Capturing design intent is critical for AI to function as a true partner in PCB design.

For the past several years, the discussion around artificial intelligence in PCB design has largely focused on capability. How fast can AI route? How many rules can it process? How well can it replicate known good designs? These are reasonable questions, but they are increasingly the wrong ones.

We are now entering a phase where AI can do many of these things quite well. It can generate layouts, evaluate electrical performance and explore solution spaces at a speed no human designer can match. That part is no longer theoretical. What remains unresolved is something far more fundamental, and it has little to do with model intelligence.

The limitation is methodology. Not methodology in the abstract sense, but the specific, often unspoken process that experienced designers use when they are confronted with a problem they have never seen before. If every board were just a variation of the last one, none of us would have jobs. Or at least, not interesting ones.

This is where the conversation begins to shift, and where the idea of agentic AI becomes relevant. Agentic AI is not simply AI that performs tasks. It is AI that operates with intent, makes decisions within a defined framework and adapts its approach as conditions change.

The question is not whether AI can be agentic. The question is what it needs in order to behave effectively in a domain as complex and continuously evolving as PCB design. The answer lies in something our industry has never formally captured: design intent.

Constraints

When we talk about constraints in PCB design, we tend to treat them as the starting point. We define trace widths, spacing, impedance targets, via structures and a lengthy list of other parameters and then expect the design process to follow. This works well when the design challenges are already understood.

But constraints are not the beginning of the process. They are the result of decisions. Before any constraint is written, a series of higher-level judgments has already been made. These include cost targets, fabricator capabilities, reliability requirements, corporate standards, usual compromises and risk tolerance. These factors define what solutions are acceptable before any geometry is ever drawn, and this layer is absent from today’s AI systems.

If AI is given a set of constraints, it can apply them. If it’s given enough data, it can even learn patterns that approximate past decisions. But when a new technology appears, or even a familiar technology used in a new way, and there is no historical data to rely on, the system has no basis for determining what the constraints should be or how to manage them.

Designer Methods

Experienced designers approach new or unfamiliar technology differently. Faced with a new problem, they structure it into a set of methods. They identify what is new, what is known and what matters most. They decompose the system, prioritize constraints based on competing objectives and iteratively converge on a solution that satisfies tradeoffs that are often not fully defined at the start. That is not supervised learning. That is methodology.

While writing the book High-Speed Constraint Values and Design Methods, I tried to formalize part of this process, particularly around how constraints are derived and applied in high-performance systems. The key idea is that constraints are not arbitrary rules. They are the encoded result of decisions about physics, manufacturability and acceptable risk. Sometimes the solution is to derive the appropriate constraint values, and other times it is to apply a different methodology altogether.

A good example is when high-speed signals are part of the design. One approach is to define constraints up front and enforce them universally. However, using the proper methodology, some of those constraints become unnecessary if the routing length is below the critical length. An experienced designer recognizes this and chooses not to apply those constraints in that context. More importantly, that decision is not static. It is revisited as the design evolves.

Another example of using methods beyond strict constraints occurs when the lengths of multiple high-speed signals need to be kept within a skew tolerance. Did the setting of the constraint consider whether skew compensation is available? It may be that skew compensation could eliminate or at least mitigate the problem. The diagram in Figure 1 illustrates numerous additional decisions that should be considered when determining what method to route the diff pairs in the context of fiber weave skew.

Figure 1. Fiber weave skew risk in differential pairs, including material conditions, routing methods and compensation strategies.

Today’s AI tools do not operate this way. They apply constraints as fixed rules. They do not continuously evaluate whether those constraints are still valid, whether their values should change or whether a different routing strategy would be more appropriate. An experienced designer does all of these things throughout the design process.

This distinction becomes critical when we consider how AI should operate. If AI is treated as a constraint executor, it will always be limited to problems that are already well understood. It will be exceptionally good at repeating the past, and increasingly irrelevant as designs move into new or unfamiliar territory.

If, however, AI is structured as a methodology execution system, it can begin to operate in unfamiliar domains. It can dynamically evaluate tradeoffs, generate appropriate constraints and adapt its approach as additional information becomes available.

This is where agentic AI becomes meaningful. An agentic system in PCB design must operate within a framework that defines not just rules, but also priorities. It must understand what is important, what is allowed and what compromises are acceptable. Without that, it has little basis for decision-making.

Design Governance and Intent

I would describe the missing layer in current AI implementations as design governance. Design governance sits above constraints and extends beyond a traditional requirements document. Requirements define what must be achieved. Governance defines what success looks like and what tradeoffs are acceptable. It captures the intent of the design in a structured way and reflects the factors experienced designers consider before they ever open a layout tool. These include cost ceilings, approved fabrication processes, reliability standards, supply chain constraints, placement and routing preferences and organizational priorities. In effect, it defines the boundaries within which methodology operates.

In the context of the governing intent, the system must interpret that intent in practical terms and be able to make workable compromises if necessary. For example, keeping a high-speed device close to its memory may be intended to minimize latency. If space makes it impossible to follow the desired placement arrangement, the system evaluates the space and applies a decision-making methodology, prioritizing what matters most. It might be able to relax placement, add layers or adjust routing if performance remains acceptable. The key is that it makes informed tradeoffs aligned with the governing intent, rather than simply enforcing constraints.

Once this layer of design governance is defined, methodology can be applied. The process is not complicated, but it is quite different from how most AI systems are currently structured. The system must first interpret the governing intent. It must then evaluate the trade space within those boundaries. From there, it applies a decision methodology to select an approach, and only then does it generate the constraints that enforce that decision.

There is a persistent belief that if we simply provide enough training data, AI will eventually learn to make the right decisions. This assumption breaks down in PCB design because the future does not look like the past. New or unfamiliar materials, packaging technologies, power architecture and performance requirements continuously require the designer to redefine the problem space.

We cannot train a model on a laminate that was introduced last week. We cannot provide historical examples for a signaling scheme that is still being characterized. And we certainly cannot anticipate every combination of constraints that will arise in future designs.

What we can do is capture how experienced designers think. Instead of asking how fast an AI system routes, we should be asking different questions. Can it adjust its strategy as conditions change? Can it explain the tradeoffs it is making? Can it make effective decisions when compromises are required? Can it operate without a historical precedent? These are the characteristics of an experienced designer. They are also the characteristics of an agentic system.

Technology is not just moving forward; it is shifting in ways that make some of our long-standing assumptions less dependable. Data rates are pushing interconnect behavior into regions where the old rules begin to break down. Packaging is blurring the once-clear boundary between the chip and the board. Material choices are evolving to support higher frequencies, and power delivery margins continue to narrow. At the same time, AI-assisted tools are beginning to influence how decisions are made, not just how layouts are executed.

Any one of these changes is manageable on its own. The difficulty is that they are all occurring at the same time. Experience is still essential, but it cannot be applied the same way as before. It has to be interpreted in the context of a continuously shifting design space.

Conclusion

Agentic AI, properly implemented, offers a path forward. But it will not emerge from larger models or more data alone. It requires a deliberate effort to capture and formalize design intent, to encode methodology and to structure decision-making in a way that machines can execute.

In practical terms, this means building systems that do more than apply rules. They must represent decision processes, evaluate tradeoffs, and derive constraints as outcomes rather than inputs.

The knowledge already exists. It is embedded in the practices of experienced designers, in the methodologies they apply, and in the constraints they derive. The challenge is to make that knowledge explicit, structured and usable by AI systems. If we can do that, AI will move from being a very capable tool to something much more interesting: a design partner.

Charles Pfeil has spent over 50 years in the PCB industry as a designer, owner of a service bureau, and in engineering management and product definition roles at Racal-Redac, ASI, Cadence, PADS, VeriBest, Mentor Graphics, and Altium. He was the original product architect of Expedition PCB, and an inventor of Team PCB, XtremePCB, XtremeAR, and the Sketch Router. He authored BGA Breakouts and Routing. He can be reached at This email address is being protected from spambots. You need JavaScript enabled to view it..