The FDA hails the recent rollout of its internal AI tool, Elsa, as a major move to tackle the crushing weight of regulatory review where documents thousands of pages long are commonplace. But as reports of a rushed, buggy rollout surface, one regulated AI expert Erez Kaminski, a former AI strategist for Amgen and now CEO of Ketryx, says the agency may be in for an uphill battle — at least initially as it builds out the infrastructure. “I think it’s just really, really hard to make regulated AI work well,” Kaminski said.
One of the challenges is the scale of regulatory documents. While it might seem simple in the abstract to use genAI to help review documents, regulatory documents are far removed from, say, a homework set or even many legal documents, which are often “much shorter,” Kaminski said. “As a result, their context window is much simpler. And as you try to scale it up for this massive scale context window, I think it’s quite hard.”
Beyond the context window
The issue, Kaminski argues, isn’t just about the size of the documents but the fundamental architecture of the AI system itself. While early reports suggest that Elsa is largely based on a large language model (LLM), Kaminski posits that a more resilient strategy would involve a “neuro-symbolic” framework, which blends the pattern-recognition power of modern neural networks with the structured, rule-based logic of traditional symbolic AI. This hybrid approach would allow the system to first break down the monolithic review process into a logical sequence of smaller, verifiable steps, much like a flowchart, and then deploy generative AI to execute those specific, smaller-context tasks where it excels.
Kaminski
Without this structured approach, even the most sophisticated LLMs can become overwhelmed by the interconnectedness and complexity of regulatory documents, where information is scattered across thousands of pages and every detail must be traceable to its source.
The documentation deluge
The other reality is that developing regulated products is an inherently complex affair. To illustrate the difference between typical AI applications and regulated environments, Kaminski contrasts a journalist’s workflow, involving perhaps 10 to 20 steps from interview to publication. The process of developing a drug, from hit to lead to regulatory approval and manufacturing has orders of magnitude more complexity.
The situation is similar in the medical device industry, where the most common pathway, the 510(k) process, hinges on the complex burden of proving “substantial equivalence” to a predicate device. This means every decision, from design to testing, creates a branching path of documentation requirements.
Kaminski presented data from McKinsey’s Numetrics R&D Analytics report showing that between 2006 and 2016, medical device software complexity grew at a 32% compound annual growth rate while productivity increased at only 2% CAGR.
A human-AI combination lock
The FDA faces significant workforce challenges, with recent reductions affecting multiple centers, exacerbating ongoing staffing pressures. Attrition at the agency has hovered around 13% since fiscal 2018. Against this backdrop, reviewers often struggle to keep up with the sheer volume of information.
Rather than relying solely on large language models, which process text sequentially through pattern recognition, Kaminski advocates for combining them with symbolic AI, the rule-based approach that dominated early artificial intelligence research, sometimes dubbed “good old-fashioned AI.” Symbolic AI is something like a chess player who knows every possible move and strategy beforehand, using explicit if-then rules to navigate complex decision trees.
Such complex rules can prove instrumental in the use of AI in regulatory science, and in creating validated AI agents that can help complement the work of humans — either at the agency or in industry. “At the end of a pharmaceutical line, for instance,there’s a quality control machine that tests every vial to see if it’s good or not good. It used to be that a person with a master’s degree or Ph.D. would check every one of them,” Kaminski said. “That is so inefficient that even biopharma companies can’t support that, so they created automated testing machines. Many of them leverage machine learning or AI.”
Toward accountable autonomy
Kaminski anticipates the FDA may need to evolve its approach as document complexity continues to outpace human review capacity. Kaminski notes that the agency, along with others working on AI in regulatory context, will need to marry neural approaches with neuro-symbolic ones. “People who’ve done this will tell you you need both,” he said.
He describes the industry goal as achieving “accountable autonomy,” In other words, “AI that can operate independently, but within clearly defined boundaries. The beauty of AI is in its potential to offload complexity, but only if we can trace its decisions, validate its actions, and ensure its safety at every step,” he said.
When implemented with the traceability infrastructure and validation protocols that regulated industries require, Kaminski reports dramatic efficiency gains: “We’ve been able to take people down from a year-long software release process to a week.”
Looking beyond immediate efficiency gains, Kaminski frames the challenge in broader terms: “I think the floodgates are open. People are seeing that there’s an immense amount of work that’s been done manually that could be done more with a combination of human and computer… The fact that there’s not better medicine, a good portion of that is because [of the volume of paperwork].”
