As manufacturers increasingly use artificial intelligence to generate procedures, quality records, and other controlled documents, attention is shifting from what AI can create to how companies govern, review, and approve those outputs.
For makers of medical devices, that shift reflects a growing recognition that speed and efficiency don’t eliminate regulatory responsibility. AI may help draft SOPs, work instructions, manufacturing records, specifications, corrective and preventive action (CAPA) documentation, quality forms, and training materials, but companies remain responsible for ensuring those outputs are accurate, controlled, and compliant.
“AI is a fantastic tool, but it’s a tool,” said Christina Arnt, QualityHub’s VP of Medical Device Compliance. “The human element becomes more important, not less.”
For device firms, that means building AI into existing quality expectations rather than creating a parallel process with lower standards.
Below are four tips from Arnt on managing AI-generated content.
Tip 1: Treat AI as Authoring Tool, Not Approval Authority
One analogy Arnt said stood out during recent discussions she’s had with industry stakeholders is the idea that AI is a “copilot.” But, she says, “you simply can’t fly a plane with only a copilot.”
Arnt said that concept captures a growing misconception around AI adoption in regulated industries.
“AI is supposed to be the assistant. It’s supposed to be the copilot. It’s not supposed to be the pilot,” she said, noting that distinction becomes especially important when AI is being used to create or modify procedures or quality records.
Many quality documents require more than complete sentences and polished formatting. Procedures reflect years of operational decisions, historical process changes, prior CAPAs, risk decisions, and regulatory interpretation. AI can generate language quickly, but it doesn’t inherently understand why a process exists or whether the proposed approach aligns with actual operations.
Instead of treating AI-generated procedures as approval-ready drafts, Arnt said companies should challenge AI outputs the same way they would challenge work produced by a new employee.
Manufacturers should ask:
- Does this procedure reflect actual practice?
- Does it align with existing controls and quality records?
- Are references to regulations accurate?
- Has anything critical been omitted?
- Would following this document consistently produce compliant outcomes?
Tip 2: Keep Experienced People in the Loop
According to Arnt, one of the more surprising trends she’s heard is the belief that AI allows organizations to reduce reliance on highly experienced personnel.
“A lot of companies have this perception that they can get AI into their systems, and then they can terminate all the expensive high-level employees,” she said. “But those are not the people to get rid of because that’s exactly who you need to review AI outputs: senior-level people.”
Arnt emphasized that she’s not referring to executive leadership positions, but rather employees with years of accumulated operational and compliance experience. Such seasoned professionals often understand why decisions were made, where processes historically failed, and what tradeoffs shaped current procedures.
“The procedure may come from AI, but the accountability still belongs to the manufacturer.” – Christina Arnt
Without that context, reviewers may approve AI outputs because they appear reasonable instead of evaluating whether they’re actually correct. “You cannot take AI output and blindly use it,” she said.
Arnt added that teams are struggling with another issue: people are being asked to review AI-supported processes before learning how the company’s underlying processes work.
“New employees in the in the medical device space need to learn about the product, learn about how it’s made and how it’s used, and then learn about the quality process or the design process that they’re going to be working in long before they should be asked to assess AI outputs in those areas,” Arnt said.
She went on: “Companies are pushing new employees to use AI as part of their day-to-day jobs, but the important thing to remember is they need the company education – the company training on the product and the processes – before they can be asked to judge the adequacy of something coming out of AI.”
Tip 3: Treat AI Like Any Other Quality-Impacting Process When Building Controls
Arnt said companies should stop thinking of AI as simply another software implementation and instead manage it like a controlled quality process.
“You have to have a robust process before leveraging or overlaying AI tools,” she said.
That includes understanding whether AI is layered over an existing workflow or embedded inside purchased software, and understanding how machine learning could influence future outputs.
“You really have to understand how that was created and how the machine learning is going to help that evolve,” Arnt said.
Manufacturers should establish governance around:
- How outputs are created,
- How updates are managed,
- Who approves final content,
- What triggers additional review,
- How performance of AI is monitored, and
- How changes are documented.
Arnt said one of the biggest concerns she sees is assuming AI outputs remain stable over time, reminding manufacturers that AI can drift and “hallucinate.” That means companies need a way to evaluate ongoing performance.
While there isn’t yet a regulatory gold standard for AI monitoring in medical devices, she said many organizations are borrowing concepts already familiar to quality professionals, including process capability and monitoring for drift.
Tip 4: Regulators Likely Won’t Care if AI is ‘Author’ of SOPs, Other Docs
From an inspection standpoint, Arnt said companies should expect questions around automation and AI-enabled processes.
“The FDA, in particular, will ask about automated processes, and if you’ve automated anything,” she said. “AI absolutely falls under that.”
She said firms should be prepared to explain where AI is used, what controls exist, and how outputs are reviewed and approved.
Arnt emphasized that investigators likely will not take issue with the use of AI. Instead, the issue becomes whether procedures contain inaccuracies or unsupported conclusions, or evidence they were never meaningfully reviewed.
That expectation was reinforced recently in an FDA warning letter issued to a drug manufacturer. The agency said the company used AI to generate specifications, procedures, and production records but failed to adequately review those outputs for compliance. Investigators also documented broader quality issues and noted the firm’s officials reportedly said it was unaware of a legal requirement because its AI tool hadn’t identified it.
For medical device firms, the message is clear: AI can help generate procedures faster, but it doesn’t transfer regulatory responsibility.
“The procedure may come from AI, but the accountability still belongs to the manufacturer,” Arnt said.
