Blog · · 5 min read
Can AI design a PCB? What works today, and where an engineer still has to look
An honest look at AI PCB design: schematics as code, pin-by-pin netlist checks against datasheets, autorouting, design rule checks and Gerbers, and the parts of a board where an engineer still has to look.
Can AI design a printed circuit board? It is a fair question, and the honest answer has two parts. Yes, an AI can take a small board from requirements to manufacturing files: a microcontroller, a few sensors, a power supply and some connectors. And no, you should not trust any of it unless every step is checked by something harder than the AI's own confidence.
This article explains how that works in practice, which checks make it trustworthy, and where a human engineer still has to look.
Why the AI writes code instead of drawing
A schematic editor is built for hands and eyes: place a symbol, drag a wire, zoom in to see whether it really snapped to the pin. Language models are bad at that kind of work. They are good at writing structured text.
So the useful approach is the same one we described for mechanical parts in CAD as code: the design lives in scripts, and the schematic and board files are generated from them. Parts and their connections are written as text, the netlist is produced from that description, and exports and checks run without anyone opening an editor.
When everything is a script, a change is a diff, a revision can be rebuilt from an empty folder, and nobody quietly "fixes" a generated file by hand.
The weakest link: pinouts
The most dangerous mistake an AI makes in electronics is not a wrong resistor value. It is a confident, wrong pin number.
A language model has seen a huge number of datasheets, and it will happily tell you that pin 4 of a voltage regulator is the enable input. That may be true for a different package of the same chip. The schematic looks perfect, the board passes every check, and the first prototype is scrap.
The rule that fixes this is simple and strict: a pinout that has not been taken from the manufacturer's datasheet is an assumption. Every pin assignment gets a primary source, written down next to the design. For anything that plugs into something else, such as a connector, a module or a cable, a second independent source is required, because a mirrored connector is a classic way to destroy a board on its first power-up.
The netlist is the contract
Once the schematic exists, the next step is not layout. It is a check.
The netlist, the list of which pin of which part connects to which net, is exported, and a script compares it with the pin tables in the design document. Every pin of every part must be connected exactly once, exactly as specified, and unused pins must be marked as intentionally unconnected.
This catches the boring errors that cost the most: swapped signal lines, a forgotten pull-up, a typo in a net name that silently splits one net into two. A human reviewer skims. A script compares every single pin.
Placement, routing and design rules
Placement is where mechanics meet electronics. The board outline, mounting holes and connector positions usually come from the mechanical design. A placement script owns every coordinate, so moving a part is a change that can be reviewed and repeated, not a drag with the mouse that nobody can reproduce.
For routing, an autorouter takes the placed board, routes it and hands the result back. For simple digital boards that is often good enough, and the AI can run it again after every change.
Then come the design rule checks. They run without a screen and end with a clear pass or fail that tells an automated process whether violations remain. That difference matters. "The AI says the board is fine" is an opinion. "The DRC finished with zero violations" is a fact the next step can rely on.
From board to order
A finished board is more than a picture of copper. A manufacturer needs Gerber files, drill files, a parts list and, for assembly, a pick-and-place file. All of these can be generated by scripts as well.
The parts list deserves its own check. A schematic full of parts that are out of stock or obsolete is a schematic you cannot build. Checking every part number against a distributor's current stock before ordering avoids the most frustrating kind of delay: waiting weeks for one small chip.
Finally, only one revision should ever be marked as orderable. When an AI iterates quickly, there can be several versions of a board in a single day. Exactly one of them is the one that goes to the factory, and that has to be written down, not remembered.
Where an engineer still has to look
Some parts of PCB design are judgment, physics and experience. This is where we would not let any AI work unsupervised:
- High-speed and RF: controlled impedance, length matching, antennas and their keep-out areas. An autorouter does not understand a 2.4 GHz antenna.
- Power and heat: current paths, copper areas, thermal vias, and what happens to a regulator at full load inside a closed housing.
- EMC and noise: ground return paths, analog signals next to switching converters, and the details that decide whether a product passes certification.
- Anything connected to mains voltage: creepage, clearance and safety standards are not a place for experiments.
- Bring-up: the first power-up, measuring the supply rails, finding the one solder bridge. The AI can write the test plan. Someone still has to hold the probe.
A good AI workflow does not hide these questions. It flags them, explains what it assumed, and asks before it guesses.
What this means for your project
If your hardware idea needs electronics, AI can now do a large part of the routine design work: turning requirements into a schematic, checking every pin against its datasheet, placing and routing a simple board, and preparing the manufacturing files. What it cannot do is take over responsibility. The checks described above exist so that a person can see, quickly and concretely, what was built and why.
On makemachine.ai the electronics module works this way, and every version shows what went in and what came out, so you can follow what changed from one revision to the next. To see how that looks for a whole product, browse the public projects on makemachine.ai, or read what AI can and cannot do when you build a hardware product.
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.
More from the blog
- Will my electronics enclosure overheat? A five-minute heat estimate for a sealed box
- Joining 3D printed parts: screws, heat-set inserts, snap fits or glue?
- Free or paid AI 3D modeling? What you really pay for, and a 30-minute test to compare tools
Open-hardware projects built with AI
- DIY Automatic Plant Watering System: An open-hardware watering system for indoor plants: a small pump, a refillable tank and a soil moisture sensor in each pot keep up to six…
- DIY BBQ Temperature Controller: A temperature controller for charcoal grills and smokers that makers can build themselves. A pit probe, two meat probes and a small fan on…
- ZählerSchau: ZählerSchau is a small, battery-powered unit that clips onto an electricity meter in Germany (water meters via a clamp adapter), takes a…
- Roadstoy: Roadstoy is a remote-controlled toy version of the new Tesla Roadster that can briefly jump and flip using compressed air. It drives on an…
- The forever Washing machine: The forever washing machine is a washing machine offered as a service instead of a product. Customers do not buy the machine but pay for…
- Bike Light: I want to build a bike light that needs no battery. Magnets sit in the spokes close to the hub, and a coil on the fork picks up their…
- Rooficle: I want to develop a light, efficient and affordable e-bike that can be used legally on public roads in Singapore. Its core is an…
- BBQ Modular Grill: I want to design a modular grill. Instead of one sealed block, it is made of separate parts that can each be bought on their own online…
- Go FANUC Go: A mobile base that moves an existing FANUC M-710iC industrial robot (about 560 kg, with a milling spindle at the wrist) out of a room and…
- Jump & Run Robot: I want to build a small, low-cost toy humanoid robot that people can put together themselves. It should be about 30 cm tall, or as small…