense management — booking, submitting receipts, and reconciling across tools

The Reliability Gap Is Closing

Early versions of browser agents — including Operator's predecessors — failed too often on real-world web complexity: CAPTCHAs, unexpected modal dialogs, session timeouts, dynamic page rendering. Enterprise teams couldn't build reliable workflows on systems that succeeded 70% of the time.

OpenAI has invested heavily in reliability improvements. The current Operator is meaningfully more robust on ambiguous UI states, and the failure modes are increasingly recoverable rather than silent. That shift — from "impressive when it works" to "reliable enough to schedule" — is what has enabled production deployments.

Where It Sits Relative to McKinsey's Agent Deployment

McKinsey's 20,000-agent Lilli deployment (see our full coverage) operates in a controlled internal knowledge environment. Operator is working in the wild — navigating the open web and third-party systems with all their inconsistency.

Both represent the same underlying shift: agents that take actions, not just generate text. The environments are different; the direction is the same.

The enterprise automation market is watching Operator closely. If browser-native agents reach the reliability threshold for mission-critical workflows, the TAM expands dramatically — every web-based process becomes automatable without an integration project.