Firms Roll Out AI Agents Without Watching Them

Ask a room of executives whether they are using AI agents and most hands now go up. Ask how they know the agents are working and the room goes quiet. That gap, between deployment and measurement, is becoming the defining feature of corporate AI in 2026, and a run of recent surveys has started to put numbers on it.

The sharpest came from a study by Liferay, which surveyed 500 professionals in the United States. It found that 54 percent of companies are running AI agents in some form, but only 25 percent have any way of measuring the impact those agents have. More than half the market has adopted the technology. Barely a quarter can tell you whether it is paying off.

Everyone's counting adoption, nobody's counting results

The headline adoption figures are genuinely high, though they vary with how the question is asked. A survey of 1,500 technology leaders described crossing a "tipping point" at around 70 percent by the middle of the year, measured as using agents in at least one business function. A separate poll of AI decision-makers found 73 percent using AI regularly, yet only about one in ten calling it central to how the business runs. The pattern holds across all of them. The looser the definition, the higher the number, and the measurement of actual results lags far behind the enthusiasm.

This is not a small accounting problem. An agent that books meetings, drafts code or answers customer queries is making decisions on a company's behalf, often with little human review in the loop. Without measurement there is no way to catch the agent that quietly does the wrong thing at scale, and no honest way to decide whether the investment was worth it. Spending races ahead while the evidence base stays thin.

Caution catching up with the hype

The unease is starting to register. Recent reporting has found enterprises growing more reluctant to let agents ship work on their own, pulling back toward arrangements that keep a person in control of the final step. That is a reasonable response to the measurement gap. If you cannot yet tell how well an autonomous system is doing, keeping a hand on the wheel is prudence, not timidity.

Some of the most aggressive deployments show both sides of this. When Shopify handed large parts of store-building and merchandising to agents, it was betting that the productivity gain would be obvious enough to justify the loss of direct oversight. For a merchant whose storefront is the whole business, that bet is only as good as the ability to notice when something breaks.

The vendors selling these tools have a commercial interest in the adoption numbers looking large, which is worth keeping in mind when reading any single figure. The more useful signal is the one they tend to bury. Companies are buying agents far faster than they are learning how to judge them, and the firms that close that gap first will be the ones that can say, with evidence rather than hope, that the technology actually worked.