Can AI help you make a physical product without losing your savings?

AI makes designing a physical product look cheap. Making it is still where the money goes. What the research says about AI-designed parts, why crowdfunding is the riskiest first step, and a first order you can place for a few dollars.

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

The agent, drawn as a hopper, pours drafted circuit boards onto a sieve labelled checks. Most pile up on the sieve; a person labelled You stands at a laptop wired into it. Two boards fall through into a small open box labelled first order, cheap to try, and its flaps close. To the right, a large crate marked with a dollar sign and labelled anything that needs capital rolls along the ground and jams against a gate labelled making, short of a second gate labelled delivery.

Drafting got cheap and checking did not, so this reading puts one person at the checks with a small first order, and away for now from anything that needs capital. A reasoned guess, not a result; the counts here are illustrative.

Contents

You have an idea for a product: a gadget, a part, something you'd sell to your own customers. AI tools now promise to design it for you in an afternoon. So why not design it, raise the money and get it made?

Because that's the path where people lose their savings. I read the research on AI-designed hardware, crowdfunding outcomes and the companies that tried to make products for others at scale. Three questions decide whether you'll be one of the people who lose.

Can AI really design it? Where does the money actually go? And what's the cheapest way to find out?

Can AI design it?

It can draft it. That's not the same thing.

On tests that check whether an AI-designed circuit board actually works, not just whether the design file opens, the best result passed 8.15% of the time. In one documented attempt, an AI nearly picked a timer chip that can't run on the coin battery it was meant to use. A person and a rule-checker caught it.

An agent, drawn as a hopper, pours 100 drafted circuit boards onto a sieve labelled design rules plus simulation. 92 boards pile up on the sieve, caught by a check. 8 fall through into a tray below, passing every check.

Drafts are cheap; most of them stop at the checks. On HWE-Bench, a set of 300 board projects, the best model passed every check 8.15% of the time, drawn here as 8 in 100.

So the obvious plan, letting the AI design it and sending it to a factory, fails at the first step.

The checks that stop most drafts are software a person can put in the loop. What looks cheap is the drawing. What still costs money is checking it.

The same agent hopper and sieve of checks. A person labelled You stands at a laptop wired into the sieve. Caught boards sit on it. One board, tagged via almost on a pad, rides a dashed chute back up into the hopper, redrops and passes. Below the sieve, 5 boards that passed sit in a small box labelled first order, 5 boards from $2.

The checks that stop most drafts are deterministic software, and a person can put them in the loop so what they catch goes back to the agent. The via is from one documented attempt; the heap and the chute are illustrative. The price is one vendor's advertised starting price.

Where does the money go?

The second obvious plan is to crowdfund it. That's where the losses are.

In one study of successful Kickstarter projects in design and technology, only 24.9% delivered on time. The founders blamed manufacturing, shipping and certification, not design. Quirky and Shapeways, two companies built to make other people's products at scale, both went bankrupt.

If you take backers' money before you've made one working unit, you're promising something nobody has checked. If it's late, it's your name and your savings on the line.

20 crates, one per funded project, roll along a road under the heading anything that needs capital. They pass a faint gate labelled design without stopping, then queue at three gates under a bracket labelled trouble clustered here: 7 at making, 5 at shipping, 3 at certification. 5 green crates reach a group of waiting backers, labelled 5 on time, about 1 in 4.

In one Kickstarter sample of successful design and technology projects, 24.9% delivered on time, drawn here as 5 of 20. Founders named manufacturing, shipping and certification; how the late crates split across those gates is illustrative.

What to skip for now

Anything that needs investors before you've made one working thing. The money in this field is going to large factory companies (one raised $1.37 billion in August 2026), not to one-person shops. I found no small team doing well at it yet.

Coins pour from a funnel into a large factory until all five of its windows are full: one factory raise, $1.37 billion. Next to it, a dashed tray for small teams stays empty, labelled none found yet. A person labelled You watches, then turns away and closes the flaps of a small green box: a small first order.

The money in this field goes to large factory companies, and no small team turned up doing well at it yet. Start with a small first order instead. The coins are illustrative.

The cheap way to find out

Make something small you'd use yourself first, not a product.

A small circuit board is advertised from $2 for five, with assembly from an $8 setup fee (one vendor's starting prices). Design it with an AI, and run the free checks before ordering: the design-rule check, a simulation and the vendor's own manufacturability check. Write down what each one caught.

Read the costs before you click. If you're in the US, tariffs on Chinese parts ran from about 35% to 92.5% as of March 2026, by one vendor's own FAQ.

Then order the smallest batch they allow. Nobody has measured how often first orders go wrong, so your notes will answer the question nobody else has.

A person labelled You stands at a bench beside a notepad titled your log. Four lines write themselves: check, timer chip, not for 3 V; check, via almost on a pad; limits, minimums, lead time, tariff; order, what went wrong, followed by a blank. A small box of 5 green boards sits on the bench. A copy of the log shrinks into a book and files into the one empty slot on a shelf of sources, under the note no source measures how often first orders fail.

The entries are from the post: one documented attempt caught the timer chip and the via, and the limits are the ones to read before ordering. The shelf is illustrative.

What this says about how we work

When a field is loud with promises, we read the evidence before the hype, and we say plainly where the money goes and where the risk sits, with the sources to check. That's the same reading we'd give your idea before you spend a dollar on it.

Three promises float in as balloons: AI designs it in an afternoon, crowdfund it, and a factory for anyone. One at a time each reaches a gauge labelled the evidence and pops. What passes through is a small stone, which drops onto the lab's bench with its number: 8.15% pass, 24.9% on time, 2 bankrupt. A lamp lights for each one, and the lab stands at the bench reading the three stones.

How we work: let the promises pop and keep what has weight, with its source. The numbers are from the research behind this post.

The full research, with every figure and source, is here. If your first order goes wrong in an interesting way, I'd like to hear about it: mackenzie@mackenziebowes.com or @_mackenziebowes.