AI Is Heating Up at FDA: What Elsa 4.0 and HALO Could Mean for the TMF

Artificial intelligence is not taking the summer off.

In May 2026, the U.S. Food and Drug Administration announced Elsa 4.0, an expanded version of its internal AI assistant, along with HALO, a centralized platform connecting more than 40 application and submission data sources, systems, and portals across the agency. Together, these developments show that AI is moving beyond experimentation and becoming part of real regulatory workflows.

For clinical research and Trial Master File professionals, that is worth paying attention to.

The FDA has not announced that Elsa or HALO will directly inspect sponsor TMFs. However, the agency is clearly investing in tools that make information easier to search, connect, analyze, and use across regulatory activities. From an LMK Clinical Research Consulting perspective, this raises an important question:

What happens when regulators can navigate information faster than the organizations submitting it?

Meet Elsa 4.0

Elsa was originally launched across the FDA in June 2025 to help employees, including scientific reviewers and investigators, work more efficiently. The FDA reported that the tool was already being used to support activities such as clinical protocol reviews, scientific evaluations, safety assessments, label comparisons, and the identification of inspection priorities.

Elsa 4.0 expands those capabilities with features that include:

• Custom AI agents
• Document generation
• Quantitative data analysis and visualization
• Secure web search
• Voice-to-text dictation
• Conversion of scanned records and images into searchable text
• Enhanced chat capabilities
• Optimized searching across large document repositories

The FDA states that Elsa operates within a secure environment, does not train on information submitted by regulated organizations, and keeps FDA subject matter experts involved throughout the process. Human reviewers remain responsible for verifying inputs, analytical processes, and how outputs are used.

That last point may be the most important one for the clinical research industry.

What Is HALO?

HALO stands for Harmonized AI & Lifecycle Operations for Data.

The platform brings together more than 40 previously separate FDA application and submission data sources, systems, and portals. The FDA is integrating HALO with Elsa so employees can search information and create workflows without repeatedly uploading records into individual AI conversations.

Think of Elsa as the assistant and HALO as the connected environment that gives the assistant controlled access to relevant information.

The goal is not simply to store more data. It is to make regulatory information easier to find, connect, and analyze across the agency.

For TMF professionals, this is where the announcement becomes especially interesting.

Why TMF Professionals Should Pay Attention

The TMF tells the story of how a clinical trial was conducted, managed, and overseen. That story is rarely contained in one record.

Evidence may be distributed across:

• Regulatory records
• Site management records
• Monitoring documentation
• Training records
• Safety information
• Vendor oversight documentation
• Essential correspondence
• System-generated metadata
• Quality review and issue-management records

Today, reviewing these relationships often requires significant manual effort. An experienced TMF professional may need to search multiple zones, compare versions, review dates, and reconstruct a sequence of events before determining whether the trial record is complete and consistent.

AI could help make some of that work faster. But faster does not automatically mean more accurate.

The real opportunity is using AI to direct human attention toward the records, relationships, and risks that require expert review.

Where AI Could Add Value to TMF Operations

Finding and Organizing Records

AI-supported eTMF technology could potentially help teams:

• Classify records
• Recommend metadata
• identify duplicate content
• Detect potential misfiling or misclassification
• Convert scanned records into searchable text
• Locate information across large record collections
• Identify records that may be missing required metadata

These capabilities could reduce the time spent on repetitive administrative tasks while allowing TMF professionals to focus on quality decisions.

Connecting Related Evidence

Some of the most valuable AI use cases may involve comparing records that should tell the same story.

Examples could include:

• Monitoring visit logs compared with monitoring visit reports
• FDA Form 1572 information compared with investigator CVs, medical licenses, and financial disclosure forms
• IRB or IEC approval dates compared with site activation dates
• Informed consent form version dates compared with participant enrollment dates
• Training completion dates compared with the first performance of a study activity
• Submission correspondence compared with regulatory, IRB, or IEC approvals
• Vendor oversight records compared with documented issues, escalations, and follow-up actions

AI may help flag inconsistencies. A qualified reviewer would still need to determine whether the difference represents a true compliance concern, an acceptable study-specific circumstance, or simply incomplete context.

Focusing Quality Review

AI may also help TMF teams identify where to look first by supporting:

• Completeness reviews
• Risk scoring
• Gap analysis
• Timeliness trending
• Recurring quality issue identification
• Audit and inspection preparation
• QC action recommendations
• Analysis of filing patterns
• Vendor oversight traceability
• Review of trends across studies, countries, sites, or service providers

The objective should not be to automate professional judgment. It should be to give professionals better information with which to make decisions.

The Summer Plot Twist: AI Does Not Own Compliance

The FDA’s own recent enforcement activity provides an important reminder.

In an April 2026 Warning Letter involving pharmaceutical manufacturing, the FDA described a company’s use of AI agents to create drug product specifications, procedures, and production or control records. The agency stated that the organization failed to adequately review the AI-generated materials for accuracy and CGMP compliance.

The FDA made its expectation clear: AI-generated output used for regulated activities must be reviewed and cleared by an authorized human representative.

This Warning Letter was related to pharmaceutical manufacturing rather than clinical trial TMF management. Still, the broader lesson is highly relevant:

An AI tool cannot accept regulatory accountability on behalf of an organization.

If AI incorrectly classifies a record, overlooks a missing approval, recommends the wrong metadata, or produces an inaccurate procedure, responsibility does not transfer to the technology.

Accountability remains with the sponsor, quality function, process owner, system owner, and qualified subject matter experts.

Before Diving In, Check the Water

Organizations exploring AI-supported TMF processes should establish clear controls before placing regulated records or decisions into the workflow.

Important considerations include:

• A clearly defined and documented context of use
• Risk-proportionate validation and performance testing
• Appropriate data privacy and security controls
• Traceable inputs, outputs, decisions, and approvals
• Defined roles for human review and escalation
• Procedures for handling incorrect or inconsistent output
• Ongoing monitoring after implementation
• Change control as tools, models, and intended uses evolve
• Training for employees responsible for reviewing AI-supported work

These concepts align with the FDA’s 2026 Good AI Practice principles, which emphasize human-centric design, risk-based approaches, data governance, documentation, multidisciplinary expertise, performance assessment, and lifecycle management.

AI governance should not be treated as a one-time technology project. It must continue throughout the life of the tool and its use within the organization.

A Smarter TMF Still Needs Strong Foundations

AI will not fix a TMF that lacks clear processes, reliable metadata, controlled terminology, defined ownership, or consistent filing practices.

In fact, AI may expose those weaknesses faster.

The value of any AI-supported analysis depends on the quality and context of the information it receives. Incomplete records, inconsistent naming conventions, poor metadata, and unclear process rules can produce misleading recommendations at greater speed.

Before asking whether an organization is ready for AI, TMF leaders should first ask:

• Are our processes clearly defined?
• Is our metadata reliable?
• Can we trace decisions and changes?
• Are responsibilities understood?
• Would a qualified reviewer trust the underlying records?

AI performs best when it is built on a strong foundation of TMF governance and quality.

Final Thoughts

The FDA’s expansion of Elsa and the introduction of HALO provide another clear signal that AI will play an increasingly important role across regulatory operations and the drug development lifecycle.

For the TMF community, this creates exciting opportunities. AI could help teams search faster, identify inconsistencies earlier, recognize emerging risks, and prepare more efficiently for audits and inspections.

But the fundamentals have not changed.

Organizations must still maintain complete, accurate, timely, and reliable trial records. They must still understand why decisions were made, who reviewed the evidence, and whether the resulting trial story can withstand regulatory scrutiny.

AI may help us find the sunscreen, check the forecast, and spot the storm clouds before they arrive. Qualified people must still decide whether it is safe to head to the beach.

The future of AI in clinical research should follow one straightforward principle:

Human reviewed. Compliance controlled.