So can AI design a PCB or not?

Partly. A language model does the judgement half well: parts, topology, trade-offs. It cannot do the measured half — placement, routing, DRC, fabrication output — because those are measurement problems, not text problems. A tool that only gives you the first half and calls it a finished board is the thing the skeptics are right to distrust.

Isn't 'it can't apply knowledge' just a moving goalpost?

No, it is the actual point, and it is worth taking seriously. Reciting what a textbook says about clearance is not the same as guaranteeing a specific trace meets it on a specific board. The honest position is to agree, and then not ask the language model to do the part it provably cannot.

FreeRouting left half my board unrouted. Is AI routing any better?

Better AI routing is not a language model writing traces — it is a search engine over physical space, the same class of tool as FreeRouting, done more carefully. The useful difference is honesty: a router should report the exact nets it could not complete, by name, instead of a completion percentage that hides the failures. Most unroutable boards are placement problems, and naming the stuck nets tells you what to move.

Does it export native KiCad files?

The design exports to standard fabrication formats — Gerber, Excellon drill, pick-and-place and BOM — the outputs a fab actually consumes, so the result is not locked inside one tool.

Can AI actually apply PCB knowledge, or just repeat it?

It is the fairest objection in every thread about AI and hardware: being able to recite what a textbook says about clearance is not the same as knowing how to apply it on a real board. That is correct. So let us start by agreeing with it.

A language model reads and repeats. It has consumed a great deal of writing about PCB design, so it can produce a confident paragraph about impedance or decoupling. But asked whether a specific trace on a specific board clears a specific pad, it has no way to measure — so it estimates, and an estimate is not a manufacturable board. The people who say “it cannot apply the knowledge” are right about the language model.

The mistake is not the skepticism. The mistake is assuming that a useful AI PCB tool must therefore be a language model pretending to be an engineer. It does not have to be, and the good ones are not.

Two different jobs, two different tools

Designing a board is really two jobs. One is judgement: what parts, what topology, what trade-offs. The other is measurement: geometry, clearance, routing, rules. They call for completely different tools, and the honest architecture keeps them apart.

Judgement — the AI is good at this

Choosing a regulator family for a given input range, load and thermal budget.
Proposing a topology, or spotting that a decoupling capacitor belongs next to a supply pin.
Reading a datasheet and drafting a plausible netlist in text form for you to check.

Measurement — the AI cannot do this

Whether this trace clears that pad by 0.15 mm — a measurement, not an opinion.
Placing parts on a board with real dimensions so the copper can actually reach.
Routing a multi-layer board: a search through physical space against hard constraints.
Running a design rule check, and exporting Gerber, drill and pick-and-place a fab will accept.

What actually does the measuring

The second column is measurement: the same class of work as a classic autorouter and design-rule checker, where every result is a calculation you could check by hand. That is what to ask of any tool that claims it: numbers you can check, which stay the same when a newer AI model comes out.

This is also why “better AI routing” does not mean a chatbot writing traces. It means a search engine over physical space that is honest about its limits — and reports the exact nets it could not complete, by name, instead of a reassuring percentage.

If FreeRouting left your board half-unrouted

You are not alone, and the fix is usually not a smarter router. Most unroutable boards are placement problems: two connectors at opposite corners with a dense IC between them cannot be recovered by any router. A router that names the stuck nets tells you what to move. That honesty — say what failed and why — is the part worth wanting from any tool, AI or not.

See how the routing is done →

Questions people actually ask

So can AI design a PCB or not?: Partly. A language model does the judgement half well: parts, topology, trade-offs. It cannot do the measured half — placement, routing, DRC, fabrication output — because those are measurement problems, not text problems. A tool that only gives you the first half and calls it a finished board is the thing the skeptics are right to distrust.
Isn't 'it can't apply knowledge' just a moving goalpost?: No, it is the actual point, and it is worth taking seriously. Reciting what a textbook says about clearance is not the same as guaranteeing a specific trace meets it on a specific board. The honest position is to agree, and then not ask the language model to do the part it provably cannot.
FreeRouting left half my board unrouted. Is AI routing any better?: Better AI routing is not a language model writing traces — it is a search engine over physical space, the same class of tool as FreeRouting, done more carefully. The useful difference is honesty: a router should report the exact nets it could not complete, by name, instead of a completion percentage that hides the failures. Most unroutable boards are placement problems, and naming the stuck nets tells you what to move.
Does it export native KiCad files?: The design exports to standard fabrication formats — Gerber, Excellon drill, pick-and-place and BOM — the outputs a fab actually consumes, so the result is not locked inside one tool.

If you have been arguing this out and want to judge it yourself: describe a circuit and watch PCBEditor choose the parts, place, route and check the board — then tell you exactly what it could not finish.

Try it in PCBEditor →