We have a three-year-old codebase maintained by two senior developers who care a lot about best practices, security, testing, and consistency. They are also equal partners in the company, so any genuine efficiency gain benefits everyone directly.
I'm a PM, and about nine months ago I asked whether I could use an AI coding assistant to contribute small changes. The general view was that reviewing AI-generated code can be harder than writing it, meaning my pull requests might simply create extra work for the developers.
The models and tooling have improved since then, so I'm looking for an honest assessment. Is it reasonable for me to submit small, AI-assisted pull requests if every change receives thorough developer review, automated testing, and QA before merging? Are there situations where bypassing a pull request would ever be acceptable, or should every change go through that process?
The motivation is that the developers are busy with other projects, and the information gap between me and them can create a lot of back-and-forth. I'd like to reduce that friction, not create more work. We're all partners, and the developers are equally free to improve PM-side work with AI or code when appropriate.
5 Answers
A good way to start would be acceptance tests, documentation, and small visual or workflow changes. Those can provide useful context and save developer time without asking them to reverse-engineer a complex implementation. Establish guardrails such as size limits, required automated checks, no self-approval, and an agreed list of areas that are off-limits.
The model version is less important than the workflow. Give it the repository’s patterns, reusable utilities, architecture notes, and coding rules instead of a vague task description. Have it stop and produce a handoff when the change becomes complex. If the resulting pull request is difficult to review or requires repeated correction, that is a signal that the task should have gone directly to a developer.
The important question is whether you can personally understand the code and take responsibility for it after it ships. AI can produce code that looks convincing while violating project conventions, duplicating existing utilities, or introducing subtle bugs. If developer feedback is just being passed back into the model without you learning what changed or why, you’re effectively outsourcing the review burden to the senior developers.
Never bypass the pull request process. Even experienced developers should have changes reviewed, tested, and merged through the normal safeguards rather than pushing directly to the main branch. AI-generated code makes this more important because plausible-looking mistakes can be easy to miss without a second set of eyes.
Small, low-risk changes can be reasonable: documentation, tests, simple UI adjustments, or clearly isolated fixes. Keep the pull requests narrow enough that a developer can understand and verify the entire change. For anything involving core business logic, security, shared abstractions, or a large amount of generated code, the review burden can easily outweigh the benefit.
Related Questions
Biggest Problem With Suno AI Audio
How to Build a Custom GPT Journalist That Posts Directly to WordPress