SumatoSoft has published an original study of AI agents in production, identifying verification costs and human review capacity as constraints on autonomous operation. The research examines 40 live deployments, distinguishing work agents produce from outcomes they can release without human approval.
The study found that 25.6% of the 39 deployments classified for autonomy operated as unattended AI agents. Separately, 32 of 40 respondents, or 80%, reported pausing, narrowing or switching off an agent after launch. Among 14 engineering respondents, the median share of agent-authored code reaching the main branch without human review was zero.
“Writing the work got cheap and checking it did not. The agents that survived were the ones whose owners paid for the verification,” said Katerina Merzlova, digital transformation consultant at SumatoSoft and author of the study.
The findings distinguish agentic AI adoption from permission to act independently. Among 32 respondents providing workflow coverage figures, median coverage was 80%, and 12 said the agent handled the entire named workflow.
Beck Wang, founder of IfShe, who runs 12 country storefronts, put both numbers side by side: “Unattended authoring is near-total. Unattended release is close to zero, deliberately.”
The failure data points the same way. Of 32 described failures, 11 involved scope or context blindness. In eight others, agents reported success despite incorrect work; three more passed automated checks while missing the intended goal.
Eleven of 14 engineering respondents reported that no agent-authored code reached the main branch unread. These respondents nevertheless reported agents writing most of their code, illustrating the distinction between authorship and release.
“Agents produce plausible code rather than bad code, which is worse, because plausible survives a skim,” said Daniel Battaglia, founder and CEO of Parksy. “Throughput without matching review capacity is not velocity; it is a queue of unverified changes.”
Gartner forecast in June 2025 that more than 40% of agentic AI projects would be canceled by the end of 2027. SumatoSoft’s study does not dispute that projection; it describes a potential mechanism for such failures: the cost of checking output going uncounted alongside the cost of producing it.
Practitioners were recruited through a public journalist query in August 2026, and five submissions were excluded. Autonomy was coded against a published five-level scale by a single analyst. The sample is self-selected and descriptive of these 40 deployments rather than an estimate of any wider population. Function-level figures are based on seven to 14 responses and should be interpreted as directional findings within this sample.
SumatoSoft is a custom AI and software development company with headquarters in Boston and a development center in Warsaw. The company has delivered more than 350 projects across more than 25 countries over 14 years and holds ISO 9001:2015 and ISO/IEC 27001:2022 certifications.