I use AI every single day in my business, and I'd be foolish not to.
If you're developing a product right now, you should be using it too, as much as you possibly can.
But there's one line AI can't cross, no matter how good the models get. Knowing where that line sits is the difference between AI saving you weeks and AI costing you months.
The line comes down to the cost of being wrong. In software, a wrong answer is cheap because you just patch it and move on.
In hardware, a confidently wrong answer can mean a board respin or a failed round of certification testing. Sometimes it means a production order you can't unwind.
And AI sounds exactly as confident when it's wrong. It can't tell you which of its answers land in the small fraction it's hallucinating.
That's why I follow one simple rule: use AI where being wrong is cheap, and use experienced engineers and product experts where being wrong is expensive.
On the cheap side of the line, AI is incredible. I use it to explain concepts I'm learning and to write first-draft code.
It'll summarize a 300-page datasheet so I know exactly where to read closely, and it'll surface options fast once I already know the question I'm asking. I even use it to pressure-test my own reasoning, like a colleague I can think out loud with.
Every one of those uses shares the same trait, a wrong answer costs me nothing because I'll catch it or iterate right past it.
Component selection sits on the other side of the line. AI is trained on the internet, and the internet is mostly hobbyist content.
So when you ask it to pick parts, it hands you the popular hobbyist answer instead of the product-grade one. That's how an Arduino ends up in a production design, or how the module every blogger loves ends up in a product that has to pass certification and ship at volume.
I've watched people arrive at my Hardware Academy with designs they built almost entirely with AI, and much of the work wasn't usable. The footprints didn't match the actual parts, and some of the components couldn't be sourced at production volumes.
One board had a hobbyist module sitting right in the middle of what was supposed to be a production-ready design.
I know some of you already use AI directly on your design, whether that's uploading a schematic or layout for review, or one of the AI copilots built into the design tools themselves. Either way, it does catch a few real problems.
But it misses more than it catches, and you have no way of knowing what it missed.
The second expensive bucket sits outside engineering entirely. The single worst question you can ask AI is "Is this a good idea?"
AI is the most agreeable advisor you'll ever have, and it's biased toward telling you yes. So a yes you went looking for works more like a mirror than a market check, the same as the yes you get from friends and family who want to see you win.
The same blind spot shows up with money, like the part that's technically fine but wrecks your margins, or the cheaper one that fails certification. Ask AI to weigh cost against your certification path and it'll give you a reasonable answer, but it's reasoning from generic knowledge, not your actual numbers.
Everything so far has been about answers, either wrong ones or flattering ones. The biggest failure is different, because it's about the questions that never get asked at all.
AI is at its best when you know exactly what your problem is. It gets much weaker when you're not sure what the problem is, and on its own it will never raise a question you didn't think to ask.
I've leaned on AI hard in my own business for years, and it's given me real help and plenty of good suggestions. Not once has it surfaced a blind spot I'd been operating with for months, because I had to find the question myself before it could help me answer it.
In hardware, those unasked questions get expensive fast. Think of the certification you didn't know applied to your product, or the manufacturing constraint you never thought to raise with your factory.
The sneakiest one is the design where every individual answer was correct, but the subsystems fail the moment they have to work together.
AI answers the prompt. A reviewer looks at the whole thing and says, "wait, this won't work."
So when you're facing one of those expensive calls, the ones where you don't know what you don't know, you have three options.
You can work through them yourself, which is fine if you know what to look for. You can hire an experienced independent engineer who's taken a product through manufacturing and certification.
Or you can bring your design inside the Hardware Academy, where experienced engineers and product experts look at your whole project and catch the questions you didn't know to ask.
Use AI for every decision that's cheap to get wrong, and push it as far as it'll go. For the expensive ones, get eyes on them that have already been there.
Talk soon,
John Teel
Predictable Designs
|
|
|
