Most "AI PCB software" content is either a vendor telling you their tool is great, or a skeptic telling you AI can't do hardware. Neither helps you when you're actually trying to pick a tool. What helps is a checklist you can run against anyone — the vendor included.
Here are the eleven questions I'd ask, with what a good answer looks like before I say what ours is.
Does it export standard Gerber files?
A real answer names the formats: Gerber RS-274X, Excellon drill, pick-and-place (CPL) and a BOM. If a tool only offers its own project format, or a PDF, or an image of a board, your fab cannot build it. Ask for a sample output and open it in a free Gerber viewer before you trust it.
PCBEditor exports Gerber, Excellon drill, pick-and-place and BOM — the same file set you'd send from any desktop EDA, with no export tier or paywall.
Can it generate fabrication-ready files, or just a picture of a board?
"Fabrication-ready" means a fab house can accept the output as-is: correct layer mapping, a drill file that matches the copper, no approximated geometry. Many AI demos produce a convincing rendering with no manufacturable data underneath. The test is simple — send the output to a fab quote tool and see whether it parses.
The board is built by a deterministic engine working in real units, so what you export is measured geometry, not an approximation of one.
Can I run DRC in my own CAD tool afterwards?
This is the question that separates an open tool from a trap. You should be able to take the output into KiCad, Altium, or your own flow and re-run design-rule checks independently. If a vendor's results can only be verified inside the vendor's own software, you have no way to audit them.
Output is standard fabrication data, so you can re-check it in whatever tool you already trust. An AI result that can't be independently verified isn't a result.
Does it explain why it placed a component or picked a part?
Ask for the reasoning, not just the outcome. A tool that can say "this regulator because of the input range and thermal budget, this capacitor here because it's the supply pin" is one you can argue with and correct. A black box that emits a board gives you nothing to review — and reviewing is the whole job.
Part choice and intent come from a language model that states its reasoning, so you can read it, disagree with it, and change it before anything becomes copper.
Which AI router is best for BGA-heavy boards?
Be skeptical of any single answer here, including ours. Fine-pitch BGA work is the hardest case in autorouting — it needs escape routing before a general router can help, and it's where every tool is weakest. The right question to a vendor isn't "can you do BGA" but "show me a BGA board you finished, and tell me which nets you couldn't route."
Dense fine-pitch parts are the area we're actively working on, and where you're most likely to see unrouted nets today. We'd rather say that than claim otherwise.
What about DFM — will it actually manufacture?
Design for manufacturability is separate from design-rule checking. Ask whether the tool enforces your fab's real capabilities — minimum trace and space, annular ring, drill sizes — rather than generic defaults. Then confirm with your fab before a first run, every time, whoever produced the design.
Rules are enforced during routing rather than flagged afterward, and the checks run against the constraints for the board you're building — but confirming against your chosen fab's capability sheet is still the right final step.
Is the result reproducible?
Run the same design twice. If you get two different boards, the layout is coming from a language model, and it'll keep drifting every time the model is updated. For hardware you want the same input to produce the same output, so a review you did yesterday still describes the board you have today.
Routing is deterministic: same design in, same board out, and it doesn't change when the model behind the judgement half is swapped or upgraded.
Can it route a multi-layer PCB automatically, or just two layers?
A board with real density needs more than two copper layers — internal ground or power planes for noise control, extra signal layers for escape routing on fine-pitch parts. Ask whether layer count is fixed at two, or whether the tool actually places vias and routes across an internal stackup, not just a single top/bottom pass.
Layer count is something you choose — 2, 4, 6, 8 and up — and the deterministic router places vias and routes across the full stackup, including stitching internal ground and power planes, not just a top/bottom trace pass.
Does it actually reduce board respins?
This is the question that matters most to a budget. Most AI PCB demos only check whether traces are far enough apart — a design-rule check. The boards that come back wrong are usually wrong somewhere DRC can't see: a shorted input, an enable pin wired backwards, a USB connector that backfeeds power. Logic-level mistakes that pass a clean DRC and still don't work.
DRC catches geometry. Separately, every generated board runs an electrical-integrity audit on the final netlist, checking for the failure modes that cause real respins — shorted connector contacts, enable pins wired to the wrong polarity, USB backfeed — before you export anything.
Can I import an existing design, or only start from scratch?
A tool that only exports is a one-way street: useful for a first draft, useless for the board you already have half-finished. Ask whether you can bring an existing project in, not just take a new one out.
PCBEditor reads and writes native KiCad .kicad_pcb files and imports Gerber archives, so an existing KiCad board can be opened and continued here. Altium, Cadence and OrCAD files are not imported.
If it gets a part or connection wrong, do I start over or can I fix it?
AI output is a first draft, and first drafts have mistakes. The real question is what fixing one costs you: do you edit the one part or connection that's wrong, or does any change mean throwing away the board and starting the whole generation over?
Corrections happen by chat, targeted at what you ask for — swap a part, fix a connection, reroute one section — without regenerating the whole board. It's still worth checking the change only touched what you asked for, the same discipline as reviewing any edit.
Run this against whoever you're evaluating
None of this is specific to one vendor. Ask it of us, ask it of whoever else you're considering, and weight the answers by what actually matters for the board you're building this week — a BGA-heavy design cares about question five more than question nine; a respin-burned team cares about question nine more than anything else on the list.
The honest version of this page lives at pcbeditor.com/evaluate-ai-pcb-software, kept current as the questions people actually ask keep changing.
Top comments (0)