AI has been progressively advancing for decades, and with tools like Google Assistant, Siri, and Alexa, people have been getting used to having conversations with non-humans in a way that we’ve never done before.
The real change for medical device design is that engineers can now interact with AI naturally, making it practical to integrate it into everyday engineering work. The true payoffs of this advance do not, however, lie in productivity gains, as many assume. With the support of AI, engineers are becoming more ambitious about the problems they’re willing to solve—and that is revolutionary.
The real value of generative AI in engineering
When we started incorporating generative AI in our work at HiArc, we anticipated that it would enable our engineers to save time. What we found is that there is definitely some of that, and some of it’s pretty substantial, depending on the nature of the work that they’re doing. It’s very helpful with things that are well-defined—like creating documentation, helping to devise test plans, or writing scripts to test the instruments and verifying that things are working as expected. There is a lot of engineering ‘busy’ work that’s either disappearing or requires less manual effort on behalf of our individual engineers. And that’s the first layer of savings.
But what’s really driving value is that our engineers are becoming more ambitious about the problems that they’re trying to solve. They’re gaining more confidence because they have AI tools to back them up, and they know they can get more work done in less time. As a result, they’re willing to engage with challenges or problems that previously they might not have.
Example: Systems modeling
For instance, we have been exploring the use of a standardized language called SysML 2.0—a systems modeling language. It’s an open standard and provides a mechanism that we can use to store some of our designs in a centralized way and apply it across many of our projects. The 2.0 version of the language is fairly recent. There’s not a big ecosystem around in terms of tooling, and the tooling that is available is expensive and proprietary.
A group of our engineers looked at it and started building our own tools using AI. Now they’re creating rendering agents to build diagrams based on these open standards. It’s clear that they never would have tried this before.
Productivity gains matter
In regulated markets, there’s a lot of work that we must do that isn’t necessarily core engineering work. We must make sure that we have the right tests in place, the right level of documentation, and the right procedures in terms of reviews. From an engineering perspective, this is quite a lot of ‘busy’ work or red tape.
AI can be leveraged very effectively here. For example, we’re adopting continuous compliance as we’re building our software. Historically, there’s always been a gap between what functionality is presented and represented within the software and what our documentation says, because there’s been some level of manual interaction that needs to happen to update documents and go through approval processes. Software may work perfectly well, but it may take several weeks until there is documentation.
By using AI and interacting with other tool sets, we can start generating draft versions of documentation in real time as the software is being built. We then have a tight coupling between the version of documentation artifacts and the version of what is being generated. Of course, we still need to go through the appropriate amount of rigor in terms of review, but it reduces a lot of the overhead and time spent waiting for the documentation to catch up.
As a lot of this type of work gets reduced, it also allows our engineers to focus on the areas that they have expertise in, to really think hard about the problems that they’re being asked to solve.
AI vs human judgement
The way to achieve most productivity gains and the most utility out of AI is via a human-AI interaction, and this is particularly the case with medical devices, which are highly regulated. You need human judgment every step of the way. Because of the way that our regulations work, we need to have accountable human decision makers. AI is not a tool that makes decisions, but a tool that provides decision support.
Ultimately, AI is going to become a commodity and, as it becomes more accessible and inexpensive, competitive advantage isn’t going to be gained by those with the smartest AI. It’s going to be those with the smartest engineers, the smartest people who know how to leverage AI to become most effective.
Human engineering judgment is absolutely critical moving forward, and we need to be careful, as an industry, that we don’t become so reliant on AI that we’re ignoring the fact that we have people with years and decades of expertise.
The changing role of engineers
As AI becomes commoditised, the most important skill for the engineers of tomorrow is critical thinking: how to take a skeptical point of view and use the right tools to validate and verify designs and decisions.
To create an environment in which engineers can thrive, they should be given the space to build trust with AI, with guidance and guardrails. The more you work with AI, the more you understand what it’s good at and what it’s not good at. People need the opportunity to learn that on their own so they can become comfortable interacting with it and maximising its potential.
AI also has the potential to enhance cross-functional collaboration, which is critical in engineering complex devices. Specialists are not as dependent for everyday questions on their counterparts in other departments. We can now create centralised knowledge repositories to help dismantle any remaining silos and allow people to work more interactively and collaboratively.
Risk vs reward
In medical and other regulated industries, there is a focus on risk, and understandably so. There are risks with AI engineering, just as there are with any other kinds of engineering. In a heavily regulated industry, we already have standards by which we must test things. We need independent peer reviews and a separate QA department that validates that we’ve done what we’ve said we’ve done. All those things are in place, and the same processes that protect us from human error can be leveraged to protect us from AI error as well.
We shouldn’t judge AI by whether it’s perfect; we should judge it by using the same engineering rigor we already apply to humans. We should be agnostic as to whether it was AI or a human who created the code, the product, and the design. We need to be just as rigorous as we’re testing it with the same set of standards. And the hope is that by combining the human and the AI agents together, we’ll end up with a better hybrid product than either of them would have generated on their own.
Example: Timing simulations
Our systems group writes a lot of timing simulations for our instruments based on the workflows that the instruments will ultimately use. Historically, this is something that’s been done manually on a project-by-project basis. One of our pilots was about using AI to create a generalized timing simulator. It took half the time of what they would have done manually before, but it’s also written in such a way that it can easily be adapted to new programs and new products.
As it’s now a more-standardised application, they’re adding functionality so they can have diagrams of the instruments, which will show where the chemicals are within the instrument based on the timing simulator. So, we now have a much more sophisticated application than we otherwise would have had that’s generally applicable.
