Could AI Be Your Design & Engineering Partner During Product Development?

Instead of taking over, AI could act as a companion to human engineers throughout the design process and help improve their work along the way. Product Creation Studio CTO Scott Thielman—and now UW professor—explains how.

May 7, 2024

15 Min Read
Scott Thielman/Product Creation Studio

Several new artificial intelligence (AI) tools have emerged for designers and engineers, but some may worry about using them. As chief technical officer for Product Creation Studio, Scott Thielman has been exploring AI’s potential for his product development team. And after he was invited to teach a class at the University of Washington Bothell on AI in product development, he came to some significant conclusions about AI.

The idea for such a course on AI came up last year when Thielman, who serves on UW’s advisory board for the mechanical engineering program, was asked for ideas on new elective courses. The inquiry came “right after I had been challenging my team to look into how we are using these tools,” he explains. It was the early days of ChatGPT, he says, when there was a groundswell of both interest and fear about what AI could mean for designers and engineers.

“We had a roundtable with our team, and I was surprised that they were already using AI in a number of ways,” he continues. “For instance, one of our engineers who does embedded software was using the tool GitHub CoPilot to help build code. So, I started playing around and developed in one afternoon a semi-functional game of checkers with a graphical interface using JavaScript, and I had never programmed anything in JavaScript in my life. I realized that if AI allows me to do that, what else could we do with AI? That got me thinking that a course on using generative AI for design could be good, because it looks like this is going to have an impact. Students could start learning about how to deploy AI to make better designs and to design faster.” His idea for the course was so well received, UW asked him to teach it. 

Thielman kicked off the course showing students a snippet from the movie Iron Man in which Jarvis is helping Tony Stark create his next Iron Man suit. “There's AI being used to develop the suit as well as AI embedded into the suit. This is where I see this potentially going—the potential for speed and quality of first articles. We're not there yet—we don't have Jarvis that magically does that—so there's still these holes.” 

Gathering User Needs 

But Thielman sees AI’s potential for product designers and engineers. “If we are good design and development engineers, what we should really care about is the impact that products are going to have for end users. If I pull an example from our history, we developed cochlear implant audio processors for different companies, such as Advanced Bionics’s first waterproof version. I'm really proud of the work that went into that product. We had a certain approach to how we pulled out user needs and worked through development toward requirements and testing. But looking back at that now 10 years ago, should we be satisfied with how we developed it? I mean, if we have better tools that could allow us to try more options or do more testing within the same time envelope or to execute that even more rapidly so that patients got access to the product even earlier, aren't those the things we should be striving for?

“So, we should be thinking about trying to have that impact with the work that we do, and the tools that we bring to do the job are really just things that help us accomplish meaningful work more efficiently,” he continues. “That doesn't necessarily mean that we'll be done and out of a job more quickly—it means that there will be more work to do. We'll see rapid benefits of the impact of the products we're working on, and the pace of demand will increase. The faster and more efficiently we can deliver quality solutions, the better it is for everybody. That's where we need to be looking at these solutions even though they're not fully baked in all cases at this point.”

AI may already be helping to shorten design cycles. “I have definitely seen cases where this is making some of this work go a little faster or helping us get to the meat of a decision faster,” Thielman says. “I can't say for sure that it's collapsed the timeline of a full holistic development effort, but I think we're on that road. Sometimes it's a question of how will we use the gifts of the tool set that we have access to now; will we use it to save time or money; do more with a smaller team; explore a broader range; or make more prototypes to have a higher level of confidence in our solution going out the door?”

Digital Personas?

Thielman is starting to see some excitement around how well some of the higher-level models like ChatGPT, Anthropic Claude, and others can act as a persona and represent a customer with certain demands and needs. “You can't replace upstream marketing or user testing by any means, and I’m not suggesting that. But if you're an organization that really does understand your users and you can build these digital personas, they can help in the early evaluation of ideas while a concept still lives on paper,” he says. “Maybe now it could be done as part of a decision matrix to help guide where R&D budgets are directed.”

He adds that “a better decision on the right product that's going to meet user preferences in the early days has a compounding effect throughout the whole R&D cycle and process. A well-informed early-stage decision increases chances of success when a product hits the market. And screening out the doomed ideas early gives us the luxury of spending our time doing things that give us greater confidence in a promising product.”

Scenarios could involve presenting a concept to a user persona played by the digital assistant, or it could involve asking the persona questions to better understand a market opportunity or how a product experience fits into their lives, Thielman explains. But he does acknowledge that it would be hard for a large language model (LLM) to fully experience a product. “I think being able to roleplay is what’s more valuable. It's almost like holding the mirror up in a different way, where the value is not really in the AI persona itself. It's in the value of the conversation that the engineering team is able to have. A team that's probably not spending enough time talking to users can suddenly do that via simulation.” 

Improving the Design Process

When asked for examples of where a design has been improved by using AI, he pointed to some of the work his team has been doing using AI Copilots like those found in GitHub to help them build code. “We've actually used that on some fairly complex life science automation projects where multiple circuit boards are talking to one another and coordinating a series of actions that have to operate effectively and safely,” he says. “Because of all the interactions, testing the boards for functionality becomes a challenge, and we have to repeat that testing every time a significant change is made. So, on both sides of that development, the operational firmware development was assisted by AI and the test software was developed using AI. This allowed us to speed up the process and get a more comprehensive test platform. In that case the benefits were more about design robustness vs. a major time savings, but it definitely benefited the product test cycles as we go forward.

Other recent examples include concept ideation, such as a recent project that entailed determining how acoustical noise exposure might be perceived by subjects. Thielman’s team was using several LLMs to understand how noise might be perceived and devise ways to evaluate and measure it effectively. “One of the models pointed out that rather than just evaluating sound pressure level, we may want what's called a loudness measure, which is, a subjective measure defined in standards that we had not been looking at previously. Taking into account how people actually perceive the sounds opens the possibility to alter perception rather than just the physical sound pressure level. This led to a rethinking of how we would test and evaluate it.

Thielman sees the potential for AI to impact the engineering development process in five areas:

  1. Problem/opportunity definition

  2. Research to expose and establish requirements

  3. Ideation or invention of solutions

  4. Development of promising solutions via build, test, learn cycles

  5. Communication of results.

“Each one of those is a place that AI can help, and the universal way that I think these tools can help is by stimulating conversations. We've seen that with our designers and engineers, where I'm having a conversation with somebody about something deeply technical that they're focused on, and I'm just there as a wall to bounce back ideas and thoughts to keep the conversation going.”

AI Challenges and Risks

There are challenges and risks to using AI, however.

“A good engineering process involves testing and learning from solutions you put forward. We're always trying to figure out how to test our solution against requirements to get confidence in it. And I think that insulates us somewhat from the potential for hallucinations that lead to wrong-headed design solutions,” Thielman says.

But I think we have to be careful,” he warns. “We can easily burn a lot of time going down the wrong path if we make a bad decision based off incorrect data. That error could compound through a development cycle in a negative way. And so, we really have to fact check the information that we're getting.

“And, of course, there's notable concerns about bias appearing in many of the language models and that could find its way into everything from those user personas,” he adds. He also says that it could be “somewhat limiting if all the information we're getting and all the thinking we're doing is bounded by what's been done before and what's being produced by these models.”

Good News for Humans

It's also important that designers and engineers continue to tap into what humans excel at: creative ideation. “A willingness to break the rules is a real strength of humans,” says Thielman, so when it comes to inventors and designers, “the challenge is figuring out how do we encourage that. The LLMs aren't necessarily helpful at all aspects of that. There are cases in which using a tool like ChatGPT in an ideation session actually limited the number and range of possible solutions that came out, and what you want in those sessions is the odd breakthrough idea that really changes the course of a company or a product line.

“So, we'll have meaning for a little while,” he continues. “Part of the reason I first started looking at this stuff that led ultimately to the chain of events for this class was that fear of missing out, but also fear that our jobs are done. That people are just going to go to ChatGPT and say, ‘Design me a new blender or whatever,” and they'll be on their way. And so far my experience with it has been that we're nowhere near that level with the tools we have. Because the real value we get out of them is still proportional to the time and thought we put into the prompts."

He adds that it's still going to be the job of “the designer or engineer in their subject area to bring the right context and discussion to the table to get any meaningful output from a tool in a way that brings out the interesting stuff. And half the magic—or maybe most of it—is still in our brains. Maybe it's a parlor trick that gets us to consider ideas more broadly and augments our capabilities more so than replacing them.”

The other very human aspect of design and engineering is that humans could be inspired by something else, perhaps in nature or a totally different industry, and then come back to their project with a fresh set of eyes or a new solution. “There are examples of art and design that are inspired by nature, by a car they loved, or by something else. We've seen our industrial design team use diffusion models to produce images that may be combined with user inputs textually in a prompt. You get something, whether or not it's useful, but you can quickly try some things when you combine some of these. The human is still such an important piece, but the tools can actually help the human cover more ground, more quickly, too. And that's the exciting thing.”


AI Tools to Consider

AI tools that interest Thielman include ChatGPT, Anthropic Claude, and Google's Gemini. He is also experimenting with, beta testing, or admiring the following programs and companies:

  • A project by Google called NotebookLM that “makes it fairly easy to put your own content data into the system and then search those documents and allow you to translate some of the output into notes,” he explains. 

  • Perplexity AI for finding references to content. “It allows you to add some of your own content in a way to put your own data papers and reference material into the system and then search. It does a good job of telling you where it's drawing the information from, and I think that's interesting in scientific fields or early work for which we're trying to back up the decisions we're making. So, it's nice to have some traceability to where we got our information from. Even though it was processed through a large language model, it was drawing from this specific content that starts to give us a little more confidence in the output.”

  • NyquistAI, “which is a model that's been trained on all the regulatory information around the world related to medtech or biotech and allows users to ask questions about that data. It allows them to make better decisions about their strategy going forward in the product category that they're looking at.”

  • for product development. “It converts a textual conversation about a product concept into a set of requirements, user personas, and project risks to be considered. It even creates initial imagery of the concept and puts it all in a draft pitch deck. If you're a product or category manager involved in product development and need to quickly evaluate a funnel of early ideas, this could provide a presentation deck that you could take to your stakeholders and your team to help with that opportunity evaluation. What would probably take a number of weeks for a team to do in the past can be put together pretty quickly.”

  • Cadstrom for developing circuit board designs. “Built for electronic design engineers, it interprets the intent of the circuit itself, evaluates all data sheets and design structures, and brings attention to areas that might be of concern. It essentially acts like a design reviewer looking over your shoulder, and we've seen the power of design views in terms of reducing errors in all sorts of design, especially in electrical circuit design.”

  • Cyrus Biotechnology provides services and software for “designing protein structures with a combination of classical physics-based modeling and AI models that help them make decisions about what sequence of amino acids should go together, such as for developing drug candidates for disease.”

  • Potato for "creating, editing, and refining wet-lab and dry-lab protocols using its AI copilot. You can incorporate lab-specific methods and relevant scientific literature to ensure reproducibility."

Thielman believes such tools can help designers and engineers identify serious ideas and concepts that meet all product requirements. “The early stuff is going to be around understanding problems and organizing requirements and then evaluating your work as you develop to see whether it actually meets the requirements,” he says. 

Future Outlook for AI

In the end, the success engineers will have with AI goes back to the time they're going to put into it, Thielman says. He found that to be the case when he used it to help prepare the UW coursework. “I think if you lovingly train the AI and give it examples of your writing, you can get something that's pretty decent. It's exciting to dump some raw notes and then get a first draft. I may rewrite the whole thing, but it got me going off the starting line in a way that sometimes is hard to do when you're staring at a blank page.”

Regarding some of the fears about AI ending humanity, Thielman believes there is a way to utilize AI’s benefits and keep it in check.

“I like to think that if we keep the goal of augmented human existence in mind, that that will be OK. Will there be some scary points along the way? Probably. Are we that close to them? Probably not as close as the fear mongers would suggest, but then again, it's hard to know. We don't do well perceiving logarithmic or exponential change, so I think it's good to be concerned, but it's not helpful to be terrified. Right now, it's just a tool that allows us to do certain things, and it hasn't allowed me to totally disengage my mind. It's just been a very strong helper in organizing information.  As with other tools we should choose to make this into something that makes human existence better and not something that puts us at risk.”

For the UW course, Thielman has explored how AI, particularly large language models, can help design things. “We learned about the design frameworks and patterns of AI that we've seen be applied and then in the second half I'm intending to focus on how you would put AI into your product,” he says. “Is there a machine learning aspect or hyper-personalized experience AI could bring to the product? Could we use lower-level machine-learning tools to add capabilities to that product? So, we’re switching over to more machine learning, which is a little more classical, about using data to train a machine to make predictions. They'll have the power of ChatGPT and Gemini at their fingertips to help them learn and execute the machine learning tasks.”

Currently serving in an affiliate professor role for up to five years, Thielman may teach the course again. “I think the teacher ratings from the current student cohort will decide my teaching career! But I plan to be a student of AI innovation regardless,” he says of UW.

Sign up for the QMED & MD+DI Daily newsletter.

You May Also Like