Developing Products With AI: The Factory Handoff

What AI actually compresses

Product development has two halves that behave nothing alike.

One half is made of language and iteration: requirements, specification drafts, industrial design variations, a first-pass bill of materials, test plans, documentation. AI has genuinely collapsed that half. Work that took a small team weeks now takes an afternoon, and the quality of a competent first draft has gone up, not down.

The other half is made of physics and lead time: steel, EMC chambers, certification queues, component allocation, the 10,000th open-close cycle of a zipper. None of that has moved. It will not move, because the constraint there is not how fast an answer can be written.

That asymmetry is the whole story. The front end got faster and the back end did not, which means the handoff between a brand and its manufacturer now happens earlier while looking more finished than it is.

Three artifacts that arrive polished and unvalidated

The specification. AI writes a fluent spec in an hour. Fluency is not the same as frozen. A specification that reads complete but still moves in month three costs exactly what an obviously incomplete one costs, and it is harder to argue with because it looks settled.

The industrial design. Renders are close to free now. A render does not know about draft angle, wall thickness, parting lines, or where the ejector pins land. A shape that is beautiful and unmoldable is generated at the same speed, with the same confidence, as one that is beautiful and moldable. Nothing in the output distinguishes them.

The bill of materials. A plausible BOM is easy to produce and every part on it will be a real part. Whether all of them can be bought, in your quantity, inside your window, is a different question. It is also the question that sets your launch date.

None of this is an argument against using AI. It is an argument about what the output is: a strong hypothesis, produced faster than ever, that has not yet been tested against anything.

Physics does not accept a confident answer

This is where a structured review earns its cost, and it earns more now than it did three years ago, because the volume of plausible-looking input has gone up.

We run a design-for-X review before any steel is cut: 33 check items, each rated red, amber or green, and a single red stops the project. Not flagged for later. Stopped.

On one portable UV-C soft-bag product, that review returned six reds. EMC pre-scan had to move to two weeks after the engineering build rather than before mass production. The waterproof zipper needed a 10,000-cycle open-close test. IEC 62471 light-leakage testing had to be scheduled. A reed switch plus firmware double interlock was required rather than a single safety path. A 275nm UV-C LED on an 8 to 12 week lead time needed a second source. And a direct USB supply had to be reworked, because it sidestepped UN38.3 and IEC 62133 entirely.

An AI-drafted design will not fail this review more often than a human-drafted one. It will fail it in the same places, for the same reasons, and it will arrive at the review with better documentation explaining why it should not have.

→ The full method and case: DFX Risk Review Before Cutting Tooling

The lead time nothing shortens

Every bill of materials gets graded for supply risk before tooling, and the grading is where an elegant design meets the calendar.

On that same UV-C project it came out as follows. The 275nm UV-C LED was red: special specification, 8 to 12 week lead time, single source. The UV-resistant reflective fabric and the fan were amber. The waterproof zipper was green, single brand but stable, ordered six weeks ahead. The PCBA and MCU were green, two suppliers across two platforms.

One red component sets the schedule for the entire product. If you compressed six weeks of specification work into two days and then selected a part with a twelve week lead time, you did not save six weeks. You moved the waiting somewhere less visible.

→ The four grading axes and what each level triggers: Component Supply Risk Matrix for OEM Projects

Where a reversible decision becomes steel

Everything upstream of tooling is cheap to change. Everything downstream is not. Changing a drawing costs an afternoon, changing steel costs weeks and five figures, and changing a product already in customers’ hands costs a recall.

Two numbers govern that moment, and they are exactly the two an AI cannot infer from your brief: expected annual volume and target market. Volume decides cavity count and cavity count decides tooling cost. Target market decides the certification path, and the certification path decides which components are even eligible for the BOM you just generated.

A team that arrives with a beautiful specification and no committed answer to those two questions has automated the easy part of the problem.

→ How cavity count, mold steel and changeover economics turn into a quote: Tooling Cost and MOQ Explained for OEM Projects

What a good AI-assisted handoff looks like

The teams who get the most out of this are not the ones who move fastest through the front end. They are the ones who spend the saved time on breadth.

Bring three directions, not one. Exploring three concepts to the point where each can be risk-reviewed used to be unaffordable. It is affordable now. Killing two of them on a review sheet costs a fraction of killing one of them after tooling.

Bring the assumptions, not just the output. The useful part of an AI-drafted specification is often the reasoning behind a choice, which is also the part most likely to be quietly wrong. An engineer can check a stated assumption. Nobody can check one that was never written down.

Freeze earlier, not later. Faster iteration makes it tempting to keep the specification open, because reopening it now feels cheap. It is cheap upstream of tooling and ruinous after it.

Let the review stay slow. The temptation is to compress the design review at the same ratio as everything before it. The review is not a bottleneck to optimise. It is the only place the hypothesis meets reality before steel does.

→ The five stages and the failure that surfaces in each: Where OEM Product Development Goes Wrong

Three questions before you hand over an AI-drafted design

1. “Which part of this specification has never been checked by a person?” Not a criticism of the tool. A map of where the risk is concentrated.
2. “Which component on this BOM worries you most, and why?” This tests whether anyone at the factory has actually read your design, and it surfaces the lead time that will set your launch date.
3. “At which stage gate would you stop this project?” A partner who has never stopped a project has never protected anyone from one.

We build under ISO 13485:2016 and ISO 9001:2015 quality systems, with engineering, tooling, PCB and PCBA design, SMT and assembly on one site in Taichung, Taiwan, where we have been since 1996. Engineering changes do not queue behind a subcontractor, and a design that needs three revisions before it is manufacturable gets three revisions.

NDA first, and an engineer, not a sales script, replies within two business days.