Which tasks do developers feel confident with releasing to AI?
_There’s a tremendous amount of production code being produced by AI at the moment. But There’s also a considerable difference between permitting AI to do something and not wanting to bother to verify the outcomes._ _Let's see what some of the most recent developer research on this topic tells us from the following sources: Stack Overflow, Microsoft, DORA, and 2026 industry studies._ More people are using than trusting Respondents answered a recent SO survey about their experiences with AI developers' work, of which 84% indicated that they have integrated AI tools into their workflow, while merely 33% trust the output to be accurate. As can be seen from the results, the developers with experience showed significantly higher distrust. Developing a few effective delegations is easy Software engineers have participated in a DORA study, and here's where AI is most helpful: software testing, code generation, code review, information retrieval, software prototyping, debugging, software documentation, and software refactoring. That is, things that you can check against something real (tests, compiler problems, prior code, documentation, or a known specification). The way people want AI to deal with craft decisions, however, is different: they want to leave the job for developers entirely. So Microsoft Research polls 860 Microsoft developers about their hopes for the next generation of development tools for AI. The one thing that was most notable was bounded delegation. That is, developers would like to give AI that the assembly aspects of software development, and allow humans to manage provenance, authority scoping, uncertainty signaling and least-privilege access. Verification is a danger zone According to a study by Black Duck released in 2026 and featuring a survey of 831 software engineers and DevOps professionals, 92% of professionals reported seeing gains in productivity or release velocity as a result of AI coding assistants. But 90 percent of the ones indicated a point of issue in the workflow as well. The greatest challenges were: Manual review (52%), Security testing (51%), and Code rework (48%). The same pattern emerges on DORA as well – AI can generate code quickly, and teams end up with more changesets to review and more change to introduce instability. Production - still of huge importance According to a September 2026 survey of some 1400 developers and software quality professionals conducted by SmartBear, 46% of the teams had already deployed AI-generated code that was ultimately not successful in production. Interestingly, almost half of the organizations reported that they reported that they were unable to explain how AI contributed to a bug or production incident when one hit production. _Therefore, as a growing percentage of codebases is written automatically by AI, the importance of traceability and review is growing._ Questions for the community: _What kind of things do you want an AI agent to deploy if you weren't going to look at the code yourself?_
- #technology
- #ai
- #software
- #developers