The promise of artificial intelligence in medical device development is often framed around speed: faster coding, faster documentation, and faster product development. But the question is often what it takes to turn that promise into reliable results. In a recent industry webinar, engineering leaders from Critical Software, IQ Inc., and Bausch + Lomb presented on the new possibilities AI is bringing to the med device industry. They broke down what actually changes when AI enters a medical device and how the right people with the right knowledge at the right time turn AI ambition into devices that reach patients.
Across the discussion, panelists describe a shift in what clients expect from development partners. Delivering functional software is no longer enough. Clients increasingly want to know where AI was used, which tools were applied, and how the resulting work was verified.
“Clients are no longer just asking us to build the software,” said Barbara Vankirk, CEO, IQ Inc. (a Critical Software Company). “They’re asking harder questions: ‘Where did you apply AI to this codebase? What did you use? How did you verify the results?’”
The expectations are shifting from delivering software products to delivering software and the evidence that it is trustworthy over its lifetime.
Start with the problem and the value
Before adopting AI, manufacturers need to identify the problem they are trying to solve and the value they expect to gain. Will the tool shorten development time, improve quality, or reduce costs?
“Treat AI as a capability you build, not a feature you’re adding,” said Vankirk.
She cautions against implementing AI simply because it is available or widely discussed.
“One of the biggest challenges is understanding its value. Does it make us go faster? Does it help us make things better or more affordable? We have to understand whether the tool provides that value,” said Birpal Sachdev, Director, Digital Systems R&D, Bausch + Lomb. “You can’t just jump into it and say, ‘We’re going to apply AI.’ You have to think about where it makes sense, experiment, and then scale up.”
He suggests asking yourself, ‘what problem are we trying to solve? Do we have the right data? And not just the right data, but the right quality of data?’
AI, of course, runs on data. So, Sachdev said that if you don’t have the right data, the right quality, and an understanding of the data, you probably aren’t going to be able to make it work.
“The most important thing is to look at the outputs and ask whether they’re meeting our needs,” said Sachdev.
Data readiness is central to that process. Having large volumes of data does not necessarily mean an organization has the right information to support AI. Scattered records, incomplete datasets, and poor collection practices can undermine results and require substantial preparation.
“Many people on this webinar probably know that 70% of the time goes into cleaning, acquiring, and managing data. Then 10% or 20% is spent actually getting results,” said Sachdev. “That 70% is critical, and that’s where a lot of things fall apart. People don’t pay attention to it or don’t do the job properly. Then the results don’t look right, and people say, ‘It’s hype.’ It’s hype because you used it for the wrong purpose, used the wrong tool, didn’t have the right data, or were missing the quality of data you needed.”
Speed depends on preparation and oversight
AI can help generate requirements, code, tests, and documentation, but learning to use the tools effectively may initially slow teams down. Developers must understand which tools fit particular tasks, establish appropriate controls, and evaluate whether outputs meet the intended requirements.
Documentation is discussed as a comparatively lower-risk starting point. Coding tools also offer potential gains, provided engineers verify that the generated code performs the intended function. Producing more material faster does not remove the need to review it. But it’s imperative to bring the right expertise into projects early. Data engineers, analysts, developers, and subject matter experts contribute different skills. Defining the problem and assembling the appropriate team at the outset can prevent costly rework.
“We need to make the right decisions at the beginning of the project,” said Gonçalo Almeida, Principal Engineer, Critical Software. “A decision in the first week might cost one day of effort. If we miss those points and make wrong decisions, then by the time things are ready, all the rebuilding, new analysis, and new documentation can have a major impact. It can take months, or we may need to discard a lot of work, whether we’re talking about hardware or software.”
Distinguish development tools from AI inside a device
A central distinction is between using AI to build a medical device and incorporating AI into the device itself.
In the first case, AI supports development activities such as coding, testing, and documentation. The resulting product still requires verification against its requirements. In the second, AI contributes to a device’s outputs, recommendations, or decisions, introducing additional questions about performance, oversight, and safety.
“The first part—using AI to write code or documentation—is being used a lot. That’s the easier part. The harder part is having AI inside the machine, where it’s advising or making decisions,” explained Sachdev.
Those are two aspects of the data models within the machine. If it’s advising, Sachdev said you’re leaving the final decision to the physician or healthcare provider. They have extra information and can make a more informed decision, and that’s easier.
“When AI is making a decision, that’s much harder,” said Sachdev.
There are two aspects to that as well. One is a static model. With a static model, you know you’ve tested the inputs and outputs. You can see that those inputs and outputs were tested, and the model is static and giving you results. You can still prove it through testing, verification, and traceability.
“But a dynamic model—which I don’t know whether anybody has done yet—is going to be the hardest part. If it’s making a decision and your parameters aren’t right, how do you handle that? You can’t have a machine make a decision and do something wrong,” said Sachdev. “I think we’ll get there over time, after going through various levels of AI… but the rules will have to be much more stringent, and the guardrails will have to be very specific so it doesn’t go out of bounds.”
Prove the system and preserve traceability
Verification must address the complete system, rather than the AI model alone. Panelists recommend defining the boundaries of AI’s role, establishing human review points, and testing complete pathways instead of relying solely on individual component tests.
Thus, accountability remains essential. Teams need to identify who reviews outputs, who approves decisions, and who signs off on the work. Confident or plausible AI responses still require scrutiny.
Traceability likewise remains a fundamental engineering responsibility. Organizations must be able to connect requirements, decisions, verification activities, and resulting actions. Safeguards and outputs should be understood and documented, with evidence maintained throughout development and release.
“Traceability doesn’t change whether you use AI tools or anything else,” said Sachdev. “You still need to know what decisions were made, by whom, and at what time. Can you trace them? Can you map everything from the requirements all the way to the actions? Are they traceable? Are they within the boundaries? Are they consistent with the reasons they were created? The evidence and verification steps have to be created, and everything has to be traced from one end to the other.”
AI is a capability that organizations must build through disciplined processes. Its practical value lies in helping teams perform well-defined work more efficiently while preserving the engineering judgment, verification, and accountability needed for safety-critical products.
Editor’s Take: AI can fall short of its promise when adopted simply for the sake of using it. With appropriate guardrails, human oversight, and quality data, AI can help medical device companies streamline testing and documentation, scale processes, and strengthen traceability.
