A purchase order is not a manufacturing plan
You send Gerbers, a BOM, a pick-and-place file and a 3D model of the enclosure. Six weeks later a box arrives. Inside are twenty units, and three of them do not boot, the paint is a shade off, and one connector is seated at an angle nobody can quite explain. You email the factory. They reply that everything matched the files.
They are usually right. That is the part founders do not expect.
Moving from prototype to production is not a handover, and the people who treat it as one lose the most time. Files go out, samples come back, the samples are wrong in some way nobody wrote down, and then you iterate. Tooling gets adjusted. Tolerances get argued about. Somebody has to look at what came off the line and be willing to say no.
What "wrong in a way nobody wrote down" means
Your drawing calls out a 4.0 mm boss. The moulded part measures 4.3 mm because the shrink rate for that resin in that tool ran differently than the simulation predicted, and the tool was cut a touch conservative on purpose, because steel comes off easier than it goes back on. Nothing in your drawing covered that. Nobody lied.
Same thing on the electrical side. A PCBA production line will happily build a board with a footprint that is technically correct and practically unassemblable, or substitute an 0402 resistor from a different manufacturer with the same value and a different temperature coefficient, or reflow a connector whose body cannot take 245 C because nobody flagged it as a hand-solder part. Every one of those is legal against the files you sent.
The work is in the gap between what you specified and what you assumed. A DFM review before the first pilot run is how you shrink that gap: clearances, panelisation, test points, fiducials, whether the stencil apertures will actually deposit paste where you want it. We would rather argue about a 0.15 mm clearance in KiCad on a Tuesday than discover it in a tray of dead boards.
Who inspects, and do they care
Manufacturing works when the person doing the QC owns the outcome. It fails when QC is a checkbox on someone else's job description.
The practical version: whoever inspects the first article must be the same person who will have to explain a field failure. That alignment is what makes them willing to reject a batch. A quality inspector who has no downstream consequence will pass a part that is marginal, because passing is faster and rejecting starts an argument. It goes wrong in the other direction too. Sometimes the tolerance itself is unreasonable, held only because a drawing said so, and that costs weeks as well. Someone has to be qualified to make that call and accountable for it.
For PCB fabrication and PCB assembly, that means a real first-article inspection before the run: AOI results reviewed by a human, a functional test fixture that exercises the actual product behaviour rather than just continuity, and a boundary-scan or programming step that confirms the firmware went on and came back with the version you expect. For injection moulding and CNC machining, it means measuring the features that matter with the tool that can measure them, not eyeballing a part against a print.
The samples will be wrong. Plan the loop, not the delivery
Founders schedule tooling as a single line item: "tooling, eight weeks". The real sequence is tool cutting, T1 samples, measurement, a tool modification, T2 samples, and sometimes T3. Each round is real calendar time - commonly two to four weeks a round once you count shipping both ways. A schedule with one sample round in it is a schedule that assumes the first samples are right, and they are not.
The same is true for boards. Rev A finds the mistakes. Rev B fixes those and finds subtler ones. Rev C is usually what you can sell. Budget for that. Funding that covers rev A is funding for a prototype.
Two things that make the loop shorter, and both are cheap. Send the factory your problem, not only your files - tell them what the part must survive and what it must cost, because they know what their machines do well. And freeze the enclosure before you cut steel. The most expensive mistake we see is not a bad circuit. It is paying for injection mould tooling before the mechanical design is finished. That is a four or five figure cheque, it is steel, and no firmware update fixes it.
Ordering the parts is a schedule decision
A five cent part with a forty week lead time stops a production line exactly as hard as the microcontroller does. Check availability before you fall in love with a component. Design in a second source where the design allows it, and if it does not allow it, know that and buy ahead.
Minimum order quantities bite in the same place. A connector with a five thousand piece minimum, when your pilot run is two hundred, is not a small problem - it is either dead stock or a redesign. Work the landed cost out before the design freezes: parts, assembly, packaging, shipping, duty. The bill of materials is the business model, and the pilot run is when you find out whether the arithmetic was honest.
EMC testing is a design input. Failing it at the lab is not fixed with a sticker; it is ground planes, filtering, cable routing and shielding, which is a respin and a new lab booking. If there is a lithium cell in the product, read UN 38.3 now rather than when the units are in a warehouse.
One point of contact, or three arguments
The failure mode we see most in this phase is not technical. It is that the electronics vendor, the mechanical shop and the assembly house each answer only for their own scope, and the problems that matter live between them. The board that does not fit. The antenna sitting next to a metal bracket. The thermal issue that is mechanical's on Monday and electrical's on Tuesday.
Someone has to own the seams. If the answer to who owns them is "everyone", the answer is nobody, and you will find out at first article.
If you want a second opinion on a design that is about to go to tooling, send it to us and we will tell you what we see.

