What is the FDA's new AI tool "Elsa" and how does it impact inspections?

The FDA's new internal generative-AI assistant ELSA (Enterprise Language Support Assistant) is an agency-wide tool launched in June 2025 to help reviewers, investigators, and scientists read, summarize, and prioritize large volumes of regulatory and safety data faster.

The agency reports use across clinical protocol review, adverse-event summarization, label comparison, and inspection targeting. Early rollout shows clear efficiency gains but also raises accuracy, oversight, and governance questions.

Pharma and medical device companies must treat outputs as investigative signals, not final regulatory decisions, and should strengthen documentation, data traceability, and proactive inspection readiness accordingly.

This information is sourced from the Atlas Compliance tool. For more details, visit their website.

What is ELSA? Short answer, architecture, and purpose

ELSA is the FDA's internally-hosted generative-AI assistant created to make routine, repetitive, and high-volume regulatory tasks faster and more consistent across the agency. The tool is designed to read, summarize, compare, and highlight items of interest in documents, for example, clinical protocols, adverse-event narratives, labeling, inspection reports, and large document libraries, enabling subject-matter experts to focus on judgment and decisions rather than clerical work. According to the FDA, ELSA is already being used across centers for tasks such as clinical protocol reviews and safety signal triage.

Technically, ELSA is a large language model (LLM)-based system integrated into a secure federal environment. The FDA has emphasized that ELSA runs within a controlled cloud environment (AWS GovCloud) and that it is configured to access internal document stores without training the base model on proprietary sponsor submissions. The agency framed ELSA as a productivity and quality tool to accelerate evidence synthesis, reduce review time, and help identify priority inspection targets.

Why the FDA built ELSA (agency motivations)

Four practical drivers explain why the FDA moved quickly to deploy ELSA:

1. Volume of work: The agency processes thousands of submissions, adverse-event reports, labeling updates, and inspection records. Summarizing and cross-checking these documents consumes skilled reviewer time. ELSA is aimed at relieving that bottleneck.
2. Speed to decision: Faster synthesis can shorten times to internal conclusions and accelerate regulatory timelines, which the agency says could shrink review tasks that used to take days into minutes.
3. Targeting inspections: By surfacing patterns and risk signals across databases, ELSA can flag facilities or products for higher-priority inspections, enabling a more data-driven field program.
4. Consistency and knowledge capture: The tool can help produce consistent summaries and preserve institutional knowledge when staff rotate or retire.

What ELSA can do today (practical capabilities)

Based on the FDA statements and early reporting, ELSA's practical capabilities include:

• Summarization of long documents (clinical study reports, adverse event narratives, inspection reports) into concise, reviewable summaries for human reviewers.
• Comparative analysis, such as checking labeling/insert language against standards or prior versions, and spotting inconsistencies faster than manual review.
• Triage and prioritization by surfacing high-priority safety signals or facilities that merit inspection attention.
• Drafting support for internal products: generating minutes, draft summaries, or database scripts to help analysts speed implementation tasks.

Important operational note from the agency: ELSA is being deployed for assistance and augmentation, not as an autonomous decision-maker. Human experts still make regulatory conclusions. The tool's goal is to reduce administrative load, not to replace regulator judgment.

Early outcomes the FDA is reporting

Public statements from the FDA include quantitative and qualitative claims: internal pilots reportedly reduced some review tasks from multi-day efforts to minutes for draft read-outs, accelerating certain scientific evaluations. The rollout was described as completed "ahead of schedule and under budget," and the agency reported that it would scale ELSA across more units. Reuters and FDA announcements confirmed the June 2025 launch and early use cases such as protocol review and adverse-event summarization.

Accuracy, hallucinations, and oversight: real concerns and reported incidents

While ELSA promises productivity, independent and trade reporting flagged accuracy and governance concerns shortly after rollout. Several outlets documented that:

• The tool has produced false citations, incorrect facts, or outputs inconsistent with source documents in some pilot instances, the classic LLM "hallucination" problem. This created worries about relying on raw ELSA outputs for formal regulatory conclusions.
• FDA communications stressed configurations intended to prevent internet access by ELSA and to avoid training the model on confidential sponsor submissions, but critics still highlighted the need for stronger validation and traceability frameworks for the AI outputs.

These issues imply the following: ELSA's outputs can produce useful signals and drafts, but must be validated and traced back to the source documents by human experts before any regulatory action or public statement is made. Regulatory staff and external stakeholders alike have urged the agency to publish comprehensive AI governance, testing, and audit protocols.

How ELSA changes inspection planning and execution

This is the part that pharma and medical device manufacturers and facility quality teams should watch closely.

1. Smarter targeting of inspections Because ELSA can analyze adverse events, recall histories, inspection histories, and other datasets at scale, it can highlight facilities or product lines that show patterns the machine considers high-risk. That means regulators may arrive with narrower, more targeted scopes and pre-prepared lines of questioning often triggered by patterns that a traditional manual review might have missed.
2. Faster identification of trends across companies ELSA can find similarity patterns, e.g., recurring deviations tied to a contract manufacturer, specific supplier lots, or a batch of stability complaints. Expect investigators to have easier access to collated trend summaries when they arrive on site. This raises the bar on how companies must demonstrate systemic controls.
3. Pre-inspection dossiers will be deeper. Investigators supported by ELSA could bring deeper, autogenerated bundles of evidence requests. Companies should assume that regulators may show up with lists of specific documents and targeted data points identified algorithmically. Be prepared to produce well-indexed, searchable records quickly.
4. Emphasis on traceability and provenance Given the risk of ELSA hallucinations, regulators will expect comprehensive provenance to back any observation. That means well-timestamped e-records, audit trails, validated electronic systems, and clear SOPs that show how conclusions were reached, especially if a flagged pattern depends on reconciled data (e.g., complaints vs. stability vs. batch records).

Practical implications for pharma and medical device companies (what to do now)

If you work in manufacturing, quality assurance, regulatory, or compliance, the arrival of ELSA means changing how you prepare for inspections and document lifecycle governance. Here's a prioritized checklist:

1. Treat any regulator-presented AI summary as a signal, not gospel. If an investigator cites AI-generated findings, politely request the source document references and show traceable evidence. Insist on document-level provenance. (ELSA outputs should be auditable; demand the supporting pages/IDs.) Maintain records that demonstrate human review and verification of your corrective actions.
2. Strengthen record indexing and searchability. Ensure batch records, change controls, CAPAs, stability reports, and complaint files are searchable and have clear metadata (lot number, dates, reviewer names). If an investigator asks for "all stability deviations for lot X since 2023," you should be able to produce them quickly.
3. Harden e-records and audit trails. Validate your e-document management, ensure time-stamps and electronic signatures are intact, and that exports show unaltered history. Automated summaries will hold less weight without verifiable provenance.
4. Review labeling and promotional material workflows. Because ELSA can compare labeling quickly, make sure label change histories, justification documents, and regulatory submissions are organized and readily accessible.
5. Practice targeted mock inspections. Simulate inspections where regulators present narrow, AI-surfaced issues. Train teams to respond with evidence mapping rather than defensive or anecdotal answers.
6. Update SOPs to include AI-driven queries. Add procedures for receiving, acknowledging, and responding to algorithmically generated queries, for example, who will map AI signals to source documents, and how the company will escalate suspected AI-driven errors.
7. Be proactive about supplier data. If patterns tie back to third-party suppliers, ensure contract manufacturers and suppliers maintain good records and agree to timely data access during regulatory inquiries.

How investigators' jobs are changing (the human investigator report)

Although the media often imagines AI replacing people, ELSA's rollout shows a different picture: it amplifies investigators' analytical bandwidth.

• Investigators will be more prepared. Field investigators may arrive with AI-aggregated summaries showing patterns across multiple datasets. This improves efficiency but also raises expectations for companies to back up claims with traceable evidence.
• Investigative work will become more analytical. Investigators will spend less time on manual collation and more time on focused interviews, systems evaluation, and judgment calls where AI-generated signals need human interpretation.
• New skill sets are needed. Investigators will increasingly need to understand data provenance, basic AI limitations (e.g., hallucination risk), and the ability to verify outputs quickly against primary sources. The agency has already used ELSA in training sessions for investigators.
• Accountability still lies with humans. The FDA has stated humans remain the decision makers; ELSA should assist, not replace, inspection judgment. Thus, investigators must document how AI outputs were used in their conclusions.

Governance, validation, and the risk of 'automation bias.'

Regulatory agencies, industry, and legal analysts have pointed out a major governance risk: automation bias, the tendency of humans to over-trust machine outputs. In a regulatory context, that could lead to:

• Over-reliance on an AI summary without checking against primary evidence.
• Misprioritization if ELSA flags false positives or misses subtle contextual factors that only humans understand.
• Legal and reputational risk if an AI-derived finding is used as the sole basis for enforcement action.

To mitigate these risks, the FDA and external commentators recommend strong validation, human-in-the-loop checkpoints, traceability of AI outputs to original files, and external auditability of the AI pipeline. Early reporting suggests that the FDA knows these risks and is working on governance frameworks, but critics urge faster transparency and published standards.

Atlas

See what the FDA knows about your next investigator.

700,000+ inspections since 2010. Search by investigator, CFR code, or supplier. Every 483 they’ve issued, in under a second.

See it live

How to respond when an investigator cites an AI finding on site

If an FDA investigator presents a summarized or AI-derived point during an inspection:

1. Request the source: politely ask for the exact documents, pages, or database query that produced the finding. AI summaries should have traceable references.
2. Show provenance: map your records to those sources (batch records, complaint forms, lab reports).
3. Document the interaction: record the investigator's statement and the references they provided. If the investigator cites an AI product, note the claim and ask whether the investigator has independent confirmation.
4. Offer to provide deeper exports: provide the original records and index pages rather than relying on a short AI summary.
5. Follow up in writing: after the inspection, respond with a curated packet that shows your validations and reconciliations of the flagged items.

Case scenarios: examples of how ELSA might influence real inspections

Scenario A: Label inconsistency flagged by ELSA
ELSA spots a discrepancy in active ingredient wording between the approved label and a recent product insert. Investigators might ask for the change history, submission correspondence, and the internal approval trail. Companies that keep clean change control records and submission transcripts will be able to resolve the query quickly.

Scenario B: Complaint cluster across sites
ELSA aggregates complaint narratives and finds a pattern of a particular defect across three contract-manufacturing sites. Investigators could open targeted site inspections at those CMOs. If you outsource, ensure your supplier quality agreements include rapid access to complaint investigations and root-cause analyses.

Scenario C: Stability outliers
ELSA highlights batches with trending outlier results in accelerated stability. Investigators may ask for raw stability chromatograms, method validation records, and sample retention logs. Labs must be ready to provide raw data with intact audit trails.

What ELSA does not (yet) replace

• Formal regulatory judgment and legal enforcement decisions.
• The need for validated, auditable primary data.
• Human clinical and scientific expertise required for nuanced assessment of benefit and risk.
• External stakeholder transparency: ELSA was designed for internal FDA use and (per the agency) is not trained on proprietary sponsor data.

What companies should ask internally (short list)

• Are our e-records and metadata searchable and auditable?
• Do our labeling change controls include a clean, retrievable submission trail?
• Can we produce raw analytical data and retain sample retentions for periods investigators might request?
• Are our suppliers contractually obligated to provide timely records?
• Have we updated mock inspection scripts to include AI-triggered document requests?

Broader industry and policy implications

ELSA's launch is a test case for modernizing regulatory agencies with AI in inspections. If the program proves reliable and governed, it could:

• Decrease time to regulatory decisions where tedious evidence synthesis is the bottleneck.
• Improve the agency's ability to spot systemic risks across industries faster than traditional manual methods.
• Force industry to modernize record keeping, data quality, and digital traceability.

If done poorly, however, it may lead to enforcement decisions built on shaky evidence or fuel public distrust if errors are publicized. The FDA's transparency about ELSA's environment (GovCloud hosting, internal-only access, non-training on sponsor files) and its stated human-in-the-loop approach are designed to reduce these risks, but independent governance and audit mechanisms will be crucial.

Quick primer for QA/RA leaders: 10 action items (checklist)

1. Inventory: map where records live (LIMS, eQMS, ERP, paper) and record owners.
2. Indexing: add searchable metadata to critical records (lot, dates, reviewers).
3. Audit trail: validate e-record systems and ensure exportable audit histories.
4. Supplier obligations: tighten clauses for timely data access.
5. Mock inspections: include AI-driven, narrow scope requests.
6. Training: teach staff how to respond to AI-sourced questions calmly and evidence-first.
7. CAPA readiness: ensure CAPA files are complete, with root cause evidence.
8. Label control: keep label histories and submission correspondence readily accessible.
9. Archival: ensure sample retention complies with expected inspection windows.
10. Legal: coordinate with counsel on responses to algorithmically-generated allegations.

Final assessment: benefits vs. risks

Benefits:

• Faster synthesis and prioritization of signals.
• More targeted inspections and efficient regulator workflows.
• Potential shortening of review timelines for sponsors.

Risks:

• Hallucination/accuracy problems that could mislead if unchecked.
• Automation bias and governance gaps that demand stronger human oversight and traceability.

How to think about ELSA as a regulated company

ELSA represents a major step in modernizing regulatory workflows. For pharma and medical device manufacturers, the practical response isn't to fight technology; it's to be inspection-ready for a world where AI helps regulators discover signals faster and in more targeted ways. That means improving data quality, traceability, indexing, and crisis response so that when regulators use tools like ELSA, you can answer every algorithmic prompt with verifiable, auditable, human-reviewed evidence.

ELSA may redefine how regulators prioritize and scope inspections. Companies that proactively upgrade their digital records and train teams to meet AI-driven questions with calm, evidence-based responses will turn a potential risk into a compliance advantage.