On September 9, UiPath published a survey of 590 technology leaders at companies with at least $1 billion in revenue. Less than one in three, 31 percent, said AI is fully embedded in the business. Eleven percent are still in a pilot. Thirty-five percent have rolled it out only to select teams.

That is not a small-business poll. The pattern still matches what we see in firms that bought seats and called it a strategy. A demo runs. A person pastes the output into the real system. ROI stays a slide.

The number that actually moves

The useful split is not "using AI" versus "not using AI." It is whether the work around the model changed.

Among respondents who said orchestration is fully embedded across the enterprise, 89 percent said their agentic implementations met or beat ROI expectations. Only 29 percent said orchestration is fully embedded in their workflows. Source: UiPath AI Adoption and Orchestration Survey, September 9, 2026. Fieldwork ran May 25 to June 8, 2026.

Read that again. Most of the sample has not built the layer that connects the model to the systems, the data, and the person who owns exceptions. The minority that did is the group reporting the return.

Where the pilot actually dies

Asked what gets in the way of scaling, leaders named three things more than model quality:

  • Data quality and readiness, 38 percent
  • Integration with existing workflows and systems, 37 percent
  • Governance and compliance, 33 percent

Integration at 37 percent is the line operators should underline. A pilot can ignore the CRM, the inbox, and the approval queue. Production cannot. If the output still waits on a human to copy it, you did not deploy AI. You added a drafting step in front of the old job.

UiPath also found 52 percent of these companies plan to apply AI to hybrid workflows in the next 12 months: part fixed rules, part judgment. Those paths are exactly where a loose pilot fails. Someone has to define the trigger, the system of record, and what happens at 11pm when a step dies.

What this means if you are not a billion-dollar firm

You do not need an orchestration platform with that name on the invoice. You need the same idea at the size of one painful path.

A ChatGPT seat is the pilot. The workflow is the rest:

  • What event starts the work, so it does not depend on someone remembering
  • Which system the result is written into, so Slack is not the database
  • Where a person is required, and where they are not
  • Who gets pinged when it fails, with enough context to fix it once

That is orchestration for a 12-person company. It is also the gap in the survey. Thirty-one percent embedded the tool. Twenty-nine percent embedded the path. The ROI number sits with the second group.

Stop funding another pilot week

If last quarter's AI experiment produced a thread of nice drafts and no change in how the week runs, do not buy a second tool to fix the first. Write the path the pilot skipped.

Pick one repeating job. Name the trigger. Name the system that must update. Name the stop. Then let the model touch only the step that needed judgment or drafting. The survey's 89 percent did not get there by prompting harder. They got there by putting the model inside a process someone owns.

A pilot is a demo. The path is the product. Build that, and the usage stat finally has somewhere to land.