If your three most-senior RF engineers retire next week, how much of your design capability would walk out the door with them? The honest answer would make most engineering leaders squirm. Today, RF and microwave design still depends heavily on institutional knowledge: Power-amplifier methodologies; device characterization flows; the on-the-fly judgment calls that never made it into a document because there was never time to write them down. Without that knowledge, onboarding is measured in months instead of afternoons. Evaluation of design spaces slows to one candidate in the time that competitors can evaluate thousands, and re-spins happen that a wider search would have caught. AI is supposed to be the answer to this. But for many teams, the AI conversation stalls before it starts, because it gets framed as an all-or-nothing bet: adopt a new platform, rebuild the flow, retrain the team. That approach is costing organizations time they don’t have. The real question is how to adopt AI “How” should be the question, rather than “whether.” As in how to build capability without forcing a disruptive rebuild of workflows that already ship products in volume. That means finding an approach that lets teams begin today, with the tools their engineers already trust and layer in capability incrementally. And the “how” falls into three categories that can run in parallel: Capture, Orchestrate, and Accelerate. Each attacks a different constraint, and none requires finishing another before moving to the next one. Let’s step through each of them. Capture: Get expertise out of people’s heads Large language models are good at code. RF expertise doesn’t live in code; it lives in tools, in equations, and in the experience of engineers who’ve spent years learning what actually correlates to measured hardware. There’s no public repository of any team’s methodology for an AI to learn from, which is exactly why this must start inside the organization. In the capture phase, you make expertise visible and reusable. That means doing things like exporting schematics and layouts to parameterized Python, recording expert workflows as executable scripts, or sketching a simulation flow as a flowchart that converts directly into documented, agent-ready code. In that way, you’re transforming knowledge in an engineer’s head into an asset the whole team can run, share, and eventually hand to an agent. Orchestrate: Stop writing the script, start stating the goal Orchestration puts to work the expertise you’ve captured. Instead of writing and maintaining scripts, engineers can increasingly just ask. Start by stating a desired outcome (“design a MMIC power amplifier with gain above 20 dB and output power above 28 dBm”) and then let a connected LLM find and call the right tools to get there. It helps to have a shared vocabulary for where a team sits on this journey. Most organizations today operate between manual scripting and early co-pilot use, leveraging natural-language assistants that grasp intent while engineers still supervise closely. But the next step isn’t full autonomy; it’s specialized agents handling multi-step tasks with partial delegation, built on the same orchestration foundation already in place. Accelerate: Explore more, wait less The third category brings faster, more realistic physics into the creative phase of design. Two engines drive it: Surrogate modeling, which replaces heavyweight EM runs with fast neural-network approximations (on some structures, two to three orders of magnitude faster). And AI-enabled optimization, which can handle far more parameters simultaneously than conventional optimizers and can find the needle in the haystack using fewer simulations. Meanwhile, trust is essential Speed only matters, though, if the answer can be trusted. That’s why the sharpest question to ask when evaluating any AI-driven design tool isn’t just speed or interface. It’s: how well do the physics correlate to measured hardware? Packaging, interconnects, coupling, power and ground integrity, and 3D currents are where fast, confident answers quietly go wrong and where silent faults become untraceable in a root-cause analysis. Grounding agentic design in measurement-correlated physics allows an organization to delegate more of the design process over time (safely), rather than trading speed for risk. Successful modeling cultures know how to close this loop. This is also where data lineage matters as much as the models themselves. Where did a surrogate’s training data come from? How was it tagged and cleaned? What are its known assumptions and limitations? Those questions need clear answers because that visibility is what keeps the speed “on the rails.” Sphere Semi, an RFIC design company, faced the wall every talent-constrained team eventually hits: manual, one-design-at-a-time exploration. Their response was a fully code-defined, Python-based flow (generate, simulate, rank, refine) running hundreds to thousands of candidates through circuit and EM co-simulation. The result: 5-10x productivity gain, 6 dB better isolation, and 30% less filter area versus traditional, manual design processes. Start from where you are Engineering leaders will adopt agentic design, but the question they need to consider now is what their workflow will look like in three years, and what of that future they choose to start building today. The elegance of Capture, Orchestrate, and Accelerate is that these steps do not require a new platform or a disruptive rebuild; they just require you to start. Want the full planning framework — including the five-level autonomy scale and a practical roadmap for capability planning? Read the complete white paper, Planning Your AI Design Journey in RF EDA. The post Planning Your AI Design Journey: Start With What You’d Lose appeared first on Semiconductor Engineering.
Source: https://semiengineering.com/planning-yo ... youd-lose/