Most AI agents still need a human to check their work

AI agents are tools that carry out tasks on their own, such as writing code or handling support requests. A new study from software firm SumatoSoft suggests most still need a person to check what they produce. Of 39 deployments rated for autonomy, 25.6% ran unattended. And 32 of 40 teams, or 80%, paused, narrowed or switched off an agent after launch.

What the study found

The study looked at 40 live deployments in engineering, sales, support and IT operations. It separated two things: work an agent produces, and work it is allowed to release without human approval. Among 32 respondents who gave coverage figures, the median agent covered 80% of its workflow. Twelve said the agent handled the entire workflow they named.

Writing is cheap, checking is not

The clearest example came from software teams. Among 14 engineering respondents, the median share of agent-written code reaching the main branch without human review was zero. The main branch is the official version of a project's code. Eleven of the 14 said no agent-written code got there unread, even though agents wrote most of their code.

Katerina Merzlova, the study's author, put it this way: "Writing the work got cheap and checking it did not." Beck Wang, founder of IfShe, said unattended authoring is near-total, while unattended release is close to zero, on purpose.

Where agents went wrong

The study described 32 failures. The most common patterns were:

• 11 involved scope or context blindness, meaning the agent missed the bigger picture.
• Eight involved an agent reporting success when the work was wrong.
• Three passed automated checks but missed the intended goal.

Daniel Battaglia, founder and CEO of Parksy, said agents produce plausible work rather than obviously bad work, and plausible work survives a quick skim.

How much weight to give it

The sample is small and self-selected. Participants answered a public journalist query in August 2026, and five submissions were excluded. A single analyst rated autonomy. Figures for each function rest on seven to 14 responses, so they are directional. SumatoSoft also sells custom AI and software development. The study does not dispute Gartner's June 2025 forecast that more than 40% of agentic AI projects will be canceled by the end of 2027. It offers one possible cause: the cost of checking output going uncounted.

What to do

Anyone planning to use an agent should count the time and money for reviewing its work, not just the price of the tool. Decide who checks the output, how often, and what the agent may do without approval.