AI in Electronic Design

A powerful tool, but is it really a designer?

AI in Electronic Design: A Powerful Tool, but Is It Really a Designer?

Over the past few months I've been getting more and more posts, emails and requests for estimates or design reviews involving electronics designed with AI. Many come from people with little or no electronics experience, some of them trying to design their first product entirely through AI.

At the same time I've been experimenting a lot with AI-assisted design myself, mainly through eda-agent, a tool I developed that lets AI work directly inside electronic design software. I've spent a lot of time testing what AI can do, where it does really well and where it still falls short.

I've already written about how AI-generated specifications are changing the way clients approach a project. This article is about the design work itself. Things are moving so fast that some of it may be out of date in a few months, but this is how I see it today.

Where AI really shines

A lot of electronic design work is well defined, repetitive and based on information that's already in a datasheet. That's exactly where AI does best.

Take component libraries. Creating schematic symbols and PCB footprints has always been one of the more tedious parts of the job. You either draw the parts yourself, copying pin names, pad dimensions and mechanical details from the datasheet, or you download them from the manufacturer or a third-party library.

Neither is ideal. Doing it by hand can mean copying hundreds of pin names and numbers and working out package drawings that aren't always clearly presented. It takes time, and one small mistake can cost a board revision. Downloading saves time but every library provider has its own conventions. The symbols don't match your style, the silkscreen needs adjusting, mechanical layers are missing. And downloaded parts still have to be checked.

With the right tools AI can read the datasheet, extract the pin table, interpret the package drawing, apply the IPC footprint guidelines and create symbols and footprints that follow your own library conventions. Eight pins or several hundred, most of the repetitive work is done in minutes. The parts still need to be verified, but the amount of manual work it removes is huge.

The same goes for assigning net labels to microcontroller pins. Anyone who has worked with a large microcontroller knows how tedious it is to turn a pin assignment plan into a schematic, especially with alternate pin functions involved. Swap two signals by accident and you're looking at another PCB revision. For AI this is a data-processing job. Give it a proper pin assignment table and the tools to edit and check the schematic, and it does the work quickly and consistently.

Drawing the schematic is another good fit. It isn't the hardest part of a design, but standard circuit blocks, repeated structures and well-defined connections all take time that's better spent on the engineering decisions.

What makes AI so strong here is that it can do these tasks, check them, correct them and repeat the whole thing far faster than a person reasonably could.

When the requirements are clear, the method is known and the result can be checked objectively, AI is an excellent engineering tool.

What about PCB layout?

PCB layout is more complicated. AI can already handle parts of it reasonably well. It can draw the board outline, position mounting holes and connectors to the mechanical constraints and often come up with a sensible first placement for the main components.

But there's a big difference between placing components on a board and producing an optimized layout. In my experience the most important part of layout is placement, not routing. I'd estimate that around 80% of the effort in a good layout goes into getting the placement right, understanding how everything needs to connect and planning the routing before any traces are drawn. Once the placement is right, routing is much more straightforward.

AI handles the obvious placement constraints but still struggles with the subtler decisions that separate a reasonable layout from an excellent one. Those mean weighing signal integrity, current paths, thermal behavior, EMC, mechanical constraints and manufacturing all at once. AI certainly knows about these subjects. Ask it about return current paths, switching loops or thermal management and it will explain them in detail. But knowing the principles and balancing them across a real physical design are two different things.

I've found AI can do a good job of routing when the placement is already optimized and the critical routing constraints are set. Without that guidance the results are much less consistent.

The overconfident junior engineer

This brings me to what I think is the biggest limitation today. Give AI a well-defined task and the tools to carry it out and check it, and the results can be remarkable. It will often run far more iterations and checks than a human engineer could justify spending time on. But give it a complex design problem with no established solution and it's a lot less convincing.

The more I use AI on difficult design problems, the more it reminds me of a very knowledgeable but overconfident junior engineer. It can recite datasheets, explain design principles, run calculations, compare components and follow instructions at impressive speed. But when a decision needs real engineering judgment, it's surprisingly easy to lead it in almost any direction.

Challenge one of its decisions and it may agree straight away that another approach is better. Challenge that one and it agrees again. Keep going and you end up in endless rounds of changes, each new recommendation delivered as confidently as the last.

The problem isn't a lack of technical knowledge. AI struggles to work out which considerations really matter, how much they matter and when a design is good enough. Keep asking it to improve a design and it will almost always find something else to change. It rarely tells you that the design is finished, that more changes bring no real benefit and that you're wasting your time.

So you get a strange contradiction. AI-generated designs can be heavily over-engineered in some areas while completely missing critical problems in others. It may add protection circuitry, filtering, monitoring or redundancy that brings little practical benefit and still miss a basic issue that makes the product fail in real use.

An experienced engineer knows that good design isn't about adding every possible precaution or maximizing every specification. It's about understanding the application, identifying the real risks and making the right compromises.

Engineering is more than following design rules

A product has to work in the real world. It has to function electrically, but it also has to be safe, reliable, manufacturable, cost-effective and right for its intended use. It may have to meet regulatory requirements, survive electrical disturbances, work across temperature extremes, put up with careless users and keep going in conditions that are hard to reproduce on the bench.

Some of these requirements can be written as clear rules and checked by simulation, calculation or testing. Others are much harder to put a number on. That's where judgment comes in, and it comes from years of designing, testing, troubleshooting and, inevitably, making mistakes. You start to see failures before they happen. You know which datasheet figures need a closer look, which choices will cause trouble in manufacturing and which minor-looking details become serious once the product is in the field.

AI has access to a huge amount of published engineering knowledge but it doesn't necessarily know how to prioritize it for a particular application. This matters most when it has no objective way to check its own decisions. A footprint can be compared against the datasheet. Pin assignments can be validated against the table. A schematic can go through electrical rule checks.

But how does it decide whether a product will survive years in a harsh environment? Whether the protection is enough? Whether the design is more expensive than it needs to be? Whether a circuit that works on paper will behave the same once it's built? Those decisions depend on practical experience and an understanding of the whole product. AI can help with all of them, but I don't think it's currently a reliable substitute for an experienced engineer making them.

Are we losing originality in electronic design?

There's another side to this that worries me, and it goes well beyond electronics into software and other engineering fields. I think we're gradually losing variety in the way we design things.

Decades ago engineering teams worked far more independently. Information was harder to get, designers had little contact with each other and different teams came up with their own ways of solving similar problems. With the internet that started to change. Manufacturers published reference designs, application notes and example code. Online communities made it easy to share knowledge. Over time engineers settled on the same established building blocks and architectures.

There are good reasons for that. Reusing proven approaches saves development time, lowers risk and makes troubleshooting easier. But it has come at a cost. Many designers now start from the same components, the same architectures and the same reference implementations. Look at hobby electronics, where a huge share of projects is built around Raspberry Pi, ESP32 and a fairly small set of well-supported modules. They're excellent platforms, but it shows how strongly designers are drawn to anything with an established ecosystem.

Now think about what happens when AI becomes the main designer. AI learns from what already exists and favors approaches that are well documented, widely discussed and often implemented. Unless you push it to look at alternatives, it goes for the same familiar solutions. If everyone designs electronics with AI and little engineering input, I expect to see more products built on the same architectures, the same circuit topologies and the same design thinking. You can already see it in AI-generated websites and apps, which tend to look remarkably alike despite all the layouts and styles that are possible.

The irony is that AI gives us more access to engineering knowledge and tools than we've ever had, and its widespread use could still reduce the variety of engineering ideas. Innovation often comes from questioning established methods and trying approaches that at first look less practical than the standard one. If we let AI make those decisions for us, we risk becoming very efficient at reproducing existing ideas and less and less inclined to come up with new ones.

AI should make engineers better, not replace their judgment

None of this means I'm against AI in electronic design. Quite the opposite. I've put a lot of effort into building tools that let AI work directly with design software, precisely because I think the potential is enormous. I don't think we should keep spending hours drawing footprints, copying pin assignments or repeating circuit blocks when AI can do it in minutes. And engineers shouldn't dismiss AI because it sometimes makes mistakes. Human engineers make mistakes too, and plenty of traditional workflows are needlessly inefficient.

But it's important to recognize what AI is actually good at, and not to confuse its ability to carry out complex tasks with the ability to make sound engineering decisions.

For an experienced engineer AI is a huge productivity multiplier. It removes repetitive work, speeds up implementation, explores alternatives and runs extensive checks, leaving the engineer to concentrate on the decisions that matter.

For someone with no electronics background it's a very different story. AI makes electronic design look deceptively easy. It produces convincing schematics, detailed explanations, component selections and even PCB layouts that look professional. But without the knowledge to challenge its decisions, spot what's missing and judge the result, it's hard to know whether the design is any good. And that, to me, is the real danger.

AI is making it easier than ever to produce an electronic design. It isn't making it any easier to recognize a bad one.

I expect AI to improve quickly, and I wouldn't be surprised if some of today's limitations disappear sooner than we think. For now, I think the approach that works best is the one where the engineer drives the design from end to end and hands AI the time-consuming tasks that add little engineering value.

The future of electronic design may well be heavily AI-driven, and I hope it becomes far more automated than it is today. But automating an engineer's work and replacing an engineer's judgment are two different things.

Not Sure Whether a Design Is Any Good?

AI-assisted or not, I review schematics and PCB layouts and tell you what holds up, what is over-engineered and what would stop the product working in real use.

Request a Design Review