AI‑Generated PCB Layout: How Far Can KiCad + GPT‑4o Go?

While AI has already transformed software development, hardware design has remained a tough nut to crack. Until recently, the KiCad ecosystem began seeing tools that let AI do actual routing and placement. As a hardware engineer who loves experimenting, I decided to run a real test: How good (or bad) is AI at drawing PCBs?

Test Subjects: Three Real Open‑Source Boards

I chose three open‑source KiCad boards with increasing complexity:

These boards were chosen because they are public, have proven reference designs, and are already in production. Anyone can verify the results in ten minutes – no “internal test” tricks.

Why does this kind of open, reproducible test matter? Because in high‑determinism fields like PCB design, there is no room for marketing magic – only hard data.

The boards were routed by two publicly available AI PCB routers (deeppcb.ai and quilter.ai) with fully automatic placement and routing – no manual intervention, no parameter tuning, no pre‑placement. Any engineer can repeat this test on public platforms.

Key Findings

I recorded completion rates and via counts for all three boards.

Metric 1: Completion Rate – Who routes more completely?

Completion rate tells you how much manual cleanup you’ll still need. If a tool finishes only 87% of the routes, you’ll be spending hours hand‑routing the rest.

On average, DeepPCB finished nearly 10 percentage points more of the routes than Quilter.

Metric 2: Via Count – Who routes more efficiently?

Completion rate alone isn’t enough – the mess left behind matters just as much. Via count is a key metric: each extra via means one more drill hole, one more impedance discontinuity, and one more potential reliability risk.

On average, DeepPCB used 44% fewer vias than Quilter.

What does that mean in practice? On the complex 414‑net board, DeepPCB left only 12 unrouted nets for manual touch‑up, while Quilter left 54.

Anatomy of a Typical AI Failure

During the deep‑dive analysis, I tracked a recurring failure pattern:

• The circuit contained a BGA package with a “left‑negative, right‑positive” special pinout. The AI misinterpreted it as a standard pinout, trying to route VCC to a ground via.

• During the final annotation step, the screen was still full of R? and C? placeholders.

• The BOM even showed a hallucination: the same PNC5513 chip was split into two separate purchase items.

• This is the classic AI “hallucination” – confident in its results, but the output is simply wrong.

As one EEVblog forum user put it: “PCB design is fundamentally different from software. It requires high determinism and physical logic. AI can push routing to 80‑90%, but the last mile – accuracy, DRC, density optimisation – still needs human expertise.”

Where Does AI Stand Today?

After the experiments, I tried to map out AI’s current capabilities:

What AI does well:

• Automatic placement and routing of simple blocks – e.g., LDO power supplies, USB‑to‑UART bridges.

• Symbol generation for common BOM items and footprints.

• Basic fan‑out and escape routing for simple rule‑based boards.

What AI still struggles with (but is improving):

• Non‑standard BGA or custom footprints – AI frequently makes mistakes.

• Cross‑vendor component substitution and alternate part management – completely out of AI’s current reach.

• High‑frequency/high‑speed signal integrity and power integrity rules – still require human oversight.

• The “R?, C?” hallucination – still common.

• DFM (Design for Manufacturing) rules – AI cannot adequately understand them; excess vias can double fabrication costs.

Practical Advice: How to Use AI Without Being Burned

After this experience, I’ve adopted a more pragmatic view:

Treat AI as an assistant, not a replacement.

• Many experienced designers now let AI handle the “grunt work” (70‑80% of routing), then spend their own time on refinement.

• Don’t expect a perfect board from AI. Embrace the “AI‑generated draft + human polish” workflow.

Use AI where it shines, save your time for what matters.

• I now limit AI to low‑complexity, loose‑constraint tasks – power conversion, I2C/SPI peripheral expansion, etc.

• I focus my own effort on critical placement, high‑speed signals, and DFM optimisation.

Build your own test suite.

• I use the STRF, PocketBeagle, and BeagleConnect Freedom as a personal benchmark. Every time a new AI tool is released, I run it through these boards to see how much it has improved.

Never skip manual design rule checking.

• The ultimate success of a PCB depends on signal integrity and manufacturability, not on routing speed.

The Road Ahead: How Far to “Fully Automated PCB Design”?

Actually, the dream of fully automated PCB design has never been closer. Systems like pcbGPT (early 2026) already generate KiCad schematics from natural language – achieving 100% pass@1 on simple tasks, 91% on medium tasks, and 72% on hard tasks.

But as the paper itself states: “It is not yet ready to replace expert review, but generating usable, reviewable early‑stage schematics is already no problem.”

In the near future, tools like Quilter and DeepPCB will likely push fully automated layout from 80% to 95%.

But the final 5% – compliance testing, SI/PI tuning, DFM sign‑off – will, I believe, remain firmly in the hands of professional engineers. In a field that demands extreme precision and high reliability, AI is still an “enthusiastic junior”, not a seasoned expert.

Finally: Would You Let AI Route Your Next Board?

If you’re a hardware engineer, take one of your low‑complexity projects and try an AI auto‑router. What would your answer be?

AnyPCBA is your partner in the electronics design‑to‑manufacturing journey, providing small‑to‑medium batch support from prototype to compliant production.

👉 AnyPCBA website: https://www.anypcba.com/
PCB Fabrication | PCBA Assembly | Small‑to‑Medium Batch | Engineering Support