What AI actually changes about hardware development — and what it doesn't

The honest gains on a hardware program are not the ones in the demos. They are unglamorous: drafting and iterating requirements documents, first-pass CAD concept generation, generating test matrices and DFMEA drafts, writing firmware scaffolding, and summarising supplier quotes into comparison tables. Together these pull one to two months out of a program phase that used to be pure desk work.

AI-assisted simulation is the second real gain. Surrogate models trained on CFD and FEA results give 80%-accurate answers in seconds instead of overnight runs, which changes behaviour: engineers explore ten design variants before lunch instead of committing to the first one because iteration was expensive. On chassis and thermal packaging, that breadth of early exploration is worth more than any single perfect analysis.

Generative design deserves the same honesty. It is excellent at exploring topologies within a defined envelope — brackets, housings, cooling structures — and nearly useless at proposing anything that respects manufacturing constraints you didn't state. The value comes from a senior engineer framing the constraints correctly, which is the part AI cannot do.

Hardware is still hardware: a bad pack layout fails a nail-penetration test the same way it did five years ago. Physical validation, supplier qualification, tooling trials and homologation are untouched by AI, and they remain the majority of the calendar on any vehicle program. Teams that assume AI compresses the whole schedule end up compressing the design phase into a validation phase they can no longer afford.

The other unchanged reality is judgement under ambiguity. Choosing whether a squeak is acceptable, whether a supplier's capability data is trustworthy, whether to hold a design freeze two weeks — these are decisions made with incomplete information and relationships, and no model is making them for you.

The programs that benefit most reorganise around AI rather than sprinkling it in. Concretely: analysis moves earlier and runs continuously instead of in review-gated batches, so design reviews happen against live simulation data. Firmware teams use AI code generation for the 70% of embedded work that is plumbing, and spend the saved time on the safety-critical state machines where generated code gets reviewed line by line, never shipped blind.

Documentation quality also rises, which matters more than it sounds. With AI drafting, specifications get updated weekly instead of quarterly, and a current spec is the difference between an outsourced workstream that works and one that doesn't.

We treat AI as leverage on engineering hours, not a substitute for engineering judgement. The stack we run day to day covers CAD concept exploration, surrogate thermal and structural models, DFMEA and test-plan drafting, firmware scaffolding, and bilingual program documentation — which on Japanese OEM programs is itself a material time saving.

The rules we hold to: generated designs get validated against physics before anyone falls in love with them, generated code gets the same review as human code, and no AI output enters a deliverable without a named engineer owning it. That last line is what keeps the acceleration real instead of creating a new class of program risk.