A few weeks ago, I decided to run an experiment. I wanted to see if AI could actually design a usable PCB – not just generate a schematic or write code, but produce something that could be fabricated, assembled, and powered on.
No cherry-picked results. No "here's what's theoretically possible." Just a straightforward test with three real open-source boards, available tools, and a willingness to be disappointed.
Here's what I learned.
The Setup: Three Boards, Two AI Tools, No Hand‑Holding
I picked three real, open-source KiCad designs with different complexity levels:
These are not toy circuits. They're real designs that have shipped in thousands of units.
I ran each board through two publicly available AI PCB routers (deeppcb.ai and quilter.ai) with fully automatic placement and routing. No manual pre-placement. No parameter tweaking. Just: import netlist → hit "go" → see what happens.
The metric? Completion rate (how much of the board did it actually route?) and via count (how messy was the result?).
The Results: Better Than I Expected – But Not Ready to Replace You
Here's what I found.
Completion rates:
DeepPCB finished nearly 10% more of the routes on average. For the complex 414-net board, it left only 12 unrouted nets for manual cleanup. Quilter left 54.
Via counts (lower = cleaner):
The AI routers got the job done, but not elegantly. The Quilter outputs especially had some questionable routing choices – traces that got thin for no reason, the occasional acute angle that would make any layout engineer wince.
Still, 97% completion is not nothing. I've seen human-designed boards with more leftover airwires.
Where AI Shines (Right Now)
AI tools are genuinely good at a few things:
1. Schematic support. KiCad's AI Assistant plugin can read your schematic, answer questions, and even place components via natural language. Need a decoupling cap near an IC? Just type it.
2. Generating symbols and footprints. KiCad Copilot can take a datasheet pinout diagram and generate a usable schematic symbol. I've tested this – it's not perfect, but it saves hours of tedious manual entry.
3. Scripting repetitive tasks. KiCad's Python API is powerful, and LLMs are actually good at generating scripts. Want to place 10 resistors in a grid? Ask GPT-4o to write the Python code. It'll work on the first try more often than you'd think.
4. Generating "good enough" first drafts. For simple boards – think test fixtures, adapter boards, or early prototypes – AI can produce a layout that's 80-90% of the way there. You still need to review and clean it up, but starting from something is faster than starting from nothing.
Where AI Still Falls Short (And Will for a While)
1. High-speed routing. AI-generated boards still need manual signal integrity optimization. DDR, PCIe, or 10G Ethernet? Not yet.
2. Understanding design intent. AI doesn't know that this trace carries a sensitive analog signal. It doesn't know that these two components need to be close for thermal reasons. It just routes.
3. The hallucination problem. I've seen AI confidently generate a BOM that split the same component into two separate line items. Or assign power pins to ground because it misread a non-standard footprint.
4. DFM awareness. AI routers love vias. Like, really love them. But every extra via adds cost, adds impedance discontinuities, and creates potential failure points. AI doesn't think about that.
Practical Advice: How to Actually Use AI for PCB Design
If you want to experiment with AI-assisted PCB design (and you should – the tools are improving fast), here's my recommendation:
Start with KiCad Copilot (free, built into KiCad 9.0.2+). Use it for:
- Generating symbols from datasheet screenshots
- Answering questions about your schematic
- Getting component suggestions and datasheet links
For layout, use AI as a starting point, not a finish line. Let it route 80% of the board, then manually clean up the critical paths – clocks, differential pairs, power delivery.
Don't trust the BOM. AI still hallucinates part numbers and splits components. Always, always verify.
Keep a checklist. Before sending to fabrication, review:
- Solder mask openings on thermal pads (AI often misses these)
- Thermal via placement (AI often under-provisions)
- High-speed signal routing (AI often ignores length matching)
- Acute angles and stub traces (AI creates weird things)
The Bottom Line
Can AI design a PCB today? Yes – sort of. For simple boards, it can produce a layout that's 95% complete with a few hours of compute time. For complex, high-speed, or mixed-signal designs? Not yet.
But the trend is clear. KiCad's AI ecosystem now includes assistants, copilots, and Python scripting that can automate repetitive tasks. Siemens even launched an enterprise AI agent for EDA workflows in March 2026.
AI isn't replacing PCB designers anytime soon. But it's becoming an incredibly powerful tool for those who learn to use it.
One More Thing: Design Still Needs to Become Hardware
Whether your PCB was generated by AI or drawn by hand, it eventually has to be fabricated, assembled, and tested. The time you saved on prompts won't matter if the board fails DFM. The hours you saved on routing might come back when a manufacturing issue pops up.
That's exactly why AnyPCBA does what we do: when we receive your design files, we don't just quote and build. We run a free DFM review – checking thermal vias, solder mask openings, panelization, drill file completeness. The exact things AI still gets wrong.
We don't sell AI routing. We help you go from file to physical board – reliably, whether your design came from an AI or from a seasoned engineer.
👉 AnyPCBA website: https://www.anypcba.com/
Small‑to‑medium batch PCB & PCBA | 5–5,000 pieces | Prototype to Production
Top comments (0)