Your best engineers aren’t slow. They’re just spending most of their time on work that doesn’t need them.
The conversation around AI in PCB design has calcified into two camps. Vendors promise boards at the push of a button. Skeptics dismiss the entire category as hype. Both miss the real question: which specific tasks benefit from AI, and which ones require human judgment that no algorithm can replicate? Answer that clearly, and you have a deployment framework. Miss it, and you waste resources on automation that doesn’t deliver.
The 80%: What AI Handles Well
Most of a PCB layout is mechanical. Fanning out BGA pads. Routing non-critical nets between passives. Matching trace lengths on a parallel bus. Filling copper pours. Connecting decoupling caps to their nearest power pins. These tasks require precision, patience, and strict adherence to design rules. They do not require creativity.
This is the work that burns hours. An engineer routing a moderately complex board will spend the majority of their time on connections that follow well-understood patterns. The constraints are explicit: trace width, clearance, via size, layer assignment. The decisions are repetitive. The risk of error comes not from the difficulty of the task but from the sheer volume of it. A thousand-net board has a thousand opportunities to fat-finger a clearance violation at 2 AM.
AI handles this class of work well because the success criteria are measurable. A routed net either passes DRC or it doesn’t. A length-matched group either hits the target tolerance or it misses. There is no ambiguity. The algorithm can attempt, evaluate, and iterate faster than any human, and it doesn’t lose focus on net 67.
The 20%: What Humans Must Keep
The remaining 20% is where engineering judgment lives. High-speed differential pairs that need specific impedance control and carefully managed return paths. Sensitive analog routing where a few millimeters of trace proximity to a switching node creates measurable noise. Power delivery networks where the copper geometry directly affects voltage drop under transient loads. EMC-critical sections where regulatory compliance depends on layout decisions that no design rule can fully capture.
These tasks share a common trait: the constraints are implicit, contextual, and often contradictory. The datasheet says keep the feedback trace short. The thermal analysis says move the component away from the heat source. The mechanical enclosure says the connector has to go on that edge. Resolving these tradeoffs requires understanding the physics behind each constraint and making a judgment call about which one matters most for this specific design in this specific application.
This is the work that justifies having experienced engineers on staff. The 20% is where decades of experience, application knowledge, and engineering intuition converge. Automating it badly is worse than not automating it at all, because a subtle error in a high-speed pair or a power delivery network doesn’t show up in DRC. It shows up in the lab, six weeks later, when the prototype doesn’t work and nobody knows why.
How to Structure the Workflow
The practical question isn’t whether to use AI. It’s how to draw the line between the 80% and the 20% on your specific board.
Start with the schematic. Identify the critical nets and signal paths that require engineering judgment. These are your 20%. High-speed interfaces, sensitive analog chains, power delivery, anything where the layout directly affects electrical performance. Mark them. Protect them. Route them yourself or review them manually with full attention.
Everything else is a candidate for automation. Fan-outs, non-critical digital buses, passive component connections, ground stitching, copper fills. Hand these to the AI. Let it grind through the mechanical work while you focus on the sections that actually determine whether the board works.
The key insight is that this split isn’t fixed. A simple IoT sensor board might be 95% automatable. A 16-layer high-speed computing platform might be 60/40. A mixed-signal medical device with regulatory constraints might need human attention on 50% of the layout. The framework adapts to the board, not the other way around.
The engineers who get the most from AI tools are the ones who think clearly about this boundary before they start. They don’t hand the entire board to an algorithm and hope for the best. They don’t refuse automation on principle and route every net by hand. They make deliberate choices about where human judgment adds value and where it’s just burning time.
The Warning Signs
Two failure modes to watch for.
Over-reliance looks like accepting AI output on critical nets without thorough review. If you’re not manually inspecting every high-speed pair, every sensitive analog section, and every power delivery path, you’re trusting the algorithm with decisions it isn’t equipped to make. The 80/20 split only works if you actually spend real time on the 20%.
Under-utilization looks like manually routing nets that follow obvious patterns with explicit constraints. If your best analog engineer is spending Tuesday afternoon connecting decoupling caps to power pins, you’re paying for judgment and getting labor. That engineer’s time has a highest and best use. It isn’t fan-outs.
Both failure modes have the same root cause: not thinking clearly about which tasks require judgment and which ones don’t.
The Opportunity
The real leverage of the 80/20 framework isn’t speed, though speed matters. It’s focus. When an engineer offloads the mechanical 80%, they don’t just finish faster. They spend more time on the decisions that determine whether the product works. More iterations on critical placement. More attention to power integrity. More time reviewing signal return paths instead of manually connecting GPIO pins.
This is what we’re building at DeepPCB. Route the critical 20% yourself. Tell the AI to handle the rest. Review it, adjust it, ship it. Not a push-button fantasy. A workflow that respects what engineers are good at and eliminates the work that wastes their time.
The 80% was never the hard part. It was just the slow part. Time to stop treating them the same.