The engineer gets the last word

The same engineer is liable as before

The same engineer as before. When a deliverable leaves an engineering firm, a named person has reviewed it, decided it is right, and put their signature on it. That person answers for the work: to the client, to the reviewing colleague, to their own profession. Bringing AI into the production of that deliverable changes none of it. The engineer still reviews. The engineer still decides. The engineer still signs. The professional chain, from the designing engineer through internal review to the sign-off toward the client, does not move an inch.

This is worth saying plainly, because it is the first question every engineering leader asks about AI, and it deserves a straight answer rather than a hedge. Responsibility stays where it has always been. The interesting question is the second one: what does that signature now stand on?

None of this is legal advice, and it does not need to be. It is a description of how engineering deliverables are actually produced and approved: a qualified person takes responsibility for the result, and everything else in the process exists to make that responsibility bearable. The question was never whether a human answers for the work. The question is how well supported that human is at the moment they sign.

Last year the signature rested on a tired skim

Be honest about the old arrangement. A calculation is produced by one engineer, usually under time pressure. It is then reviewed by a colleague: a competent, conscientious colleague, reading someone else's numbers on a Thursday afternoon, between two deadlines of their own. The busier the project, the later the review lands, and the less time it gets. Perhaps a second reviewer follows, working under the same conditions. Serial human review, each pair of eyes assuming the previous pair caught what mattered.

None of these people is careless. The problem is the arithmetic of attention. A reviewer cannot re-derive every figure in a hundred-page deliverable, so they sample: they check the parts experience says are risky and trust the rest. Most of the time that works. Sometimes it does not, and the error that slips through does so past three signatures, not one.

The quality system is not the villain either. It was built from the only reviewing instrument available: human reading time. That instrument is scarce, expensive, and tired by Thursday afternoon, and no quality plan can conjure more reading hours out of a fully booked week.

With machine review first, the signature is better supported

Two things sit beneath a signature: who gives it, and what has been verified before it is given. The first does not change. The second changes completely. When the calculation chain behind a deliverable has been recomputed by a machine, every figure rather than a sample, the reviewing engineer is no longer the first line of defence against arithmetic errors. They are the judge of work that has already been checked line by line, and their attention goes where it earns the most: the assumptions, the boundary conditions, the judgment calls a machine cannot make.

The pattern is familiar to anyone who has shipped a deliverable under deadline: a calculation can pass one author and two reviewers and still carry a small error, not because anyone failed at their job but because serial human review has a ceiling and a busy project reaches it. We took that mechanism apart in the error that slipped past two reviewers. A machine that recomputes every figure and flags deviations does not replace the reviewers; it hands them a shorter, sharper list of the things that deserve their judgment.

That is the whole shift, in two verdicts. The signature: still human. The catch-rate: superhuman.

Liability is a reason to start, not to wait

It is usually put that way: the rules are unclear, so hold off. But look at what the choice actually consists of. Wait, and the deliverables you sign next year rest on the same serial human review as the ones you signed last year. Put machine review underneath, and the same signature, given by the same engineer under the same responsibility, sits on top of a calculation chain that has been verified in full.

Nothing in the professional process has to move for that to happen. No new liability arrangement, no renegotiated contracts, no waiting for a regulation to name the technology. The engineer gets the last word, exactly as before. The difference is how much has been checked before the last word is spoken. The signature you give with machine review underneath is safer than the one you gave last year. Liability is not a reason to wait. It may be the best reason to start.