DESPITE HIS LONG-STANDING career and role as a consultant to some of Siemens‘ biggest EDA clients, Steve Gascoigne still thinks like an applications engineer. There is no one better placed to share what’s actually happening at the sharp end of AI in PCB design tools.

This conversation has been edited for length and clarity.

You’ve been in PCB design since 1988. How do you describe what you do to people outside the industry?

I always reinforce that I’m an engineer. But depending on people’s perspective, I may need to go into a little more detail. In Ireland, where I live, if you say you’re an engineer, they instantly presume you’re in construction, because that’s the dominant industry. So, typically, I’d reference something they’re familiar with. The most common one now is your smartphone.

Any modern device has a mechanical form, something we’re used to looking at. But it has to have functionality, and that’s going to be the electronics. And the PCB is the backbone of the electronics. Irrespective of how complex or clever the electronics get, or even how simple, at some point, the components have to be mounted onto something, and that thing is a printed circuit board.

“We were the youngsters in the drawing office, the ones who just weren’t scared of the computer. We’d switch it on and try and make it do something.”

It’s not just three or four wires connected to a plug. There’s a phenomenal amount of complexity. Many hundreds of thousands of hours go into designing a PCB, and it’s core to the functionality of any electronic device. Creating one is a mixture of legacy knowledge, absolute understanding of electronics, and an eye for detail.

When I started, CAD was a brand-new thing. I’m from the first home-computer generation: Sinclair Spectrums and VIC-20s. We were the first kids to have computers in school, and we had this natural inquisition to get into them. You’ll find there’s a whole load of 57- to 60-year-old guys in PCB for that very reason. We were the youngsters in the drawing office – the ones who just weren’t scared of the computer. We’d switch it on and try and make it do something. Some of the more mature members of staff were resisting the move to CAD. Their comfort zone was paper documentation.

What distinguishes the tools that genuinely help PCB designers now from the ones that don’t?

I think we’re now on the cusp of automation becoming genuinely useful. We’re moving from an era where automation was seen as cumbersome, a blocker or something that forced the engineer to lose control. And I think that’s why, until recently, many designers resisted it. Either the technology wasn’t able to make complex decisions, or the outcomes were just undesirable.

If a tool has automation or AI – whatever we want to call it – and it guides the engineer, helps them arrive at a decision more quickly… and if it’s a decision they understand and value, they will embrace it. What they won’t embrace is an outcome they don’t recognise, or one they fundamentally don’t understand how the tool arrived at.

The automation my customers value is anything that removes repetitive, tedious tasks. That can be many things, not just clicks, but sometimes hours of sitting at a machine pushing buttons. If I can automate a task and get to the same point more quickly, or have it run overnight, work in parallel, that’s where the interest is. If they can correlate the outcome to something they recognise, they’ll take advantage. But if they see it as a constraining force, something removing their innovation, they don’t want it.

Can you give a concrete example?

One thing I know we’re taking advantage of now, and customers value too, is actually a small, subtle change. It’s the recording of behaviour. The ability for the tool to record your typical actions, and then when you go back through a similar task, it prompts you with the same behaviour – fundamentally, taking away clicks.

The technology behind it is far superior to the days of Clippy in Word, where it would just make random suggestions. Now it’s, ‘this task is reasonably complex, but if I can do 10 clicks instead of 20, it makes me more comfortable and more productive.’ And it possibly removes the classic human error, the mistake you make because it’s an annoying, tedious task.

“If they see it as a constraining force, something removing their innovation, they don’t want it.”

It’s the same thing as when you get in your car, and it says, ‘Do you want to drive to the office?’ Because for the last five days, that’s where you went. I don’t have to go, ‘search satnav’ and ‘office’. It just says, ‘Do you want to go to the office?’ There’s probably a lot of power behind making that happen. It seems simplistic, but it absolutely has value. And those are the things users immediately value in a tool.

What about simulation? How is AI changing that specifically?

If somebody said to me, ‘AI is a load of nonsense, you can’t use it,’ I’d say, ‘No. I can show you two or three places where it is absolutely used today, if we accept that there are different sorts of AI.’ Simulation is one of them.

If I have a simulation with 10 parameters and I sweep one of them, that’s not too bad. But if I sweep all 10, I can very quickly get up to tens of thousands of permutations. It’s quite easy to create a simulation that won’t finish in your lifetime. Even if each run only takes two minutes, well, that’s fine, unless you’ve got 32,000 permutations to try.

So the first step is that AI still lets me run it thousands of times, but helps me figure out which is the best outcome. The next stage, and this is kind of where we are now, is AI helping me minimise the task by avoiding the permutations which make no sense.

That’s quite a significant shift, from using AI to run calculations faster, to trusting AI to decide which calculations are worth running.

Absolutely. And there is a comfort barrier to get over. If I say to a tool, ‘there are many permutations, find me a solution without running them all,’ then there’s a risk I missed the ideal solution because we never simulated it. It also means I have to define exactly what I find acceptable. Parameters I might previously have left floating, I now have to pin down precisely. As I’d say to the tool, ‘if you’re going to help me, I need to tell you exactly what it is I want you to arrive at.’

But we do have customers adopting this already. And for some simulations, it’s genuinely the only viable route, purely because of the time involved. And I should be clear: there is no AI where I can just say, ‘Please design me a circuit board,’ and the perfect circuit board pops out. I don’t expect to see that in my lifetime, never mind my career. But that’s not what this is about.

We had a line at a C-suite presentation recently: AI is only valid if it’s verifiable. You can do anything, but if you can’t verify the outcome, if you can’t look at what it produced and say, ‘Yes, it did what I wanted,’ then it has no value.

Are there other areas where you’re seeing AI used in practical, everyday ways?

One use case that’s not specific to our tools but is used across the industry is scripting. You can automate and run the product yourself if you want to, writing scripts in Visual Basic, C, or Python. And I include myself in this, I find it very tedious. I make mistakes. I went on a scripting course once, and there was a test at the end where you had to write a script. The instructor looked at mine and said, ‘Yours works, but you did it in 180 lines. The answer was 25 lines.’

So, I’m just bad at scripting. However, AI can be precise, and it’s quick. If I want a script that counts the number of parts on a board and tells me how much they cost, AI can do that in seconds. Whereas it might take me an hour to write that script, and two hours to figure out why it doesn’t work. With AI, the quality can be 80–90%. It’s fine, it happens quickly, and it did a task I find tedious and troublesome. I know customers are already using generic AI for exactly that.

What about routing – is AI playing a role there yet?

There’s a small category of products that lend themselves to autorouting and are fully auto-routed. But the vast majority of designs now use what we call interactive routing. We still need the visual quality, the inspection, and the knowledge of the layout person. In the background, the tool ensures they don’t deviate or break any rules, but the engineer is still driving.

One place we’ve added an extra level of intelligence is in making autorouting not just electrically correct, but aesthetically pleasing. Because that’s what you do manually, and in some cases, if it’s aesthetically pleasing, it’s probably a better solution. If you’ve routed a bus nicely, like a river on one side, what looks better works better, versus just a mathematical outcome that is technically correct.

“If AI can help re-consume valuable IP that’s already been generated, you can focus on the new stuff.”

In the past, we didn’t have a way for the software to determine what looked good. There was no way to complete a net if it looks good – what does that even mean? With AI, because you can overlay large language models and interrogate trends, we now have the capability to run a process that also says, ‘Yeah, that looks good.’ We’ve had that ability for several years. I’m not saying it’s perfect, but we know from feedback that’s where layout engineers want to go.

Looking ahead, where do you see AI making the biggest difference to how PCB designers actually work?

Within any company, you might have 20 years of designs. Your next product will have some innovation, but it will also consume things you’ve developed in the past. Rather than redrawing a schematic and doing the layout again in the exact same way, when you know it worked before, what if AI could just help you implement that?

It’s not just a case of doing more faster, you’ve got to do it differently. If AI can handle the tasks that could be automated but need more intelligence, engineers can focus on innovation. If AI can help re-consume valuable IP that’s already been generated, you can focus on the new stuff.

You mentioned reusing IP? Can you explain what you mean by that?

We used to talk about reusing blocks. But it’s not as simple as that anymore. It’s more like a block with fuzzy edges that needs to connect to other blocks with fuzzy edges. Understanding how those fuzzy edges connect is where AI actually helps. Because that’s the long, repetitive, tedious research and clicking that engineers are trying to avoid.

For some companies, that value is entirely in what they’ve already proven. Go to a high-end hi-fi company, and there’s some very secret sauce in those printed circuit boards – that’s what makes their product better than anybody else’s, and they want to make sure that’s maintained. So the question becomes, ‘Do I only consume IP I’ve generated myself, or do I look elsewhere for similar technology?’ For many companies, the answer is firmly the former.

There’s a thread running through everything you’ve just described: the behaviour recording, the car satnav, the IP reuse. It’s almost like making a copy of yourself. That’s quite different from how something like ChatGPT works, which doesn’t copy one person – it copies the world. What you’re suggesting is that you actually want AI to copy your expertise?

Yes. But the fear of opening it too wide is that you incorporate something without full understanding, without it being proven. Like I mentioned previously, AI is only valid if it’s verifiable. And I think that’s where we are with PCB design more broadly. Two or three years ago, you’d Google something and speed-read a thousand results to find what you were looking for. Now you get a paragraph that tells you exactly what you wanted to know. We’re crossing that same bridge with automation, where the outcome finally satisfies the user.

You work directly with engineers day to day. Is what they want from AI different from what their organisations think they want?

I deal with engineers on the factory floor, so I get to see what they’re actually trying to achieve. Higher in the organisation, there are probably more long-term, blue-sky ideas about what AI may be able to do. But at the engineering level, it’s simpler than that: ‘Help me re-consume what I’ve already done, and remove the repetitive tasks.’ That’s what they want. That’s what would make a difference today.

The pressure to do more with less seems to be intensifying. What does that mean for the industry going forward?

One of the companies I’m working with at the moment is in automotive. They tell me that historically, they typically had three years to develop a product, such as some sort of electronics for a safety or driving feature. Now, because the market is moving so fast, they’ve been told it’s got to be 18 months. And shortly, one year maximum.

If I’m going to do that, I need automation to help me. But it’s not just a case of doing more faster, I’ve got to do it differently. I’ve got to be able to develop things differently. If AI can do the tasks which could be automated but need more intelligence, the engineers can focus on innovation. If AI can help me re-consume valuable IP I’ve already generated, I can focus on the new stuff.

It purely seems that everybody needs to do more, in a shorter time, with fewer people. And we’re short of hardware engineers. If somebody gave me half a day off every week, I’d go to every secondary school and tell every kid doing physics, ‘You want to become an electronics engineer. Don’t believe anything else. That’s what the world wants’.