Rethinking Regulatory Submissions in the AI Era: How FDA’s Elsa May Reshape Sponsor Strategy

FDA Introduces “Elsa,” an AI-powered Tool

The FDA recently introduced “Elsa,” an AI-powered tool designed to assist regulatory reviewers by reading, summarizing, and transforming content into structured formats like tables and databases.¹ The implications are significant: inventors must now rethink how they craft regulatory submissions, not just for compliance, but also for clarity and machine interpretability.

Why Elsa Changes the Game

1. Elevated scrutiny on subtle findings
   Elsa is picking up signals that human reviewers might overlook or dismiss.² For example, “normal background findings” in animal studies are now flagged more frequently, prompting the need for more detailed explanations up front.* Sponsors may need to preemptively address these in submissions, rather than assuming they are understood.
2. Ambiguity becomes a liability
   Unlike humans, Elsa doesn’t interpret meaning; it restructures information. If the submission text includes vague language, Elsa may misrepresent or misclassify content. As one expert cautions, “if there are ambiguities in the source information, that can trip up the models”.² That means clarity of writing, i.e., precision in terminology and structure, is more essential than ever.
3. Humans still retain the final say
   Despite AI’s growing role, regulatory professionals confirm humans remain central to decision making: “final fact checking and decision making is still going to need a human.”² AI tools like Elsa augment but do not replace the rigor and judgment of human reviewers.

Strategic Takeaways for Regulatory Affairs Teams

Audit language proactively

Revise submissions to remove vague phrasing and enhance structure. Consider writing with Elsa in mind: think in tables, metadata tags, and explicit statements.

Suboptimal phrasing:
“The observed elevations were not considered clinically significant.”

Refined example:
“Observed elevations in liver enzymes were transient, non-dose dependent, and consistent with historical baseline ranges; they were not considered clinically significant.”

Why it works: The refined version removes ambiguity by quantifying duration, dose relationship, and historical context, elements Elsa can parse more accurately.

Anticipate AI flags

If Elsa is prone to challenge typical nondrug-related findings, clearly call out these observations in your submission narratives and contextualize them as background, baseline, or non-adverse.

Suboptimal observations:
“Several rats showed spontaneous tumors, which are common in this strain.”

Refined observations:
“Spontaneous tumors were observed in 3 of 50 rats. These tumors are characteristic of Fischer 344 rats, have been reported at similar incidence rates in historical control databases, and are unrelated to compound administration.”

Why it works: By naming the strain, incidence rate, and historical data, the submission preempts Elsa from flagging these findings as treatment-related.

Bridge AI-readable and human-centric content

Use dual-layer authoring: one layer optimized for AI (structured, templated), another for human reviewers (narrative, interpretive). This approach meets both audiences without compromise.

Suboptimal (narrative only):
“Patients receiving the investigational drug experienced improved lung function over 12 weeks.”

Refined (dual-layer):
• Table for Elsa: Mean FEV1 improvement at week 12: +320 mL (Drug A, n=75) vs +110 mL (Placebo, n=74).
• Narrative for humans: “Patients treated with Drug A showed a clinically meaningful improvement in lung function, averaging a 320 mL gain in FEV1 at week 12, compared with 110 mL in the placebo group.”

Why it works: Structured data satisfies AI parsing, while narrative interpretation supports human review.

Build internal AI validation workflows

Create “pre-flight” checks that mirror Elsa’s logic: validate consistency across content and data, confirm clarity in language, and examine potential interpretation pitfalls.

Suboptimal:
In the safety section, one paragraph states: “No cases of arrhythmia were reported.” Later, in a summary table: “Two arrhythmia cases noted.”

Refined:
Cross-check and reconcile before submission:
• Table: “Two arrhythmia cases reported; both were mild, transient, and adjudicated as unrelated to study drug.”
• Narrative: “Two mild cases of arrhythmia occurred but were judged unrelated to treatment, consistent with baseline cardiac risk factors in the study population.”

Why it works: The refined version eliminates internal contradictions, which Elsa would otherwise flag as inconsistencies.

Upskill your team

Invest in regulatory science professionals who understand data structuring, metadata tagging, and AI-driven QC processes. This is not AI jargon but the strategic readiness that sets you apart.

Suboptimal:
Team writes submissions entirely in free-form narrative, with limited knowledge of data structuring or tagging.

Refined:
Your team incorporates metadata tags (e.g., “” or “”), builds modular text libraries aligned with CDISC standards, and trains staff to run mock “AI parses” of draft documents to spot weak points before submission.

Why it works: Practical skills in structuring and validation prepare teams for AI-driven review, reducing downstream delays.

Looking Ahead: AI as a Catalyst, Not a Crutch

Elsa’s rollout signals a paradigm shift from document-centric to data-centric, AI interpretable regulatory science.* The industry must adapt accordingly:

• Speed and precision will define competitive advantage. Submission back-and-forth cycles may shrink, so first-time quality becomes a strategic goal.
• Governance frameworks must evolve. Establish clear protocols on AI usage from audit trails to model validation. They should be included in the submissions quality metrics.
• Regulatory engagement grows in importance. Take advantage of FDA’s emerging technology pathways, such as pre-IND consultations, to field these new submission strategies early.

Final Thought

This is about approaching a fundamental change in regulatory practice with cautious curiosity and strategic intent. Regulatory affairs professionals now face both a challenge and an opportunity: adapt workflows to AI-driven review, and you’ll be the team building future-proof compliance frameworks.

Please let me know if you’d like to explore tools or methodologies that help teams operationalize these changes or if you’d like to frame this in your preferred content structure.