93% of Developers Have Merged AI Code They Did Not Fully Understand, and Only 13% of Teams Have a Plan for When It Breaks

The headline finding in TestMu AI's State of AI in Quality Engineering 2026 report is a number that most engineering leaders will recognize and few will enjoy reading: 81% of software teams have merged AI-generated code or tests they did not fully understand. The survey was released on November 14 by TestMu AI, the company formerly known as LambdaTest, and it describes a gap between how much code AI now writes and how little of it is verified independently.

The scale of the shift shows up first. At 40% of teams, AI now writes more than half of production code. Ninety-seven percent of teams either use AI in testing or plan to. And 44% report more escaped defects or production incidents as AI-written code has grown, which is the part that turns a productivity story into a quality story.

Verification has not kept pace. Only 41% of teams run any automated check on AI-written code before merge. Most rely on manual review, at 55%, or on a hand-off to a quality assurance team after the code is written, at 51%. The report finds a measurable difference between teams that run separate continuous integration tests and teams that do not: 49% of the former group report rising escaped defects, against 38% of the latter.

Developers are the most exposed group in the data. Ninety-three percent say they have merged AI output they did not fully understand, a higher share than the team-level figure, which suggests the practice is routine rather than exceptional. The report also finds that using one model to grade another widens the gap instead of closing it: teams relying on LLM-as-a-judge merge AI output they do not understand 29% of the time, roughly twice the 15% rate of other teams. For AI-generated tests, the median share described as almost right but not quite is 65%.

Agents get shipped on weaker evidence still. Seventy-seven percent of teams are building AI agents that need testing, but 37% of those teams have no systematic way to evaluate them, and only 28% test agents against users who deliberately try to break them. Budgets follow the same bias toward production over proof: 50% of teams spend most of their AI testing budget on generating AI output, while 17% spend most of it on verifying that output. Only 13% have an incident-response plan for AI errors.

One caution belongs in any reading of these figures. TestMu AI sells AI testing and quality engineering tooling, so the report is a vendor-commissioned survey with an interest in the conclusion, and the release does not disclose the sample size, the sampling method or the regional breakdown. The specific percentages should be treated as directional rather than as precise population estimates. What survives that caveat is the shape of the finding, which lines up with independent work on the same problem.

Asad Khan, co-founder and chief executive of TestMu AI, frames the fix as structural: any output AI produces should be validated by a system independent of the model that produced it, before it merges and every time it changes. That is a reasonable reading of the data, since the teams reporting fewer escaped defects are the ones running separate checks rather than asking the same model family to grade itself.

The commercial consequence is that verification is becoming the constraint on how much AI-written code an organization can safely absorb. Generation is cheap and getting cheaper, and every team now has access to it. The teams that pull ahead will be the ones that can answer, for any given change, which independent check ran, what it covered and what happened when it failed. On this survey's numbers, most organizations cannot answer that question yet, and a tenth of them have not written down what they would do if an agent broke something in production.