The hype around AI agents, software that carries out tasks on its own, says they win by acting alone. In production, agents succeed by stopping. The difference comes down to the cost of failure in real systems and the cost of controlling it.
A 2025 study documents what teams run once the demo is over. It draws on a survey of 306 people who build and run agents, plus 20 in-depth case studies. The paper also notes the doubts raised in leadership discussions, citing another study, which reports that 95% of agent deployments fail.
The reasons for using agents are practical. The top benefit, cited by 72.7% of those surveyed, is higher productivity. Next come fewer hours of human work, at 63.6%, and automating routine labor, at 50.0%. So the case for agents rests on cost and output inside clearly defined work.
Teams also cap the downside. In 68% of systems, an agent takes ten steps or fewer before a person has to step in. Nearly half make fewer than five requests to the AI model before they stop and hand back. Those limits work as a budget, set by reliability targets, computing costs and the cost of reversing a bad action.
Human oversight is built into the design. According to the study, 74% of production agents depend mainly on people to judge their work. Reliability is the top challenge in building them, because correctness is hard both to ensure and to measure. That human checking is an operating cost, which companies either plan for or pay later in incidents, rework and lost trust. Success means fewer human hours per outcome, with people still involved.
Build choices follow the same logic. Seventy percent of the agents run on ready-made models steered by written instructions, rather than models retrained for the job. That keeps updates fast, change control simpler and mistakes reversible when behavior shifts after launch.
Control is where the discipline shows most. The paper finds that teams favor predefined workflows over open-ended freedom, and 80% of the detailed case studies follow a fixed sequence of steps. That is governance built into the software. Every action can be listed and logged, so operations, risk and audit teams can follow how a system fails. The same structure lets a team test it, monitor it and shut it down cleanly when reliability slips.
