RAPS Convergence 2026 Takeaways: Why Human Judgment Still Matters in AI-Assisted Regulatory Writing

Across sessions, exhibit hall conversations, and discussions with regulatory professionals at RAPS Convergence 2026, artificial intelligence was impossible to miss. But what stood out to me was not simply how often AI came up. It was how much the conversation had matured.

A few years ago, discussions often began with, “Can we use AI for this?” Today, the questions are much more practical: Where does the evidence come from? What remains under human control? How is generated content reviewed? And can the final output withstand regulatory scrutiny?

That shift was at the center of the session I spoke at, “Can AI Actually Write a Defensible State of the Art? Lessons from Building AI for Regulatory Evidence Synthesis.”

Drawing on what we have learned while developing AI-assisted Clinical Evaluation Report authoring capabilities within CAPTIS®, we explored a simple but important idea: the value of AI in regulatory writing cannot be measured by speed alone.

For clinical evaluation teams, productivity matters. But regulatory outputs must also be traceable, reviewable, and defensible. A polished paragraph is not enough if the reviewer cannot determine where the information came from, how it was interpreted, or whether the conclusion is supported by the underlying evidence.

The AI Conversation Has Moved from Possibility to Practice

The growing maturity of the AI conversation was one of my clearest takeaways from RAPS. Regulatory professionals were less interested in broad claims about what AI might do and more interested in what actually happens inside the workflow.

The questions I heard focused on implementation:

• Can AI support literature review while leaving inclusion and exclusion decisions with the human reviewer?
• How can a generated statement be traced back to its source evidence?
• Where does expert review take place, and which steps remain under human control?
• How should AI-assisted outputs be validated before they are used in regulatory documentation?
• Can AI support the ongoing maintenance of a CER, not only its initial authoring?

These are the right questions. They show that AI adoption in regulatory affairs is moving beyond experimentation. Organizations are now trying to understand how AI can fit into real regulatory processes without weakening the controls, judgment, and traceability those processes require.

Regulatory Writing Is Not the Same as Text Generation

This distinction becomes especially important in Clinical Evaluation Reports.

Some CER sections are more structured and repetitive than others. They may follow established templates, predefined data structures, and relatively consistent authoring patterns. In our experience with AI-assisted CER authoring in CAPTIS®, we have seen approximately 75 to 80% reduction in effort across selected structured sections of the CER.

Those gains are meaningful. Automating or accelerating repetitive writing activities can give regulatory professionals more time to focus on complex datasets, nuanced interpretation, and decisions that require subject matter expertise.

State of the Art, however, presents a different challenge.

A SOTA section is not simply a summary of scientific literature. It requires an understanding of current clinical practice, available technologies, treatment alternatives, risks, expected performance, and the broader evidence landscape relevant to the device being evaluated.

The task is not only to retrieve information. The evidence must be interpreted within the appropriate clinical and regulatory context.

This is why AI-assisted evidence synthesis is fundamentally different from asking a general-purpose model to summarize a set of documents. Readability is only the starting point. The regulatory professional must be able to see how the content was derived, identify the evidence that supports it, and assess whether the interpretation is appropriate for the intended regulatory purpose.

Human-in-the-Loop Is Not a Safeguard Added at the End

One of the strongest themes from the session and the conversations that followed was the need to preserve a clear role for regulatory experts throughout the process.

AI can accelerate evidence discovery, organize information, support drafting, and prepare content for review. It can reduce repetitive work and make large volumes of information easier to manage. But it should not remove the regulatory professional from the decision-making process.

Regulatory experts still need to determine whether evidence is relevant, whether an interpretation is appropriate, whether a conclusion is adequately supported, and whether the final output is suitable for its intended use.

The most effective model is therefore not one in which AI makes regulatory decisions independently. It is a structured human-in-the-loop workflow in which technology

supports the work that can be accelerated, while experts retain responsibility for the decisions that carry regulatory significance.

Traceability Will Be Central to Trust

Reliability was another recurring concern at RAPS.

If AI generates a statement in a CER, the regulatory professional needs to know where that statement came from. The reviewer should be able to trace it back to its supporting evidence and verify that the interpretation is appropriate.

Without that level of traceability, AI can shift work rather than remove it. Drafting may happen faster, but the review team may then spend additional time reconstructing the evidence behind the generated content. That does not solve the underlying productivity challenge.

For AI-assisted authoring to deliver meaningful value, both authoring and review need to become more efficient. Evidence-level traceability is therefore not an optional feature. It is part of what makes the output usable in a regulatory context.

This has been a core principle in the development of CAPTIS®. The objective is not simply to generate more content, faster. It is to support a workflow in which content remains connected to the evidence and can be reviewed efficiently by the experts responsible for the final regulatory assessment.

CER Maintenance May Be One of the Most Practical Opportunities

Another topic that came up repeatedly at RAPS was CER maintenance.

Many medical device organizations already have established clinical evaluation processes and years of existing documentation. Their most immediate challenge is not necessarily creating a CER from the beginning. It is keeping the document current as new literature, clinical information, and post-market evidence become available.

Maintenance requires teams to identify what has changed, determine whether new information affects existing conclusions, and update the documentation accordingly. This can involve substantial effort, particularly when the evidence base is large or changing frequently.

AI can help identify, organize, and surface potentially relevant information more efficiently. It can support comparison and review preparation. But the final assessment still depends on expert interpretation.

A regulatory professional must determine whether the new evidence is meaningful, whether it changes the existing evaluation, and what action should follow.

This is a good example of where the combination of technology and expertise can create value. AI can reduce the burden of finding and organizing change. Human judgment determines its regulatory significance.

What Regulatory Leaders Should Focus on Now

My biggest takeaway from RAPS was that AI adoption in regulatory affairs is no longer a future discussion. The industry is moving from “Should we use AI?” to “How do we implement it responsibly?”

For medical device organizations, the focus should extend beyond selecting a technology or automating an isolated task. Regulatory leaders should consider how AI fits into the broader operating model, including:

• Where human oversight occurs and how decisions are documented
• How evidence-level traceability is maintained
• How subject matter expertise is captured and applied
• What validation and governance are required for AI-assisted workflows
• Which activities are appropriate for acceleration and which must remain expert-led

The opportunity is significant. Regulatory teams are managing growing workloads, expanding evidence volumes, and increasingly complex documentation requirements.

AI can help reduce effort, improve consistency, accelerate evidence synthesis, and allow regulatory professionals to focus more attention on the complex areas of clinical evaluation.

But regulatory defensibility must remain central to the design of these workflows. Efficiency is valuable only when the resulting work can still be understood, reviewed, and trusted.

Moving Faster Without Compromising Regulatory Rigor

So, can AI help regulatory teams move faster without compromising regulatory rigor?

Yes, but success depends on how it is implemented.

AI can support evidence discovery, organization, drafting, and review preparation. It can reduce the burden of repetitive activities and help teams work more efficiently across complex documentation processes. Regulatory experts, however, must remain accountable for interpretation, judgment, and final decisions.

Leaving RAPS, I came away with a clear impression: regulatory professionals are not looking for AI that replaces expertise. They are looking for AI that amplifies it.

The organizations that succeed will not necessarily be the ones that automate the most. They will be the ones that find the right balance between efficiency, traceability, and expert judgment.

In regulatory affairs, trust remains the ultimate deliverable. When AI-assisted outputs are traceable, reviewable, and defensible, speed and rigor do not have to compete.

That, to me, is the real opportunity for AI in clinical evaluation.