Blog · · 4 min read

What AI can and cannot do when you build a hardware product

An honest look at where AI already helps turn an idea into a physical product, from research to CAD and electronics, and where engineers still have to check every step.

Software founders can now go from an idea to a working prototype in an afternoon. Hardware is different. A product has to exist in the physical world: parts need dimensions, tolerances and suppliers, a circuit board has to pass its design rules, and somebody has to pay for the first batch. So how far can AI take you today?

We build makemachine.ai to find that out in public. Every step of a project – research, patents, concept, mechanical CAD, electronics, simulation, sourcing, manufacturing files and a sales deck – is done by AI modules, and the results show up on our live projects page as they come in. Here is what we have learned so far.

Where AI is already strong

Research and briefs. Language models are excellent at turning a vague wish into a structured brief: who is it for, what does it have to do, which constraints matter, which existing products already solve part of the problem. They read datasheets, standards and patents faster than any person and summarise them with sources.

Concepts and variants. Asking for three different concepts with their pros and cons used to take a design workshop. Now it takes minutes. The value is not that the AI picks the right one, but that it lays out options a team can argue about.

Code that produces geometry. Modern models are good programmers, and that matters, because CAD can be code. A model writes a parametric script, runs it, renders the result and fixes its own mistakes. What comes out is a real STEP file, not a picture of one.

Essential but boring documents. Bills of materials, assembly notes, test plans, risk assessments and pitch decks are structured writing. AI drafts them well and keeps them consistent when the design changes.

Where AI still struggles

Spatial reasoning. A model can write the script for a bracket, but it does not "see" whether two parts collide or whether a screw can actually be reached with a screwdriver. It needs renders, collision checks and simulation to catch its own errors, and even then it misses things a mechanical engineer notices at a glance.

Real-world numbers. Prices, lead times and stock levels change every day and live behind supplier websites, logins and captchas. Estimates are useful early on, but real quotes still come from real suppliers.

Tolerances and manufacturability. Whether a part can be milled, moulded or printed at a sensible cost depends on details that rarely show up in training data: tool radii, draft angles, minimum wall thickness, what a particular shop can do. Checklists help. Experience still wins.

Long projects. A product is hundreds of decisions that depend on each other, and models have a limited working memory. A useful system has to keep the state outside the model: versions, inputs, results and open questions, all written down where the next step can find them.

How we work around the gaps

Three principles have made the biggest difference for us.

  1. Real tools, not just chat. Each module runs actual engineering software – CAD kernels, circuit design tools, finite element solvers, renderers – and has to deliver files that open in standard programs. If a script fails, the module sees the error and tries again.
  2. Small steps with checkpoints. A module only starts after a person has checked what goes into it. Every version records what went in (which earlier results, which requests) and what came out. Nothing is ordered or sent to a manufacturer without explicit approval.
  3. Building in the open. Public projects show their results as they arrive, including the ones that turn out to be wrong. That keeps us honest and turns mistakes into input for the next version.

You can watch this happen right now: a modular charcoal grill, a small jumping robot and a project around an industrial robot arm are being built module by module.

What this means if you want to build something

AI will not replace the engineer who signs off on a design. But it changes the economics of the first 80 percent: the research, the options, the first CAD model, the first circuit, the first cost estimate. Work that used to take weeks and a team now takes days, and it is cheap enough to explore ideas that would never have been worth a workshop.

The last 20 percent – testing, certification and the details that make a product safe and manufacturable – still needs people who know what they are doing. The trick is to let AI do the first part quickly and in the open, so those people can spend their time where it really matters.

Curious how far it gets? Follow the live projects or sign up free to get new results and articles by email.

Watch hardware get built by AI. Follow public projects on makemachine.ai, read every result in full and get an email when there is something new.

Get started See the live projects

More from the blog

Open-hardware projects built with AI