FDA's HALO Platform And Elsa 4.0: Five Critical Risks For Sponsors

By Kimberly Chew and Odette Hauke

On May 6, 2026, the FDA announced two linked changes to its internal technology infrastructure: the launch of Elsa 4.0 and the consolidation of more than 40 application and submission data sources across FDA centers into HALO, the agency's Harmonized AI & Lifecycle Operations for Data platform. The FDA stated that Elsa is built within a FedRAMP High secure Google Cloud Platform environment, does not train on input data or data submitted by regulated industry, and keeps FDA staff involved so human subject matter experts verify inputs, analytic processes, and output implementation.¹

Those assurances matter. But they do not eliminate the practical questions HALO raises for sponsors. As FDA Chief AI Officer Jeremy Walsh put it, Elsa will soon become the agency's main entry point into FDA systems and data: previously, staff brought data to Elsa; now, Elsa sits on top of FDA data.¹

This represents a fundamental architectural shift from document-centric regulatory review to data-centric review, where the FDA has unified, cross-center access to historical and current submission data. Earlier analysis of FDA's migration from Anthropic's Claude to Google's Gemini identified heightened exposure for confidential data, administrative record integrity, and regulatory predictability.² HALO's consolidation has amplified those structural risks. Recent analysis of FDA's real-time clinical trial (RTCT) pilot identified concerns about data quality and confidentiality in continuous data flows.³ HALO compounds those concerns by making real-time data potentially queryable across all FDA centers.

This article examines five critical risks created by HALO's architecture and Elsa 4.0's expanded capabilities and concludes with practical safeguards sponsors should implement immediately.

Risk 1: Cross-Center Data Access

HALO consolidates more than 40 application and submission data sources across all centers.¹ Traditional FDA workflows created natural data silos. HALO eliminates these silos by consolidating submission data, adverse event databases, inspection records, and application systems into a unified platform accessible through Elsa's query interface.¹

The FDA has not yet publicly detailed HALO's role-based access controls, retrieval boundaries, or audit log practices. This operational consolidation could potentially expose sponsor data across all FDA centers simultaneously (e.g., a CDER reviewer querying Elsa about Drug A could inadvertently surface confidential data from a CBER Biologic B submission if both share a common API or indication).

The legal framework protecting trade secrets and confidential commercial information (CCI) remains unchanged. The Federal Food, Drug, and Cosmetic Act § 301(j) prohibits FDA from revealing protected trade secret methods or processes; the Trade Secrets Act criminalizes unauthorized disclosure; and FOIA Exemption 4 shields trade secrets and confidential commercial information.⁴ FDA regulations at 21 CFR 20.61, 314.430, and 601.51 protect specified trade secrets and CCI in drug and biologics files.⁵

One citation deserves precision: 21 CFR 20.85 addresses disclosure of otherwise exempt FDA records to other federal departments and agencies — it does not define cross-center access within the FDA. Contractor access is addressed separately in 21 CFR 20.90.⁵ Out of an abundance of caution, sponsors should assume that any data submitted to FDA is potentially queryable across all centers unless explicitly segregated and protected by documented access controls.

Immediate Actions:

• Request written confirmation: "Which centers, offices, review teams, contractors, or administrative personnel can access our submission data via HALO?"
• Request a description of role-based access controls and data segregation rules.
• Mark CCI and trade secret material clearly and cite 21 CFR 20.61, 314.430, and/or 601.51 in cover letters.⁵
• If contractor or system integrator access is material, ask how FDA is applying 21 CFR 20.90's written agreement and security precaution requirements.
• Request access to HALO audit logs showing which FDA personnel queried your data.

Risk 2: Custom Agents Fragment The Administrative Record

Elsa 4.0 introduces "custom agents" — agentic AI tools that reviewers can tailor for specific tasks or therapeutic areas.¹ The FDA also launched an Agentic AI Challenge for staff to build AI solutions for agency workflows.⁶ At the time of this writing, it is unknown how custom agents are configured, validated, monitored, or preserved in records for specific reviews.

Custom agents could potentially create non-standardized AI outputs across reviews. If Reviewer A uses a custom oncology agent and Reviewer B uses a different agent for the same indication, sponsors may receive inconsistent queries. This could undermine the administrative record integrity; the basis for FDA's decision varies depending on which agent was used.

The Administrative Procedure Act (APA) requires that agency decisions be based on reasoned explanation supported by the administrative record. Courts review agency action for arbitrariness and capriciousness, grounded in the record before the agency. State Farm emphasized that an agency must examine relevant data and articulate a satisfactory explanation for its action.⁷

If a custom agent generates a query or flags a concern, both the agent's output and the human reviewer's rationale for relying on it should be preserved in the record.⁸ Without dual documentation, sponsors cannot effectively challenge adverse findings.

Immediate Actions:

• Request disclosure of whether a custom agent materially informed the review and request preservation of the AI-generated output and human validation.
• Where an FDA request appears highly specific, unexpected, or inconsistent with prior communications, ask for the human reviewer's scientific rationale and the source materials supporting the request.
• Request that the custom agent's output, validation documentation, and the human reviewer's independent rationale be included in the administrative record.⁸
• Maintain timestamped records of FDA questions, sponsor responses, and clarifications.

Risk 3: OCR And Document Generation Create Metadata And Authorship Risks

Elsa 4.0 can convert scanned documents and images into searchable text using optical character recognition (OCR) and can generate documents.¹ OCR can make previously hard-to-search material machine-readable: handwritten notes, scanned protocols, slide annotations, visible redactions, and watermarks. While this improves efficiency, scanned documents may contain annotations or redactions that sponsors assumed were not machine-readable.²

Document generation raises questions about authorship. If a review memo contains AI-generated text, is it the reviewer's official position or an unvalidated draft? Under the APA, the administrative record should reflect the actual basis for the agency's decision.⁷ AI-generated drafts that were not validated should not be treated as official findings.

Immediate Actions:

• Avoid submitting scans with proprietary annotations, handwritten notes, or visible redactions.² Submit clean text versions.
• Ask how OCR-extracted text and related metadata are retained, access controlled, and used in search or review workflows.
• If FDA provides an AI-generated summary or chart, request the underlying human-authored or human-validated analysis.⁸

Risk 4: Web Search Creates Precedent Bias For Novel Therapies

Elsa 4.0 has "web search through a secure web access feature" that allows it to access "refreshed secure web data."¹ The FDA has not yet publicly identified the sources, scope, or curation standards for that data. This may mean Elsa can query curated external databases such as PubMed and Clinicaltrials.gov in real time.

For novel modalities such as psychedelics, gene therapies, digital therapeutics, and rare disease interventions, external databases have limited or outdated precedents. Elsa may surface older, lower-quality studies that do not reflect current best practices, leading to inappropriate comparisons.⁸ For example, searching "psilocybin clinical trials" might surface 1960s-era studies with poor methodology, creating unfair expectations for modern trials.

Prior analysis has noted that AI-assisted review may underrepresent novel or precedent-sparse modalities.⁸ Psychedelic-assisted therapy outcomes, patient reported outcomes in rare diseases, and digital biomarkers may lack robust external precedent, and Elsa's web search could impose standards from unrelated therapeutic areas.⁸

Immediate Actions:

• Provide a curated literature package that distinguishes high-quality current evidence from outdated precedent.
• Include concise bridging rationales explaining why historical precedent may not apply to your novel modality.⁸
• Ask FDA to disclose which external databases Elsa accessed during your review and which were materially relied on in a question, deficiency, or review conclusion.
• When a question appears based on an outdated or mismatched comparator, respond with a clear scientific distinction.

Risk 5: Quantitative Analysis On Unverified Data

Elsa 4.0 can perform "quantitative data analysis and visualization, including chart/graph creation."¹ The FDA states that human subject matter experts verify inputs and outputs.¹ That verification is essential because AI-generated visualizations can look definitive even when they depend on incomplete data or incorrect assumptions.

Elsa may apply standard statistical methods to novel endpoints, such as generating a dose-response curve that assumes linearity when the true relationship is nonlinear. AI-generated charts may appear authoritative even if based on incomplete or misinterpreted data.

The issue also intersects FDA's real-time clinical trial initiative. The FDA has announced proof-of-concept real-time clinical trials with AstraZeneca and Amgen.⁹ Traditional trials rely on data review, query resolution, source data verification, and database lock before key analyses reach regulators.³ RTCT could compress that quality control sequence, and a miscoded lab value or site entry error could look like a meaningful safety signal before basic review is complete.³

If real-time trial data flows into HALO and Elsa's quantitative analysis is applied before human verification, a site entry error could be visualized as a meaningful safety signal, triggering regulatory action based on uncorrected data.³

Immediate Actions:

• Provide your own prespecified statistical analysis plan and request that FDA validate any AI-generated analyses against it.
• Label preliminary, queried, corrected, and locked data states clearly.
• If FDA provides an AI-generated chart, request the underlying source data set, version date, assumptions, and human validation basis.²
• For RTCTs, clarify whether Elsa's quantitative analysis will be applied to real-time data before human verification and request explicit data segregation in HALO.³

Essential Sponsor Checklist

Before Submission:

• Map CCI and trade secret content; mark it clearly and cite 21 CFR 20.61, 314.430, 601.51.⁵
• Submit clean machine-readable documents; avoid scans with annotations or redactions.
• Request written confirmation: "Which FDA centers/personnel can access our data via HALO?"
• For novel modalities, include a curated precedent package and ask FDA to disclose which external databases Elsa accessed during your review and which were materially relied on in a question, deficiency, or review conclusion.

During Review:

• Maintain timestamped records of all FDA queries and responses.²
• If queries appear AI-generated, request clarification in writing.
• Ask which AI tools were used (custom agents, web search, quantitative analysis).
• Request that AI-generated outputs and human validation be included in the administrative record.⁸

After Review:

• Request the complete administrative record, including all AI-generated artifacts.⁸
• Verify that HALO access logs are available if needed for appeal.
• Preserve all records for potential APA challenge if the administrative record appears incomplete.⁸

Conclusion

HALO and Elsa 4.0 may improve FDA efficiency. Sponsors should support FDA modernization that improves review quality while ensuring their procedural rights are protected. But modernization also changes the sponsor risk profile. Cross-center data access, custom agents, OCR metadata, web search precedent bias, and unvalidated quantitative analysis all require proactive sponsor safeguards.

The legal protections for trade secrets and confidential commercial information remain unchanged, but sponsors must actively enforce them in this new architecture. Earlier analysis of FDA's migration from Claude to Gemini identified heightened risks for data confidentiality and administrative record integrity.² HALO has amplified them, making proactive engagement and meticulous documentation essential. Sponsors should mark confidential information, submit clean materials, document FDA interactions, ask precise questions about AI use when it matters, and insist that material agency positions remain traceable to human scientific judgment. The sponsors that adapt fastest to this architecture will be best positioned to protect confidential information while engaging productively with FDA's AI-enabled review process.

References:

1. U.S. Food & Drug Administration, FDA Expands AI Capabilities and Completes Data Platform Consolidation, News Release (May 6, 2026). https://www.fda.gov/news-events/press-announcements/fda-expands-ai-capabilities-and-completes-data-platform-consolidation
2. Kimberly Chew & Michael Yang, FDA's Elsa AI Switches From Claude To Gemini: What Sponsors Need To Know, Clinical Leader (March 12, 2026), https://www.clinicalleader.com/doc/fda-s-elsa-ai-switches-from-claude-to-gemini-what-sponsors-need-to-know-0001; Kimberly Chew & Michael Yang, Elsa's AI Model Migration: Technical, Compliance, And Regulatory Risks For Sponsors (Part 2), Clinical Leader (March 26, 2026), https://www.clinicalleader.com/doc/elsa-s-ai-model-migration-technical-compliance-and-regulatory-risks-for-sponsors-part-0001; Kimberly Chew & Michael Yang, Navigating Elsa's AI Transition: Practical Guidance To Safeguard Confidential Information (Part 3), Clinical Leader (April 22, 2026), https://www.clinicalleader.com/doc/navigating-elsa-s-ai-transition-practical-guidance-to-safeguard-confidential-information-part-3-0001
3. Kimberly Chew & Odette Hauke, 5 Critical Risks With The FDA's Real-Time Trial Monitoring, Clinical Leader (May 8, 2026), https://www.clinicalleader.com/doc/critical-risks-with-the-fda-s-real-time-trial-monitoring-0001?utm_source=linkedin&utm_medium=social&utm_campaign=AP
4. 21 U.S.C. § 331(j); 18 U.S.C. § 1905; 5 U.S.C. § 552(b)(4).
5. 21 C.F.R. §§ 20.61, 20.85, 20.90, 314.430, 601.51.
6. U.S. Food & Drug Administration, FDA Expands Artificial Intelligence Capabilities with Agentic AI Deployment, News Release (Dec. 1, 2025), https://www.fda.gov/news-events/press-announcements/fda-expands-artificial-intelligence-capabilities-agentic-ai-deployment
7. 5 U.S.C. §§ 551 et seq., 706; Motor Vehicle Mfrs. Ass'n v. State Farm Mut. Auto. Ins. Co., 463 U.S. 29 (1983).
8. Kimberly Chew, Odette Hauke & Kathleen Snyder, AI At The FDA: Legal Implications And Strategic Considerations For Drug Developers, Clinical Leader (Jan. 19, 2026), https://www.clinicalleader.com/doc/ai-at-the-fda-legal-implications-and-strategic-considerations-for-drug-developers-0001; Kimberly Chew, Odette Hauke & Kathleen Snyder, Navigating FDA's New AI Systems: Practical Tips For Regulatory Success, Clinical Leader (Jan. 19, 2026), https://www.clinicalleader.com/doc/navigating-fda-s-new-ai-systems-practical-tips-for-regulatory-success-0001
9. U.S. Food & Drug Administration, FDA Announces Major Steps to Implement Real-Time Clinical Trials, News Release (April 28, 2026), https://www.fda.gov/news-events/press-announcements/fda-announces-major-steps-implement-real-time-clinical-trials; FDA, AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program; Request for Information, 91 Fed. Reg. 23100 (April 29, 2026), https://regulations.justia.com/regulations/fedreg/2026/04/29/2026-08281.html

About The Authors:

Kimberly Chew is senior counsel in Husch Blackwell LLP’s virtual office, The Link. Chew is a seasoned professional with a background in biotech research, leveraging her experience to guide clients through the intricate landscape of clinical trials, FDA regulations, and academic research compliance. As the cofounder and co-lead of the firm’s Psychedelic and Emerging Therapies practice group, Kimberly is inspired by the potential of psychedelic therapeutics to address mental health conditions like PTSD. Her practice encompasses regulatory due diligence and intellectual property enforcement, particularly in patent infringement and validity. She can be reached at kimberly.chew@huschblackwell.com.

Odette Hauke is a global regulatory affairs consultant supporting regulatory strategy across clinical development and registration, with an emphasis on clear regulatory narratives and submission strategies that meet heightened evidentiary expectations. She has 12+ years’ experience directing IND/CTA/NDA/BLA/MAA work across the U.S., EU, U.K., Japan, Canada, APAC, and Latin America, and is experienced in integrating AI/ML-enabled regulatory intelligence into decision-making. Previously, she served as associate director of regulatory affairs at AtaiBeckley, leading global regulatory strategy for first-in-class psychedelic and neuropsychiatric programs including VLS-01 (DMT) and EMP-01 (MDMA), navigating novel endpoints, complex trial operations, Schedule I requirements, and evolving global guidance. Earlier, at Memorial Sloan Kettering Cancer Center, she managed 200+ oncology IND submissions and maintained regulatory documentation for 30+ clinical trials, including pediatric research. She holds an M.S. in regulatory affairs and a B.S. in epidemiology.