In short
Nobody should hand an AI authority over a tape-out date on day one, and nobody should have to. Authority on a program is earned the way a new engineer earns it: by being right, visibly, about small things first. We built that into the product as a rule, not a setting.
The word I distrust
Every AI pitch I hear now uses the word autonomous. I wrote two papers with that word in the title, so I am not against it. But in a chip program the word is doing a lot of work it has not earned. A VP of Engineering is not going to let software re-sequence a tape-out because a vendor said so. Nor should they. The question is not whether an AI can act on a program. It is how it earns the right to.
How a person earns it
Think about how a new program manager earns authority. First they observe: they learn where the program's facts live and what normal looks like. Then they diagnose: when something moves, they say what it breaks, and they are right often enough that people start to ask them. Then they recommend: they bring options with costs, and their recommendation is usually the one taken. Only then do they act on their own judgement, within limits everyone understands, and the limits widen as the record of being right grows.
That is four stages, and the step from each to the next is a decision someone else makes about them. We built Tymeline's AI to work exactly that way, because it is the only way I would accept it on a program I was responsible for.
The envelope, not the switch
Concretely: every kind of action an AI Employee can take sits inside an authority envelope that a named person set. Posting to the engineering channel might be inside it. Re-sequencing a sprint in Jira might be inside it after approval. Approving a tape-out gate is never inside it; that is a rule enforced at runtime, not a preference. Each action leaves a signed receipt: what was done, on what evidence, under whose approval.
The envelope is widened one kind of action at a time, by the customer, on the strength of the receipts. There is no switch marked autonomous. There is a record, and a person reading it, and a decision to trust the next thing.
Why this is the fast path, not the slow one
People assume the cautious design is the slow one. It is the opposite. A program that has watched an AI be right about small corrections for a month will let it make them without asking, and that is when the time starts coming back: the chasing, the re-planning, the first draft of the customer note, all done the hour they are needed. Trust that was earned compounds. Trust that was asserted gets switched off the first time it is wrong.