The best results come when AI gets the project's context, a clearly stated task, and verifiable criteria for "done." Then it helps you explore unfamiliar code faster, prepare changes, and document decisions.
The engineer stays in control
Architecture, security, data handling, and verifying the result all require human accountability. AI strengthens the practice -- it doesn't replace it.
Where AI genuinely helps
The biggest wins come on tasks with a high "reading cost": making sense of an unfamiliar part of the codebase, drafting a migration from a description, proposing a few implementation options, writing routine tests, turning scattered notes into readable documentation. These are tasks where a mistake is cheap and easy to catch.
Where AI won't replace an engineer
Decisions about architecture, data models, where a service's responsibilities end, and what can be simplified without losing reliability -- these are decisions with long-term consequences. AI can suggest options, but the choice, and the accountability for it, stays with the person who understands the business context and the cost of getting it wrong.
How to set up the process
The working pattern is simple: frame the task so it can be objectively verified; give the AI access to the project's real context, not an abstract description; always read and test the proposed changes before they reach production; and treat the result as a colleague's draft, not a finished solution.